Quick Guide

Quick Guide to OpenStat
By
William G. Miller, PhD
July, 2007
Table of Contents
QUICK GUIDE TO OPENSTAT............................................................................................................... 1
TABLE OF CONTENTS............................................................................................................................. 2
FIGURES...................................................................................................................................................... 6
INTRODUCTION...................................................................................................................................... 10
INSTALLING AND STARTING OPENSTAT ................................................................................................... 10
DATA CHARACTERISTICS AND RELATED ANALYSES............................................................... 12
TYPES OF VARIABLES............................................................................................................................... 12
Nominal Variables .............................................................................................................................. 12
Ordinal Variables ............................................................................................................................... 12
Interval Variables ............................................................................................................................... 12
Ratio Variables ................................................................................................................................... 13
Discrete and Continuous Variables .................................................................................................... 13
RECORDING DATA................................................................................................................................. 15
VARIABLE LABELS ................................................................................................................................... 15
ENTERING DATA IN THE GRID .................................................................................................................. 16
SAVING YOUR DATA................................................................................................................................ 16
EDITING DATA.......................................................................................................................................... 17
GENERATING DATA.................................................................................................................................. 18
IMPORTING AND EXPORTING DATA.......................................................................................................... 18
Copy and Paste Method of Importing Excel Data .............................................................................. 18
Importing a Tab File........................................................................................................................... 18
Exporting Data from OpenStat ........................................................................................................... 19
THE OUTPUT FORM.................................................................................................................................. 20
TESTING HYPOTHESES ........................................................................................................................ 23
SAMPLE STATISTICS AND POPULATION PARAMETERS............................................................................... 23
DISTRIBUTION OF SAMPLING ERRORS: TYPE I AND TYPE II ERRORS......................................................... 23
DATA DESCRIPTION.............................................................................................................................. 25
DISTRIBUTION OF VALUES ........................................................................................................................ 25
CENTRAL TENDENCY AND VARIABILITY................................................................................................... 26
COMPARISONS WITH THEORETICAL DISTRIBUTIONS ................................................................................. 26
TRANSFORMATIONS.................................................................................................................................. 27
BOX PLOTS............................................................................................................................................... 28
THE BREAKDOWN PROCEDURE ................................................................................................................ 30
STEM AND LEAF PLOTS ............................................................................................................................ 33
THE QQ OR PP PLOT................................................................................................................................ 35
THREE DIMENSIONAL ROTATION ............................................................................................................. 37
GROUP FREQUENCIES ANALYSIS.............................................................................................................. 38
PARAMETRIC AND NON-PARAMETRIC ANALYSES .................................................................... 42
PARAMETRIC ANALYSES .......................................................................................................................... 42
Univariate Analyses............................................................................................................................ 42
Multivariate Analyses ......................................................................................................................... 42
NON-PARAMETRIC ANALYSES.................................................................................................................. 42
Contingency tables.............................................................................................................................. 43
Relationship among rankings.............................................................................................................. 43
Rater agreement.................................................................................................................................. 44
MEASUREMENT INSTRUMENTS........................................................................................................ 45
RELIABILITY AND VALIDITY..................................................................................................................... 45
EXAMPLES OF PARAMETRIC ANALYSES WITH OPENSTAT.................................................... 46
COMPARING SAMPLE MEANS ................................................................................................................... 46
The t-test ............................................................................................................................................. 46
ANALYSIS OF VARIANCE .......................................................................................................................... 48
One, Two or Three Way ANOVA........................................................................................................ 48
Repeated Measures ANOVAs.............................................................................................................. 55
Nested ANOVA Designs...................................................................................................................... 61
Latin Square Designs ANOVA............................................................................................................ 65
ANALYZING RELATIONSHIPS AND PREDICTION CORRELATION AND MULTIPLE REGRESSION................ 69
Correlation.......................................................................................................................................... 69
Multiple Regression and Prediction ................................................................................................... 71
Using Multiple Regression for Analysis of Variance.......................................................................... 74
Regression with a Dichotomous Dependent Variable......................................................................... 80
Simultaneous Multiple Regression...................................................................................................... 82
Cox Proportional Hazards Survival Regression................................................................................. 85
Non-Linear Regression ....................................................................................................................... 88
Weighted Least-Squares Regression................................................................................................... 94
2-Stage Least-Squares Regression.....................................................................................................103
MULTIVARIATE PROCEDURES.........................................................................................................110
IDENTIFYING LATENT VARIABLES...........................................................................................................110
Factor Analysis ..................................................................................................................................110
IDENTIFYING NATURAL GROUPS - CLUSTER ANALYSIS ..........................................................................117
PREDICTING GROUP CLASSIFICATION......................................................................................................121
PATH ANALYSIS: EXAMINING THE PLAUSIBILITY OF CAUSE AND EFFECT MODELS ................................127
Example of a Path Analysis ...............................................................................................................128
NON-PARAMETRIC...............................................................................................................................134
CONTINGENCY CHI-SQUARE ...................................................................................................................134
Example Contingency Chi Square .....................................................................................................135
SPEARMAN RANK CORRELATION ............................................................................................................136
Example Spearman Rank Correlation ...............................................................................................137
MANN-WHITNEY U TEST........................................................................................................................138
FISHERS EXACT TEST.............................................................................................................................139
KENDALLS COEFFICIENT OF CONCORDANCE .........................................................................................141
KRUSKAL-WALLIS ONE-WAY ANOVA..................................................................................................142
WILCOXON MATCHED-PAIRS SIGNED RANKS TEST ................................................................................144
COCHRAN Q TEST ...................................................................................................................................145
SIGN TEST ...............................................................................................................................................147
FRIEDMAN TWO WAY ANOVA..............................................................................................................148
PROBABILITY OF A BINOMIAL EVENT......................................................................................................150
RUNS TEST ..............................................................................................................................................151
KENDALL'S TAU AND PARTIAL TAU........................................................................................................153
KAPLAN-MEIER SURVIVAL TEST.............................................................................................................155
MEASUREMENT RELATED PROCEDURES.....................................................................................164
CLASSICAL ITEM ANALYSIS METHODS....................................................................................................164
Item Discrimination ...........................................................................................................................164
Item difficulty .....................................................................................................................................165
The Item Analysis Program ...............................................................................................................165
Example Item Analysis.......................................................................................................................166
RASCH ONE PARAMETER LOGISTIC SCALING..........................................................................................173
The One Parameter Logistic Model...................................................................................................173
Estimating Parameters in the Rasch Model: Prox. Method...............................................................174
SPEARMAN-BROWN RELIABILITY PROPHECY..........................................................................................183
THREE PARAMETER LOGISTIC SCALING..................................................................................................184
KUDER-RICHARDSON #21 RELIABILITY..................................................................................................184
RELIABILITY ADJUSTMENT FOR SCORE VARIABILITY .............................................................................185
WEIGHTED COMPOSITE TEST RELIABILITY .............................................................................................185
SUCCESSIVE INTERVAL SCALING.............................................................................................................187
GUTTMAN SCALOGRAM ANALYSIS .........................................................................................................189
DIFFERENTIAL ITEM FUNCTIONING .........................................................................................................190
POLYTOMOUS DIF ANALYSIS .................................................................................................................201
COURSE GRADES SYSTEM.......................................................................................................................203
GENERAL UNFOLDING MODEL ANALYSIS...............................................................................................208
ITEM BANKING SYSTEM..........................................................................................................................213
HOYT AND CRONBACH RELIABILITY.......................................................................................................219
THREE PARAMETER LOGISTIC SCALING..................................................................................................220
STATISTICAL PROCESS CONTROL..................................................................................................221
XBAR CHART.........................................................................................................................................221
RANGE CHART.........................................................................................................................................223
THE S CHART ..........................................................................................................................................225
CUSUM CHART......................................................................................................................................227
P CHART..................................................................................................................................................230
DEFECT (NONCONFORMITY) C CHART ....................................................................................................232
DEFECTS PER UNIT U CHART...................................................................................................................234
FINANCIAL PROCEDURES..................................................................................................................237
LOAN AMORTIZATION.............................................................................................................................237
FUTURE VALUE .......................................................................................................................................239
PAYMENT ................................................................................................................................................240
PERIOD PAYMENT....................................................................................................................................241
INTEREST PAYMENT ................................................................................................................................242
INTEREST RATE .......................................................................................................................................244
NUMBER OF PERIODS...............................................................................................................................245
PRESENT VALUE OF AN INVESTMENT ......................................................................................................246
DOUBLE DECLINING BALANCE................................................................................................................247
STRAIGHT LINE DEPRECIATION...............................................................................................................248
ACCELERATED DEPRECIATION................................................................................................................249
INTERNAL RATE OF RETURN ...................................................................................................................250
NET PRESENT VALUE ..............................................................................................................................252
SIMULATION PROCEDURES ..............................................................................................................254
CORRELATION.........................................................................................................................................254
GENERATING MULTIVARIATE NORMAL DISTRIBUTION DATA.................................................................255
ESTIMATING SAMPLE SIZES NEEDED FOR TYPE I AND TYPE II ERROR CONTROL ...................................258
GENERATING POWER CURVES FOR Z-TESTS............................................................................................260
THEORETICAL DISTRIBUTIONS ................................................................................................................263
GENERATE SEQUENTIAL VALUES............................................................................................................264
RANDOM NUMBER GENERATION.............................................................................................................265
SIMULATE 2-WAY ANALYSIS OF VARIANCE ...........................................................................................267
PROJECT EVALUATION GENERATOR........................................................................................................270
PLOT A Z DISTRIBUTION ..........................................................................................................................273
PLOT A CHI SQUARE DISTRIBUTION ........................................................................................................275
PLOT AN F DISTRIBUTION........................................................................................................................276
RESAMPLING AND BOOTSTRAP DEMO.....................................................................................................277
MATRIX MANIPULATION...................................................................................................................282
KEYBOARD INPUT ...................................................................................................................................283
CLEARING A GRID ...................................................................................................................................283
USING THE COMBINATION BOXES ...........................................................................................................283
OPERATIONS AND OPERANDS..................................................................................................................284
MENUS ....................................................................................................................................................284
THE OPERATIONS SCRIPT ........................................................................................................................285
GETTING HELP ON A TOPIC......................................................................................................................286
MATRIX OPERATIONS..............................................................................................................................287
REFERENCES..........................................................................................................................................288
INDEX........................................................................................................................................................294
Figures
Figure 1 The Main Form and Data Grid.......................................................................... 11
Figure 2 The Data Dictionary Form................................................................................. 15
Figure 3 Editing Options.................................................................................................. 17
Figure 4 The Output Form.............................................................................................. 20
Figure 5 Copy and Paste of a Graphic Image ................................................................. 22
Figure 6 Options for Descriptive Analyses..................................................................... 25
Figure 7 Comparison of Observed and Theoretical Distribution.................................... 27
Figure 8 Transformation of a Variable. .......................................................................... 28
Figure 9 Box Plot Form................................................................................................... 29
Figure 10 Sample Box Plot .............................................................................................. 30
Figure 11 The Breakdown Procedure Form..................................................................... 31
Figure 12 Stem and Leaf Plot Form................................................................................. 33
Figure 13 PP or QQ Plot Form........................................................................................ 35
Figure 14 Expected and Observed Values in the QQ Plot of weight in cansas.txt ...... 36
Figure 15 The PP Plot of weight in the cansas.txt file. ................................................ 37
Figure 16 Three Dimension Plot...................................................................................... 38
Figure 17 Group Frequencies Analysis Form. ................................................................. 39
Figure 18 Frequency Distribution of Defects in the Defects.txt file................................ 40
Figure 19 A Pie Chart of Defect Frequencies from the Defects.txt file. ......................... 41
Figure 20. Specification Form for the Student t-test....................................................... 47
Figure 21 Two Way Analysis of Variance Specification ............................................... 49
Figure 22 Means for Row 1 and Row 2. .......................................................................... 53
Figure 23 Means for Column 1 and Column 2 ................................................................ 54
Figure 24 Row and Column Interaction........................................................................... 55
Figure 25 A Repeated Measures Analysis of Variance Form.......................................... 56
Figure 26 One Within and One Between Analysis of Variance Form. ........................... 57
Figure 27 Two Between and One Within Analysis of Variance Form............................ 58
Figure 28 Two Repeated Measures ANOVA Form. ....................................................... 60
Figure 29 Factor B Nested in Factor A Specification Form. ........................................... 62
Figure 30 Three Treatment Design with B Treatment Nested in the A Treatment. ........ 64
Figure 31 Form to Plot GPA versus GREV..................................................................... 70
Figure 32 Plot of GREV with GPA Including Regression Line and Confidence............ 71
Figure 33 Block Entry Multiple Regression Form. ......................................................... 72
Figure 34 Analysis of Covariance Procedure Form......................................................... 75
Figure 35 Binary Logistic Regression Form.................................................................... 80
Figure 36 Form for Specifying a Simultaneous Multiple Regression Analysis .............. 82
Figure 37 Cox Regression Specification Form. ............................................................... 87
Figure 38 Non-Linear Regression Specification Form. ................................................... 90
Figure 39 X Versus Observed Y Plot From the Non-Linear Regression Analysis. ........ 91
Figure 40 X Versus the Predicted Y Plot from the Non-Linear Regression Analysis..... 92
Figure 41 The Non-Linear Regression Form Following the Analysis. ........................... 94
Figure 42 Weighted Least Squares Regression Specification Form................................ 96
Figure 43 Regression Plot from the Weighted Least Squares Analysis. ......................... 98
Figure 44 Residual Plot from the Weighted Least Squares Procedure. ........................... 99
Figure 45 Plot of the Pulse Versus Squared Residual from Weighted Least Squares. .. 100
Figure 46 Chins Variable Plot from the Weighted Least Squares Analysis. ................. 101
Figure 47 Situps Variable Plot from the Weighted Least Squares Analysis. ................ 102
Figure 48 Predicted Values and Weights from the Weighted Least Squares Analysis. 103
Figure 49 Two Stage Least Squares Regression Specification Form. ........................... 105
Figure 50 Specification of a Sample Factor Analysis.................................................... 112
Figure 51 Hierarchical Cluster Analysis Specification Form. ....................................... 118
Figure 52 Discriminant Analysis and Multivariate Analysis of Variance Form. .......... 122
Figure 53 Path Analysis Dialogue Form........................................................................ 128
Figure 54 Contingency Chi-Square Dialogue Form. ..................................................... 135
Figure 55 Spearman Rank Correlation Dialogue Form. ................................................ 137
Figure 56 Mann-Whitney U Test Dialogue Form........................................................... 138
Figure 57 Fischer's Exact Test Dialogue Form............................................................... 140
Figure 58 Kendal's Coefficient of Concordance............................................................ 142
Figure 59Kruskal-Wallis One Way ANOVA on Ranks Dialogue Form........................ 143
Figure 60 Wicoxon Matched Pairs Signed Ranks Test Dialogue Form. ....................... 145
Figure 61 Cochran Q Test Dialogue Form. ................................................................... 146
Figure 62 The Matched Pairs Sign Test Dialogue Form. .............................................. 147
Figure 63 The Friedman Two-Way ANOVA Dialogue Form....................................... 149
Figure 64 The Binomial Probability Dialogue Form. .................................................... 151
Figure 65 A Sample File for the Runs Test ................................................................... 152
Figure 66 The Runs Dialogue Form. .............................................................................. 153
Figure 67 Kendal's Tau and Partial Tau Dialog Form. .................................................. 154
Figure 68 The Kaplan-Meier Dialog.............................................................................. 159
Figure 69 Experimental and Control Curves ................................................................. 163
Figure 70 Classical Item Analysis Form........................................................................ 167
Figure 71 Example Item Characteristic Curve............................................................... 170
Figure 72 Distribution of Test Scores (Classical Test Analysis Procedure.)................. 171
Figure 73 Test Item Difficulty Plot (Classical Test Analysis Procedure.) .................... 172
Figure 74 Rasch One Parameter Logistic Procedure Form. .......................................... 178
Figure 75 Log Item Difficulty - Rasch One Parameter Logistic Analysis. ................... 179
Figure 76 Log Ability Scores - Rasch One Parameter Logistic Analysis...................... 180
Figure 77 Sample Item Information Function Plot - Rasch Analysis. ........................... 181
Figure 78 Kuder-Richardson Formula 21 Reliability Calculation................................. 184
Figure 79 Reliability Changes With Changes in Variance. ........................................... 185
Figure 80 Estimation of Reliability for Composite Tests. ............................................. 186
Figure 81 Successive Interval Scaling. .......................................................................... 187
Figure 82 Guttman Scaling of Affective Items.............................................................. 189
Figure 83 Differential Item Functioning........................................................................ 191
Figure 84 Frequency Plot of Subject Item Scores - Differential Item Functioning....... 194
Figure 85 Form for Differential Item Function of Polytomous Items. .......................... 201
Figure 86 The Grading System Initial Form.................................................................. 204
Figure 87 Grading System Dialog Form for Test Information. ..................................... 204
Figure 88 Example Test Data Entered in the Grading System Form. ........................... 205
Figure 89 Dialog Form for the Grading System to Create a Grade Book. .................... 205
Figure 90 Specifying Grading Procedures in the Grading System. ............................... 206
Figure 91 Test Results Message from the Grading System. .......................................... 206
Figure 92 Example Results of the Grading System Score Conversions. ....................... 207
Figure 93 A Plot of Test Scores from the Grading System. .......................................... 207
Figure 94 Generalized Unfolding Model Specification Form. ...................................... 209
Figure 95 GUM Theta Plot. ........................................................................................... 212
Figure 96 GUM Item Delta Values Plot. ....................................................................... 213
Figure 97 Form for Creating or Editing an Item Bank .................................................. 214
Figure 98 Sample Item Bank Opened for Editing ......................................................... 215
Figure 99 Entry Form for an Item Stem in the Item Bank System................................ 216
Figure 100 Item Bank Test Specification Form............................................................. 217
Figure 101 Entry of a Student ID in the Item Bank System Test Specification............ 217
Figure 102 Item Bank System Test Administration Form............................................. 218
Figure 103 Sample On-Screen Test Administration...................................................... 218
Figure 104 Results Shown for an On-Screen Test......................................................... 219
Figure 105 Reliability Analysis Using Repeated Measures Analysis of Variance........ 219
Figure 106 X Bar Chart Form........................................................................................ 222
Figure 107 X Bar Control Chart for BoltSize.txt file .................................................... 223
Figure 108 Specifications for an SPC Range Chart....................................................... 224
Figure 109 SPC Range Chart for BoltSize.txt ............................................................... 225
Figure 110 Specification Form for the SPC Sigma Chart.............................................. 226
Figure 111 S Control Chart for the BoltSize.txt Data.................................................... 227
Figure 112 CUMSUM Control Chart Specification Form............................................ 229
Figure 113 SPC CUMSUM Chart for BoltSize.txt Data. .............................................. 230
Figure 114 SPC p Control Chart Specification Form. ................................................... 231
Figure 115 SPC p Chart ................................................................................................. 232
Figure 116 Defects Conformity Chart (c Chart) Form.................................................. 233
Figure 117 SPC c Chart for Defects. ............................................................................. 234
Figure 118 Defects Per Unit SPC Form......................................................................... 235
Figure 119 Defects Per Unit SPC Chart for the file Defects.txt. ................................... 236
Figure 120 Loan Amortization Specification Form....................................................... 237
Figure 121 Future Value Form. ..................................................................................... 240
Figure 122 Payment Calculation Form. ......................................................................... 241
Figure 123 Period Payment Form. ................................................................................. 242
Figure 124 Interest Paid On A Loan Form. ................................................................... 243
Figure 125 Interest Rate Calculation Form.................................................................... 244
Figure 126 Number of Payment Periods Form. ............................................................. 245
Figure 127 Present Value Form. .................................................................................... 247
Figure 128 Double Declining Value Form. ................................................................... 248
Figure 129 Straight Line Depreciation Form................................................................. 249
Figure 130 Accelerated Depreciation Schedule Form. .................................................. 250
Figure 131 Internal Rate of Return Form. ..................................................................... 251
Figure 132 Net Present Value Form. ............................................................................. 253
Figure 133 Simulated Correlation Between Two Variables. ......................................... 255
Figure 134 Simulation of a Multivariate Distribution. .................................................. 256
Figure 135 Simulated Multivariate Data in the Grid . ................................................... 257
Figure 136 Form for Sample Size of a z-Test................................................................ 258
Figure 137 Sample Size Estimate for a z-Test............................................................... 259
Figure 138 Sample Size for a z-Test (Smaller Difference Between Means.)................ 260
Figure 139 Form for Simulation of Power Curves. ....................................................... 261
Figure 140 Power Curves for a Mean of 100 and N = 50 with Levels of Alpha........... 262
Figure 141 Simulated Distribution Cumulative Distribution Frequency....................... 263
Figure 142 Probability Distribution of a Simulated F Statistic. .................................... 264
Figure 143 Generation of Sequential Values. ................................................................ 264
Figure 144 Simulation of Scores from a Theoretical Distribution. ............................... 266
Figure 145 Form for the Degrees of Freedom in Simulation Chi-Square Values ........ 266
Figure 146 Generated Chi-Square Values in the Grid. .................................................. 267
Figure 147 Simulation Form for a Two Way Analysis of Variance.............................. 268
Figure 148 Simulated Grid Data for a Two Way Analysis of Variance. ....................... 269
Figure 149 Plot of Cell Means for the Simulated Two Way ANOVA Data. ................ 270
Figure 150 Form for Specifying a Simulated Evaluation Report. ................................. 271
Figure 151 Form for Specifying Alpha for a Simulated z Score Distribution............... 273
Figure 152 Simulated z-Score Distribution. .................................................................. 274
Figure 153 Form for Specifying Type I Error in Simulating Chi-Square Values. ........ 275
Figure 154 Simulated Chi-Square Distribution. ............................................................ 275
Figure 155 Type I Error Rate for Simulation of F Statistics.......................................... 276
Figure 156 Numerator Degrees of Freedom for Simulated F Values. ........................... 276
Figure 157 Denominator Degrees of Freedom for Simulated F Values. ....................... 276
Figure 158 Simulated F Distribution. ............................................................................ 277
Figure 159 Bootstrap Demonstration for the Weight Variable from the cansas.txt File.
................................................................................................................................. 279
Figure 160 Distribution of Resampling Estimates of the Mean. ................................... 280
Figure 161 Distribution of Means Estimated by the Bootstrap Method. ....................... 281
Figure 162 The Matrix Manipulation Procedure Main Form. ....................................... 282
Figure 163 Matrix Operations Available in MatMan. ................................................... 287
Introduction
Installing and Starting OpenStat
If you purchased the CD insert it in your computer's CD reader. With the
Windows Explorer, open the CD to examine the files on it. In the Explorer window you
should see a folder with the title PROGRAM. Double click on this folder and it should
open. In this folder you will see a file labeled "Setup.exe". When you double click this
file it should automatically begin the installation of the program. Follow any directions
and respond to any requests. Once the installation has been completed, there should be a
new entry in the start menu labeled OpenStat.
If you downloaded the free version from a web site, there are two possible setup
procedures. One uses the InstallShield software to install a Windows product and the
other uses a free setup package called Inno. Both will work. The InstallShield has the
advantage that it also contains a data directory of sample data files and registers the
software in your Windows system. The Inno setup is advantageous if you are using a
network station and want to install without modifying the server or Windows register on
your machine. The InstallShield version is downloaded as a .zip file and must be
unzipped before the Setup.exe file can be double-clicked to start the installation. The
Inno setup is an .exe file and you can simply double-click it to start the installation.
To start OpenStat, click on the Windows start button, move the mouse to the "all
programs" area and click on the OpenStat option. The first time you run the program,
there will be a "disclaimer" form displayed. When you accept the conditions of this
disclaimer, you will then be asked to select program options that are recorded
automatically in a file labeled Options.Txt. You may simply click on the OK button to
accept the default values. Once that task is completed, you should see the figure below:
Figure 1 The Main Form and Data Grid
The top row of labels (under the blue program caption) are options of the "Main Menu".
When your mouse clicks one of these menu labels a list of sub-options "drops down".
These sub-options may have additional options from which to choose. When you click
on an option, it begins a procedure. You are encouraged to explore the Main Menu items
and their lists of options.
Once you have collected data, you will be entering it into the grid that you see.
Initially, there are only two rows and columns shown in the grid. The first row (top) is an
area reserved for the label of variables. The first column is reserved for labels of cases
(observations.) When you start, there is only one cell available for entry of a variable's
value. In the following section we will describe how you can enter and edit additional
data values.
Data Characteristics and Related Analyses
Types of Variables
A variable is simply a characteristic of an object or person that we observe and
which differs in some way from object to object. For people, the color of hair may vary
or the height may vary or the weight may vary or their skill in learning may vary. All of
these may be considered "variables". You probably noticed however that some variables
are simply categories like red hair, black hair and blond hair. Others may simply indicate
an order such as top in her class, bottom in her class or middle of her class. Other
variables indicate an actual quantity such as inches or pounds. Researchers have
typically described variables like those above as belonging to one of four types: nominal,
ordinal, interval or ratio. We will describe each. It is important to know the type of your
variable because some types are only appropriate for certain analyses.
Nominal Variables
A nominal variable is one that simply describes a category or classification of the
object observed. The researcher may assign letters such as M or F for male and female or
numbers such as 1 for male and 0 for female. The code assigned for the characteristic
observed does not imply any quantity or order but simply serves as a name for the
observation. Examples of such variables are hair color, ethnic origin, political party, etc.
Ordinal Variables
Sometimes observers can be consistent in judging greater or less attributes of an
observed variable. For example, we might assign values of 1, 2, 3, etc. to runners as they
finish the crossing line in a race. The numbers we assign only indicate the order in which
they were observed and not the distance among them. Many teacher tests assign student
scores based on the number of questions correctly answered. Unless the difficulty of the
various items are known to have equal increments of increasing difficulty, the student
scores are really only an indication of order and not how much they have learned.
Interval Variables
Some measurement scales yield values that indicate a distance among objects
observed as well as their order. For example, a standardized test of intelligence may
yield scores where the distance from one value to another has some interpretation related
to the proportion of individuals in the population expected to lay between those scores.
These scales do not have a meaningful zero point indicating an absence of the attribute
but rather, an arbitrary zero point. The values of nominal variables are often expressed
on a scale of measurement where the average score is zero and the individual
observations are expressed in standardized deviations or distances from the mean.
Among these measurement scales are the z score, the T score and the Stanine score. The
research may often assume that the observed scores are distributed according to some
theoretical curve.
Ratio Variables
When a scale of measurement has a zero point indicating an absence of the
attribute and yields scores with equal increment interpretations it is called a ratio scale. A
variable with ratio scale properties is common among the physical sciences. For
example, the "absolute temperature" scale indicates the amount of heat in an object. An
object with a measure of 8 can be said to have twice as much heat as an object with a
measure of four. The ratio of measures equally distant from each other are equal.
Therefore the ratio of 8 to 4 is the same as the ratio of 16 to 8. In both cases, the
interpretation is that the first has twice the attribute as the second. Notice that we cannot
say that about intelligence test scores. That is, we cannot say that the child with a score
of 100 is not twice as intelligent as the child with a score of 50 and the child with a score
of 180 is not twice as intelligent as the child with a score of 90. And we cannot say that
the ratio of 100 to 50 is equal to the ratio of 180 to 90 in terms of an amount of
intelligence. Common ratio measures are distances in inches, meters, feet, yards, miles,
etc. Measures of speed, weight, volume, intensity, etc. produce variable observations that
are ratio measures.
Discrete and Continuous Variables
In addition to the scale of measurement used to observe an object, we need to also
be aware of whether our variable values can be considered continuous or discrete.
Clearly, variables measured on a nominal scale are discrete values, that is, numbers
representing categories with no possibility of values in between those. Ordinal values are
also typically considered discrete since the numbers represent a position of the object in
an ordering of the objects. Interval and ratio measurements usually are considered
continuous in that we can expect values at any point on a continuum. The values we
collect will be, of course, finite values but if we had infinite precision of measurement,
we could theoretically have values at any point on our scale. Therefore while we may
record weight in whole pounds, we can still consider a variable of weights to be
continuous since, with more precise measurement, we could record values between whole
pounds (or kilograms.) Most of the analyses found in OpenStat under the Analyses menu
assume a continuous variable as the "dependent" variable (with several important
exceptions.) Under the non-parametric menu, there are a variety of procedures more
typically used with discrete measures (nominal and interval data.) Discrete values will
often be coded as integers while continuous variables will be coded as floating-point
values. Most procedures will specify the type of variable coding expected. For example,
in completing an analysis of variance, the groups analyzed will be coded as integers
while the dependent variable will be coded as a floating-point value.
Recording Data
Variable Labels
The user of OpenStat will typically want to have column labels for his or her
variables other than the default Var. 1, Var. 2, etc. The label should be descriptive of the
variable such as "IQ z Score", "Hair Color", "Ability", etc. The labels should be short
enough to easily fit in the default column width of the grid (typically 8 or fewer
characters.) You can define both a short label and a longer, more descriptive label by
clicking on the Variables option of the Main Menu and clicking on the "Define" option.
The dialogue form shown below will appear in which you can enter the short and long
labels as well as indicate the kind of data you will be entering (alphabetic, integer,
floating-point, etc.)
Figure 2 The Data Dictionary Form
Entering Data in the Grid
You can click on the grid at the cell where you want to enter a value. Initially,
there is only the one cell for the first variable. Type in the value. If you then press the
down arrow, another cell will be added in that column for you to enter a value. If instead
you press the return key, you will move to the right to the next column. If a variable has
not yet been defined for that column, nothing will happen. At the bottom left corner
there is a "button" that you can click to add another variable. If you have added another
variable and press the return key in a previous variable's column, the next cell selected
will be the one in the next column. You can, of course, elect to define this variable at a
later time if you wish by selecting the "Define Variables" option in the menu. You may
also notice that characters entered from the keyboard appear in a box labeled "Cell Edit"
that lies below the menu options. If needed, you can edit a cell entry directly in this box
and press the enter key to finish the editing. You will also notice other boxes that inform
you about the location of the current row and column, the number of current cases (rows)
that have been entered, and the number of variables (columns) that have been created. A
box labeled "ASCII" actually shows the keyboard computer code for each typed
character. Also shown is a box containing the name of the last file to be saved or one that
has been loaded from the disk.
Once you have entered a number of values, you will probably already have
discovered that you can use the up, down, left and right arrow keys to navigate among the
cells. You can also use the page up, page down and home keys to make larger moves.
When you make an error in entering a value, use the backspace key to delete the incorrect
characters and then enter the correct ones.
Saving Your Data
Before you continue to enter a lot of data, you should learn to periodically save
your data in a disk file. Under the "Files" option you can click on the "Save as a text-type
file" option the first time you save your file. You need to type the name of the file and
click the save button. Notice that one can save a file in several different formats. When
you save a file as a text-type file, it will have an extension of ".TEX" added to the name.
The .BIN files are binary files and can not be easily viewed with a standard word
processing program. The text-type files can be viewed with any word processor such as
the Windows Notepad editor. These files contain the number of variables and cases, the
definition of each variable, and the values in the grid. You can also export your data to a
"bare-bones" file where the grid values are separated by tab, comma or space characters.
These may be useful for transporting your data to another program.
Editing Data
There may be occasions when you need to delete an entire row or column of the
grid data. For example, if you pressed the return key after entering a value and created a
new variable that you do not need, you can delete that column. In a similar manner, if
you pressed the down arrow key while on the last row of your data, you will have added
an extraneous row which also can be deleted. The Main Menu Edit option contains a
variety of editing functions as shown below:
Figure 3 Editing Options
Not only can you add and delete rows or columns of the grid, you can copy and paste
values from one section to another, delete individual values, re-code values of a variable,
exchange rows and columns of the grid and switch between European and US decimal
formatting. OpenStat also creates all of its output on a form which can be used as a small
word processor to add, delete, highlight, and color text entered in it. With the Edit menu,
you can open this output form, write letters, save them, print them, etc.
Generating Data
You may want to create artificial data to use in testing some of the procedures of
OpenStat or to create data that may resemble data you plan to collect in order to practice
the analyses you plan to make. The Simulation option of the Main Menu provides
several procedures to create simulated data. For example, you can create data
representing a sample from a population of normally distributed scores with a mean of
100 and standard deviation of 15 that might represent a sample of intelligence scores.
You can even generate a sample of data from a multivariate population that exhibits
common factors among the variables. It is easy to explore the simulation options. Just
click on one and follow the instructions on the dialogue that is displayed.
Importing and Exporting Data
A large number of people utilize the Microsoft Excel program to create and
analyze data. While it is an excellent program it may be limited for the type of statistical
analyses you wish to perform on data you have entered. It is possible to import an Excel
file into OpenStat. There are two methods you may employ.
Copy and Paste Method of Importing Excel Data
Start Excel and open the file you wish to transport to OpenStat. Now start
OpenStat. Select the Excel program. With the left mouse button down, drag your mouse
over the data to be copied to the OpenStat data grid. The data to be copied will usually
be highlighted in a blue color. Go to the Excel menu and click the copy option. Now
select the OpenStat program. Click the empty cell at row 1 and column 1. Go to the edit
menu and click paste. The data should now appear in your data grid. You may need to
define the variables for the data that was copied.
Importing a Tab File
First, you save your Excel file as a tab-character separated file. Excel will create
this file (including the variable labels in the first row) as a ".txt" file. Once the file is
saved, use the Windows Explorer program to find the saved file and rename it so that the
extension is ".tab" instead of ".txt". You do that by clicking once on the file name that
will cause the name of the file to be highlighted in blue. Click on this highlighted name
and it becomes a small edit box in which you can modify the name. When finished, press
the return key. Now that the name has been changed, you can import it directly into
OpenStat as tab-separated values file.
Exporting Data from OpenStat
You can export data that is loaded into the grid in a manner similar to importing
data. For example, you can drag your mouse over the data to be copied with the left
mouse button held down. The highlighted cells can now be copied to the Microsoft
clipboard by clicking the OpenStat copy option. Now select the Excel program (or other
program that can paste), select a cell to begin the pasting (top left cell of the data), and
click that programs paste option.
You can also save a OpenStat file as a tab-separated values file. Go to the File menu and
select the option to save as a tab file. Be sure the extension is .TAB. Excel and SPSS are
both programs that let you import such files.
Output and Printing
The Output Form
When you complete an analysis of your data, the results are placed on an Output
Form. As an example, we have loaded the file labeled cansas.txt and completed a
multiple regression analysis to predict the variable "chins" from measures of weight,
waist and pulse. The results are shown below:
Figure 4 The Output Form
At the top of this output form there are a series of "icons". The first one
represents a folder for opening a previously saved output file. The second one which
appears as a diskette may be clicked to save the output in a disk file as a "rich-text" file.
The third icon represents a printer. Clicking on this icon will result in the printing of the
ouput on the printer attached to your computer. The next four icons represent ways in
which you can edit the output. For example, you may drag your mouse over a section of
text and cut that text by clicking the sissors icon. That text can then be pasted to a new
position by clicking the position to begin the pasting and click the paste icon. You can
also modify selected text font and color.
You may wish to drag your mouse while holding down the left mouse button over
text that you want to paste into a Microsoft Word document. To do so, first start Word
and begin a new document or open an existing document. Next, select the text from the
output form that you wish to copy into the Word document. Notice that the color of the
text and background are reversed when you have selected the text. Now copy this
selected text to the "clipboard" by pressing the Control key and the C key at the same
time. Click on the Word document at the point where you wish to paste the copied text
and press the Control key and the V key at the same time. We have done just that to
paste the above text below:
CORRELATION MATRIX
VARIABLE
weight waist pulse chins
weight 1.000 0.870 -0.366 -0.390
waist 0.870 1.000 -0.353 -0.552
pulse -0.366 -0.353 1.000 0.151
chins -0.390 -0.552 0.151 1.000
Dependent variable: chins
Variable Beta B Std.Err. t Prob.>t VIF TOL
weight 0.368 0.079 0.089 0.886 0.389 4.189 0.239
waist -0.882 -1.456 0.683 -2.132 0.049 4.144 0.241
pulse -0.026 -0.019 0.160 -0.118 0.907 1.161 0.861
Intercept 0.000 47.968 18.285 2.623 0.018
SOURCE DF SS MS F Prob.>F
Regression 3 180.296 60.099 2.742 0.0774
Residual 16 350.654 21.916
Total 19 530.950
R2 = 0.3396, F = 2.74, D.F. = 3 16, Prob>F = 0.0774
Adjusted R2 = 0.2157
Standard Error of Estimate = 4.68
F = 2.742 with probability = 0.077
Block 1 did not meet entry requirements
Notice that the pasted text may "wrap-around" due to the output width being
greater than the width of your Word document. You can, of course, highlight this copied
text and selected a font that is smaller or further edit the output.
You can perform similar copy and paste of graphical output forms that are
displayed in OpenStat. As an example, we use the same data file cansas.txt to obtain the
frequency distribution of the chin variable. We elected to plot the normal distribution
along with the distribution of the observed data. Once we obtained the graphical image
as output, we clicked on the graphical image, pressed the alternate (alt) and prt scr (print
screen) keys at the same time, clicked on this Word document and pasted the image. The
result is shown below:
Figure 5 Copy and Paste of a Graphic Image
Testing Hypotheses
Sample statistics and Population parameters
The professional area labeled Statistics receives its name due to the predominant
interest in obtaining sample values that estimate characteristics of a population. These
population characteristics are called parameters and the estimates of them obtained from
the sample are called statistics. If we collect data on the entire population, there is no
need for statistical analysis - only the reporting of the population parameters. In most
cases however, the population is considered infinitely large or so large that it is
impractical to attempt to collect all of the population values. When data are sampled
from a population, it is important that the researcher take steps to insure that the sample
is, in fact, representative of the population. This is typically done by the method of
random sampling. Of course, if we should draw multiple random samples of the same
size, the sample statistics will vary from one another. It is the distribution of these
sample statistics that forms the basis for making inferences about the population from
which they are obtained.

Distribution of sampling errors: Type I and Type II errors
A researcher has typically collected a sample of data to support or question some
theory about the objects observed. For example, a teacher may theorize that students
performing practical laboratory experiments in a physics class will result in greater
learning gains than if the equivalent student time is spent in only observing someone else
perform the experiments. The teacher might then collect data from a sample of students
from each experimental treatment: practice and observe. The teacher would probably
make the traditional "null hypothesis" that the two methods are equally effective although
his or her prior experience is that one (practice?) treatment is superior to the other. Of
course, the teacher needs some measure of the learning achieved as well as the degree to
which the students already knew the material to be learned in order to assess the gain in
knowledge. Using carefully constructed pre and post-tests that are carefully administered
by an impartial observer, the data would be collected. The variables entered in the grid
might be labeled "GROUP", "PRETEST", and "POSTEST". The analysis performed
might be the "Analysis of Covariance" found in the Analyses menu under Analyses of
Variance.
In examining the results of the ANCOVA, the teacher has the possibility of
making one of several types of errors. If the samples obtained differ by a large amount
simply because of random sampling error and the teacher infers that the difference is
"significant" then a Type I error will be made. On the other hand, if the samples obtained
are very similar due to random sampling error when in fact the populations are really
quite different, the teacher may infer there is no difference and make a Type II error. If
we have a reasonable estimate of the distribution of sampling errors, we can actually state
the likely probability that each error might occur. Here is where we must make some
assumptions about the distribution of sampling errors. In most parametric analyses, the
errors are assumed to be normally distributed! If they are not, we are more likely to make
one of the two errors than we thought. The researcher can actually control the
probabilities of making both of these errors. In general, if our assumptions about the
distribution of errors is correct and the researcher can specify the difference between the
samples that would be meaningful, then the larger our samples are the less likely we are
to make an inferential Type II error! The simulation procedure "Power Curves for
Hypotheses" is helpful for estimating the sample size needed to control both Type I and
Type II error.
The table below summarizes these two major types of errors in hypothesis testing:
Null Hypothesis Rejected Null Hypothesis Accepted
No difference in Population Type I Error (Prob. Alpha) No error
Difference in Population No error Type II Error (Prob. Beta)
Data Description
Distribution of values
Now we are ready to start analyzing our data. Most researchers will want to
examine several aspects of the data. If the subsequent analyses assume the sample is
obtained from a normally distributed population of values, a check on this assumption is
pursued. If the data are categorical, for example, values of a "Likert-type" of
questionnaire then the researcher may want to examine the frequency of responses in
each of the categories. In the figure below you can see there are a variety of analyses
under the Descriptive option of the Analyses menu. If you have not yet collected your
data, you can generate some simulated data and examine it with one or more of these
procedures.
Figure 6 Options for Descriptive Analyses
Central tendency and variability
If you are analyzing a continuous variable, you probably will be interested in
obtaining an estimate of the mean and standard deviation of the values in that variable.
In some cases, the median and the percentile ranks will be useful information. Most
reported research of interval or ratio data provides the mean, variance and standard
deviation. In addition, the skew and kurtosis may be reported. A test of normality may
also be reported. If working with two or more continuous variables, it is helpful to do an
X versus Y type of plot and report the product-moment correlation between the variables,
that is, a measure of the co-variability of the variables.
Comparisons with theoretical distributions
There exist a wide variety of theoretical distributions that are used in statistics.
The most common ones are the Normal (Gaussian), t (Student), Chi-square, F,
rectangular, Binomial, Poisson, and Wilk's Lambda distributions. A graphic comparison
of your data's distribution with a theoretical distribution may indicate conformity or non-
conformity to the distribution you expected to obtain. As an example, we will load one
of the sample files distributed with your program. It is labeled cansas.txt. We will open
it into the grid and compare the variable "Weight" with the normal distribution using the
"descriptive" option "Compare Observed to Theoretical Distribution". The results are
shown in the figure below:
Figure 7 Comparison of Observed and Theoretical Distribution
You can see in the above plot that the observed data seem to "fit" a normal distribution
fairly well. The test for normality confirms this result.
Transformations
In some cases, the data collected needs to be transformed prior to analysis.
Highly skewed data, for example, might be transformed to "z" scores that correspond to
the proportions observed in the original data. The figure below represents the dialogue
form which appears when you select the transform option under the Main Menu
Variables menu:
Figure 8 Transformation of a Variable.
You can see that there are a wide variety of ways to create a new variable from one or
more existing variables. You select the type of transformation from the list on the right
by clicking on the desired one, enter a label for the new variable and then enter the
arguments required for that transformation. When you click the OK button, the new
variable will be created and entered in the grid.
Box Plots
Box plots are useful for obtaining a visual picture of the distribution of a
continuous variable that has been observed in two or more samples. To demonstrate, we
will load a file labeled OneWay.txt that contains data for a one way (or two way) analysis
of variance. We then select the Box Plots option from within the Descriptive Analyses
menu. Below is the form and our selections:
Figure 9 Box Plot Form
When we click the Compute button we obtain:
Box Plot of Groups
Results for group 1, mean = 105.483
Centile Value
Ten 93.096
Twenty five 97.843
Median 105.512
Seventy five 112.816
Ninety 119.389
Score Range Frequency Cum.Freq. Percentile Rank
______________ _________ _________ _______________
58.04 - 58.77 0.00 0.00 0.00
58.77 - 59.50 0.00 0.00 0.00
59.50 - 60.23 0.00 0.00 0.00
60.23 - 60.96 0.00 0.00 0.00
.
.
.
146.41 - 147.14 0.00 40.00 100.00
147.14 - 147.87 0.00 40.00 100.00
Results for group 2, mean = 99.413
Centile Value
Ten 85.062
Twenty five 90.540
Median 100.400
Seventy five 109.529
Ninety 114.763
Score Range Frequency Cum.Freq. Percentile Rank
______________ _________ _________ _______________
58.04 - 58.77 1.00 1.00 1.25
.
.
etc. for the frequency data and the plot below:
Figure 10 Sample Box Plot
The boxes show the range of scores falling in the 25
th
to 75
th
percentile (lower
and upper quartiles) and the whiskers extend down to the 10
th
percentile and up to the
90
th
percentile scores. Examining the plot immediately shows possible skew in the
distribution of scores and variability of scores in each group.
The Breakdown Procedure
Subjects may have been observed on one or more categorical variables as well as
a continuous variable. This is common in doing analyses of variance where the
categorical (nominal) variables represent treatment groups and the continuous variable
represents a criterion or dependent variable. If the research design is balanced, that is,
has equal or proportional group sizes across the treatment levels, then there is less likely a
confounding of the significance tests for one treatment variable with another. The
Breakdown Procedure lets the researcher obtain a count of the cases in each category
combination. It also will provide a one-way analysis of variance for each of the
categorical variables (e.g. treatment variables.) To demonstrate we will use the same file
(OneWay.txt) that was used for the box plots above. Select the breakdown option and
complete the form as shown below:
Figure 11 The Breakdown Procedure Form
Notice that your first move the variables of the analysis to the right box and then click the
name of the variable that is continuous. When you click the OK button the results as
shown below are obtained.
BREAKDOWN ANALYSIS PROGRAM
VARIABLE SEQUENCE FOR THE BREAKDOWN:
A_Treat. (Variable 1) Lowest level = 1 Highest level = 3
B_Treat (Variable 2) Lowest level = 1 Highest level = 2
Variable levels:
A_Treat. level = 1
B_Treat level = 1
Freq. Mean Std. Dev.
20 102.117 12.771
Variable levels:
A_Treat. level = 1
B_Treat level = 2
20 108.850 13.022
Number of observations across levels = 40
Mean across levels = 105.484
Std. Dev. across levels = 13.179
Variable levels:
A_Treat. level = 2
B_Treat level = 1
20 97.469 14.481
Variable levels:
A_Treat. level = 2
B_Treat level = 2
20 101.358 14.323
Variable levels:
A_Treat. level = 3
B_Treat level = 1
20 95.258 15.104
Variable levels:
A_Treat. level = 3
B_Treat level = 2
20 100.714 14.385
Grand number of observations across all categories = 120
Overall Mean = 100.961
Overall standard deviation = 14.389
================================================================================
ANALYSES OF VARIANCE SUMMARY TABLES
Variable levels:
A_Treat. level = 1
B_Treat level = 1
Variable levels:
A_Treat. level = 1
B_Treat level = 2
SOURCE D.F. SS MS F Prob.>F
GROUPS 1 453.25 453.25 2.725 0.1070
WITHIN 38 6320.92 166.34
TOTAL 39 6774.17
Variable levels:
A_Treat. level = 2
B_Treat level = 1
Variable levels:
A_Treat. level = 2
B_Treat level = 2
GROUPS 1 151.23 151.23 0.729 0.3985
WITHIN 38 7881.96 207.42
TOTAL 39 8033.19
Variable levels:
A_Treat. level = 3
B_Treat level = 1
Variable levels:
A_Treat. level = 3
B_Treat level = 2
GROUPS 1 297.61 297.61 1.368 0.2494
WITHIN 38 8266.37 217.54
TOTAL 39 8563.98
ANOVA FOR ALL CELLS
GROUPS 5 2170.08 434.02 2.202 0.0588
WITHIN 114 22469.24 197.10
TOTAL 119 24639.32
FINISHED
Stem and Leaf Plots
One of the older methods for visualizing the distribution of collected data is the
stem and leaf plot procedure. Using the cansas.txt data, we will examine the distribution
of the continuous variables with the Stem and Leaf procedure. Show below is the dialog
form for specifying the analysis:
Figure 12 Stem and Leaf Plot Form
When the Compute button is clicked, the results shown below are obtained:
STEM AND LEAF PLOTS
Stem and Leaf Plot for variable: weight
Frequency Stem & Leaf
1 1 3
16 1 5555666677888999
3 2 014
Stem width = 100.00, max. leaf depth = 1
Min. value = 138.000, Max. value = 247.000
No. of good cases = 20
Stem and Leaf Plot for variable: waist
9 3 123333444
10 3 5566677788
1 4 6
Stem and Leaf Plot for variable: pulse
2 4 66
8 5 00022244
4 5 6668
4 6 0224
1 6 8
1 7 4
Stem and Leaf Plot for variable: chins
5 0 12244
4 0 5668
7 1 1222334
4 1 5577
Stem and Leaf Plot for variable: situps
3 0 567
8 1 00001122
2 1 56
6 2 011123
1 2 5
Stem and Leaf Plot for variable: jumps
9 0 233334444
6 0 556678
4 1 0012
1 2 5
The procedure first determines the largest value, for example 250.0 in the jumps variable.
It then treats each observed value as a string of characters the width of the largest value
(in this example 3 digits or a stem width of 100.) Each leading character is a stem and
the leaves are the next digit from the leading digit. In our example there are 15 values (9
+ 6) that have a leading 3-digit value of 0 and 5 values (4 + 1) that have a leading digit of
1 or 2. Additional digits beyond the first two are ignored. The result is one gets a quick
view of the distribution of values using the leading two digits of the variable.
The QQ or PP Plot
The researcher sometimes needs to compare a distribution of values against one or
more theoretical distributions to determine the appropriate method of analysis. For
example, one often needs to assume that scores are approximately distributed as a normal
disttribution. The QQ or PP Plots let you compare your data to one or more theoretical
distributions. We will use the cansas.txt file to demonstrate. Shown below is the form
for specifying the variable to analyze, the theoretical distribution for comparison and the
type of plot:
Figure 13 PP or QQ Plot Form
We obtain the following for the weight variable:
Figure 14 Expected and Observed Values in the QQ Plot of weight in cansas.txt
If we had chosen the PP plot instead we would obtain:
Figure 15 The PP Plot of weight in the cansas.txt file.
Three Dimensional Rotation
It is helpful to visualize the relationship among three variables by plotting the
variable values and rotating them about each of the axis in the three dimension space. To
demonstrate, load the cansas.txt. Then load the Three Dimension Rotation procedure file
and select the first three variables. You will see the figure below:
Figure 16 Three Dimension Plot.
Notice the three vertical sliders to the right of the plot. You can drag each one up or
down to rotate the figure. For finer increments of rotation, click one of the arrows at the
end of a slider. The degree of rotation is shown in the boxes below the sliders.
Group Frequencies Analysis
Yet another way to look at a distribution of scores is by use of the Group
Frequencies Analysis procedure. Load the Defects.txt file and create a three-dimension
bar graph of the defects by selecting this procedure. Shown below is the form for
specifying your analysis:
Figure 17 Group Frequencies Analysis Form.
When you click the OK button you should obtain the following:
Figure 18 Frequency Distribution of Defects in the Defects.txt file.
If instead you select the Group Frequencies Pie Chart procedure and select the three-
dimension pie chart you obtain:
Figure 19 A Pie Chart of Defect Frequencies from the Defects.txt file.
Parametric and Non-parametric Analyses
Parametric Analyses
Parametric Analyses involve analyzing single variables or multiple variables in
order to estimate population parameters from the sample statistics. Hypotheses may
concern one or more populations and involve estimation of means, variances,
correlations, centroids, etc. Some of these analyses are described in the sections below.
Univariate Analyses
Univariate analyses typically are those that involve a single variable sampled
from one or more populations. Many of these analyses are regarding hypotheses that two
or more samples are obtained from the same population, that is, that the sample means do
not differ between or among themselves beyond what is expected due to random
sampling variability. We include in this category the following analyses:
1. The student "t" test for mean differences.
2. The one-way, two-way and three-way analyses of variance.
3. Latin square analyses of variance.
4. Mixed within and between subject analyses of variance.
5. Tests of equal proportions.
6. Tests of equal correlations.
The tests above are found under the Main Menu Analyses topic in two areas, the
univariate and the analyses of variance sub-menus.
Multivariate Analyses
Multivariate analyses involve estimation of parameters for two or more variables
in one or more populations. We would include the following analyses for this category:
1. Product-moment correlation.
2. Multiple regression analyses.
3. Cluster analyses.
4. Factor analysis.
5. Discriminate function analysis.
6. Binary logistic regression analysis.
7. Path analysis.
8. Logistic classification analyses.
9. Canonical correlation analysis.
Non-parametric Analyses
Non-parametric analyses are typically those that analyze variables that are
nominal or ordinal in nature. Parameters of the population sampled are not estimated.
Rather, the tests of hypotheses are based on the computation or estimate of the number of
ways that various orders or combinations can occur by chance alone. For example, we
may ask if a coin that is repeatedly tossed in the air to obtain the number of times each
side is shown at rest is a "fair" coin. We would hypothesize that a fair coin would yield
about 50% of the counts for each side. If we know the number of times the coin is
tossed, we can use what is called the "binomial distribution" to estimate how frequently
we would expect each count to occur by chance alone. The binomial distribution is based
on the number of combinations of values and can be computed exactly. We describe
some of the most common non-parametric analyses in the following paragraphs.
Contingency tables
One of the most frequently used analysis is called the contingency chi-square
analysis. This analysis asks whether or not two classification variables are independent
or related. For example, if people are classified into one of, say, 6 regions of the country
and also classified into one of three categories for their preferred political party, would
we expect to find the same proportion of people in each cross-classification. In this
example we could construct a 6 by 3 table and record the number of sampled subjects
belonging to each combination of the categories in the variables. The expected count in
each "cell" of the table is easily calculated by multiplying the row total times the column
total of a particular cell and dividing by the total sample size. The chi-square statistic is
obtained from the sum of squared differences between the observed and expected cell
frequencies with each difference divided by the expected value for the cell. This statistic
can then be compared to the theoretical distribution of the statistic when no relationship
exists.
If the table above contains only two categories for each variable thus producing a
2 by 2 table, the Fisher's Exact test yields more exact results. It utilizes what is called the
hyper-geometric distribution that can be calculated exactly.
Relationship among rankings
Analyses available in OpenStat that examine relationships and differences among
ordinal measures are:
1. Wilcoxon Matched-Pairs, Signed Ranks Test.
2. Kruskal-Wallace Test.
3. Friedman's Test.
4. Mann-Whitney U Test.
5. Kendall's Tau and Partial Tau.
6. Q Test.
7. Spearman Rank Correlation.
Rater agreement
Researchers often need to know the degree to which judges or subjects agree with
each other in ranking objects on some scale. For example, in athletic competition there
are often 5 judges that assign values, say 1 to 10, to the competitors. The question may
be asked, to what degree are they consistent in their judgements and is it better than
random assignment? Among the analysis in this area are:
1. Kendall's Coefficient of Concordance.
2. Kappa and Generalized Kappa.
Measurement Instruments
Reliability and validity
Research results are highly dependent on the degree to which the measures
(observations) that are made are consistent when applied to the same objects (reliable)
and actually measure what they are supposed to measure (valid.) Clearly, a meter stick
that expands and contracts with small changes in the temperature would not yield
consistent results in measuring objects. Additionally, we would not use a meter stick to
measure someone's weight. The results obtained from unreliable and invalid measures
result in greater Type II error than might be expected for more precise measures. A
variety of methods have been developed to assess reliability and validity of measurement
instruments. Many of these are routinely used in the development of educational and
psychological measures but are not necessarily restricted to those areas. Among the
analyses used are:
1. Classical test reliability analysis providing reliability indices such as the Cronbach
Alpha reliability, Kuder-Richardson formulas 20 and 21 reliability and Hoyt's
intraclass reliability index.
2. One, two or three parameter logistic scaling analysis such as the Rasch one-
parameter logistic model.
3. Composite test reliability analysis for estimating the reliability of a weighted
combination of instruments of known reliability.
4. Estimating the reliability of an instrument if the number of items in the instrument
is changed or the variability of the sample changes.
5. Estimating the scale value of items in an attitude type of instrument using
techniques such as equal appearing interval scaling or Guttman scaling.
Examples of Parametric Analyses with OpenStat
Comparing Sample Means
One of the most common questions asked in research is "Do the means of these
groups differ?" One is really asking if the samples obtained are from the same
population. The hypothesis being tested when there are only two groups is that the
population mean
1
is equal to the population mean
2
. If the difference is zero (null
difference) then one accepts the hypothesis that the two samples come from the same
population. If the hypothesis is rejected (in either direction or in a pre-specified
direction) then one infers that the samples have been obtained from populations that are
different. Often the samples obtained are those that represent subjects that have been
given treatments. For example, a physician may have two alternative treatments for a
given diagnosis. After treating randomly selected subjects from a pool of those with the
same diagnosis, the physician measures the outcome with a ratio or interval measure of
the effectiveness of the treatments. If the treatments are equally effective but differ in
cost or intrusiveness of administration, one treatment might still be favored over the
other. If one is more effective, then the question of how much more effective is it than
the alternative. Ideally, the physician researcher has specified before hand the degree of
difference that can be considered practical, has decided on the probabilities of accepting
or rejecting the hypothesis by chance random sampling error, and has determined the
sample sizes needed to insure control of these errors.
The t-test
We will illustrate the use of the t-test to test hypotheses like that described above
using a demonstration file included with your distribution of OpenStat. The name of the
file is ancova3.txt. The variables in the file are labeled row, col, slice, X, cov1 and cov2.
The X variable is the dependent variable, that is, the outcome measure of an experiment.
The row variable represents two different treatments applied to the 72 subjects where
there is 36 subjects in each treatment group. To complete the analysis, we select the "test
of two means" under the Univariate sub-menu under the Analyses menu. The form as we
will use it is shown below:
Figure 20. Specification Form for the Student t-test.
Notice that we have clicked the button for Values in the Data Grid. Also selected is the
assumption that these are independent samples, that is, that subjects were selected at
random independent of one another. We first clicked the X variable from the list of
available variables, then clicked the Row variable as the variable that represents the two
treatment groups. The two groups of interest have been coded with 1's and 2's so we
have entered those values for the Group 1 Code and Group 2 Code. Finally, we have left
the Percent Confidence Interval at 95 which means we have set our probability of a Type
I error rate at 0.05 (1.00 - 0.95).
When we click the Continue button on the form, the following results are
obtained:
COMPARISON OF TWO MEANS
Variable Mean Variance Std.Dev. S.E.Mean N
Group 1 3.50 1.63 1.28 0.21 36
Group 2 4.67 5.37 2.32 0.39 36
Assuming = variances, t = -2.646 with probability = 0.0101 and 70 degrees of freedom
Difference = -1.17 and Standard Error of difference = 0.44
Confidence interval = (-2.05, -0.29)
Assuming unequal variances, t = -2.646 with probability = 0.0106 and 54.44 degrees of
freedom
Difference = -1.17 and Standard Error of difference = 0.44
Confidence interval = (-2.05, -0.28)
F test for equal variances = 3.298, Probability = 0.0003
NOTE: t-tests are two-tailed tests.
We notice that the probability of obtaining the difference between the means of -
1.17 is approximately 0.0101 by random sampling error alone if one can assume equal
variability in the two groups. However, the test for equality of variance indicates that the
two variances are quite different (1.63 and 5.37) so we should use the more conservative
test that does not assume equal variance. This still leads to the inference that the
treatments were not equally effective and that group 2 achieved a higher mean. If a
higher mean indicated a higher level of white cells in the blood count, then we would
infer that treatment 1 was more effective in reducing the illness. We may also want to
pursue the investigation of why the second treatment might have produced wider
variability in the outcome measure. Remember, of course, random sampling Iwill give
differences some times that lead to a Type II error!
Analysis of Variance
One, Two or Three Way ANOVA
If there are more than two groups involved, then the t-test is not the best method
to employ in testing the equality of means among those groups. When multiple t-tests are
performed, the probability of chance rejection is increased beyond the nominal value
previously selected. A preferred method is the analysis of variance.
We will continue our above medical example using the same data file
(ancova3.txt.) Let us assume our physician has used two different drugs for treating an
infection that results in a heightened white cell count (our variable X.) We will also
assume that the 36 subjects randomly selected for these two groups have been further
divided into two random groups of 18 each that receive or do not receive a prescription
for an anti-inflamatory drug to be taken along with the primary anti-biotic drug. These
new groups are coded 1 and 2 in the variable labeled col. Our researcher considers the
possibility that these two drugs may, in fact, interact with one another and will be
interested in testing for "significant" interactions. To complete the analysis, we will
select the option "one, two or three way anova" from within the Analysis of Variance
sub-menu of the Analyses menu. The form used to specify our analysis is shown below:
Figure 21 Two Way Analysis of Variance Specification
Note that the dependent variable is still X but that we have now selected two factors (row
and col) as our independent variables. These factors are considered "Fixed Levels"
factors and not randomly selected levels of treatment. We have also elected to report
some "post-hoc" comparisons and to plot group means using three-dimensional vertical
bars. Our alpha levels for testing the hypotheses (Type I error rates) are left at 0.05 (95%
confidence level.) When we click the Continue button, we obtain the results shown
below:
Two Way Analysis of Variance
Variable analyzed: X
Factor A (rows) variable: Row (Fixed Levels)
Factor B (columns) variable: Col (Fixed Levels)
SOURCE D.F. SS MS F PROB.> F Omega Squared
Among Rows 1 24.500 24.500 12.250 0.001 0.083
Among Columns 1 84.500 84.500 42.250 0.000 0.304
Interaction 1 24.500 24.500 12.250 0.001 0.083
Within Groups 68 136.000 2.000
Total 71 269.500 3.796
Omega squared for combined effects = 0.470
Note: Denominator of F ratio is MSErr
Descriptive Statistics
GROUP Row Col. N MEAN VARIANCE STD.DEV.
Cell 1 1 18 3.000 1.412 1.188
Cell 1 2 18 4.000 1.412 1.188
Cell 2 1 18 3.000 2.824 1.680
Cell 2 2 18 6.333 2.353 1.534
Row 1 36 3.500 1.629 1.276
Row 2 36 4.667 5.371 2.318
Col 1 36 3.000 2.057 1.434
Col 2 36 5.167 3.229 1.797
TOTAL 72 4.083 3.796 1.948
TESTS FOR HOMOGENEITY OF VARIANCE
---------------------------------------------------------------------
Hartley Fmax test statistic = 2.00 with deg.s freem: 4 and 17.
Cochran C statistic = 0.35 with deg.s freem: 4 and 17.
Bartlett Chi-square statistic = 7.26 with 3 D.F. Prob. larger value = 0.064
---------------------------------------------------------------------
COMPARISONS AMONG COLUMNS WITHIN EACH ROW
ROW 1 COMPARISONS
----------------------------------------------------------------
Scheffe contrasts among pairs of means.
alpha selected = 0.05
Group vs Group Difference Scheffe Critical Significant?
Statistic Value
----------------------------------------------------------------
1 2 -1.00 2.12 2.074 YES
----------------------------------------------------------------
---------------------------------------------------------------
Tukey HSD Test for (Differences Between Means
Groups Difference Statistic Probability Significant?
---------------------------------------------------------------
1 - 2 -1.000 q = 3.000 0.0376 YES
---------------------------------------------------------------
---------------------------------------------------------------
Tukey B Test for (Contrasts on Ordered Means
---------------------------------------------------------------
Groups Difference Statistic d.f. Prob.>value Significant?
1 - 2 -1.000 3.000 2, 68 0.038 YES
----------------------------------------------------------------------
Neuman-Keuls Test for (Contrasts on Ordered Means
Group Mean
1 3.000
2 4.000
Groups Difference Statistic d.f. Probability Significant?
----------------------------------------------------------------------
1 - 2 -1.000 q = 3.000 2 68 0.0376 YES
----------------------------------------------------------------------
ROW 2 COMPARISONS
----------------------------------------------------------------
Statistic Value
----------------------------------------------------------------
1 2 -3.33 7.07 2.074 YES
----------------------------------------------------------------
---------------------------------------------------------------
---------------------------------------------------------------
1 - 2 -3.333 q = 10.000 0.0001 YES
---------------------------------------------------------------
---------------------------------------------------------------
---------------------------------------------------------------
1 - 2 -3.333 10.000 2, 68 0.000 YES
----------------------------------------------------------------------
Group Mean
1 3.000
2 6.333
----------------------------------------------------------------------
1 - 2 -3.333 q = 10.000 2 68 0.0001 YES
----------------------------------------------------------------------
COMPARISONS AMONG ROWS WITHIN EACH COLUMN
COLUMN 1 COMPARISONS
----------------------------------------------------------------
Statistic Value
----------------------------------------------------------------
1 2 0.00 0.00 2.074 NO
----------------------------------------------------------------
---------------------------------------------------------------
---------------------------------------------------------------
1 - 2 0.000 q = 0.000 1.0000 NO
---------------------------------------------------------------
---------------------------------------------------------------
---------------------------------------------------------------
1 - 2 0.000 0.000 2, 68 1.000 NO
----------------------------------------------------------------------
Group Mean
1 3.000
2 3.000
----------------------------------------------------------------------
1 - 2 0.000 q = 0.000 2 68 1.0000 NO
----------------------------------------------------------------------
COLUMN 2 COMPARISONS
----------------------------------------------------------------
Statistic Value
----------------------------------------------------------------
1 2 -2.33 4.95 2.074 YES
----------------------------------------------------------------
---------------------------------------------------------------
---------------------------------------------------------------
1 - 2 -2.333 q = 7.000 0.0001 YES
---------------------------------------------------------------
---------------------------------------------------------------
---------------------------------------------------------------
1 - 2 -2.333 7.000 2, 68 0.000 YES
----------------------------------------------------------------------
Group Mean
1 4.000
2 6.333
----------------------------------------------------------------------
1 - 2 -2.333 q = 7.000 2 68 0.0001 YES
----------------------------------------------------------------------
When we examine the above results we note several things. First, we find "significant"
differences among the rows (anti-biotic drugs), differences among the columns (anti-
inflammatory) and also significant interaction between the two drugs! Of course the
purist statistician would not use the term "significant" but rather state that we failed to
accept the null hypothesis!
Examining the plot of the sample means helps us to further identify the treatment
effects and their interactions. These graphs are presented below:
Figure 22 Means for Row 1 and Row 2.
Figure 23 Means for Column 1 and Column 2
Figure 24 Row and Column Interaction.
The last plot best demonstrates how the two drugs interacted. Note that within the
anti-inflammatory treatment 2 that the anti-biotic treatment 2 mean is considerably larger
than the other means. As indicated by the post-hoc tests previously presented, the means
in column 1 did not differ for the two rows but they did for the second column. It appears
that anti-biotic 1 is most effective (lowest white count) and does not interact to an
appreciable degree with the anti-inflammatory drug.
Repeated Measures ANOVAs
Several procedures involve research designs in which subjects or objects are
observed under repeated criterion measures. For example, test item responses of students
are repeated measures of the student. One might also examine reaction time of subjects
under varying levels of alcohol consumption. The question of the study is typically
something like Are the means of the measures different? We will load a file labeled
type1.txt that includes three replications of a measure for each case. When we invoke
the Within Subjects and Hoyt Reliability option under the Analysis of Variance menu
option, the following dialog form is shown:
Figure 25 A Repeated Measures Analysis of Variance Form.
Pressing the Compute button yields:
Treatments by Subjects (AxS) ANOVA Results.
-----------------------------------------------------------
SOURCE DF SS MS F Prob. > F
-----------------------------------------------------------
SUBJECTS 11 34.556 3.141
WITHIN SUBJECTS 24 30.667 1.278
TREATMENTS 2 1.722 0.861 0.655 0.530
RESIDUAL 22 28.944 1.316
-----------------------------------------------------------
TOTAL 35 65.222 1.863
-----------------------------------------------------------
TREATMENT (COLUMN) MEANS AND STANDARD DEVIATIONS
VARIABLE MEAN STD.DEV.
B1 3.167 1.528
B2 3.083 1.240
B3 3.583 1.379
In this case, the treatment variables do not appear to be different.
A repeated measures design may be combined with one or more between
subjects variables. Using the same data file as above, we will select the One Between
and One Within Analysis of variance procedure. The specifications dialog and output of
the analysis are shown below:
Figure 26 One Within and One Between Analysis of Variance Form.
ANOVA With One Between Subjects and One Within Subjects Treatments
------------------------------------------------------------------
Source df SS MS F Prob.
------------------------------------------------------------------
Between 11 34.556
Groups (A) 2 3.556 1.778 0.516 0.6135
Subjects w.g. 9 31.000 3.444
Within Subjects 24 30.667
B Treatments 2 1.722 0.861 1.632 0.2233
A X B inter. 4 19.444 4.861 9.211 0.0003
B X S w.g. 18 9.500 0.528
TOTAL 35 65.222
------------------------------------------------------------------
Means
TRT. B 1 B 2 B 3 TOTAL
A
1 2.500 3.500 4.500 3.500
2 2.500 2.250 3.750 2.833
3 4.500 3.500 2.500 3.500
TOTAL 3.167 3.083 3.583 3.278
Standard Deviations
TRT. B 1 B 2 B 3 TOTAL
A
1 1.291 1.291 1.291 1.446
2 1.291 0.957 0.957 1.193
3 1.291 1.291 1.291 1.446
TOTAL 1.528 1.240 1.379 1.365
In this analysis, it appears there may be a significant interaction between the repeated
measures and the between subjects treatment.
If we have two between subjects treatment variables and one within subjects
treatment variable, we can elect the Two Between and One Within ANOVA procedure.
This is demonstrated by loading a file labeled ABRDATA.txt. The specification for the
analysis and results are shown below:
Figure 27 Two Between and One Within Analysis of Variance Form.
SOURCE DF SS MS F PROB.
Between Subjects 11 181.000
A Effects 1 10.083 10.083 0.978 0.352
B Effects 1 8.333 8.333 0.808 0.395
AB Effects 1 80.083 80.083 7.766 0.024
Error Between 8 82.500 10.312
Within Subjects 36 1077.000
C Replications 3 991.500 330.500 152.051 0.000
AC Effects 3 8.417 2.806 1.291 0.300
BC Effects 3 12.167 4.056 1.866 0.162
ABC Effects 3 12.750 4.250 1.955 0.148
Error Within 24 52.167 2.174
Total 47 1258.000
ABR Means Table
Repeated Measures
C1 C2 C3 C4
A1 B1 17.000 12.000 8.667 4.000
A1 B2 15.333 10.000 7.000 2.333
A2 B1 16.667 10.000 6.000 2.333
A2 B2 17.000 14.000 9.333 8.333
AB Means Table
B Levels
B 1 B 2
A1 10.417 8.667
A2 8.750 12.167
AC Means Table
C Levels
C 1 C 2 C 3 C 4
A1 16.167 11.000 7.833 3.167
A2 16.833 12.000 7.667 5.333
BC Means Table
C Levels
C 1 C 2 C 3 C 4
B1 16.833 11.000 7.333 3.167
B2 16.167 12.000 8.167 5.333
Variance-Covariance Matrix for A2 B2
C Levels
C1 C2 C3 C4
C1 3.000 3.000 1.000 -0.500
C2 3.000 4.000 2.000 0.000
C3 1.000 2.000 1.333 0.333
C4 -0.500 0.000 0.333 0.333
Pooled Variance-Covariance Matrix
C Levels
C1 C2 C3 C4
C1 5.167 2.250 1.000 0.375
C2 2.250 4.000 3.500 1.750
C3 1.000 3.500 5.417 3.333
C4 0.375 1.750 3.333 2.250
Test that sample covariances are from same population:
Chi-Squared = 299.246 with 30 degrees of freedom.
Probability of Chi-Squared = 0.000
Test that variance-covariances matrix has equal variances and equal covariances:
Chi-Squared = 21.736 with 8 degrees of freedom.
Probability of Chi-Squared = 0.005
Using the file labeled tworepeated.txt, we can also demonstrate an analysis in
which there are two repeated measures variables but no between subjects treatments. In
this case, the data file consists of a subject identification integer, two integer variables
giving the levels of the two within treatments and a variable for the dependent measure.
Shown below is the dialog form for specifying the analysis using the Two Repeated
analysis of variance and the results obtained for the example analysis:
Figure 28 Two Repeated Measures ANOVA Form.
-------------------------------------------------------------------
-------------------------------------------------------------------
Factor A 1 204.167 204.167 9.853 0.052
Factor B 2 8039.083 4019.542 24.994 0.001
Subjects 3 1302.833 434.278
A x B Interaction 2 46.583 23.292 0.803 0.491
A x S Interaction 3 62.167 20.722
B x S Interaction 6 964.917 160.819
A x B x S Inter. 6 174.083 29.01
-------------------------------------------------------------------
Total 23 10793.833
-------------------------------------------------------------------
Group 1 : Mean for cell A 1 and B 1 = 17.250
Nested ANOVA Designs
In the Nested ANOVA design, one factor (B) is completely nested within levels
of another factor (A). Thus unlike the AxB Fixed Effects ANOVA in which all levels of
B are crossed with all levels of A, each level of B is found in only one level of A in the
nested design. The design may be graphically depicted as below:
_________________________________________________________
A Factor | Treatment 1 | Treatment j | Treatment M |
| | | |
B Factor | Level 1 Level 2 | ... Level k | .... Level L |
| | | |
Obser- | X
111
X
112
| .... X
1jk
| .... X
1ML
|
vations | X
211
X
212
| .... X
2jk
| .... X
2ML
|
| . . | .... . | .... . |
| . . | .... . | .... . |
| X
n11
X
n12
| .... X
njk
| .... X
nML
|
_________|_________________|______________|______________|
| _ _ | _ | _ |
B Means | X
.11
X
.12
| .... X
.jk
| .... X
.ML
|
| _ | _ | _ |
A Means | X
.1.
| X
.j.
| X
.M.
|
_________________________________________________________
We partition the total squared deviations of X scores from the grand mean of
scores into sources of variation. The independent sources may be used to form F ratios
for the hypothesis that the treatment means of A levels are equal and the hypothesis that
the treatment levels of B are equal.
Shown below is an example of a nested analysis using the file ABNested.OS4. When
you select this analysis, you see the dialog below:
Figure 29 Factor B Nested in Factor A Specification Form.
The results are shown below:
NESTED ANOVA by Bill Miller
File Analyzed: C:\Documents and Settings\Owner\My
Documents\Projects\Clanguage\OpenStat4\ABNested.OS4
CELL MEANS
A LEVEL B LEVEL MEAN STD.DEV.
1 1 2.667 1.528
1 2 3.333 1.528
1 3 4.000 1.732
2 4 3.667 1.528
2 5 4.000 1.000
2 6 5.000 1.000
3 7 3.667 1.155
3 8 5.000 1.000
3 9 6.333 0.577
A MARGIN MEANS
A LEVEL MEAN STD.DEV.
1 3.333 1.500
2 4.222 1.202
3 5.000 1.414
GRAND MEAN = 4.185
ANOVA TABLE
SOURCE D.F. SS MS F PROB.
A 2 12.519 6.259 3.841 0.041
B(A) 6 16.222 2.704 1.659 0.189
w.cells 18 29.333 1.630
Total 26 58.074
Of course, if you elect to plot the means, additional graphical output is included.
Assume that an experiment involves the use of two different teaching methods,
one which involves instruction for 1 consecutive hour and another that involves two half-
hours of instruction 4 hours apart during a given day. Three schools are randomly
selected to provide method 1 and three schools are selected to provide method 2. Note
that school is nested within method of instruction. Now assume that n subjects are
randomly selected for each of two categories of students in each school. Category 1
students are males and category 2 students are female. This design may be illustrated in
the table below:
________________________________________________________________________
| Instruction Method 1 | Instruction Method 2
------------------------------------------------------------------------------------------------------------
| School 1 | School 2 | School 3 | School 4 | School 5 | School 6
------------------------------------------------------------------------------------------------------------
Category 1 | n | n | n | n n n
Category 2 | n | n | n | n n n
________________________________________________________________________
Notice that without School, the Categories are crossed with method and therefore are
NOT nested. The expected values of the mean squares is:
________________________________________________________________________
Source of Variation df Expected Value
------------------------------------------------------------------------------------------------------------
A (Method) p - 1
2
e
+ nD
q
D
r
+ nqD
r
+ nrD
q
+ nqr
2
B within A p(q-1)
2
e
+ nD
r
2

+ nr
2
C (Category) r - 1
2
e
+ nD
q
+ nqD
p
2

+ npq
2
AC (p-1)(r-1)
2
e
+ nD
q
+ nq
2
(B within A)C p(q-1)(r-1)

2
e
+ n
2
Within Cell pqr(n-1)

2
e
________________________________________________________________________
where there are p methods of A, q nested treatments B (Schools) and r C treatments
(Categories). The D's with subscripts q, r or p have the value of 0 if the source is fixed
and a value of 1 if the source is random. In this version of the analysis, all effects are
considered fixed (D's are all zero) and therefore the F tests all use the Within Cell mean
square as the denominator. If you use random treatment levels, you may need to
calculate a more appropriate F test.
Shown below is the dialog for this ANOVA design and the results of analyzing the file
ABCNested.txt:
Figure 30 Three Treatment Design with B Treatment Nested in the A Treatment.
NESTED ANOVA by Bill Miller
File Analyzed:
CELL MEANS
A LEVEL B LEVEL C LEVEL MEAN STD.DEV.
1 1 1 2.667 1.528
1 1 2 3.333 1.155
1 2 1 3.333 1.528
1 2 2 3.667 2.082
1 3 1 4.000 1.732
1 3 2 5.000 1.732
2 4 1 3.667 1.528
2 4 2 4.667 1.528
2 5 1 4.000 1.000
2 5 2 4.667 0.577
2 6 1 5.000 1.000
2 6 2 3.000 1.000
3 7 1 3.667 1.155
3 7 2 2.667 1.155
3 8 1 5.000 1.000
3 8 2 6.000 1.000
3 9 1 6.667 1.155
3 9 2 6.333 0.577
A MARGIN MEANS
A LEVEL MEAN STD.DEV.
1 3.667 1.572
2 4.167 1.200
3 5.056 1.731
B MARGIN MEANS
B LEVEL MEAN STD.DEV.
1 3.000 1.265
2 3.500 1.643
3 4.500 1.643
4 4.167 1.472
5 4.333 0.816
6 4.000 1.414
7 3.167 1.169
8 5.500 1.049
9 6.500 0.837
C MARGIN MEANS
C LEVEL MEAN STD.DEV.
1 4.222 1.577
2 4.370 1.644
AB MEANS
A LEVEL B LEVEL MEAN STD.DEV.
1 1 3.000 1.265
1 2 3.500 1.643
1 3 4.500 1.643
2 4 4.167 1.472
2 5 4.333 0.816
2 6 4.000 1.414
3 7 3.167 1.169
3 8 5.500 1.049
3 9 6.500 0.837
AC MEANS
A LEVEL C LEVEL MEAN STD.DEV.
1 1 3.333 1.500
1 2 4.000 1.658
2 1 4.222 1.202
2 2 4.111 1.269
3 1 5.111 1.616
3 2 5.000 1.936
GRAND MEAN = 4.296
ANOVA TABLE
SOURCE D.F. SS MS F PROB.
A 2 17.815 8.907 5.203 0.010
B(A) 6 42.444 7.074 4.132 0.003
C 1 0.296 0.296 0.173 0.680
AxC 2 1.815 0.907 0.530 0.593
B(A) x C 6 11.556 1.926 1.125 0.368
w.cells 36 61.630 1.712
Total 53 135.259
Latin Square Designs ANOVA
In a typical 2 or 3-way analysis of variance design, there are independent groups
assigned to each combination of the A, B (and C) treatment levels. For example, if one is
designing an experiment with 3 levels of Factor A, 4 levels of Factor B and 2 levels of
Factor C, then a total of 24 groups of randomly selected subjects would be used in the
experiment (with random assignment of the groups to the treatment combinations.) With
only 4 observations (subjects) per group, this would require 96 subjects in total. In such a
design, one can obtain the main effects of A, B and C independent of the AxB, AxC, BxC
and AxBxC interaction effects of the treatments. Often however, one may know before
hand by previous research or by logical reasoning that the interactions should be minimal
or would not exist. When such a situation exists, one can use a design which confounds
or partially confounds such interactions with the main effects and drastically reduces the
number of treatment groups required for the analysis. If the subjects can be repeatedly
observed under various treatment conditions as in some of the previously discussed
repeated-measures designs, then one can even further reduce the number of subjects
required in the experiment. The designs to be discussed in this section utilize what are
known as Latin Squares.
The Latin Square
A Latin square is a balanced two-way classification scheme. In the following
arrangement of letters, each letter occurs just once in each row and once in each column:
A B C
B C A
C A B
If we interchange the first and second row we obtain a similar arrangement with the same
characteristics:
B C A
A B C
C A B
Two Latin squares are orthogonal if, when they are combined, the same pair of symbols
occurs no more than once in the composite squares. For example, if the two Latin
squares labeled Factor A and Factor B are combined to produce the composite shown
below those squares the combination is NOT orthogonal because treatment combinations
A1B2, A2B3, and A3B1 occur in more than one cell. However, if we combine Factor A
and Factor C we obtain a combination that IS orthogonal.
FACTOR A FACTOR B FACTOR C
A1 A2 A3 B2 B3 B1 C1 C2 C3
A2 A3 A1 B3 B1 B2 C3 C1 C2
A3 A1 A2 B1 B2 B3 C2 C3 C1
COMBINED A and B
A1B2 A2B3 A3B1
A2B3 A3B1 A1B2
A3B1 A1B2 A2B3
COMBINED A and C
A1C1 A2C2 A3C3
A2C3 A3C1 A1C2
A3C2 A1C3 A2C1
Notice that the 3 levels of treatment A and the 3 levels of treatment C are combined in
such a way that no one combination is found in more than one cell. When two Latin
squares are combined to form an orthogonal combination of the two treatment factors, the
combination is referred to as a Greco-Latin square. Notice that the number of levels of
both the treatment factors must be the same to form a square. Extensive tables of
orthogonal Latin squares have been compiled by Cochran and Cox in Experimental
Designs, New York, Wiley, 1957.
Typically, the Greco-Latin square is represented using only the number (subscripts)
combinations such as:
11 22 33
23 31 12
32 13 21
One can obtain additional squares by interchanging any two rows or columns of a Greco-
Latin square. Not all Latin squares can be combined to form a Greco-Latin square. For
example, there are no orthogonal squares for 6 by 6 or for 10 by 10 Latin squares. If the
dimensions of a Latin square can be expressed as a prime number raised to the power of
any integer n, then orthogonal squares exist. For example, orthogonal Latin squares exist
of dimension 3, 4, 5, 8 and 9 from the relationships 3 from 3
1
, 4 from 2
2
, 5 from 5
1
, 8
from 2
3
, 9 from 3
2
, etc.
Latin squares are often tabled in only standard form. A square in standard form is one
in which the letters of the first row and column are in sequence. For example, the
following is a standard form for a 4-dimension square:
A B C D
B A D C
C D B A
D C A B
There are potentially a large number of standard forms for a Latin square of dimension n.
There are 4 standard forms for a 4 by 4 square, and 9,408 standard forms for a 6 by 6
square. By interchanging rows and columns of the standard forms, one can create
additional non-standard forms. For a 4 by 4 there are a total of 576 Latin squares and for
a 6 by 6 there are a total of 812,851,200 squares! One can select at random a standard
form for his or her design and then randomly select rows and columns to interchange to
create a randomized combination of treatments.
Assume you are interested in the achievement of students under three methods of
instruction for a required course in biology (self, computer, and classroom), interested in
differences of these instruction modes for three colleges within a university (agriculture,
education, engineering) and three types of students (in-state, out-of-state, out-of-country).
We could use a completely balanced 3-way analysis of variance design with Factor A =
instructional mode, Factor B = College and Factor C = type of student. There would be
27 experimental units (samples of subjects) in this design. On the other hand we might
employ the following design:
FACTOR A (Instruction)
Self Computer Classroom
FACTOR B
(College)
Agriculture C2 C1 C3
Education C1 C3 C2
Engineering C3 C2 C1
In this design C1 is the in-state student unit, C2 is the out-of-state student unit and C3 is
the out-of-country student unit. There are only 9 units in this design as contrasted with
27 units in the completely balanced design. Note that each type of student receives each
type of instruction. Also note however that, wthin a college, students of each type do
NOT receive each type of instruction. We will have to assume that the interaction of
college and type of instruction, the interaction of college and type of student, the
interaction of type of instruction and type of student and the triple interaction of College,
instruction and student are small or do not exist. We are primarily interested in the main
effects, that is differences among student types, types of instruction and colleges on the
achievement scores obtained in the biology course. We might use Plan 1.
In his book Statistical Principles in Experimental Design, New York, McGraw-
Hill, 1962, Winer outlines a number of experimental designs that utilize Latin squares.
He refers to these designs as Plans 1 through 13 (with some variations in several plans.)
Not all plans have been included in OpenStat. Eight have been selected for inclusion at
this time. The most simple design is that which provides the following model and
estimates:
MODEL: X
ijkm
= +
i(s)
+
j(s)
+
k(s)
+ res
(s)
+
m(ijk)
Where i, j, k refer to levels of Factors A, B and C and m the individual subject in the unit.
The (s) indicates this is a model from a Latin (s)quare design.
Source of Variation Degrees of Freedom Expected Mean Square
A p 1
2
+ np
2
B p 1
2
+ np
2
C p 1
2
+ np
2
Residual (p 1)(p 2)
2
+ np
2
res
Within cell p2(n 1)
2
In the above, p is the dimension of the square and n is the number of observations per
unit.
Analyzing Relationships and Prediction Correlation and Multiple
Regression
Real-life experiences teach us that often two or more things appear to be related.
Our experience with weather tells us that as the hours of daylight vary, so does the
temperature. Our body weight varies with our age as well as our gender. Success in
college appears to be related to previous academic success as well as other factors (socio-
economic, motivation, attitudes, etc.) Researchers often wish to quantify these
relationships and, if possible, utilize the relationships observed to predict behavior or
outcomes in the future. By studying the relationships among temperature, humidity,
wind speed, and barometric pressure at one time with subsequent events such as rain, the
weather forecaster attempts to predict future weather. The admissions office of a college
attempts to predict student applicant success based on the relationship with prior success,
test scores, etc.
We will examine some examples of relationships and their use in prediction in the
sections below.
Correlation
First we will examine a relationship between college grade point average (GPA)
and a standardized test score used to predict college success (GREV.) The data we will
use is found in a demonstration file labeled PedH200.txt. Our initial analysis will be to
simply plot the relationship between these two variables and obtain the "regression line"
between them. To do this analysis, we will use the X versus Y Plot procedure in the
descriptive sub-menu. Below is the form as we will use it:
Figure 31 Form to Plot GPA versus GREV.
When we press the OK button we obtain the results below:
X versus Y Plot
X = GREV, Y = GPA from file: C:\Documents and
Settings\Owner\My
Documents\Projects\Clanguage\Stat4U\PedH200.txt
Variable Mean Variance Std.Dev.
GREV 575.33 6894.71 83.03
GPA 3.31 0.36 0.60
Correlation = 0.5815, Slope = 0.00, Intercept =
0.90
Number of good cases = 30
Figure 32 Plot of GREV with GPA Including Regression Line and Confidence.
The product-moment correlation of .58 indicates a relatively low degree of
relationship between the two variables. The square of the correlation is an estimate of the
proportion of variability (variance) of one variable "explained" by the other.
Multiple Regression and Prediction
In the above analysis, one is using only one variable's relationship with another.
We can, however, examine the relationship between one variable and a combination of
other variables. An index of this relationship is known as the coefficient of determination
or the squared multiple correlation coefficient. We will continue our example from
above and use the GREV, GREQ, MAT and AR test scores to predict GPA and obtain an
index of the degree to which the combination of the tests relate to the GPA. We will first
go to the Multiple Regression sub-menu and select one of the several multiple regression
methods to use. Initially, let us select the method described as "Block Entry Multiple
Regression". The form that appears will be completed as shown below:
Figure 33 Block Entry Multiple Regression Form.
Once you have clicked and entered the dependent variable and independent variables for
a block of analysis, you can click the Compute button. We have elected to print several
things as well as obtain predicted scores and their confidence intervals. Below are the
results:
Means
Variables GREQ GREV MAT AR GPA
565.333 575.333 67.000 3.567 3.313
Standard Deviations
Variables GREQ GREV MAT AR GPA
48.618 83.034 9.248 0.838 0.600
No. of valid cases = 30
CORRELATION MATRIX
VARIABLE
GREQ GREV MAT AR GPA
GREQ 1.000 0.468 0.267 0.508 0.611
GREV 0.468 1.000 0.426 0.405 0.581
MAT 0.267 0.426 1.000 0.525 0.604
AR 0.508 0.405 0.525 1.000 0.621
GPA 0.611 0.581 0.604 0.621 1.000
Dependent variable: GPA
GREQ 0.324 0.004 0.002 2.184 0.039 1.531 0.653
GREV 0.211 0.002 0.001 1.451 0.159 1.469 0.681
MAT 0.322 0.021 0.010 2.188 0.038 1.507 0.664
AR 0.202 0.144 0.113 1.276 0.214 1.735 0.576
Intercept 0.000 -1.738 0.951 -1.828 0.079
Regression 4 6.683 1.671 11.134 0.0000
Residual 25 3.752 0.150
Total 29 10.435
R2 = 0.6405, F = 11.13, D.F. = 4 25, Prob>F = 0.0000
F = 11.134 with probability = 0.000
Block 1 met entry requirements
Pred.z zResid. Pred.Raw RawResid.
0.0299 -0.2188 3.331 0.131
0.8334 0.4780 3.813 -0.287
-0.8705 0.3481 2.791 -0.209
-0.8063 -0.3829 2.830 0.230
-0.2322 0.8768 3.174 -0.526
0.6019 0.5429 3.674 -0.326
1.3256 0.3192 4.108 -0.192
-0.4844 -0.5381 3.023 0.323
0.4664 0.0115 3.593 -0.007
0.4610 0.8505 3.590 -0.510
-0.8652 -0.1573 2.794 0.094
-0.9726 0.2835 2.730 -0.170
-0.9730 -0.3829 2.730 0.230
0.1951 -0.7174 3.430 0.430
1.2908 -1.3130 4.088 0.788
0.0299 -0.2188 3.331 0.131
0.8334 0.4780 3.813 -0.287
-0.8705 0.3481 2.791 -0.209
-0.8063 -0.3829 2.830 0.230
-0.2322 0.8768 3.174 -0.526
0.6019 0.5429 3.674 -0.326
1.3256 0.3192 4.108 -0.192
-0.4844 -0.5381 3.023 0.323
0.4664 0.0115 3.593 -0.007
0.4610 0.8505 3.590 -0.510
-0.8652 -0.1573 2.794 0.094
-0.9726 0.2835 2.730 -0.170
-0.9730 -0.3829 2.730 0.230
0.1951 -0.7174 3.430 0.430
1.2908 -1.3130 4.088 0.788
In examining the above output we find a square multiple correlation that indicates nearly
60% of the GPA is "explained" by the weighted combination of test scores. An F test is
provided indicating the chance of the F value occurring by random sampling error is less
than 0.00001. We can write a model for the prediction of GPA scores as
GPA = 0.004 GREQ + 0.002 GREV + 0.021 MAT + 0.144 AR - 1.738
When the above equation is applied to the data of this sample, the predicted scores (both
raw and standardized) are obtained as well as the residual (error of prediction) for each
case. Note that in some cases, a fairly large error can occur, some over a half a grade
point!
Multiple regression analysis is a method for examining the relationship between
one continuous variable of interest (the dependent or criterion variable) and one or more
independent (predictor) variables. Typically we assume a linear relationship of the type
Y
i
= B
1
X
i1
+ B
2
X
i2
+ ... + B
k
X
ik
+ B
0
+ E
i
(1)
where
Y
i
is the score obtained for individual i on the dependent variable,
X
i1
... X
ik
are scores obtained on k independent variables,
B
1
... B
k
are weights (regression coefficients) of the k independent variables which
maximize the relationship with the Y scores,
B
0
is a constant (intercept) and E
i
is the error for individual i.
In the above equation, the error score E
i
reflects the difference between the
subject's actual score Yi and the score which is predicted on the basis of the weighted
combination of the independent variables. That is,

Y'
i
- Y
i
= E
i
.
where Y'
i
is predicted from
Y'i = B
1
X
i1
+ B
2
X
i2
+ ... + B
k
X
ik
+ B
0
(2)
In addition to assuming the above general linear model relating the Y scores to the
X scores, we usually assume that the

E
i
scores are normally distributed.
When we complete a multiple regression analysis, we typically draw a sample
from a population of subjects and observe the Y and X scores for the subjects of that
sample. We use that sample data to estimate the weights (B's) that will permit us the
"best" prediction of Y scores for other individuals in the population for which we only
have knowledge of their X scores.
Using Multiple Regression for Analysis of Variance
Students in statistics courses are often surprised to learn that the analyses of
variance they learned in a previous course can be performed using multiple regression
procedures. In fact, both ANOVA and MR use a linear equation to express the variability
of a dependent variable that can be explained by a weighted combination of independent
variables. The underlying assumptions are basically the same! To demonstrate this we
will return to the physician's research that was used in the analysis of variance sections
earlier. We will again load the file labeled ancova3.txt. As before, the row codes 1 and 2
correspond to two anti-biotic drug treatments, the column codes 1 and 2 correspond to
two anti-inflammatory drugs and X corresponds to a blood white cell count. In this
example however we will add a third treatment factor labeled "slice" which corresponds
to three dosage levels (once, twice or three times per day.) In addition, we will be using
two "covariates" that correspond to patient information regarding the number of family
members in the household (cov1) and number of months since last appointment. The use
of covariates can sometimes reduce the error of our analysis. Now select the ANCOVA
procedure from the Analysis of Variance sub-menu. The form should be completed as
shown below:
Figure 34 Analysis of Covariance Procedure Form.
When the Compute button is clicked, you will obtain the following output:
Part 1. Full Model ==========================================
ANALYSIS OF COVARIANCE USING MULTIPLE REGRESSION
File Analyzed: C:\Documents and Settings\Owner\My
Documents\Projects\Clanguage\Stat4U\ANCOVA3.TXT
Model for Testing Assumption of Zero Interactions with Covariates
Dependent variable: X
Cov1 -0.349 -0.406 0.049 -8.203 0.000 7.276 0.137
Cov2 0.205 0.247 0.069 3.574 0.001 15.273 0.065
A1 1.064 2.059 0.446 4.621 0.000 224.656 0.004
B1 -1.721 -3.330 0.446 -7.472 0.000 226.187 0.004
C1 -0.835 -1.978 0.511 -3.870 0.000 199.725 0.005
C2 0.246 0.584 0.648 0.900 0.374 317.678 0.003
A1xB1 -0.133 -0.258 0.446 -0.580 0.566 222.911 0.004
A1xC1 -0.078 -0.184 0.511 -0.360 0.721 197.526 0.005
A1xC2 -0.514 -1.218 0.648 -1.878 0.068 315.305 0.003
B1xC1 -0.195 -0.462 0.511 -0.904 0.372 198.094 0.005
B1xC2 0.494 1.171 0.648 1.806 0.079 320.527 0.003
A1xB1xC1 -0.577 -1.367 0.511 -2.674 0.011 195.836 0.005
A1xB1xC2 0.645 1.528 0.648 2.357 0.024 318.199 0.003
Cov1xA1 -0.275 -0.124 0.049 -2.509 0.017 50.419 0.020
Cov1xB1 1.198 0.542 0.049 10.963 0.000 46.621 0.021
Cov1xC1 1.272 0.739 0.066 11.197 0.000 52.138 0.019
Cov1xC2 -0.129 -0.072 0.072 -1.006 0.321 69.654 0.014
Cov1xA1xB1 -0.160 -0.073 0.049 -1.467 0.151 49.392 0.020
Cov1xA1xC1 -0.504 -0.292 0.066 -4.430 0.000 54.581 0.018
Cov1xA1xC2 -0.763 -0.425 0.072 -5.924 0.000 74.016 0.014
Cov1xB1xC1 -0.359 -0.209 0.066 -3.168 0.003 53.142 0.019
Cov1xB1xC2 -0.614 -0.342 0.072 -4.763 0.000 64.597 0.015
Cov1xA1xB1xC1 0.268 0.156 0.066 2.362 0.024 55.332 0.018
Cov1xA1xB1xC2 0.619 0.344 0.072 4.796 0.000 68.478 0.015
Cov2xA1 -0.968 -0.444 0.069 -6.422 0.000 94.979 0.011
Cov2xB1 0.273 0.126 0.069 1.822 0.077 95.088 0.011
Cov2xC1 -0.529 -0.309 0.077 -4.040 0.000 78.584 0.013
Cov2xC2 0.090 0.049 0.098 0.498 0.622 138.534 0.007
Cov2xA1xB1 0.570 0.261 0.069 3.775 0.001 87.249 0.011
Cov2xA1xC1 0.928 0.548 0.077 7.153 0.000 72.137 0.014
Cov2xA1xC2 0.833 0.454 0.098 4.630 0.000 131.178 0.008
Cov2xB1xC1 0.319 0.187 0.077 2.438 0.020 73.617 0.014
Cov2xB1xC2 0.247 0.134 0.098 1.369 0.180 144.204 0.007
Cov2xA1xB1xC1 0.373 0.218 0.077 2.852 0.007 67.766 0.015
Cov2xA1xB1xC2 -1.318 -0.715 0.098 -7.300 0.000 137.926 0.007
Intercept 0.000 4.686 0.446 10.516 0.000
Regression 35 267.214 7.635 120.246 0.0000
Residual 36 2.286 0.063
Total 71 269.500
R2 = 0.9915, F = 120.25, D.F. = 35 36, Prob>F = 0.0000
Part 2. Covariance Model ======================================
Model for Analysis of Covariance
Dependent variable: X
Cov1 -0.083 -0.096 0.076 -1.265 0.211 1.445 0.692
Cov2 -0.067 -0.081 0.082 -0.992 0.326 1.563 0.640
A1 -0.323 -0.624 0.109 -5.711 0.000 1.080 0.926
B1 -0.571 -1.104 0.116 -9.504 0.000 1.220 0.820
C1 -0.249 -0.590 0.150 -3.927 0.000 1.361 0.735
C2 0.206 0.489 0.156 3.128 0.003 1.473 0.679
A1xB1 0.287 0.554 0.107 5.164 0.000 1.042 0.960
A1xC1 0.225 0.532 0.155 3.429 0.001 1.453 0.688
A1xC2 -0.195 -0.463 0.152 -3.046 0.003 1.394 0.717
B1xC1 -0.201 -0.477 0.158 -3.024 0.004 1.497 0.668
B1xC2 0.265 0.629 0.153 4.109 0.000 1.411 0.709
A1xB1xC1 -0.264 -0.626 0.153 -4.096 0.000 1.410 0.709
A1xB1xC2 -0.229 -0.542 0.152 -3.578 0.001 1.384 0.723
Intercept 0.000 4.780 0.482 9.924 0.000
Regression 13 223.302 17.177 21.565 0.0000
Residual 58 46.198 0.797
Total 71 269.500
R2 = 0.8286, F = 21.57, D.F. = 13 58, Prob>F = 0.0000
Part 3. Difference Between Full and Covariance Model: Test for Equal Slopes
=========================================================
Test for Homogeneity of Group Regression Coefficients
Change in R2 = 0.1629. F = 31.437 Prob.> F = 0.0000 with d.f. 22 and 36
Part 4. Unadjusted Means and Means Adjusted for Covariates ============
Unadjusted Group Means for Group Variables Row
Means
Variables Group 1 Group 2
3.500 4.667
Intercepts for Each Group Regression Equation for Variable: Row
Intercepts
4.156 5.404
Adjusted Group Means for Group Variables Row
Means
3.459 4.707
Unadjusted Group Means for Group Variables Col
Means
3.000 5.167
Intercepts for Each Group Regression Equation for Variable: Col
Intercepts
4.156 5.404
Adjusted Group Means for Group Variables Col
Means
2.979 5.187
Unadjusted Group Means for Group Variables Slice
Means
Variables Group 1 Group 2 Group 3
3.500 4.500 4.250
Intercepts for Each Group Regression Equation for Variable: Slice
Intercepts
4.156 3.676 6.508
Adjusted Group Means for Group Variables Slice
Means
3.493 4.572 4.185
Part 5. Comparisons Among Means (2 x 2 x 3 = 12 Means) =============
Multiple Comparisons Among Group Means
Comparison of Group 1 with Group 2
F = 8.243, probability = 0.006 with degrees of freedom 1 and 58
.
.
.
Part 6. Total Variance Partitioned by Source =====================
Test for Each Source of Variance Obtained by Eliminating
from the Regression Model for ANCOVA the Vectors Associated
with Each Fixed Effect.
----------------------------------------------------------------------
SOURCE Deg.F. SS MS F Prob>F
----------------------------------------------------------------------
Cov1 1 1.27 1.27 1.600 0.2109
Cov2 1 0.78 0.78 0.983 0.3255
A 1 25.98 25.98 32.620 0.0000
B 1 71.95 71.95 90.335 0.0000
C 2 13.87 6.94 8.709 0.0005
AxB 1 21.24 21.24 26.666 0.0000
AxC 2 11.89 5.95 7.466 0.0013
BxC 2 14.20 7.10 8.916 0.0004
AxBxC 2 47.25 23.62 29.659 0.0000
----------------------------------------------------------------------
ERROR 58 46.20 0.80
----------------------------------------------------------------------
TOTAL 71 269.50
----------------------------------------------------------------------
----------------------------------------------------------------------
ANALYSIS FOR COVARIATES ONLY
Covariates 2 6.99 3.49 0.918 0.4041
----------------------------------------------------------------------
While the above output is quite voluminous, it provides considerable information. There
are basically six parts to the output.
Part 1.
First, the procedure tests assumptions about the model for ANCOVA itself. It is
assumed that the correlation of the covariates with the dependent variable is essentially
the same within each treatment group. A regression model is created that contains
treatments, covariates and interaction of treatments with covariates.
Part 2.
This "Full" model is then contrasted with a model containing just the treatment
and covariates.
Part 3.
If the difference in the squared multiple correlation coefficient is significant, this
indicates a violation of the assumption. The rest of the analysis provides the analysis of
covariance results.
Part 4.
The treatment group means can be "adjusted" for the effects of the covariates. In
this part, the dependent variable for the individual cases within each treatment group is
predicted using the covariates as predictors. The mean of those predicted scores are the
adjusted means. In analysis of covariance, we are essentially completing an analysis of
variance of these predicted scores, that is, scores adjusted for the effects of the covariates.
Part 5.
In this particular example, there are two row treatment levels, two column
treatment levels and three slice treatment levels which results in a total of 12 treatment
groups in all. Individual comparisons can be made among these treatment groups using
an F statistic.
Part 6.
The total variance of the adjusted dependent variable scores can be "partitioned"
into portions attributed to row, column, slice, row and column interactions, row and slice
interactions, column and slice interactions and row, column and slice interactions. Each
portion can be tested for significance and the results displayed in this part.
When the user returns to the Main Form, it will be apparent that a number of new
variables have been entered in the grid. The values in each of these new columns
represents a coding of the treatment groups and the product of the treatment groups
(interactions.) A method called "effect" coding is used to generate variables representing
the treatment groups. Using this method, one creates independent variables that represent
the treatments and their interactions that can be used in a multiple regression model.
Don't worry! We haven't changed the original disk file! Of course, if you want to save
this modified grid of data, you can do so. We suggest you should save it under a different
file name if you do save it.
Regression with a Dichotomous Dependent Variable
It is not unusual that a researcher desires to predict a variable that has only two
possible values, that is, a dichotomous criterion. For example, one might want to predict
success or failure for a task. A physician might want to predict life or death at some
point following diagnosis of an illness. Often, the independent variables (predictors) are
continuous variables. For example, one might predict college dropout versus non-
dropout using standardized test scores, high school grade point average, etc..
The method used to perform regression analysis with a dichotomous dependent
variable is called "logistic" regression. This is based on the transformation of the 0 and 1
codes used for the dependent variable into logarithms of the proportion of 0's and 1's.
To illustrate this procedure, we will load a file labeled BINARYREG.txt. In this
file you will see three variables labeled Var1, var2 and Y. The Y variable is the
dependent variable. You can see that it contains either a 0 or a 1 for each case. The
values for the predictors range from 1.0 to 10.0. To complete the analysis, you select the
"binary logistic regression" procedure from within the Multiple Regression sub-menu.
The following form appears on which you specify your analysis:
Figure 35 Binary Logistic Regression Form.
As for previous analyses, you select the appropriate variables from the list of available
variables and enter them by clicking a right-arrow. You will notice that for this
procedure, that an iterative process is used to obtain the results. In most cases, the final
solution will be obtained in less that 20 iterations. If necessary however, you can
increase the maximum iterations required to obtain a stable solution. We have elected to
show the iterations as well as obtain descriptive statistics and probabilities for the
dependent variable. When you click the OK button, you obtain the following results:
Logistic Regression Adapted from John C. Pezzullo
Java program at http://members.aol.com/johnp71/logistic.html
6 cases have Y=0; 4 cases have Y=1.
Variable Label Average Std.Dev.
1 VAR1 5.5000 2.8723
2 var2 5.5000 2.8723
Iteration History
-2 Log Likelihood = 13.4602 (Null Model)
-2 Log Likelihood = 8.7491
Converged
Overall Model Fit... Chi Square = 5.1302 with df = 2 and prob. = 0.0769
Coefficients and Standard Errors...
Variable Label Coeff. StdErr p
1 VAR1 0.3498 0.6737 0.6036
2 var2 0.3628 0.6801 0.5937
Intercept -4.6669
Odds Ratios and 95% Confidence Intervals...
Variable O.R. Low -- High
VAR1 1.4187 0.3788 5.3135
var2 1.4373 0.3790 5.4506
X X Y Prob
1.0000 2.0000 0 0.0268
2.0000 1.0000 0 0.0265
3.0000 5.0000 0 0.1414
4.0000 3.0000 0 0.1016
5.0000 4.0000 1 0.1874
6.0000 7.0000 0 0.4929
7.0000 8.0000 1 0.6646
8.0000 6.0000 0 0.5764
9.0000 10.0000 1 0.8918
10.0000 9.0000 1 0.8905
Classification Table
Predicted
---------------
Observed 0 1 Total
---------------
0 | 5 | 1 | 6 |
1 | 1 | 3 | 4 |
---------------
Total | 6 | 4 | 10
---------------
In the above analysis it took 6 iterations to obtain a stable solution. The regression model
of Y = 0.3498 Var1 + 0.3628 var2 - 4.6669 was not significant at the 0.05 level. Of
course, if one is will to accept a higher risk of Type I error, you might consider it
significant. By and large the sample is just too small to draw any definitive conclusions.
Examining the classification table (8 correctly predicted, 2 incorrectly predicted cases) it
appears that the model does a fairly good job. You have to remember that by randomly
selecting subjects from a population in which there is no actual relationship between the
dependent and independent variables, you might still get these results fairly often for such
a small sample.
Simultaneous Multiple Regression
The simultaneous multiple regression procedure obtains the regression of each
variable in a set of variables on the remaining variables. If the set of variables happen to
be a set being considered for use as independent variables in another analysis, the
simultaneous multiple regression procedure may help identify dependencies among the
variables that can produce serious co-linearity problems in a subsequent analysis.
Using the cansas.txt data, we can obtain equations for each variable. The dialog
form for this analysis is shown below:
Figure 36 Form for Specifying a Simultaneous Multiple Regression Analysis
The output from a sample analysis is shown below:
Simultaneous Multiple Regression by Bill Miller
Product-Moment Correlations Matrix with 20 cases.
Variables
weight waist pulse chins situps
weight 1.000 0.870 -0.366 -0.390 -0.493
waist 0.870 1.000 -0.353 -0.552 -0.646
pulse -0.366 -0.353 1.000 0.151 0.225
chins -0.390 -0.552 0.151 1.000 0.696
situps -0.493 -0.646 0.225 0.696 1.000
jumps -0.226 -0.191 0.035 0.496 0.669
Variables
jumps
weight -0.226
waist -0.191
pulse 0.035
chins 0.496
situps 0.669
jumps 1.000
Means with 20 valid cases.
Variables weight waist pulse chins situps
178.600 35.400 56.100 9.450 145.550
Variables jumps
70.300
Standard Deviations with 20 valid cases.
24.691 3.202 7.210 5.286 62.567
Variables jumps
51.277
Determinant of correlation matrix = 0.0208
Multiple Correlation Coefficients for Each Variable
Variable R R2 F Prob.>F DF1 DF2
weight 0.902 0.814 12.249 0.000 5 14
waist 0.939 0.882 21.017 0.000 5 14
pulse 0.386 0.149 0.490 0.778 5 14
chins 0.734 0.539 3.275 0.036 5 14
situps 0.884 0.782 10.026 0.000 5 14
jumps 0.798 0.636 4.901 0.008 5 14
Betas in Columns with 20 cases.
Variables
weight -1.000 0.676 -0.321 0.347 0.372
waist 1.070 -1.000 0.004 -0.616 -0.771
pulse -0.070 0.000 -1.000 -0.017 0.049
chins 0.140 -0.157 -0.031 -1.000 0.143
situps 0.317 -0.415 0.191 0.303 -1.000
jumps -0.301 0.317 -0.149 0.254 0.533
Variables
jumps
weight -0.588
waist 0.982
pulse -0.064
chins 0.201
situps 0.888
jumps -1.000
Standard Errors of Prediction
Variable Std.Error
weight 12.407
waist 1.279
pulse 7.749
chins 4.181
situps 34.056
jumps 36.020
Raw Regression Coefficients with 20 cases.
Variables
weight -1.000 0.088 -0.094 0.074 0.944
waist 8.252 -1.000 0.008 -1.017 -15.069
pulse -0.240 0.000 -1.000 -0.012 0.424
chins 0.655 -0.095 -0.042 -1.000 1.697
situps 0.125 -0.021 0.022 0.026 -1.000
jumps -0.145 0.020 -0.021 0.026 0.650
Variables
jumps
weight -1.221
waist 15.718
pulse -0.453
chins 1.947
situps 0.728
jumps -1.000
Variable Constant
weight -114.302
waist 22.326
pulse 71.223
chins 27.313
situps 424.896
jumps -366.967
Partial Correlations with 20 cases.
Variables
weight -1.000 0.851 -0.150 0.221 0.344
waist 0.851 -1.000 0.001 -0.311 -0.566
pulse -0.150 0.001 -1.000 -0.023 0.097
chins 0.221 -0.311 -0.023 -1.000 0.208
situps 0.344 -0.566 0.097 0.208 -1.000
jumps -0.420 0.558 -0.097 0.226 0.688
Variables
jumps
weight -0.420
waist 0.558
pulse -0.097
chins 0.226
situps 0.688
jumps -1.000
Notice that the beta and B weights (standardized and raw regression coefficients)
are reported in columns. A minus one (-1) is shown in the position of the dependent
variable (the one being predicted.) The last matrix shows the partial correlation between
the dependent and each independent variable with the remaining independent variables
partialled out. These results are often used to determine a variable which may be
eliminated due to its small contribution to the prediction beyond the other independent
variables. For example, if one were predicting waist, it appears that pulse contributes
little unique variance prediction beyond the remaining variables.
Cox Proportional Hazards Survival Regression
This program, originally written in the Java language by John Pezzullo, analyzes
survival-time data by the method of Proportional Hazards regression (Cox). Given
survival times, final status (alive or dead) , and one or more covariates, it produces a
baseline survival curve, covariate coefficient estimates with their standard errors, risk
ratios, 95% confidence intervals, and significance levels.
Survival analysis takes the survival times of a group of subjects (usually with
some kind of medical condition) and generates a survival curve, which shows how many
of the members remain alive over time. Survival time is usually defined as the length of
the interval between diagnosis and death, although other "start" events (such as surgery
instead of diagnosis), and other "end" events (such as recurrence instead of death) are
sometimes used.
The major mathematical complication with survival analysis is that you usually
do not have the luxury of waiting until the very last subject has died of old age; you
normally have to analyze the data while some subjects are still alive. Also, some subjects
may have moved away, and may be lost to follow-up. In both cases, the subjects were
known to have survived for some amount of time (up until the time you last saw them),
but you don't know how much longer they might ultimately have survived. Several
methods have been developed for using this "at least this long" information to preparing
unbiased survival curve estimates, the most common being the Life Table method and the
method of Kaplan and Meier.
We often need to know whether survival is influenced by one or more factors,
called "predictors" or "covariates", which may be categorical (such as the kind of
treatment a patient received) or continuous (such as the patient's age, weight, or the
dosage of a drug). For simple situations involving a single factor with just two values
(such as drug vs placebo), there are methods for comparing the survival curves for the
two groups of subjects. But for more complicated situations we need a special kind of
regression that lets us assess the effect of each predictor on the shape of the survival
curve.
To understand the method of proportional hazards, first consider a "baseline"
survival curve. This can be thought of as the survival curve of a hypothetical "completely
average" subject -- someone for whom each predictor variable is equal to the average
value of that variable for the entire set of subjects in the study. This baseline survival
curve doesn't have to have any particular formula representation; it can have any shape
whatever, as long as it starts at 1.0 at time 0 and descends steadily with increasing
survival time.
The baseline survival curve is then systematically "flexed" up or down by each of
the predictor variables, while still keeping its general shape. The proportional hazards
method computes a coefficient for each predictor variable that indicates the direction and
degree of flexing that the predictor has on the survival curve. Zero means that a variable
has no effect on the curve -- it is not a predictor at all; a positive variable indicates that
larger values of the variable are associated with greater mortality. Knowing these
coefficients, we could construct a "customized" survival curve for any particular
combination of predictor values. More importantly, the method provides a measure of the
sampling error associated with each predictor's coefficient. This lets us assess which
variables' coefficients are significantly different from zero; that is: which variables are
significantly related to survival.
The log-likelihood function is minimized by Newton's method, with a very simple
elimination algorithm to invert and solve the simultaneous equations. Central-limit
estimates of parameter standard errors are obtained from the diagonal terms of the inverse
matrix. 95% confidence intervals around the parameter estimates are obtained by a
normal approximation. Risk ratios (and their confidence limits) are computed as
exponential functions of the parameters (and their confidence limits). The baseline
survival function is generated for each time point at which an event (death) occurred.
No special convergence-acceleration techniques are used. For improved precision,
the independent variables are temporarily converted to "standard scores" ( value - Mean )
/ StdDev. The Null Model (all parameters = 0 )is used as the starting guess for the
iterations. Convergence is not guaranteed, but this method should work properly with
most real-world data.
There are no predefined limits to the number of variables or cases this procedure
can handle. The actual limits are probably dependent on your computer's available
memory.
To demonstrate, a file COXREG.txt has been loaded that contains the following
information:
VAR1 TIME STATUS
50.00 1.00 0.00
70.00 2.00 1.00
45.00 3.00 0.00
35.00 5.00 0.00
62.00 7.00 1.00
50.00 11.00 0.00
45.00 4.00 0.00
57.00 6.00 0.00
32.00 8.00 0.00
57.00 9.00 1.00
60.00 10.00 1.00
The specification form for this analysis is shown below with variables entered for
a sample file:
Figure 37 Cox Regression Specification Form.
Results for the above sample are as follows:
Cox Proportional Hazards Survival Regression Adapted from John C. Pezzullo's Java program
at http://members.aol.com/johnp71/prophaz.html
Variable Label Average Std.Dev.
1 VAR1 51.1818 10.9778
Iteration History...
-2 Log Likelihood = 11.4076 (Null Model)
Converged
Overall Model Fit...
Chi Square = 7.3570 with d.f. 1 and probability = 0.0067
Coefficients, Std Errs, Signif, and Confidence Intervals
Var Coeff. StdErr p Lo95% Hi95%
VAR1 0.3770 0.2542 0.1379 -0.1211 0.8752
Risk Ratios and Confidence Intervals
Variable Risk Ratio Lo95% Hi95%
VAR1 1.4580 0.8859 2.3993
Baseline Survivor Function (at predictor means)...
2.0000 0.9979
7.0000 0.9820
9.0000 0.9525
10.0000 0.8310
Non-Linear Regression
(Contributed From John Pezzullo's Non-Linear Regression page at
http://members.aol.com/johnp71/nonlin.html )
Background Info (just what is nonlinear curve-fitting, anyway?):
Simple linear curve fitting deals with functions that are linear in the parameters, even
though they may be nonlinear in the variables. For example, a parabola
y=a+b*x+c*x*x is a nonlinear function of x (because of the x-squared term), but fitting
a parabola to a set of data is a relatively simple linear curve-fitting problem because the
parameters enter into the formula as simple multipliers of terms that are added together.
Another example of a linear curve-fitting problem is y= a+b*Log(x)+c/x; the terms
involve nonlinear functions of the independent variable x, but the parameters enter into
the formula in a simple, linear way.
Unfortunately, many functions that arise in real world situations are nonlinear in
the parameters, like the curve for exponential decay y=a*Exp(-b*x), where b is
"wrapped up" inside the exponential function. Some nonlinear functions can be linearized
by transforming the independent and/or dependent variables. The exponential decay
curve, for example, can be linearized by taking logarithms: Log(y)=a'-b*x . The a'
parameter in this new equation is the logarithm of a in the original equation, so once a'
has been determined by a simple linear curve-fit, we can just take its antilog to get a.
But we often encounter functions that cannot be linearized by any such tricks, a simple
example being exponential decay that levels off to some unknown value: y=a*Exp(-
b*x)+c. Applying a logarithmic transformation in this case produces Log(y-c)=a'-b*x.
This linearizes b, but now c appears inside the logarithm; either way, we're stuck with an
intrinsically nonlinear parameter estimation problem, which is considerably more
difficult than linear curve-fitting. That's the situation this web page was designed to
handle.
For a more in-depth treatment of this topic, check out Dr. Harvey Motulsky's new web
site: Curvefit.com -- a complete guide to nonlinear regression. Most of the information
here is excerpted from Analyzing Data with GraphPad Prism, a book that accompanies
the program GraphPad Prism.
This procedure involves expanding the function to be fitted in a Taylor series
around current estimates of the parameters, retaining first-order (linear) terms, and
solving the resulting linear system for incremental changes to the parameters. The
program computes finite-difference approximations to the required partial derivatives,
then uses a simple elimination algorithm to invert and solve the simultaneous equations.
Central-limit estimates of parameter standard errors are obtained from the diagonal terms
of the inverse of the normal equations matrix. The covariance matrix is computed by
multiplying each term of the inverse normal matrix by the weighted error-variance. It is
used to estimate parameter error correlations and to compute confidence bands around the
fitted curve. These show the uncertainty in the fitted curve arising from sampling errors
in the estimated parameters, and do not include the effects of errors in the independent
and dependent variables. The page also computes a generalized correlation coefficient,
defined as the square root of the fraction of total y variance explainable by the fitted
function.
Unequal weighting is accomplished by specifying the standard error associated
with the y variable. Constant errors, proportional errors, or Poisson (square root) errors
can be specified by a menu, and don't have to be entered with the data. Standard errors
can also be entered along with the x and y variables. Finally, replicate y measurements
can be entered; the program will compute the average and standard error of the mean.
Also available are a number of simple variable transformations (log, reciprocal, square
root), which might simplify the function to be fitted and improve the convergence,
stability and precision of the iterative algorithm. If a transformation is applied to the y
variable, the program will adjust the weights appropriately.
The procedure also fits least-absolute-value curves by applying an iterative re-
weighting scheme by which each point is given a standard error equal to the distance of
that point from the fitted curve. An option allows differential weighting of above-curve
points vs. below-curve points to achieve a specified split of points above and below the
curve (a percentile curve fit).
No special goal-seeking methods, precision-preserving techniques (such as
pivoting), convergence-acceleration, or iteration-stabilizing techniques (other than a
simple, user-specified fractional adjustment), are used. This method may not succeed
with extremely ill-conditioned systems, but it should work with most practical problems
that arise in real-world situations.
As an example, I have created a "parabola" function data set labeled parabola.txt. To
generate this file I used the equation y = a + b * x + c * x * x. I let a = 0, b = 5 and c = 2
for the parameters and used a sequence of x values for the independent variables in the
data file that was generated. To test the non-linear fit program, I initiated the procedure
and entered the values shown below:
Figure 38 Non-Linear Regression Specification Form.
You can see that y is the dependent variable and x is the independent variable.
Values of 1 have been entered for the initial estimates of a, b and c. The equation model
was selected by clicking the parabola model from the drop-down models box. I could
have entered the same equation by clicking on the equation box and typing the equation
into that box or clicking parameters, math functions and variables from the drop-down
boxes on the right side of the form. Notice that I selected to plot the x versus y values
and also the predicted versus observed y values. I also chose to save the predicted scores
and residuals (y - predicted y.) The results are as follows:
Figure 39 X Versus Observed Y Plot From the Non-Linear Regression Analysis.
Figure 40 X Versus the Predicted Y Plot from the Non-Linear Regression Analysis.
The printed output shown below gives the model selected followed by the
individual data points observed, their predicted scores, the residual, the standard error of
estimate of the predicted score and the 95% confidence interval of the predicted score.
These are followed by the obtained correlation coefficient and its square, root mean
square of the y scores, the parameter estimates with their confidence limits and t
probability for testing the significance of difference from zero.
y = a + b * x1 + c * x1 * x1
x y yc y-yc SEest YcLo YcHi
0.39800 2.31000 2.30863 0.00137 0.00161 2.30582 2.31143
-1.19700 -3.13000 -3.12160 -0.00840 0.00251 -3.12597 -3.11723
-0.48600 -1.95000 -1.95878 0.00878 0.00195 -1.96218 -1.95538
-1.90800 -2.26000 -2.26113 0.00113 0.00522 -2.27020 -2.25205
-0.84100 -2.79000 -2.79228 0.00228 0.00206 -2.79586 -2.78871
-0.30100 -1.32000 -1.32450 0.00450 0.00192 -1.32784 -1.32115
0.69600 4.44000 4.45208 -0.01208 0.00168 4.44917 4.45500
1.11600 8.08000 8.07654 0.00346 0.00264 8.07195 8.08112
0.47900 2.86000 2.85607 0.00393 0.00159 2.85330 2.85884
1.09900 7.92000 7.91612 0.00388 0.00258 7.91164 7.92061
-0.94400 -2.94000 -2.93971 -0.00029 0.00214 -2.94343 -2.93600
-0.21800 -0.99000 -0.99541 0.00541 0.00190 -0.99872 -0.99211
0.81000 5.37000 5.36605 0.00395 0.00183 5.36288 5.36923
-0.06200 -0.31000 -0.30228 -0.00772 0.00185 -0.30549 -0.29907
0.67200 4.26000 4.26629 -0.00629 0.00165 4.26342 4.26917
-0.01900 -0.10000 -0.09410 -0.00590 0.00183 -0.09728 -0.09093
0.00100 0.01000 0.00525 0.00475 0.00182 0.00209 0.00841
0.01600 0.08000 0.08081 -0.00081 0.00181 0.07766 0.08396
1.19900 8.88000 8.87635 0.00365 0.00295 8.87122 8.88148
0.98000 6.82000 6.82561 -0.00561 0.00221 6.82177 6.82945
Corr. Coeff. = 1.00000 R2 = 1.00000
RMS Error = 5.99831, d.f. = 17 SSq = 611.65460
Parameter Estimates ...
p1= 0.00024 +/- 0.00182 p= 0.89626
p2= 5.00349 +/- 0.00171 p= 0.00000
p3= 2.00120 +/- 0.00170 p= 0.00000
Covariance Matrix Terms and Error-Correlations...
B(1,1)= 0.00000; r= 1.00000
B(1,2)=B(2,1)= -0.00000; r= -0.28318
B(1,3)=B(3,1)= -0.00000; r= -0.67166
B(2,2)= 0.00000; r= 1.00000
B(2,3)=B(3,2)= 0.00000; r= 0.32845
B(3,3)= 0.00000; r= 1.00000
X versus Y Plot
X = Y, Y = Y' from file: C:\Documents and Settings\Owner\My
Documents\Projects\Clanguage\OpenStat4\Parabola.OS4
Y 1.76 16.29 4.04
Y' 1.76 16.29 4.04
Correlation = 1.0000, Slope = 1.00, Intercept = 0.00
You can see that the fit is quite good between the observed and predicted scores.
Once you have obtained the results you will notice that the parameters, their standard
errors and the t probabilities are also entered in the dialog form. Had you elected to
proceed in a step-fashion, these results would be updated at each step so you can observe
the convergence to the best fit (the root mean square shown in the lower left corner.)
Figure 41 The Non-Linear Regression Form Following the Analysis.
The data grid now contains the predicted values we requested.
x1 y Pred_Y Residual
0.398 2.310 2.309 -0.001373
-1.197 -3.13 -3.121602 0.008398
-0.486 -1.95 -1.958781 -0.008781
-1.908 -2.26 -2.261127 -0.001127
-0.841 -2.79 -2.792285 -0.002285
-0.301 -1.32 -1.324499 -0.004499
0.696 4.44 4.452082 0.012082
etc.
Weighted Least-Squares Regression
For regressions with cross-section data (where the subscript "i" denotes a
particular individual or firm at a point in time), it is usually safe to assume the errors are
Un-correlated, but often their variances are not constant across individuals. This is known
as the problem of heteroskedasticity (for "unequal scatter"); the usual assumption of
constant error variance is referred to as homoskedasticity. Although the mean of the
dependent variable might be a linear function of the regressors, the variance of the error
terms might also depend on those same regressors, so that the observations might "fan
out" in a scatter diagram.
Approaches to Dealing with Heteroskedasticity:
. For known heteroskedasticity (e.g., grouped data with known group sizes), use weighted
least squares (WLS) to obtain efficient unbiased estimates;
. Test for heteroskedasticity of a special form using a squared residual regression;
. Estimate the unknown heteroskedasticity parameters using this squared residual
regression, then use the estimated variances in the WLS formula to get efficient
estimates of regression coefficients (known as feasible WLS); or
. Stick with the (inefficient) least squares estimators, but get estimates of standard errors
which are correct under arbitrary heteroskedasticity.
In this procedure, the "residualization" method is used to obtain weights that will
reduce the effect of heteroskedastic values. The method consists of four stages:
Step 1. Perform an Ordinary Least Squares (OLS) regression and obtain the residuals and
squared residuals where the residual is the difference between the observed dependent
variable and the predicted dependent variable value for each case.
Step 2. Regress the values of the squared residuals on the independent variables using
OLS. The F test for the model is an indication of heteroskedasticity in the data.
Step 3. Obtain the reciprocal of the square root of the absolute squared residuals. These
weights are then multiplied times all of the variables of the regression model.
Step 4. Obtain the OLS regression of the weighted dependent variable on the weighted
independent variables. One can obtain the regression through the origin. If elected, each
variable's values are converted to deviations from their respective mean before the OLS
analysis is performed.
As an alternative, the user may use weights he or she has derived. These should
be similar to the reciprocal values obtained in step 3 above. When these weights are
used, they are multiplied times the values of each variable and step 4 above is completed.
Shown below is the dialog box for the Weighted Least Squares Analysis and an
analysis of the cansas.txt data file.
Figure 42 Weighted Least Squares Regression Specification Form.
When the Compute button is clicked you would obtain the following output:
OLS REGRESSION RESULTS
Means
Variables weight waist pulse chins situps jumps
178.600 35.400 56.100 9.450 145.550 70.300
Standard Deviations
24.691 3.202 7.210 5.286 62.567 51.277
CORRELATION MATRIX
VARIABLE
weight waist pulse chins situps jumps
weight 1.000 0.870 -0.366 -0.390 -0.493 -0.226
waist 0.870 1.000 -0.353 -0.552 -0.646 -0.191
pulse -0.366 -0.353 1.000 0.151 0.225 0.035
chins -0.390 -0.552 0.151 1.000 0.696 0.496
situps -0.493 -0.646 0.225 0.696 1.000 0.669
jumps -0.226 -0.191 0.035 0.496 0.669 1.000
Dependent variable: jumps
weight -0.588 -1.221 0.704 -1.734 0.105 4.424 0.226
waist 0.982 15.718 6.246 2.517 0.025 5.857 0.171
pulse -0.064 -0.453 1.236 -0.366 0.720 1.164 0.859
chins 0.201 1.947 2.243 0.868 0.400 2.059 0.486
situps 0.888 0.728 0.205 3.546 0.003 2.413 0.414
Intercept 0.000 -366.967 183.214 -2.003 0.065
Regression 5 31793.741 6358.748 4.901 0.0084
Residual 14 18164.459 1297.461
Total 19 49958.200
R2 = 0.6364, F = 4.90, D.F. = 5 14, Prob>F = 0.0084
REGRESSION OF SQUARED RESIDUALS ON INDEPENDENT VARIABLES
Means
Variables weight waist pulse chins situps ResidSqr
178.600 35.400 56.100 9.450 145.550 908.196
Standard Deviations
Variables weight waist pulse chins situps ResidSqr
24.691 3.202 7.210 5.286 62.567 2086.828
CORRELATION MATRIX
VARIABLE
weight waist pulse chins situps ResidSqr
weight 1.000 0.870 -0.366 -0.390 -0.493 -0.297
waist 0.870 1.000 -0.353 -0.552 -0.646 -0.211
pulse -0.366 -0.353 1.000 0.151 0.225 -0.049
chins -0.390 -0.552 0.151 1.000 0.696 0.441
situps -0.493 -0.646 0.225 0.696 1.000 0.478
ResidSqr -0.297 -0.211 -0.049 0.441 0.478 1.000
Dependent variable: ResidSqr
weight -0.768 -64.916 36.077 -1.799 0.094 4.424 0.226
waist 0.887 578.259 320.075 1.807 0.092 5.857 0.171
pulse -0.175 -50.564 63.367 -0.798 0.438 1.164 0.859
chins 0.316 124.826 114.955 1.086 0.296 2.059 0.486
situps 0.491 16.375 10.515 1.557 0.142 2.413 0.414
Intercept 0.000 -8694.402 9389.303 -0.926 0.370
Regression 5 35036253.363 7007250.673 2.056 0.1323
Residual 14 47705927.542 3407566.253
Total 19 82742180.905
R2 = 0.4234, F = 2.06, D.F. = 5 14, Prob>F = 0.1323
X versus Y Plot
X = ResidSqr, Y = weight from file: C:\Program Files\WGM Consulting\OpenStat\cansas.txt
ResidSqr 908.20 4354851.63 2086.83
weight 178.60 609.62 24.69
Correlation = -0.2973, Slope = -0.00, Intercept = 181.79
Figure 43 Regression Plot from the Weighted Least Squares Analysis.
X versus Y Plot
X = ResidSqr, Y = waist from file: C:\Program Files\WGM Consulting\OpenStat\cansas.txt
ResidSqr 908.20 4354851.63 2086.83
waist 35.40 10.25 3.20
Figure 44 Residual Plot from the Weighted Least Squares Procedure.
X versus Y Plot
X = ResidSqr, Y = pulse from file: C:\Program Files\WGM Consulting\OpenStat\cansas.txt
ResidSqr 908.20 4354851.63 2086.83
pulse 56.10 51.99 7.21
Figure 45 Plot of the Pulse Versus Squared Residual from Weighted Least Squares.
X versus Y Plot
X = ResidSqr, Y = chins from file: C:\Program Files\WGM Consulting\OpenStat\cansas.txt
ResidSqr 908.20 4354851.63 2086.83
chins 9.45 27.94 5.29
Figure 46 Chins Variable Plot from the Weighted Least Squares Analysis.
X versus Y Plot
X = ResidSqr, Y = situps from file: C:\Program Files\WGM Consulting\OpenStat\cansas.txt
ResidSqr 908.20 4354851.63 2086.83
situps 145.55 3914.58 62.57
Figure 47 Situps Variable Plot from the Weighted Least Squares Analysis.
WLS REGRESSION RESULTS
Means
9.126 1.848 2.943 0.329 6.005 2.695
Standard Deviations
7.774 1.685 2.816 0.157 3.729 1.525
CORRELATION MATRIX
VARIABLE
weight 1.000 0.994 0.936 0.442 0.742 0.697
waist 0.994 1.000 0.965 0.446 0.783 0.729
pulse 0.936 0.965 1.000 0.468 0.889 0.769
chins 0.442 0.446 0.468 1.000 0.395 0.119
situps 0.742 0.783 0.889 0.395 1.000 0.797
jumps 0.697 0.729 0.769 0.119 0.797 1.000
weight -2.281 -0.448 0.414 -1.082 0.298 253.984 0.004
waist 3.772 3.415 2.736 1.248 0.232 521.557 0.002
pulse -1.409 -0.763 0.737 -1.035 0.318 105.841 0.009
chins -0.246 -2.389 1.498 -1.594 0.133 1.363 0.734
situps 0.887 0.363 0.165 2.202 0.045 9.258 0.108
Intercept 0.000 1.326 0.669 1.982 0.067
Regression 5 33.376 6.675 8.624 0.0007
Residual 14 10.837 0.774
Total 19 44.212
R2 = 0.7549, F = 8.62, D.F. = 5 14, Prob>F = 0.0007
Figure 48 Predicted Values and Weights from the Weighted Least Squares
Analysis.
2-Stage Least-Squares Regression
Two Stage Least Squares regression may be used in the situation where the errors
of independent and dependent variables are known (or likely) to be correlated. For
example, the market price of a commodity and the demand for that commodity are non-
recursive, that is, demand affects price and price affects demand. Prediction variables are
"explanatory" variables to explain variability of the dependent variable. However, there
may be other "instrumental" variables that predict one or more of these explanatory
variables in which the errors are not correlated. If we first predict the explanatory
variables with these instrumental variables and use the predicted values, we reduce the
correlation of the errors with the dependent variable.
In this procedure, the user first selects the dependent variable of the study. Next,
the explanatory variables (predictors) are entered. Finally, the instrumental variables
AND the explanatory variables affected by these instrumental variables are entered into
the instrumental variables list.
The two stages of this procedure are performed as follows:
Stage 1. The instrumental variables are identified as those in the instrumental list that are
not in the explanatory list. The explanatory variables that are listed in both the
explanatory and the instrumental lists are those for which predictions are to be obtained.
These predicted scores are referred to as "proxy" scores. The predictions are obtained by
regressing each explanatory variable listed in both lists with all of the remaining
explanatory variables and instrumental variables. The predicted scores are obtained and
stored in the data grid with a "P_" appended to the beginning of the original predictor
name.
Stage 2. Once the predicted values are obtained, an OLS regression is performed with
the dependent variable regressed on the proxy variables and the other explanatory
variables not predicted in the previous stage.
In the following example, the cansas.OS4 file is analyzed. The dependent variable is the
height of individual jumps. The explanatory (predictor) variables are pulse rate, no. of
chinups and no. of situps the individual completes. These explanatory variables are
thought to be related to the instrumental variables of weight and waist size. In the dialog
box for the analysis, the option has been selected to show the regression for each of the
explanatory variables that produces the predicted variables to be used in the final
analysis. Results are shown below:
Figure 49 Two Stage Least Squares Regression Specification Form.
The output obtained is shown below:
FILE: C:\Program Files\WGM Consulting\OpenStat\cansas.txt
Dependent = jumps
Explanatory Variables:
weight
waist
pulse
Instrumental Variables:
weight
waist
pulse
chins
situps
Proxy Variables:
P_weight
P_waist
P_pulse
============================================================
Analysis for the P_Weight
Analysis for P_weight
Dependent: weight
Independent:
waist
pulse
chins
situps
Means
Variables waist pulse chins situps weight
35.400 56.100 9.450 145.550 178.600
Standard Deviations
Variables waist pulse chins situps weight
3.202 7.210 5.286 62.567 24.691
CORRELATION MATRIX
VARIABLE
waist pulse chins situps weight
waist 1.000 -0.353 -0.552 -0.646 0.870
pulse -0.353 1.000 0.151 0.225 -0.366
chins -0.552 0.151 1.000 0.696 -0.390
situps -0.646 0.225 0.696 1.000 -0.493
weight 0.870 -0.366 -0.390 -0.493 1.000
Dependent variable: weight
waist 0.941 7.259 1.317 5.513 0.000 1.935 0.517
pulse -0.062 -0.212 0.450 -0.471 0.644 1.147 0.872
chins 0.097 0.453 0.814 0.556 0.586 2.017 0.496
situps 0.061 0.024 0.075 0.322 0.752 2.397 0.417
Intercept 0.000 -74.286 64.402 -1.153 0.267
Regression 4 8964.778 2241.194 12.841 0.0001
Residual 15 2618.022 174.535
Total 19 11582.800
R2 = 0.7740, F = 12.84, D.F. = 4 15, Prob>F = 0.0001
Analysis for the P_Waist
Analysis for P_waist
Dependent: waist
Independent:
weight
pulse
chins
situps
Means
Variables weight pulse chins situps waist
178.600 56.100 9.450 145.550 35.400
Standard Deviations
Variables weight pulse chins situps waist
24.691 7.210 5.286 62.567 3.202
CORRELATION MATRIX
VARIABLE
weight pulse chins situps waist
weight 1.000 -0.366 -0.390 -0.493 0.870
pulse -0.366 1.000 0.151 0.225 -0.353
chins -0.390 0.151 1.000 0.696 -0.552
situps -0.493 0.225 0.696 1.000 -0.646
waist 0.870 -0.353 -0.552 -0.646 1.000
Dependent variable: waist
weight 0.711 0.092 0.017 5.513 0.000 1.462 0.684
pulse -0.029 -0.013 0.051 -0.249 0.806 1.159 0.863
chins -0.136 -0.082 0.090 -0.911 0.377 1.951 0.513
situps -0.194 -0.010 0.008 -1.228 0.238 2.193 0.456
Intercept 0.000 21.864 5.050 4.330 0.001
Regression 4 161.539 40.385 18.212 0.0000
Residual 15 33.261 2.217
Total 19 194.800
R2 = 0.8293, F = 18.21, D.F. = 4 15, Prob>F = 0.0000
=============================================================
Analysis for the P_pulse
Analysis for P_pulse
Dependent: pulse
Independent:
weight
waist
chins
situps
Means
Variables weight waist chins situps pulse
178.600 35.400 9.450 145.550 56.100
Standard Deviations
Variables weight waist chins situps pulse
24.691 3.202 5.286 62.567 7.210
CORRELATION MATRIX
VARIABLE
weight waist chins situps pulse
weight 1.000 0.870 -0.390 -0.493 -0.366
waist 0.870 1.000 -0.552 -0.646 -0.353
chins -0.390 -0.552 1.000 0.696 0.151
situps -0.493 -0.646 0.696 1.000 0.225
pulse -0.366 -0.353 0.151 0.225 1.000
Dependent variable: pulse
weight -0.235 -0.069 0.146 -0.471 0.644 4.360 0.229
waist -0.144 -0.325 1.301 -0.249 0.806 5.832 0.171
chins -0.062 -0.084 0.468 -0.179 0.860 2.055 0.487
situps 0.059 0.007 0.043 0.158 0.876 2.409 0.415
Intercept 0.000 79.673 32.257 2.470 0.026
Regression 4 139.176 34.794 0.615 0.6584
Residual 15 848.624 56.575
Total 19 987.800
R2 = 0.1409, F = 0.62, D.F. = 4 15, Prob>F = 0.6584
Adjusted R2 = -0.0882
Means
Variables P_weight P_waist P_pulse jumps
178.600 35.400 56.100 70.300
Standard Deviations
Variables P_weight P_waist P_pulse jumps
21.722 2.916 2.706 51.277
Final Step Stage 2
CORRELATION MATRIX
VARIABLE
P_weight P_waist P_pulse jumps
P_weight 1.000 0.886 -0.947 -0.106
P_waist 0.886 1.000 -0.960 -0.394
P_pulse -0.947 -0.960 1.000 0.239
jumps -0.106 -0.394 0.239 1.000
P_weight 0.723 1.707 1.395 1.223 0.239 10.313 0.097
P_waist -1.906 -33.511 12.032 -2.785 0.013 13.819 0.072
P_pulse -0.907 -17.181 18.664 -0.921 0.371 28.648 0.035
Intercept 0.000 1915.587 1584.184 1.209 0.244
Regression 3 22880.606 7626.869 4.507 0.0179
Residual 16 27077.594 1692.350
Total 19 49958.200
R2 = 0.4580, F = 4.51, D.F. = 3 16, Prob>F = 0.0179
Data in the Grid Following the Analysis:
P_weight P_waist P_pulse Pred.z zResid. Pred.Raw
RawResid. ResidSqr
182.63079 36.82400 55.53800 -0.6081 0.4073 39.118
-20.882 436.058
186.85215 37.37713 55.25055 -0.7328 0.5319 32.724
-27.276 743.980
197.15099 36.93646 53.75071 0.4005 0.1982 90.837
-10.163 103.287
174.62117 33.98673 56.88201 0.5291 -1.1785 97.431
60.431 3651.906
179.67327 36.10186 55.28117 -0.1486 -0.0913 62.680
4.680 21.902
179.43312 36.60546 55.82741 -0.6687 0.1168 36.011
-5.989 35.868
195.76352 38.95108 52.84951 -0.6603 0.0304 36.441
-1.559 2.430
165.55139 34.76832 57.50210 -0.4913 -0.0996 45.107
5.107 26.081
146.69161 33.93533 57.60987 -0.6108 0.0199 38.980
-1.020 1.040
167.16365 31.46446 58.64984 1.3369 2.1675 138.853
-111.147 12353.656
etc.
Multivariate Procedures
Identifying Latent Variables
We will now turn our attention to an area of research that involves "theoretical" or
implied "underlying" variables. As an example, we might sight the early research by
psychologists that theorized the existence of "intelligence". These psychologists
reasoned that variability in the success on a variety of different tests and academic
subjects must be due to some "underlying" variable of general intelligence. That is, those
individuals with higher intelligence would perform better than those of lower
intelligence. It was theorized that if one examines the correlation among a group of
measures obtained on a large sample of subjects, that these inter-correlations can be
"explained" by a single, or perhaps few, "latent" variables. A square array (matrix) of
values with certain properties can be "transformed" into a product of independent
(uncorrelated) "roots" and "vectors" sometimes referred to as eigenvalues and
eigenvectors. If the matrix is obtained via random sampling from a population, some of
the roots and vectors may be attributable to random sampling error. Thus, for example, a
matrix of correlation among 10 variables might be reduced to 5 "real" roots and vectors
and error components. Principal components analysis and factor analysis are thus based
on some relatively simple matrix algebra concepts. Principal components analysis is
basically just the transformation of an observed matrix into independent roots and
vectors. Factor analyses, on the other hand, are based not only on the transformation to
roots and vectors but also the "rotation" of the scaled vectors to enhance interpretability.
Factor Analysis
Factor analysis is based on the procedure for obtaining a new set of uncorrelated
(orthogonal) variables, usually fewer in number than the original set, that reproduces the
co-variability observed among a set or original variables. Two models are commonly
utilized:
1. The principal components model wherein the observed score of an individual i on
the jth variable X
i,j
is given as:
X
i,j
= A
j,1
S
i,1
+ A
j,2
S
i,2
+ ....+ A
j,k
S
i,k
+ C
where A
j,k
is a loading of the kth factor on variable j,
S
i,k
is the factor score of the ith individual on the kth factor and
C is a constant.

The A
j,k
loadings are typically least-squares regression coefficients.
2. The common factor model assumes each variable X may contain some unique
component of variability among subjects as well as components in common with
other variables. The model is:
X
i,j
= A
j,1
S
i,1
+ .... + A
j,k
S
i,k
+ A
j,u
S
i,u
The above equation may also be expressed in terms of standard z scores as:
z
i,j
= a
j,1
S
i,1
+ .... + a
j,k
S
i,k
+ a
j,u
S
i,u
Since the average of standard score products for the n cases is the product-
moment correlation coefficient, the correlation matrix among the j observed variables
may be expressed in matrix form as:
[R] = [F] [F]' - [U]
2
jxj jxk kxj jxj (array sizes k <= j)
The matrix [F] is the matrix of factor loadings or correlations of the k theoretical
orthogonal variables with the j observed variables. The [U] matrix is a diagonal matrix
with unique loadings on the diagonal.
The factor loadings above are the result of calculating the eigenvalues and
associated vectors of the characteristic equation:
| [R] - [U]
2
- [I] |

where the lambda values are eigenvalues (roots) of the equation.
When you execute the Factor Analysis Program in OpenStat, you are asked to
provide information necessary to complete an analysis of your data file. You enter the
name of your file and identify the variables to analyze. If you elect to send output to the
printer, be sure the printer is on when you start. You will also be asked to specify the
type of analysis to perform. The principle components method, a partial image analysis,
a Guttman Image Analysis, a Harris Scaled Image Analysis, a Canonical Factor Analysis
or an Alpha Factor Analysis may be elected. Selection of the method depends on the
assumptions you make concerning sampling of variables and sampling of subjects as well
as the theory on which you view your variables. You may request a rotation of the
resulting factors which follows completion of the analysis of the data,. The most
common rotation performed is the Varimax rotation. This method rotates the orthogonal
factor loadings so that the loadings within each factor are most variable on the average.
This tends to produce "simple structure", that is, factors which have very high or very
low loadings for the original variables and thus simplifies the interpretation of the
resulting factors. One may also elect to perform a Procrustean rotation whereby the
obtained factor loadings are rotated to be maximally congruent with another factor
loading matrix. This second set of loadings which is entered by the user is typically a set
which represents some theoretical structure of the original variables. One might,
however, obtain factor loadings for males and females separately and then rotate one
solution against the other to see if the structures are highly similar for both sexes.
We will demonstrate factor analysis using a sample data file used in a previous
example. The file is labeled cansas.txt and contains three physiological measures (height,
weight and pulse rate) and three performance measures (chin-ups, pull-ups and jumps)
obtained on 20 subjects. We might theorize that all 6 variables actually reflect only one
underlying variable which we might call "general health". So lets select the Principal
Components and Factor Analysis option from within the Multivariate sub-menu. Shown
below is the form which appears:
Figure 50 Specification of a Sample Factor Analysis
Notice the options that we have elected as well as the type of analysis and rotation we
have selected. Clearly, there are a variety of analyses available. One typically has
completed a course on factor analysis or read extensively about these techniques prior to
completing a factor analysis. One may discover however that often very similar results
are obtained across these different (theoretical) approaches!
When we click the OK button for our above example, the results we obtain are
shown below:
Factor Analysis
See Rummel, R.J., Applied Factor Analysis
Northwestern University Press, 1970
CORRELATION MATRIX
Correlations
weight 1.000 0.870 -0.366 -0.390 -0.493 -0.226
waist 0.870 1.000 -0.353 -0.552 -0.646 -0.191
pulse -0.366 -0.353 1.000 0.151 0.225 0.035
chins -0.390 -0.552 0.151 1.000 0.696 0.496
situps -0.493 -0.646 0.225 0.696 1.000 0.669
jumps -0.226 -0.191 0.035 0.496 0.669 1.000
Kaiser-Meyer-Olkin MSA statistic = 0.591
Harris Scaled Image Analysis
Harris Scaled Image Covariance Matrix
weight 4.375 5.033 -0.769 -1.551 -2.790 -0.450
waist 5.033 7.506 -1.117 -2.061 -3.464 -1.484
pulse -0.769 -1.117 0.175 0.263 0.426 0.160
chins -1.551 -2.061 0.263 1.170 1.985 0.985
situps -2.790 -3.464 0.426 1.985 3.581 1.687
jumps -0.450 -1.484 0.160 0.985 1.687 1.750
Original matrix trace = 18.56
SCATTERPLOT - Scree Plot
Values
. | |- 15.30
| |- 14.57
| |- 13.84
| |- 13.12
| |- 12.39
| |- 11.66
| |- 10.93
| |- 10.20
| |- 9.47
| |- 8.74
------------------------------------------------------------|- 8.01
| |- 7.29
| |- 6.56
| |- 5.83
| |- 5.10
| |- 4.37
. | |- 3.64
| |- 2.91
. | |- 2.19
_______________________________________________________________
| | | | | | | | | |
Sequence Number of Eigenvalues
1.00 1.56 2.11 2.67 3.22 3.78 4.33 4.89 5.44 6.00
Roots (Eigenvalues) Extracted:
1 14.573
2 2.680
3 1.146
4 0.117
5 0.042
6 0.000
Unrotated Factor Loadings
FACTORS
Factor 1 Factor 2 Factor 3 Factor 4 Factor 5 Factor 6
weight 0.832 0.252 -0.225 0.084 -0.017 0.001
waist 0.899 0.177 0.202 -0.035 0.011 0.000
pulse -0.349 -0.131 -0.088 -0.023 0.043 0.009
chins -0.624 0.336 0.120 0.105 0.108 -0.001
situps -0.747 0.416 0.220 0.031 -0.049 0.001
jumps -0.396 0.607 -0.311 -0.119 0.016 -0.001
Percent of Trace In Each Root:
1 Root = 14.573 Trace = 18.557 Percent = 78.531
COMMUNALITY ESTIMATES
1 weight 0.814
2 waist 0.882
3 pulse 0.149
4 chins 0.539
5 situps 0.782
6 jumps 0.636
Proportion of variance in unrotated factors
1 45.435
2 12.777
3 4.304
Communality Estimates as percentages:
1 80.657
2 88.107
3 14.657
4 51.654
5 77.833
6 62.191
Varimax Rotated Loadings
FACTORS
Factor 1 Factor 2 Factor 3
weight 0.726 -0.520 0.098
waist 0.863 -0.278 -0.242
pulse -0.372 0.075 0.054
chins -0.274 0.570 0.342
situps -0.298 0.743 0.371
jumps -0.073 0.243 0.747
Percent of Variation in Rotated Factors
Factor 1 26.317
Factor 2 21.482
Factor 3 14.717
Total Percent of Variance in Factors : 62.516
Communalities as Percentages
1 for weight 80.657
2 for waist 88.107
3 for pulse 14.657
4 for chins 51.654
5 for situps 77.833
6 for jumps 62.191
SCATTERPLOT - FACTOR PLOT
Factor 2
| |- 1.10
| |- 1.00
. | |- 0.90
| |- 0.80
. | |- 0.70
| |- 0.60
| |- 0.50
. | |- 0.40
| |- 0.30
. | |- 0.20
------------------------------------------------------------|- 0.10
| |- -0.00
| . |- -0.10
| |- -0.20
| |- -0.30
| . |- -0.40
| |- -0.50
| |- -0.60
| |- -0.70
| |- -0.80
_______________________________________________________________
| | | | | | | | | |
Factor 1
-1.00 -0.78 -0.56 -0.33 -0.11 0.11 0.33 0.56 0.78 1.00
Factor 3
| |- 1.10
| |- 1.00
. | |- 0.90
| |- 0.80
| |- 0.70
| |- 0.60
* | |- 0.50
| |- 0.40
| |- 0.30
. | . |- 0.20
------------------------------------------------------------|- 0.10
| |- -0.00
| . |- -0.10
| |- -0.20
| |- -0.30
| |- -0.40
| |- -0.50
| |- -0.60
| |- -0.70
| |- -0.80
_______________________________________________________________
| | | | | | | | | |
Factor 1
-1.00 -0.78 -0.56 -0.33 -0.11 0.11 0.33 0.56 0.78 1.00
Factor 3
| |- 1.10
| |- 1.00
| . |- 0.90
| |- 0.80
| |- 0.70
| |- 0.60
| . . |- 0.50
| |- 0.40
| |- 0.30
. | . |- 0.20
------------------------------------------------------------|- 0.10
| |- -0.00
. | |- -0.10
| |- -0.20
| |- -0.30
| |- -0.40
| |- -0.50
| |- -0.60
| |- -0.70
| |- -0.80
_______________________________________________________________
| | | | | | | | | |
Factor 2
-1.00 -0.78 -0.56 -0.33 -0.11 0.11 0.33 0.56 0.78 1.00
COORDINATES OF THE PLOTS (LOADINGS)
VARIABLE FACTOR LOADING
weight 1 0.7258
waist 1 0.8633
pulse 1 -0.3716
chins 1 -0.2736
situps 1 -0.2976
jumps 1 -0.0733
weight 2 -0.5198
waist 2 -0.2782
pulse 2 0.0750
chins 2 0.5699
situps 2 0.7428
jumps 2 0.2434
weight 3 0.0979
waist 3 -0.2416
pulse 3 0.0538
chins 3 0.3420
situps 3 0.3714
jumps 3 0.7465
SUBJECT FACTOR SCORE RESULTS:
Regression Coefficients
FACTORS
Factor 1 Factor 2 Factor 3
weight -0.609 -2.468 2.251
waist 1.951 3.352 -2.864
pulse -0.123 -0.210 0.161
chins 0.236 0.637 -0.309
situps 0.890 2.321 -1.378
jumps -0.546 -1.535 1.777
Standard Error of Factor Scores:
Factor 1 0.999
Factor 2 1.431
Factor 3 1.277
In examining the above results, it appears that there is not just one, but potentially
three latent variables. If we examine the rotated factor loadings (correlations of the
observed variables with the latent variables) it appears there is one latent variable
consisting primarily of two body characteristics: weight and waist size. A second latent
variable appears to be composed of primarily performance abilities: chins and sit-ups.
Finally, there seems to be a latent variable that is unique to jump distances.
You can see from the above, that factor analysis is often a rather subjective
procedure, at least in the interpretation of the results. If we select another method, we
would not be surprised to find two major "latent" variables, one for the three physical
measures and one for the three performance measures. Again we remind ourselves that a
sample of 20 subjects with 6 measures is not nearly large enough to obtain definitive
results. We often need about 20 to 30 cases for each variable. Thus a sample of 120 to
180 would be more acceptable for supporting or rejecting our initial hypothesis of a
single latent variable!
Identifying Natural Groups - Cluster Analysis
Whenever we observe objects like people, dogs, cars, fish, etc. we tend to see
similarities and differences among the objects. There often appear to be "natural" sub-
groups that share more in common with one another than with others in the larger
population. Several methods for identifying these "clusters" or sub-groups have been
developed. Within the procedure called "Cluster Analysis" the OpenStat provides several
alternatives. We will use one called "hierarchical cluster analysis" to demonstrate these
methods. The hierarchical method examines the "distance" among all the objects studied
and creates groupings based on those that are "closest" to each other. Typically, one has
observed the objects using a variety of interval or ratio measures. Each object is then
considered a point in the Euclidean space defined by these measures. We can measure
the distance (or squared distance) between the objects in this space. Usually this
clustering is done in an "iterative" fashion, that is, we combine the two objects that are a
least distance from one another and treat this as a new object that replaces the previous
two in the space. The process of obtaining distances is then repeated and again the next
two objects of minimum distance are combined. We can continue to combine objects (or
their replacements) until only a single group remains.
To demonstrate the clustering procedure we will load the Hierarchical Clustering
procedure from the Clustering sub-menu within the Multivariate sub-menu. The file that
we will analyze is one we have used before - the cansas.txt file which contains 20
subjects that have been measured on three physiological variables and three performance
variables. When this procedure is selected, the form below is shown:
Figure 51 Hierarchical Cluster Analysis Specification Form.
Notice that we have selected all variables to include in the analysis. We have elected to
standardize the measures since they have quite different means and standard deviations.
We will plot the "error" as more groups are created. The error reflects the growth in the
distances among those within the groupings. We have also specified the maximum
number of groups to show in our results and to record up to five group codes in the data
grid for the subjects grouped. Notice that we are "clustering" the subjects and NOT the
variables as is done in factor analysis. When we click the OK button we obtain the
following results:
Hierarchical Cluster Analysis
Number of object to cluster = 20 on 6 variables.
Variable Means
178.600 35.400 56.100 9.450 145.550 70.300
Variable Variances
609.621 10.253 51.989 27.945 3914.576 2629.379
Variable Standard Deviations
24.691 3.202 7.210 5.286 62.567 51.277
19 groups after combining group 1 (n = 1 ) and group 5 (n = 1) error = 0.386
Group 1 (n= 5)
Object = 1
Object = 2
Object = 6
Object = 15
Object = 17
Group 3 (n= 2)
Object = 3
Object = 7
Group 4 (n= 2)
Object = 4
Object = 8
Group 5 (n= 2)
Object = 5
Object = 11
Group 9 (n= 1)
Object = 9
Group 10 (n= 1)
Object = 10
Group 12 (n= 4)
Object = 12
Object = 13
Object = 18
Object = 19
Group 14 (n= 1)
Object = 14
Group 16 (n= 1)
Object = 16
Group 20 (n= 1)
Object = 20
Group 1 (n= 5)
Object = 1
Object = 2
Object = 6
Object = 15
Object = 17
Group 3 (n= 2)
Object = 3
Object = 7
Group 4 (n= 3)
Object = 4
Object = 8
Object = 20
Group 5 (n= 2)
Object = 5
Object = 11
Group 9 (n= 1)
Object = 9
Group 10 (n= 1)
Object = 10
Group 12 (n= 4)
Object = 12
Object = 13
Object = 18
Object = 19
Group 14 (n= 1)
Object = 14
Group 16 (n= 1)
Object = 16
.
.
etc.
.
.
Group 1 (n= 9)
Object = 1
Object = 2
Object = 3
Object = 6
Object = 7
Object = 14
Object = 15
Object = 16
Object = 17
Group 4 (n= 11)
Object = 4
Object = 5
Object = 8
Object = 9
Object = 10
Object = 11
Object = 12
Object = 13
Object = 18
Object = 19
Object = 20
SCATTERPLOT - Plot of Error vs No. of Groups
Size of Error
| |- 18.06
| |- 17.20
| |- 16.34
. | |- 15.48
| |- 14.62
| |- 13.76
| |- 12.90
. | |- 12.04
| |- 11.18
| |- 10.32
------------------------------------------------------------|- 9.46
. | |- 8.60
. . | |- 7.74
| |- 6.88
| |- 6.02
| |- 5.16
. . | |- 4.30
. | |- 3.44
* . . . |- 2.58
| . . . . . |- 1.72
_______________________________________________________________
| | | | | | | | | |
No. of Groups
2.00 4.00 6.00 8.00 10.00 12.00 14.00 16.00 18.00 20.00
In examining the above error plot, it appears that 6 groups might provide some "natural"
grouping. By examining the characteristics of objects in each of those 6 groups, the
researcher may begin to understand those characteristics that best define those groups.
If you elect to save the group numbers in the grid, you should be aware that they
are NOT sequential numbers such as 1, 2, 3, 4, 5 but rather "group" numbers that reflect
current and previous group numbers. The number of unique values will be the number
you specified on the form but simply not sequential values. You can, of course, re-code
those values to be sequential if needed.
Predicting Group Classification
If you need to predict the class to which a particular object belongs, you may want
to consider the use of the discriminant function analysis procedure. For example, assume
you have a group of subjects that are happily employed in various occupations such as
Engineering, teaching, farming, etc. Also assume you have measured these subjects on a
variety of measures such as their interests, their intelligence, their socio-economic level,
etc. You can then develop linear equations that predict their occupation as a function of
the obtained measures. Of course, the better that the measures differentiate among the
groups, the better the classification prediction. Discriminant analysis is closely related to
multivariate analysis of variance (MANOVA). With multivariate analysis of variance,
we test the hypothesis that the observed groups are from the same (multivariate)
population. If they are different, then discriminant analysis will use the differences to
classify subjects.
To demonstrate the prediction of classifications, we will load the file labeled
"manodiscrim.txt". This file contains 15 objects from three groups that have been
measured on two variables. Select the option Discriminant Function and MANOVA
from the Multivariate sub-menu. You will see the form below (which we have completed
for the analysis):
Figure 52 Discriminant Analysis and Multivariate Analysis of Variance Form.
When the OK button on the form above is clicked, you obtain the following output:
MULTIVARIATE ANOVA / DISCRIMINANT FUNCTION
Reference: Multiple Regression in Behavioral Research
Elazar J. Pedhazur, 1997, Chapters 20-21
Harcourt Brace College Publishers
Total Cases = 15, Number of Groups = 3
MEANS FOR GROUP 1, N = 5
Variables Y1 Y2
4.600 8.200
VARIANCES FOR GROUP 1
Variables Y1 Y2
1.300 1.700
STANDARD DEVIATIONS FOR GROUP 1
Variables Y1 Y2
1.140 1.304
Variables Y1 Y2
5.000 6.600
Variables Y1 Y2
1.000 1.300
Variables Y1 Y2
1.000 1.140
Variables Y1 Y2
6.200 6.200
Variables Y1 Y2
0.700 1.700
Variables Y1 Y2
0.837 1.304
OVERALL
MEANS
Variables Y1 Y2
5.267 7.000
VARIANCES
Variables Y1 Y2
1.352 2.143
STANDARD DEVIATIONS
Variables Y1 Y2
1.163 1.464
UNIVARIATE ANOVA FOR VARIABLE Y1
SOURCE DF SS MS F PROB > F
BETWEEN 2 6.933 3.467 3.467 0.065
ERROR 12 12.000 1.000
TOTAL 14 18.933
UNIVARIATE ANOVA FOR VARIABLE Y2
SOURCE DF SS MS F PROB > F
BETWEEN 2 11.200 5.600 3.574 0.061
ERROR 12 18.800 1.567
TOTAL 14 30.000
Number of roots extracted = 2
Percent of trace extracted = 100.0000
Roots of the W inverse time B Matrix
No. Root Proportion Canonical R Chi-Squared D.F. Prob.
1 8.7985 0.9935 0.9476 26.8845 4 0.000
2 0.0571 0.0065 0.2325 0.6388 1 0.424
Eigenvectors of the W inverse x B Matrix
Root No
1 2
Y1 -2.316 0.188
Y2 1.853 0.148
Raw Discriminant Function Coefficients from Pooled Covariance
Root No
1 2
Y1 -2.030 0.520
Y2 1.624 0.409
Raw Discriminant Function Constants
Variables 1 2
-0.674 -5.601
Fisher Discriminant Functions
Group 1 Constant = -24.402
Variable Coefficient
1 -5.084
2 8.804
1 1.607
2 3.084
1 8.112
2 -1.738
CLASSIFICATION OF CASES
SUBJECT ACTUAL HIGH PROBABILITY SEC'D HIGH DISCRIM
ID NO. GROUP IN GROUP P(G/D) GROUP P(G/D) SCORE
1 1 1 0.9999 2 0.0001 4.6019
-1.1792
2 1 1 0.9554 2 0.0446 2.5716
-0.6590
3 1 1 0.8903 2 0.1097 2.1652
0.2699
4 1 1 0.9996 2 0.0004 3.7890
0.6786
5 1 1 0.9989 2 0.0011 3.3826
1.6075
6 2 2 0.9746 3 0.0252 -0.6760
-1.4763
7 2 2 0.9341 1 0.0657 0.9478
-1.0676
8 2 2 0.9730 1 0.0259 0.5414
-0.1387
9 2 2 0.5724 3 0.4276 -1.4888
0.3815
10 2 2 0.9842 1 0.0099 0.1350
0.7902
11 3 3 0.9452 2 0.0548 -2.7062
-0.9560
12 3 3 0.9999 2 0.0001 -4.7365
-0.4358
13 3 3 0.9893 2 0.0107 -3.1126
-0.0271
14 3 3 0.9980 2 0.0020 -3.5191
0.9018
15 3 3 0.8007 2 0.1993 -1.8953
1.3104
CLASSIFICATION TABLE
PREDICTED GROUP
O 1 2 3 TOTAL
B
S 1 5 0 0 5
E 2 0 5 0 5
R 3 0 0 5 5
V TOTAL 5 5 5 15
E
D
Standardized Discriminant Function Coefficients from Pooled Covariance
Root No
1 2
Y1 -2.030 0.520
Y2 2.032 0.511
Centroids
Root No
Group 1 2
1 3.302 0.144
2 -0.108 -0.302
3 -3.194 0.159
Raw Discriminant Function Coefficients from Total Covariance
Root No
Var. 1 2
Y1 -0.701 0.547
Y2 0.560 0.429
Raw Discriminant Function Constants
Variables 1 2
-0.674 -5.601
Standardized Discriminant Function Coefficients from Total Covariance
Root No.
Var. 1 2
Y1 -0.815 0.636
Y2 0.820 0.628
Correlations Between Variables and Functions
Root No
Var. 1 2
Y1 -0.608 0.794
Y2 0.615 0.788
Wilkes Lambda = 0.0965.
F = 12.2013 with D.F. 4 and 22 . Prob > F = 0.0000
Bartlett Chi-Squared = 26.8845 with 4 D.F. and prob. = 0.0000
Pillia Trace = 0.9520
The typical user will probably first examine the classification table in the above
output. In this example, all current subjects were correctly classified. The F-test and
Bartlett Chi-square test both confirm that the probability of obtaining the observed
discrimination by chance alone is highly unlikely (probabilities less than 0.00001. The
plot of the subjects in the space of the two discriminant variables also demonstrates the
within group similarity and between group differences. It is interesting to note that the
analysis of variance for each variable might not be considered "significant" but that the
combination of the two variables would be considered significant! There was, in fact,
one "root" that is significant. The roots reflect the canonical correlation between the two
observed variables and group code vectors. The correlation of the discriminant functions
and the observed variables indicated that first function seems to play the major role in the
separation of the groups. Notice that the standardized coefficients of the first function are
quite different when compared to the second function.
While the two discriminant functions can be used to predict the classification of
the objects, Fisher developed an alternative set of functions, one for each group, for
classifying objects. With these functions one classifies the objects with all three
functions and places the subject in the group with the largest predicted score. We have
not used these functions but rather the two discriminant functions previously discussed to
obtain the probability of group membership. The table produced for predicting each
subject's classification reports the highest and second highest probabilities for group
membership. In addition, each discriminant score is shown for the objects.
Of course, one will typically do much better in predicting group membership for
those objects on which the discriminant functions are obtained. The real question is
whether or not those same function coefficients will accurately predict group membership
for a "cross-validation" sample, that is, a sample of similar objects not in the original
study for whom the group memberships are also known.
Path Analysis: Examining the Plausibility of Cause and Effect Models
Path analysis is a procedure for examining the inter-correlations among a
set of variables to see if they are consistent with a model of causation. A causal model is
one in which the observed scores (events) of an object are assumed to be directly or
indirectly caused by one or more preceding events. For example, entrance to college may
be hypothesized to be a result of high achievement in high school. High achievement in
high school may be the result of parent expectations and the student's intelligence.
Intelligence may be a result of parent intelligence, early nutrition, and early
environmental stimulation, etc., etc. . Causing and resultant variables may be described
in a set of equations. Using standardized z scores, the above example might be described
by the following equations:
(1) z
1
= e
1
Parent intelligence
(2) z
2
= P
21
z
1
+ e
2
Child's nutrition
(3) z
3
= P
31
z
1
+ P
32
z
2
+ e
3
Child's intelligence
(4) z
4
= P
41
z
1
+ e
4
Parent expectations
(5) z
5
= P
53
z
3
+ p
54
z
4
+ e
5
School achievement
(6) z
6
= P
63
z
3
+ P
64
z
4
+ P
65
z
5
+ e
6
College GPA
In the above equations, the P's represent path coefficients measuring the strength
of causal effect on the resultant due to the causing variable z. In the above example, z
1
has no causing variable and path coefficient. It is called an exogenous variable and is
assumed to have only external causes unknown in this model. The "e" values represent
contributions that are external and unknown for each variable. These external causes are
assumed to be uncorrelated and dropped from further computations in this model. By
substituting the definitions of each z score in a model like the above, the correlation
between the observed variables can be expressed as in the following examples:
r
12
= 3z
1
z
2
/ n = P
21
3z
1
z
1
/ n = P
21
r
23
= 3z
2
z
3
/ n = P
31
P
21
+ P
32
etc.
In other words, the correlations are estimated to be the sum of direct path coefficients and
products of indirect path coefficients. The path coefficients are estimated by the
standardized partial regression coefficients (betas) of each resultant variable on its
causing variables. For example, coefficients P
31
and P
32
above would be estimated by
31.2
and
32.1
in the multiple regression equation
z
3
=
31.2
z
1
+
32.1
z
2
+ e
3
If the hypothesized causal flow model sufficiently describes the interrelationships
among the observed variables, the reproduced correlation matrix using the path
coefficients should deviate only by sampling error from the original correlations among
the variables.
When you execute the Path Analysis procedure in OpenStat, you will be asked to
specify the exogenous and endogenous variables in your analysis. The program then asks
you to specify, for each resultant (endogenous) variable, the causing variables. In this
manner you specify your total path model. The program then completes the number of
multiple regression analyses required to estimate the path coefficients, estimate the
correlations that would be obtained using the model path coefficients and compare the
reproduced correlation matrix with the actual correlations among the variables.
You may discover in your reading that this is but one causal model method. More
complex methods include models involving latent variables (such as those identified
through factor analysis), correlated errors, adjustments for reliability of the variables, etc.
Structural model equations of these types are often analyzed using the LISREL
package found in commercial packages such as SPSS or SAS.
Example of a Path Analysis
In this example we will use the file cansas.txt that we have used in other analyses.
The user begins by selecting the Path Analysis option of the Analyses / Multivariate
menu. In the figure below we have selected all variables to analyze and have entered our
first path indicating that waist size is caused by weight:
Figure 53 Path Analysis Dialogue Form.
We will hypothesize that pulse rate is caused by weight, chin-ups are caused by
weight, waist and pulse, that the number of sit-ups is caused by weight, waist and pulse
and that jumps are caused by weight, waist and pulse. Each time we enter a new causal
relationship we click the scroll bar to move to a new model number prior to entering the
caused and causing variables. Once we have entered each model, we then click on
the Compute button. Note we have elected to print descriptive statistics, each models
correlation matrix, and the reproduced correlation matrix which will be our measure of
how well the models fit the data. The results are shown below:
PATH ANALYSIS RESULTS
CAUSED VARIABLE: waist
Causing Variables:
weight
CAUSED VARIABLE: pulse
Causing Variables:
weight
CAUSED VARIABLE: chins
Causing Variables:
weight
waist
pulse
CAUSED VARIABLE: situps
Causing Variables:
weight
waist
pulse
CAUSED VARIABLE: jumps
Causing Variables:
weight
waist
pulse
Correlation Matrix with 20 valid cases.
Variables
weight 1.000 0.870 -0.366 -0.390 -0.493
waist 0.870 1.000 -0.353 -0.552 -0.646
pulse -0.366 -0.353 1.000 0.151 0.225
chins -0.390 -0.552 0.151 1.000 0.696
situps -0.493 -0.646 0.225 0.696 1.000
jumps -0.226 -0.191 0.035 0.496 0.669
Variables
jumps
weight -0.226
waist -0.191
pulse 0.035
chins 0.496
situps 0.669
jumps 1.000
MEANS with 20 valid cases.
178.600 35.400 56.100 9.450 145.550
Variables jumps
70.300
VARIANCES with 20 valid cases.
609.621 10.253 51.989 27.945 3914.576
Variables jumps
2629.379
STANDARD DEVIATIONS with 20 valid cases.
24.691 3.202 7.210 5.286 62.567
Variables jumps
51.277
Dependent Variable = waist
Variables
weight waist
weight 1.000 0.870
waist 0.870 1.000
Variables weight waist
178.600 35.400
609.621 10.253
24.691 3.202
Dependent Variable = waist
R R2 F Prob.>F DF1 DF2
0.870 0.757 56.173 0.000 1 18
Adjusted R Squared = 0.744
Std. Error of Estimate = 1.621
Variable Beta B Std.Error t Prob.>t
weight 0.870 0.113 0.015 7.495 0.000
Constant = 15.244
Dependent Variable = pulse
Variables
weight pulse
weight 1.000 -0.366
pulse -0.366 1.000
Variables weight pulse
178.600 56.100
609.621 51.989
24.691 7.210
Dependent Variable = pulse
0.366 0.134 2.780 0.113 1 18
weight -0.366 -0.107 0.064 -1.667 0.113
Constant = 75.177
Dependent Variable = chins
Variables
weight waist pulse chins
weight 1.000 0.870 -0.366 -0.390
waist 0.870 1.000 -0.353 -0.552
pulse -0.366 -0.353 1.000 0.151
chins -0.390 -0.552 0.151 1.000
Variables weight waist pulse chins
178.600 35.400 56.100 9.450
609.621 10.253 51.989 27.945
24.691 3.202 7.210 5.286
Dependent Variable = chins
0.583 0.340 2.742 0.077 3 16
weight 0.368 0.079 0.089 0.886 0.389
waist -0.882 -1.456 0.683 -2.132 0.049
pulse -0.026 -0.019 0.160 -0.118 0.907
Constant = 47.968
Dependent Variable = situps
Variables
weight waist pulse situps
weight 1.000 0.870 -0.366 -0.493
waist 0.870 1.000 -0.353 -0.646
pulse -0.366 -0.353 1.000 0.225
situps -0.493 -0.646 0.225 1.000
Variables weight waist pulse situps
178.600 35.400 56.100 145.550
609.621 10.253 51.989 3914.576
24.691 3.202 7.210 62.567
Dependent Variable = situps
0.661 0.436 4.131 0.024 3 16
weight 0.287 0.728 0.973 0.748 0.466
waist -0.890 -17.387 7.465 -2.329 0.033
pulse 0.016 0.139 1.755 0.079 0.938
Constant = 623.282
Dependent Variable = jumps
Variables
weight waist pulse jumps
weight 1.000 0.870 -0.366 -0.226
waist 0.870 1.000 -0.353 -0.191
pulse -0.366 -0.353 1.000 0.035
jumps -0.226 -0.191 0.035 1.000
Variables weight waist pulse jumps
178.600 35.400 56.100 70.300
609.621 10.253 51.989 2629.379
24.691 3.202 7.210 51.277
Dependent Variable = jumps
0.232 0.054 0.304 0.822 3 16
Adjusted R Squared = -0.123
weight -0.259 -0.538 1.034 -0.520 0.610
waist 0.015 0.234 7.928 0.029 0.977
pulse -0.055 -0.389 1.863 -0.209 0.837
Constant = 179.887
Matrix of Path Coefficients with 20 valid cases.
Variables
weight 0.000 0.870 -0.366 0.368 0.287
waist 0.870 0.000 0.000 -0.882 -0.890
pulse -0.366 0.000 0.000 -0.026 0.016
chins 0.368 -0.882 -0.026 0.000 0.000
situps 0.287 -0.890 0.016 0.000 0.000
jumps -0.259 0.015 -0.055 0.000 0.000
Variables
jumps
weight -0.259
waist 0.015
pulse -0.055
chins 0.000
situps 0.000
jumps 0.000
SUMMARY OF CAUSAL MODELS
Var. Caused Causing Var. Path Coefficient
waist weight 0.870
pulse weight -0.366
chins weight 0.368
chins waist -0.882
chins pulse -0.026
situps weight 0.287
situps waist -0.890
situps pulse 0.016
jumps weight -0.259
jumps waist 0.015
jumps pulse -0.055
Reproduced Correlation Matrix with 20 valid cases.
Variables
weight 1.000 0.870 -0.366 -0.390 -0.493
waist 0.870 1.000 -0.318 -0.553 -0.645
pulse -0.366 -0.318 1.000 0.120 0.194
chins -0.390 -0.553 0.120 1.000 0.382
situps -0.493 -0.645 0.194 0.382 1.000
jumps -0.226 -0.193 0.035 0.086 0.108
Variables
jumps
weight -0.226
waist -0.193
pulse 0.035
chins 0.086
situps 0.108
jumps 1.000
Average absolute difference between observed and reproduced
coefficients := 0.077
Maximum difference found := 0.562
We note that pulse is not a particularly important predictor of chin-ups or sit-ups.
The largest discrepancy of 0.562 between an original correlation and a correlation
reproduced using the path coefficients indicates our model of causation may have been
inadequate.
Non-Parametric
Beginning statistics students are usually introduced to what are called
"parametric" statistics methods. Those methods utilize "models" of score distributions
such as the normal (Gaussian) distribution, Poisson distribution, binomial distribution,
etc. The emphasis in parametric statistical methods is estimating population parameters
from sample statistics when the distribution of the population scores can be assumed to
be one of these theoretical models. The observations made are typically based on
continuous variables that utilize an interval or ratio scale of measurement. Frequently the
measurement scales available yield only nominal or ordinal values and nothing can be
assumed about the distribution of such values in the population sampled. If however,
random sampling has been utilized in selecting subjects, one can still make inferences
about relationships and differences similar to those made with parametric statistics. For
example, if students enrolled in two courses are assigned a rank on their achievement in
each of the two courses, it is reasonable to expect that students that rank high in one
course would tend to rank high in the other course. Since a rank only indicates order
however and not "how much" was achieved, we cannot use the usual product-moment
correlation to indicate the relationship between the ranks. We can estimate, however,
what the product of rank values in a group of n subjects where the ranks are randomly
assigned would tend to be and estimate the variability of these sums or rank products for
repeated samples. This would lead to a test of significance of the departure of our rank
product sum (or average) from a value expected when there is no relationship.
A variety of non-parametric methods have been developed for nominal and
ordinal measures to indicate congruence or similarity among independent groups or
repeated measures on subjects in a group.
Contingency Chi-Square
The frequency chi-square statistic is used to accept or reject hypotheses
concerning the degree to which observed frequencies depart from theoretical frequencies
in a row by column contingency table with fixed marginal frequencies. It therefore tests
the independence of the categorical variables defining the rows and columns. As an
example, assume 50 males and 50 females are randomly assigned to each of three types
of instructional methods to learn beginning French, (a) using a language laboratory, (b)
using a computer with voice synthesizer and (c) using an advanced student tutor.
Following a treatment period, a test is administered to each student with scoring results
being pass or fail. The frequency of passing is then recorded for each cell in the 2 by 3
array (gender by treatment). If gender is independent of the treatment variable, the
expected frequency of males that pass in each treatment would be the same as the
expected frequency for females. The chi-squared statistic is obtained as
row col
(f
ij
- F
ij
)
2
i=1 j=1
2
= -----------------------
F
ij
where f
ij
is the observed frequency, F
ij
the expected frequency, and
2
is the chi-
squared statistic with degrees of freedom (rows - 1) times (columns - 1).
Example Contingency Chi Square
In this example we will use the data file ChiData.txt which consists of two
columns of data representing the row and column of a three by three contingency table.
The rows represent each observation with the row and column of that observation
recorded in columns one and two. We begin by selecting the Analyses / Non Parametric /
Contingency Chi Square option of the menu. The following figure demonstrates that the
row and column labels have been selected for the option of reading a data file containing
individual cases. We have also elected all options except saving the frequency file.
Figure 54 Contingency Chi-Square Dialogue Form.
When we click the compute button, we obtain the results shown below:
Chi-square Analysis Results
OBSERVED FREQUENCIES
Rows
Variables
COL.1 COL.2 COL.3 COL.4 Total
Row 1 5 5 5 5 20
Row 2 10 4 7 3 24
Row 3 5 10 10 2 27
Total 20 19 22 10 71
EXPECTED FREQUENCIES with 71 valid cases.
Variables
COL.1 COL.2 COL.3 COL.4
Row 1 5.634 5.352 6.197 2.817
Row 2 6.761 6.423 7.437 3.380
Row 3 7.606 7.225 8.366 3.803
ROW PROPORTIONS with 71 valid cases.
Variables
Row 1 0.250 0.250 0.250 0.250 1.000
Row 2 0.417 0.167 0.292 0.125 1.000
Row 3 0.185 0.370 0.370 0.074 1.000
Total 0.282 0.268 0.310 0.141 1.000
COLUMN PROPORTIONS with 71 valid cases.
Variables
Row 1 0.250 0.263 0.227 0.500 0.282
Row 2 0.500 0.211 0.318 0.300 0.338
Row 3 0.250 0.526 0.455 0.200 0.380
Total 1.000 1.000 1.000 1.000 1.000
PROPORTIONS OF TOTAL N with 71 valid cases.
Variables
Row 1 0.070 0.070 0.070 0.070 0.282
Row 2 0.141 0.056 0.099 0.042 0.338
Row 3 0.070 0.141 0.141 0.028 0.380
Total 0.282 0.268 0.310 0.141 1.000
CHI-SQUARED VALUE FOR CELLS with 71 valid cases.
Variables
COL.1 COL.2 COL.3 COL.4
Row 1 0.071 0.023 0.231 1.692
Row 2 1.552 0.914 0.026 0.043
Row 3 0.893 1.066 0.319 0.855
Chi-square = 7.684 with D.F. = 6. Prob. > value = 0.262
It should be noted that the user has the option of reading data in three different
formats. We have shown the first format where individual cases are classified by row
and column. It is sometimes more convenient to record the actual frequencies in each
row and cell combination. Examine the file labeled ChiSquareOne.TXT for such an
example. Sometimes the investigator may only know the cell proportions and the total
number of observations. In this case the third file format may be used where the
proportion in each row and column combination are recorded. See the example file
labeled ChiSquareTwo.TXT.
Spearman Rank Correlation
When the researchers data represent ordinal measures such as ranks with some
observations being tied for the same rank, the Rank Correlation may be the appropriate
statistic to calculate. While the computation for the case of untied cases is the same as
that for the Pearson Product-Moment correlation, the correction for tied ranks is found
only in the Spearman correlation. In addition, the interpretation of the significance of the
Rank Correlation may differ from that of the Pearson Correlation where bivariate
normalcy is assumed.
Example Spearman Rank Correlation
We will use the file labeled Spearman.txt for our example. The third variable
represents rank data with ties. Select the Statistics / Non Parametric / Spearman Rank
Correlation option from the menu. Shown below is the specification form for the
analysis:
Figure 55 Spearman Rank Correlation Dialogue Form.
Spearman Rank Correlation Between VAR2 & VAR3
Observed scores, their ranks and differences between ranks
VAR2 Ranks VAR3 Ranks Rank Difference
42.00 3.00 0.00 1.50 1.50
46.00 4.00 0.00 1.50 2.50
39.00 2.00 1.00 3.50 -1.50
37.00 1.00 1.00 3.50 -2.50
65.00 8.00 3.00 5.00 3.00
88.00 11.00 4.00 6.00 5.00
86.00 10.00 5.00 7.00 3.00
56.00 6.00 6.00 8.00 -2.00
62.00 7.00 7.00 9.00 -2.00
92.00 12.00 8.00 10.50 1.50
54.00 5.00 8.00 10.50 -5.50
81.00 9.00 12.00 12.00 -3.00
Spearman Rank Correlation = 0.615
t-test value for hypothesis r = 0 is 2.467
Probability > t = 0.0333
Notice that the original scores have been converted to ranks and where ties exist they
have been averaged.
Mann-Whitney U Test
An alternative to the Student t-test when the scale of measurement cannot be
assumed to be interval or ratio and the distribution of errors is unknown is a non-
parametric test known as the Mann-Whitney test. In this test, the dependent variable
scores for both groups are ranked and the number of times that one groups scores exceed
the rank of scores in the other group are recorded. This total number of times scores in
one group exceed those of the other is named U. The sampling distribution of U is
known and forms the basis for the hypothesis that the scores come from the same
population.
As an example, load the file labeled MannWhitU.txt and then select the option
Analyses / Non Parametric / Mann-Whitney U Test from the menu. Shown below is the
specification form in which we have indicated the analysis to perform:

Figure 56 Mann-Whitney U Test Dialogue Form.
Upon clicking the Compute button you obtain:
Mann-Whitney U Test
See pages 116-127 in S. Siegel: Nonparametric Statistics for the Behavioral Sciences
Score Rank Group
6.00 1.50 1
6.00 1.50 2
7.00 5.00 1
7.00 5.00 1
7.00 5.00 1
7.00 5.00 1
7.00 5.00 1
8.00 9.50 1
8.00 9.50 2
8.00 9.50 2
8.00 9.50 1
9.00 12.00 1
10.00 16.00 1
10.00 16.00 2
10.00 16.00 2
10.00 16.00 2
10.00 16.00 1
10.00 16.00 1
10.00 16.00 1
11.00 20.50 2
11.00 20.50 2
12.00 24.50 2
12.00 24.50 2
12.00 24.50 2
12.00 24.50 2
12.00 24.50 1
12.00 24.50 1
13.00 29.50 1
13.00 29.50 2
13.00 29.50 2
13.00 29.50 2
14.00 33.00 2
14.00 33.00 2
14.00 33.00 2
15.00 36.00 2
15.00 36.00 2
15.00 36.00 2
16.00 38.00 2
17.00 39.00 2
Sum of Ranks in each Group
Group Sum No. in Group
1 200.00 16
2 580.00 23
No. of tied rank groups = 9
Statistic U = 304.0000
z Statistic (corrected for ties) = 3.4262, Prob. > z = 0.0003
Fishers Exact Test
Assume you have collected data on principals and superintendents concerning
their agreement or disagreement to the statement "high school athletes observed drinking
or using drugs should be barred from further athletic competition". You record their
responses in a table as below:
Disagree Agree
Superintendents 2 8
Principals 4 5
You ask, are the responses of superintendents and principals significantly different?
Another way to ask the question is, "what is the probability of getting the pattern of
responses observed or a more extreme pattern?". The probability of any given pattern of
responses in this 2 by 2 table may be calculated from the hypergeometric probability
distribution as
(A+B)!(C+D)!(A+C)!(B+D)!
P = --------------------------------------
N!A!B!C!D!
where A, B, C, and D correspond to the frequencies in the four quadrants of the table and
N corresponds to the total number of individuals sampled.
When you elect the Analyses / NonParametric / Fishers Exact Test option from
the menu, you are shown a specification form which provides for four different formats
for entering data. We have elected the last format (entry of frequencies on the form
itself):
Figure 57 Fischer's Exact Test Dialogue Form.
Fisher Exact Probability Test
Contingency Table for Fisher Exact Test
Column
Row 1 2
1 2 8
2 4 5
Probability := 0.2090
Cumulative Probability := 0.2090
Column
Row 1 2
1 1 9
2 5 4
Column
Row 1 2
1 0 10
2 6 3
Tocher ratio computed: 0.002
A random value of 0.893 selected was greater than the Tocher value.
Conclusion: Accept the null Hypothesis
Notice that the probability of each combination of cell values as extreme or more extreme
than that observed is computed and the probabilities summed.
Alternative formats for data files are the same as for the Contingency Chi Square
analysis discussed in the previous section.
Kendalls Coefficient of Concordance
It is not uncommon that a group of people are asked to judge a group of persons
or objects by rank ordering them from highest to lowest. It is then desirable to have some
index of the degree to which the various judges agreed, that is, ranked the objects in the
same order. The Coefficient of Concordance is a measure varying between 0 and 1 that
indicates the degree of agreement among judges. It is defined as:
W = Variance of rank sums / maximum variance of rank sums.
The coefficient W may also be used to obtain the average rank correlation among the
judges by the formula:
M
r
= (mW - 1) / (m - 1)
where M
r
is the average (Spearman) rank correlation, m is the number of judges and W
is the Coefficient of Concordance.
Our example analysis will use the file labeled Concord2.txt . Load the file and
select the Analyses / NonParametric / Coefficient of Concordance option. Shown below
is the form completed for the analysis:
Figure 58 Kendal's Coefficient of Concordance.
Clicking the Compute button results in the following output:
Kendall Coefficient of Concordance Analysis
Ranks Assigned to Judge Ratings of Objects
Judge 1 Objects
VAR1 VAR2 VAR3 VAR4 VAR5 VAR6 VAR7 VAR8
1.5000 3.5000 3.5000 5.5000 5.5000 7.0000 8.0000
Judge 2 Objects
2.0000 3.0000 4.0000 5.0000 6.0000 7.0000 8.0000
Judge 3 Objects
2.5000 2.5000 2.5000 6.5000 6.5000 6.5000 6.5000
Sum of Ranks for Each Object Judged
Objects
6.0000 9.0000 10.0000 17.0000 18.0000 20.5000 22.5000
Coefficient of concordance = 0.942
Average Spearman Rank Correlation = 0.913
Chi-Square Statistic = 19.777
Probability of a larger Chi-Square = 0.0061
If you are observing competition in the Olympics or other athletic competitions, it is fun
to record the judges scores and examine the degree to which there is agreement among
them!
Kruskal-Wallis One-Way ANOVA
One-Way, Fixed-Effects Analysis of Variance assumes that error (residual) scores
are normally distributed, that subjects are randomly selected from the population and
assigned to treatments, and that the error scores are equally distributed in the populations
representing the treatments. The scale of measurement for the dependent variable is
assumed to be interval or ratio. But what can you do if, in fact, your measure is only
ordinal (for example like most classroom tests), and you cannot assume normally
distributed, homoscedastic error distributions?
Why, of course, you convert the scores to ranks and ask if the sum of rank scores
in each treatment group are the same within sampling error! The Kruskal-Wallis One-
Way Analysis of variance converts the dependent score for each subject in the study to a
rank from 1 to N. It then examines the ranks attained by subjects in each of the treatment
groups. Then a test statistic which is distributed as Chi-Squared with degrees of freedom
equal to the number of treatment groups minus one is obtained from:
12 K
H = ---------- 3 R
j
2
/ n
j
- 3(N + 1)
N(N + 1) j=1
where N is the total number of subjects in the experiment, n
j
is the number of subjects in
the jth treatment, K is the number of treatments and R
j
is the sum of ranks in the jth
treatment.
The Analyses / NonParametric / Kruskal-Wallis One-Way ANOVA option on
your menu will permit analysis of data in a data file. At least two variables should be
defined for each case - a variable recording the treatment group for the case and a
variable containing the dependent variable score.
As an example, load the file labeled kwanova.txt into the data grid and select the
menu option for the analysis. Below is the form and the results of the analysis:
Figure 59Kruskal-Wallis One Way ANOVA on Ranks Dialogue Form
Kruskal - Wallis One-Way Analysis of Variance
See pages 184-194 in S. Siegel's Nonparametric Statistics for the Behavioral Sciences
Score Rank Group
61.00 1.00 1
82.00 2.00 2
83.00 3.00 1
96.00 4.00 1
101.00 5.00 1
109.00 6.00 2
115.00 7.00 3
124.00 8.00 2
128.00 9.00 1
132.00 10.00 2
135.00 11.00 2
147.00 12.00 3
149.00 13.00 3
166.00 14.00 3
Sum of Ranks in each Group
Group Sum No. in Group
1 22.00 5
2 37.00 5
3 46.00 4
No. of tied rank groups = 0
Statisic H uncorrected for ties = 6.4057
Correction for Ties = 1.0000
Statistic H corrected for ties = 6.4057
Corrected H is approx. chi-square with 2 D.F. and probability = 0.0406
Wilcoxon Matched-Pairs Signed Ranks Test
This test provides an alternative to the student t-test for matched score data where
the assumptions for the parametric t-test cannot be met. In using this test, the difference
is obtained between each of N pairs of scores observed on matched objects, for example,
the difference between pretest and post-test scores for a group of students. The
difference scores obtained are then ranked. The ranks of negative score differences are
summed and the ranks of positive score differences are summed. The test statistic T is
the smaller of these two sums. Difference scores of 0 are eliminated since a rank cannot
be assigned. If the null hypothesis of no difference between the groups of scores is true,
the sum of positive ranks should not differ from the sum of negative ranks beyond that
expected by chance. Given N ranks, there is a finite number of ways of obtaining a given
sum T. There are a total of 2 raised to the N ways of assigning positive and negative
differences to N ranks. In a sample of 5 pairs, for example, there are 2 to the fifth power
= 32 ways. Each rank sign would occur with probability of 1/32. The probability of
getting a particular total T is
Ways of getting T
P
T
= -----------------------
2
N
The cumulative probabilities for T, T-1,,0 are obtained for the observed T value and
reported. For large samples, a normally distributed z score is approximated and used.
Our example uses the file labeled Wilcoxon.txt. Load this file and select the
Analyses / NonParametric / Wilcoxon Matched-Pairs Signed Ranks Test option from the
menu. The specification form and results are shown below:
Figure 60 Wicoxon Matched Pairs Signed Ranks Test Dialogue Form.
The Wilcoxon Matched-Pairs Signed-Ranks Test
See pages 75-83 in S. Seigel's Nonparametric Statistics for the Social Sciences
Ordered Cases with cases having 0 differences eliminated:
Number of cases with absolute differences greater than 0 = 8
CASE VAR1 VAR2 Difference Signed Rank
3 73.00 74.00 -1.00 -1.00
8 65.00 62.00 3.00 2.00
7 76.00 80.00 -4.00 -3.00
4 43.00 37.00 6.00 4.00
5 58.00 51.00 7.00 5.00
6 56.00 43.00 13.00 6.00
1 82.00 63.00 19.00 7.00
2 69.00 42.00 27.00 8.00
Smaller sum of ranks (T) = 4.00
Approximately normal z for test statistic T = 1.960
Probability (1-tailed) of greater z = 0.0250
NOTE: For N < 25 use tabled values for Wilcoxon Test
Cochran Q Test
The Cochran Q test is used to test whether or not two or more matched sets of
frequencies or proportions, measured on a nominal or ordinal scale, differ significantly
among themselves. Typically, observations are dichotomous, that is, scored as 0 or 1
depending on whether or not the subject falls into one or the other criterion group. An
example of research for which the Q test may be applied might be the agreement or
disagreement to the question "Should abortions be legal?". The research design might
call for a sample of n subjects answering the question prior to a debate and following a
debate on the topic and subsequently six months later. The Q test applied to these data
would test whether or not the proportion agreeing was the same under these three time
periods. The Q statistic is obtained as
K K
(K - 1) G
j
2
- ( G
j
)
2
j=1 j=1
Q = ----------------------------------
n n
K L
i
- L
i
2
i=1 i=1
where K is the number of treatments (groups of scores, G
j
is the sum with
the jth treatment group, and L
i
is the sum within case i (across groups). The Q statistic is
distributed approximately as Chi-squared with degrees of freedom K-1. If Q exceeds the
Chi-Squared value corresponding to the 1-3 cumulative probability value, the hypothesis
of equal proportions for the K groups is rejected.
Load the file labeled Qtest.txt and select the Analyses / NonParametric / Cochran
Q Test option from the menu. Shown below is the specification form completed for the
analysis of the file data and the results obtained when you click the Compute button:
Figure 61 Cochran Q Test Dialogue Form.
Cochran Q Test for Related Samples
See pages 161-166 in S. Siegel's Nonparametric Statistics for the Behavioral Sciences
McGraw-Hill Book Company, New York, 1956
Cochran Q Statistic = 16.667
which is distributed as chi-square with 2 D.F. and probability = 0.0002
Sign Test
Imagine a counseling psychologist who sees, over a period of months, a number
of clients with personal problems. Suppose the psychologist routinely contacts each
client for a six-month follow-up to see how they are doing. The counselor could make an
estimate of client "adjustment" before treatment and at the follow-up time (or better still,
have another person independently estimate adjustment at these two time periods). We
may assume some underlying continuous "adjustment" variable even though we have no
idea about the population distribution of the variable. We are interested in knowing, of
course, whether or not people are better adjusted six months after therapy than before.
Note that we are only comparing the "before" and "after" state of the individuals with
each other, not with other subjects. If we assign a + to the situation of improved
adjustment and a - to the situation of same or poorer adjustment, we have the data
required for a Sign Test. If treatment has had no effect, we would expect approximately
one half the subjects would receive plus signs and the others negative signs. The
sampling distribution of the proportion of plus signs is given by the binomial probability
distribution with parameter of .5 and the number of events equal to n, the number of pairs
of observations.
The file labeled SignTest.txt contains male and female cases in which have been
matched on relevant criteria and observations have been made on a 5-point Likert-type
instrument. The program will compare the two scores for each pair and assign a positive
or negative difference indicator. Load the file into the data grid and select the Analyses /
NonParametric / Sign Test option. Shown below is the specification form which appears
and the results obtained when clicking the Compute button:
Figure 62 The Matched Pairs Sign Test Dialogue Form.
Results for the Sign Test
Frequency of 11 out of 17 observed + sign differences.
Frequency of 3 out of 17 observed - sign differences.
Frequency of 3 out of 17 observed no differences.
The theoretical proportion expected for +'s or -'s is 0.5
The test is for the probability of the +'s or -'s (which ever is fewer)
as small or smaller than that observed given the expected proportion.
Binary Probability of 0 = 0.0001
Binomial Probability of 3.00 or smaller out of 14.00 = 0.0287
Friedman Two Way ANOVA
Imagine an experiment using, say, ten groups of subjects with four subjects in
each group that have been matched on some relevant variables (or even using the same
subjects). The matched subjects in each group are exposed to four different treatments
such as teaching methods, dosages of medicine, proportion of positive responses to
statements or questions, etc. Assume that some criterion measure on at least a nominal
scale is available to measure the effect of each treatment. Now rank the subjects in each
group on the basis of their scores on the criterion. We may now ask whether the ranks in
each treatment come from the same population. Had we been able to assume an interval
or ratio measure and normally distributed errors, we might have used a repeated measures
analysis of variance. Failing to meet the parametric test assumptions, we instead examine
the sum of ranks obtained under each of the treatment conditions and ask whether they
differ significantly. The test statistic is distributed as Chi-squared with degrees of
freedom equal to the number of treatments minus one. It is obtained as where N is the
number of groups, K the number of treatments (or number of subjects in each group), and
R
j
is the sum of ranks in each treatment.
For an example analysis, load the file labeled Friedman.txt and select Analyses /
NonParametric / Friedman Two Way ANOVA from the menu. The data represent four
treatments or repeated measures for three groups, each containing one subject. Shown
below is the specification form and the results following a click of the Compute button:
Figure 63 The Friedman Two-Way ANOVA Dialogue Form.
FRIEDMAN TWO-WAY ANOVA ON RANKS
See pages 166-173 in S. Siegels Nonparametric Statistics
for the Behavioral Sciences, McGraw-Hill Book Co., New York, 1956
Treatment means - values to be ranked.

Cond.1 Cond.2 Cond.3 Cond.4
Group 1 9.000 4.000 1.000 7.000
Group 2 6.000 5.000 2.000 8.000
Group 3 9.000 1.000 2.000 6.000
Number in each group's treatment.
GROUP
Group 1 1 1 1 1
Group 2 1 1 1 1
Group 3 1 1 1 1
Score Rankings Within Groups
Treatment
Group 1 4.000 2.000 1.000 3.000
Group 2 3.000 2.000 1.000 4.000
Group 3 4.000 1.000 2.000 3.000
TOTAL RANKS
Variables Cond.1 Cond.2 Cond.3 Cond.4
11.000 5.000 4.000 10.000
Chi-square with 3 D.F. = 7.400 with probability = 0.0602
Chi-square too approximate-use exact table (TABLE N)
page 280-281 in Siegel
Probability of a Binomial Event
The BINOMIAL program is a short program to calculate the probability of
obtaining k or fewer occurrences of a dichotomous variable out of a total of n
observations when the probability of an occurrence is known. For example, assume a test
consists of 5 multiple choice items with each item scored correct or incorrect. Also
assume that there are five equally plausible choices for a student with no knowledge
concerning any item. In this case, the probability of a student guessing the correct
answer to a single item is 1/5 or .20 . We may use the binomial program to obtain the
probabilities that a student guessing on each item of the test gets a score of 0, 1, 2, 3, 4, or
5 items correct by chance alone.
The formula for the probability of a dichotomous event k where the probability of
a single event is p (and the probability of a non-event is q = 1 - p is given as:
N!
P(k) = ------------- p
(N-k)
q
k
(N - k)! k!
For example, if a fair coin is tossed three times with the probabilities of heads is p = .5
(and q = .5) then the probabilty of observing 2 heads is
3!
P(2) = ------------0.5
1
x 0.5
2
(3-2)! 2!
3 x 2 x 1
= -------------- x 0.5 x 0.25
1 x (2 x 1)
6
= --------- x 0.125 = .375
2
Similarly, the probability of getting one toss turn up heads is
3! 6
P(1) = ------------ 0.5
2
x 0.5 = ----------- x 0.25 x 0.5 = .375
(3-1)! 1! 2
and the probability of getting zero heads turn up in three tosses is
3! 6
P(0) = ------------ 0.5
0
x 0.5
3
= ----------x 1.0 x 0.125 = 0.125
(3-0)! 0! 6
The probability of getting 2 or fewer heads in three tosses is the sum of the three
probabilities, that is, 0.375 + 0.375 + 0.125 = 0.875 .
As a check on the above, select the Analyses / NonParametric / Binomial Probability
option from the menu. Enter the values as shown in the specification form below and
press the Compute button to obtain the shown results.
Figure 64 The Binomial Probability Dialogue Form.
Binomial Probability Test
Frequency of 2 out of 3 observed
The theoretical proportion expected in category A is 0.500
The test is for the probability of a value in category A as small or smaller
than that observed given the expected proportion.
Probability of 0 = 0.1250
Binomial Probability of 2 or less out of 3 = 0.8750
Runs Test
Random sampling is a major assumption of nearly all statistical tests of
hypotheses. The Runs test is one method available for testing whether or not an obtained
sample is likely to have been drawn at random. It is based on the order of the values in
the sample and the number of values increasing or decreasing in a sequence. For
example, if a variable is composed of dichotomous values such as zeros (0) and ones (1)
then a run of values such as 0,0,0,0,1,1,1,1 would not likely to have been selected at
random. As another example, the values 0,1,0,1,0,1,0,1 show a definite cyclic pattern
and also would not likely be found by random sampling. The test involves finding the
mean of the values and examining values above and below the mean (excluding values at
the mean.) The values falling above or below the mean should occur in a random
fashion. A run consists of a series of values above the mean or below the mean. The
expected value for the total number of runs is known and is a function of the sample size
(N) and the numbers of values above (N1) and below (N2) the mean. This test may be
applied to nominal through ratio variable types.
EXAMPLE:
The figure below shows a data set with 14 values in a file labeled
"RunsTest.OS4". The Runs Test option was selected from the NonParametric sub-menu
under the Analyses menu. The next figure displays the dialogue box used for specifying
the variable to analyze and the results of clicking the compute button.
Figure 65 A Sample File for the Runs Test
Figure 66 The Runs Dialogue Form.
In the above, we selected the first variable to test for randomness. There was little
evidence that the sample was not obtained by random selection.
Kendall's Tau and Partial Tau
When two variables are at least ordinal, the tau correlation may be obtained as a
measure of the relationship between the two variables. The values of the two variables
are ranked. The method involves ordering the values using one of the variables. If the
values of the other variable are in the same order, the correlation would be 1.0. If the
order is exactly the opposite for this second variable, the correlation would be -1.0 just as
if we had used the Pearson Product-Moment correlation method. Each pair of ranks for
the second variable are compared. If the order (from low to high) is correct for a pair it is
assigned a value of +1. If the pair is in reverse order, it is assigned a value of -1. These
values are summed. If there are N values then we can obtain the number of pairs of
scores for one variable as the number of combinations of N things taken 2 at a time which
is N(N-1). The tau statistic is the ratio of the sum of 1's and -1's to the total number of
pairs. Adjustments are made in the case of tied scores. For samples larger than 10, tau is
approximately normally distributed.
Whenever two variables are correlated, the relationship observed may, in part, be
due to their common relationship to a third variable. We may be interested in knowing
what the relationship is if we partial out this third variable. The Partial Tau provides this.
Since the distribution of the partial tau is not known, no test of significance is included.
Figure 67 Kendal's Tau and Partial Tau Dialog Form.
RANKS
X Y Z
1 3.000 2.000 1.500
2 4.000 6.000 1.500
3 2.000 5.000 3.500
4 1.000 1.000 3.500
5 8.000 10.000 5.000
6 11.000 9.000 6.000
7 10.000 8.000 7.000
8 6.000 3.000 8.000
9 7.000 4.000 9.000
10 12.000 12.000 10.500
11 5.000 7.000 10.500
12 9.000 11.000 12.000
Kendall Tau for File: C:\Owner\My Documents\Stat4U\TAUDATA.txt
Kendall Tau for variables X and Y
Tau = 0.6667 z = 3.017 probability > |z| = 0.001
Kendall Tau for variables X and Z
Kendall Tau for variables Y and Z
Partial Tau = 0.6136
NOTE: Probabilities are for large N (>10)
At the time this program was written, the distribution of the Partial Tau was
unknown (see Siegel 1956, page 228).
Kaplan-Meier Survival Test
Survival analysis is concerned with studying the occurrence of an event such as death or
change in a subject or object at various times following the beginning of the study.
Survival curves show the percentage of subjects surviving at various times as the study
progresses. In many cases, it is desired to compare survival of an experimental treatment
with a control treatment. This method is heavily used in medical research but is not
restricted to that field. For example, one might compare the rate of college failure among
students in an experimental versus a control group.
To obtain a survival curve you need only two items of information in your data file for
each subject: the survival time and a code for whether or not an event occurred or the
subject has been lost from the study moved, disappeared, etc. (censored.) If an event
such as death occurred, it is coded as a 1. If censored it is coded as a 2.
CASES FOR FILE C:\OpenStat\KaplanMeier1.TEX
0 Time Event_Censored
1 1 2
2 3 2
3 5 2
4 6 1
5 6 1
6 6 1
7 6 1
8 6 1
9 6 1
10 8 1
11 8 1
12 9 2
13 10 1
14 10 1
15 10 2
16 12 1
17 12 1
18 12 1
19 12 1
20 12 1
21 12 1
22 12 2
23 12 2
24 13 2
25 15 2
26 15 2
27 16 2
28 16 2
29 18 2
30 18 2
31 20 1
32 20 2
33 22 2
34 24 1
35 24 1
36 24 2
37 27 2
38 28 2
39 28 2
40 28 2
41 30 1
42 30 2
43 32 1
44 33 2
45 34 2
46 36 2
47 36 2
48 42 1
49 44 2
We are really recording data for the "Time" variable that is sequential through the data
file. We are concerned with the percent of survivors at any given time period as we
progress through the observation times of the study. We record the "drop-outs" or
censored subjects at each time period also. A unit cannot be censored and be one of the
deaths - these are mutually exclusive.
Next we show a data file that contains both experimental and control subjects:
CASES FOR FILE C:\OpenStat\KaplanMeier2.TEX
0 Time Group Event_Censored
1 1 1 2
2 3 2 2
3 5 1 2
4 6 1 1
5 6 1 1
6 6 2 1
7 6 2 1
8 6 2 1
9 6 2 1
10 8 2 1
11 8 2 1
12 9 1 2
13 10 1 1
14 10 1 1
15 10 1 2
16 12 1 1
17 12 1 1
18 12 1 1
19 12 1 1
20 12 2 1
21 12 2 1
22 12 1 2
23 12 2 2
24 13 1 2
25 15 1 2
26 15 2 2
27 16 1 2
28 16 2 2
29 18 2 2
30 18 2 2
31 20 2 1
32 20 1 2
33 22 2 2
34 24 1 1
35 24 2 1
36 24 1 2
37 27 1 2
38 28 2 2
39 28 2 2
40 28 2 2
41 30 2 1
42 30 2 2
43 32 1 1
44 33 2 2
45 34 1 2
46 36 1 2
47 36 1 2
48 42 2 1
49 44 1 2
In this data we code the groups as 1 or 2. Censored cases are always coded 2 and Events
are coded 1. This data is, in fact, the same data as shown in the previous data file. Note
that in time period 6 there were 6 deaths (cases 4-9.) Again, notice that the time periods
are in ascending order.
Shown below is the specification dialog for this second data file. This is followed by the
output obtained when you click the compute button.
Figure 68 The Kaplan-Meier Dialog
Comparison of Two Groups Methd
TIME GROUP CENSORED TOTAL AT EVENTS AT RISK IN EXPECTED NO. AT RISK IN EXPECTED NO.
RISK GROUP 1 EVENTS IN 1 GROUP 2 EVENTS IN 2
0 0 0 49 0 25 0.0000 24 0.0000
1 1 1 49 0 25 0.0000 24 0.0000
3 2 1 48 0 24 0.0000 24 0.0000
5 1 1 47 0 24 0.0000 23 0.0000
6 1 0 46 6 23 3.0000 23 3.0000
6 1 0 40 0 21 0.0000 19 0.0000
6 2 0 40 0 21 0.0000 19 0.0000
6 2 0 40 0 21 0.0000 19 0.0000
6 2 0 40 0 21 0.0000 19 0.0000
6 2 0 40 0 21 0.0000 19 0.0000
8 2 0 40 2 21 1.0500 19 0.9500
8 2 0 38 0 21 0.0000 17 0.0000
9 1 1 38 0 21 0.0000 17 0.0000
10 1 0 37 2 20 1.0811 17 0.9189
10 1 0 35 0 18 0.0000 17 0.0000
10 1 1 35 0 18 0.0000 17 0.0000
12 1 0 34 6 17 3.0000 17 3.0000
12 1 0 28 0 13 0.0000 15 0.0000
12 1 0 28 0 13 0.0000 15 0.0000
12 1 0 28 0 13 0.0000 15 0.0000
12 2 0 28 0 13 0.0000 15 0.0000
12 2 0 28 0 13 0.0000 15 0.0000
12 1 1 28 0 13 0.0000 15 0.0000
12 2 1 27 0 12 0.0000 15 0.0000
13 1 1 26 0 12 0.0000 14 0.0000
15 1 1 25 0 11 0.0000 14 0.0000
15 2 1 24 0 10 0.0000 14 0.0000
16 1 1 23 0 10 0.0000 13 0.0000
16 2 1 22 0 9 0.0000 13 0.0000
18 2 1 21 0 9 0.0000 12 0.0000
18 2 1 20 0 9 0.0000 11 0.0000
20 2 0 19 1 9 0.4737 10 0.5263
20 1 1 18 0 9 0.0000 9 0.0000
22 2 1 17 0 8 0.0000 9 0.0000
24 1 0 16 2 8 1.0000 8 1.0000
24 2 0 14 0 7 0.0000 7 0.0000
24 1 1 14 0 7 0.0000 7 0.0000
27 1 1 13 0 6 0.0000 7 0.0000
28 2 1 12 0 5 0.0000 7 0.0000
28 2 1 11 0 5 0.0000 6 0.0000
28 2 1 10 0 5 0.0000 5 0.0000
30 2 0 9 1 5 0.5556 4 0.4444
30 2 1 8 0 5 0.0000 3 0.0000
32 1 0 7 1 5 0.7143 2 0.2857
33 2 1 6 0 4 0.0000 2 0.0000
34 1 1 5 0 4 0.0000 1 0.0000
36 1 1 4 0 3 0.0000 1 0.0000
36 1 1 3 0 2 0.0000 1 0.0000
42 2 0 2 1 1 0.5000 1 0.5000
44 1 1 0 0 1 0.0000 0 0.0000
TIME DEATHS GROUP AT RISK PROPORTION CUMULATIVE
SURVIVING PROP.SURVIVING
1 0 1 25 0.0000 1.0000
3 0 2 24 0.0000 1.0000
5 0 1 24 0.0000 1.0000
6 6 1 23 0.9130 0.9130
6 0 1 21 0.0000 0.9130
6 0 2 19 0.0000 0.8261
6 0 2 19 0.0000 0.8261
6 0 2 19 0.0000 0.8261
6 0 2 19 0.0000 0.8261
8 2 2 19 0.8947 0.7391
8 0 2 17 0.0000 0.7391
9 0 1 21 0.0000 0.9130
10 2 1 20 0.9000 0.8217
10 0 1 18 0.0000 0.8217
10 0 1 18 0.0000 0.8217
12 6 1 17 0.7647 0.6284
12 0 1 13 0.0000 0.6284
12 0 1 13 0.0000 0.6284
12 0 1 13 0.0000 0.6284
12 0 2 15 0.0000 0.6522
12 0 2 15 0.0000 0.6522
12 0 1 13 0.0000 0.6284
12 0 2 15 0.0000 0.6522
13 0 1 12 0.0000 0.6284
15 0 1 11 0.0000 0.6284
15 0 2 14 0.0000 0.6522
16 0 1 10 0.0000 0.6284
16 0 2 13 0.0000 0.6522
18 0 2 12 0.0000 0.6522
18 0 2 11 0.0000 0.6522
20 1 2 10 0.9000 0.5870
20 0 1 9 0.0000 0.6284
22 0 2 9 0.0000 0.5870
24 2 1 8 0.8750 0.5498
24 0 2 7 0.0000 0.5136
24 0 1 7 0.0000 0.5498
27 0 1 6 0.0000 0.5498
28 0 2 7 0.0000 0.5136
28 0 2 6 0.0000 0.5136
28 0 2 5 0.0000 0.5136
30 1 2 4 0.7500 0.3852
30 0 2 3 0.0000 0.3852
32 1 1 5 0.8000 0.4399
33 0 2 2 0.0000 0.3852
34 0 1 4 0.0000 0.4399
36 0 1 3 0.0000 0.4399
36 0 1 2 0.0000 0.4399
42 1 2 1 0.0000 0.0000
44 0 1 1 0.0000 0.4399
Total Expected Events for Experimental Group = 11.375
Observed Events for Experimental Group = 10.000
Total Expected Events for Control Group = 10.625
Observed Events for Control Group = 12.000
Chisquare = 0.344 with probability = 0.442
Risk = 0.778, Log Risk = -0.250, Std.Err. Log Risk =
0.427
95 Percent Confidence interval for Log Risk = (-
1.087,0.586)
95 Percent Confidence interval for Risk = (0.337,1.796)
EXPERIMENTAL GROUP CUMULATIVE PROBABILITY
CASE TIME DEATHS CENSORED CUM.PROB.
1 1 0 1 1.000
3 5 0 1 1.000
4 6 6 0 0.913
5 6 0 0 0.913
12 9 0 1 0.913
13 10 2 0 0.822
14 10 0 0 0.822
15 10 0 1 0.822
16 12 6 0 0.628
17 12 0 0 0.628
18 12 0 0 0.628
19 12 0 0 0.628
22 12 0 1 0.628
24 13 0 1 0.628
25 15 0 1 0.628
27 16 0 1 0.628
32 20 0 1 0.628
34 24 2 0 0.550
36 24 0 1 0.550
37 27 0 1 0.550
43 32 1 0 0.440
45 34 0 1 0.440
46 36 0 1 0.440
47 36 0 1 0.440
49 44 0 1 0.440
CONTROL GROUP CUMULATIVE PROBABILITY
CASE TIME DEATHS CENSORED CUM.PROB.
2 3 0 1 1.000
6 6 0 0 0.826
7 6 0 0 0.826
8 6 0 0 0.826
9 6 0 0 0.826
10 8 2 0 0.739
11 8 0 0 0.739
20 12 0 0 0.652
21 12 0 0 0.652
23 12 0 1 0.652
26 15 0 1 0.652
28 16 0 1 0.652
29 18 0 1 0.652
30 18 0 1 0.652
31 20 1 0 0.587
33 22 0 1 0.587
35 24 0 0 0.514
38 28 0 1 0.514
39 28 0 1 0.514
40 28 0 1 0.514
41 30 1 0 0.385
42 30 0 1 0.385
44 33 0 1 0.385
48 42 1 0 0.000
The chi-square coefficient as well as the graph indicates no difference was found between
the experimental and control group beyond what is reasonably expected through random
selection from the same population.
Figure 69 Experimental and Control Curves
Measurement Related Procedures
OpenStat contains a variety of procedures related to the general area of
measurement. Each is described briefly in the following sections.
Classical Item Analysis Methods
Item Discrimination
If a test is constructed to test one predominant domain or area of achievement or
knowledge then each item of the test should correlate positively with a total score on the
test. The total score on the test is usually obtained by awarding a value of 1 to a student
if they get an item correct and a 0 if they miss it and summing across all items. On a 47
item test, a student that gets all items correct would therefore have a total score of 47
while the student that missed all items would have a score of 0.
We can correlate each item with the total score obtained by the students. We
may use the Pearson Product-Moment correlation formula (see the section on simple
correlation and regression) to do our calculations. We note however that we are
correlating a dichotomous variable (our item is scored 0 or 1) with a continuous variable
(total scores vary from 0 to the number of items in the test). This type of correlation is
also called a "Point-Biserial Correlation". Unfortunately, when one of the variables in
the product-moment correlation is dichotomous, the correlation is affected by the
proportion of scores in the dichotomous variable. If the proportion of 0 and 1 scores is
about the same (50% for each), the correlation may approach 1.0. When the split of the
dichotomous variable is quite disproportionate, say .2 and .8, then the correlation is
restricted to much lower values. This certainly makes interpretation of the point-biserial
correlation difficult. Nevertheless, a "good" test item will have positive correlations with
the total score of the test. If the correlation is negative, it implies that more knowledgable
students are more likely to have missed the item and less knowledgeable students likely
to have gotten the item correct! Clearly, such an item is inconsistent with the
measurement of the remaining items. Remember that the total score contains, as part of
the total, the score of each item. For that reason, the point-biserial correlation will tend to
be positive. A "corrected" point-biseral correlation can be obtained by first subtracting
the individual item score from the total score before calculating the correlation between
the item and total. If a test has many items, say more than 30, the correction will make
little difference in the correlation. When only a few items are administered however, the
correction should be applied.
The point-biserial correlation between test item and test total score is a measure of
how well the item discriminates between low and high achievement students. It is a
measure of item discrimination potential. Other item discrimination indices may also be
used. For example, one may simply use the difference between the proportion passing
the item in students ranking in the top half on the total score and the proportion passing
the item among students in the bottom half of the class. Another index, the biserial
correlation, may be calculated if one assumes that the dichotomously scored item is
actually an imprecise measure of a continuous variable. The biserial correlaiton may be
obtained using the formula:
______
r
bis
= r
pbis
p
i
q
i
/ y
i
where r
pbis
is the point-biserial correlation, p
i
and q
i
are the proportions passing and
failing the item, and yi is the ordinate of the normal curve corresponding to the
cumulative proportion pi.
Item difficulty
In classical test analysis, the difficulty of an item is indicated by the proportion of
subjects passing the item. An easy item therefore has values closer to 1.0 while more
difficult items have values closer to 0.0 . Since the mean of an item scored either 0 or 1
is the same as the proportion of subjects receiving scores of 1, the mean is the difficulty
of the item. An ideal yardstick has markings equally spaced across the ruler. This
permits its use to measure objects varying widely in length. Similarly, a test composed of
items equally spaced in difficulty permits reasonable precision in measuring subjects that
vary widely in their knowledge. With item difficulties known, one can select items along
the continuum from 0 to 1.0 so that the revised instrument has approximately equal
interval measurement. Unfortunately, the sample of subjects on which the item difficulty
estimates are based must adequately represent all of the subjects for which the instrument
is to be used. If another group of subjects that differs considerably in their knowledge is
used to estimate the item difficulties, quite different estimates can be obtained. In other
words, the item difficulty estimates obtained in classical test analysis methods are
dependent on the sample from which they are obtained. It would clearly be desirable to
have item parameter estimates that are invariant from group to group, that is, independent
of the subjects being measured by those items.
In our discussion we have not mentioned errors of measurement for individual
items. In classical test analysis procedures we must assume that each item measures with
the same precision and reliability as all other items. We usually assume that errors of
measurement for single items are normally distributed with a mean of zero and that these
errors contribute proportionally to the error of measurement of the total test score. Hence
the standard error of measurement is assumed equal for subjects scoring from 1 to 50 on a
50 item test!
The Item Analysis Program
The OpenStat package includes a program for item analysis using the Classical
test theory. The program provides for scoring test items that have been entered as 0's and
1's or as item choices coded as numbers or letters. If item choices are in your data file,
you will be asked to enter the correct choice for each item so that the program may
convert to 0 or 1 score values for each item. A set of items may consist of several
independent sub-tests. If more than one sub-test exists, you will be asked to enter the
sequence number of each item in the sub-tests. You may also elect to correct for
guessing in obtaining total scores for each subject. Either rights-wrongs or rights - 1/4
wrongs may be elected. Finally, you may weigh the items of a test to give more or less
credit in the total score to various items. An option is provided for printing the item
scores and sub-score totals. You may elect one of several methods to estimate the
reliability of the scored items. The sub-test means and standard deviations are computed
for the total scores and for each item. In addition, the point-biserial correlation of each
item with each sub-score total is obtained. Item characteristic curves are optionally
printed for each item. The curves are based on the sub-score in which the item is
included. The proportion of subjects at each decile on the sub-score that pass the item is
plotted. If a reasonably large number of subjects are analyzed, this will typically result in
an approximate "ogive" curve for each item with positive point-biserial correlations.
Examination of the plots will reveal the item difficulty and discrimination characteristics
for various ability score groups.
Example Item Analysis
A file labeled "ITEMREL.txt" will be loaded into the main grid. This file
contains observations on 17 subjects using a test of 8 items. The subject responses have
been coded 0 for wrong and 1 for a correct response. To analyze the data we select the
Analysis / Measurement / Classical Item Analysis and Reliability procedure. The form
for specifying the analysis is shown below:
Figure 70 Classical Item Analysis Form.
You can see by the above form that you can also include last and first names and /
or subject identification numbers in addition to item data. This is useful if you are
creating a list of subject scores (see the output options.) You will also note that there is
the capability of obtaining five different estimates of reliability. We have elected all of
them to demonstrate the similarities and differences among them. When you click the
Compute button, the following results are obtained:
Classical Test Analysis Results
Intercorrelations of items and scores
Items and Subscores
VAR1 VAR2 VAR3 VAR4 VAR5 VAR6
VAR1 1.000 0.537 0.383 0.293 0.228 0.174
VAR2 0.537 1.000 0.713 0.545 0.424 0.324
VAR3 0.383 0.713 1.000 0.764 0.595 0.455
VAR4 0.293 0.545 0.764 1.000 0.778 0.595
VAR5 0.228 0.424 0.595 0.778 1.000 0.764
VAR6 0.174 0.324 0.455 0.595 0.764 1.000
VAR7 0.124 0.231 0.324 0.424 0.545 0.713
VAR8 0.067 0.124 0.174 0.228 0.293 0.383
TOTAL 0.440 0.683 0.805 0.860 0.860 0.805
Items and Subscores
VAR7 VAR8 TOTAL
VAR1 0.124 0.067 0.440
VAR2 0.231 0.124 0.683
VAR3 0.324 0.174 0.805
VAR4 0.424 0.228 0.860
VAR5 0.545 0.293 0.860
VAR6 0.713 0.383 0.805
VAR7 1.000 0.537 0.683
VAR8 0.537 1.000 0.440
TOTAL 0.683 0.440 1.000
Means for Items, Sub-Scores, Total Score
Variables VAR1 VAR2 VAR3 VAR4 VAR5 VAR6
0.938 0.812 0.688 0.562 0.438 0.312
Variables VAR7 VAR8 TOTAL
0.188 0.062 4.000
Standard Deviations for Items, Sub-Scores and Total Score
Variables VAR1 VAR2 VAR3 VAR4 VAR5 VAR6
0.250 0.403 0.479 0.512 0.512 0.479
Variables VAR7 VAR8 TOTAL
0.403 0.250 2.422
Alpha Reliability Estimate for scale TOTAL = 0.8636 S.E. of Measurement = 0.894
KR#21 Reliability Estimate for scale TOTAL = 0.7532 S.E. of Measurement = 1.203
KR#20 Reliability Estimate for scale TOTAL = 0.8636 S.E. of Measurement = 0.894
Hoyt Unadjusted Test Rel. for scale TOTAL = 0.2559 S.E. of Measurement = 2.089
Hoyt Adjusted Test Rel. for scale TOTAL = 0.8636 S.E. of Measurement = 0.894
Hoyt Unadjusted Item Rel. for scale TOTAL = 0.0412 S.E. of Measurement = 2.372
Hoyt Adjusted Item Rel. for scale TOTAL = 0.1596 S.E. of Measurement = 2.220
STEPWISE KR#20 FOR SCALE TOTAL
KR#20 = 0.875, Mean = 1.000, Variance = 0.933
ITEM MEAN VARIANCE Pt.Bis. Correlation
VAR4 0.562 0.263 0.9428
VAR5 0.438 0.263 0.9428
KR#20 = 0.882, Mean = 1.688, Variance = 1.829
VAR4 0.562 0.263 0.9441
VAR5 0.438 0.263 0.8839
VAR3 0.688 0.229 0.8688
KR#20 = 0.886, Mean = 2.000, Variance = 2.933
VAR4 0.562 0.263 0.9117
VAR5 0.438 0.263 0.9117
VAR3 0.688 0.229 0.8131
VAR6 0.312 0.229 0.8131
KR#20 = 0.882, Mean = 2.812, Variance = 3.896
VAR4 0.562 0.263 0.9023
VAR5 0.438 0.263 0.8776
VAR3 0.688 0.229 0.8511
VAR6 0.312 0.229 0.7717
VAR2 0.812 0.163 0.7070
KR#20 = 0.882, Mean = 3.000, Variance = 4.933
VAR4 0.562 0.263 0.8787
VAR5 0.438 0.263 0.8787
VAR3 0.688 0.229 0.8151
VAR6 0.312 0.229 0.8151
VAR2 0.812 0.163 0.6701
VAR7 0.188 0.163 0.6701
KR#20 = 0.870, Mean = 3.938, Variance = 5.396
VAR4 0.562 0.263 0.8718
VAR5 0.438 0.263 0.8648
VAR3 0.688 0.229 0.8206
VAR6 0.312 0.229 0.7981
VAR2 0.812 0.163 0.6986
VAR7 0.188 0.163 0.6541
VAR1 0.938 0.062 0.4520
KR#20 = 0.864, Mean = 4.000, Variance = 5.867
VAR4 0.562 0.263 0.8595
VAR5 0.438 0.263 0.8595
VAR3 0.688 0.229 0.8049
VAR6 0.312 0.229 0.8049
VAR2 0.812 0.163 0.6828
VAR7 0.188 0.163 0.6828
VAR1 0.938 0.062 0.4404
VAR8 0.062 0.062 0.4404
Figure 71 Example Item Characteristic Curve.
Determinant of correlation matrix = 0.0084
Multiple Correlation Coefficients for Each Variable
Variable R R2 F Prob.>F DF1 DF2
VAR1 0.537 0.289 0.464 0.836 7 8
VAR2 0.768 0.590 1.643 0.251 7 8
VAR3 0.842 0.709 2.786 0.087 7 8
VAR4 0.864 0.746 3.357 0.056 7 8
VAR5 0.864 0.746 3.357 0.056 7 8
VAR6 0.842 0.709 2.786 0.087 7 8
VAR7 0.768 0.590 1.643 0.251 7 8
VAR8 0.537 0.289 0.464 0.836 7 8
Figure 72 Distribution of Test Scores (Classical Test Analysis Procedure.)
Figure 73 Test Item Difficulty Plot (Classical Test Analysis Procedure.)
Rasch One Parameter Logistic Scaling
The One Parameter Logistic Model
In the classical approach to test theory, the item difficulty parameter is estimated
by the proportion of subjects in some "norming" group that passes the item. Other
methods may be used however, to estimate item difficulty parameters. George Rasch
developed one such method. In his model, all items are assumed to have equal item
characteristic slopes and little relevant chance probabilities. The probability of a subject
answering an item correctly is given by the formula
D(d
j
- b
i
)
e
P(X=1|b
i
) = _____________
D(d
j
- b
i
)
1.0 - e
where b
i
is the ability of an individual,
d
j
is the difficulty of item j, D is an arbitrary scaling or expansion factor, and
e is the constant 2.7182818.....( the base of the natural logarithm system).
An individual's ability b
i
is estimated by the product of the expansion factor D and
the natural log odds of obtaining a score X out of K items, i.e.,
X
b
i
= D log ____
K - X
The above equation may also be solved for X, the subject's raw score X expected
given his or her ability, that is
(b
i
/ D)
K e
X
i
= ______________
(b
i
/ D)
1 + e
The expansion factor D is a value which reflects the variability of both item difficulties d
j
and abilities b
i
. When scores are approximately normally distributed, this value is
frequently about 1.7.
The Rasch one-parameter logistic model assumes that all items in the test that are
analyzed measure a common latent variable. Researchers sometimes will complete a
factor analysis of their test to ascertain this unidimensional property prior to estimating
item difficulties using the Rasch model. Items may be selected from a larger group of
items that "load" predominantly on the first factor of the set of items.
The OpenStat package includes a program to analyze subject responses to a set of
items. The results include estimates of item difficulties in log units and their standard
errors. Ability estimates in log units and errors of estimate are also obtained for subjects
in each raw total score group. One cannot estimate abilities for subjects that miss all
items or correctly answer all items. In addition, items that all subjects miss or get correct
cannot be scaled. Such subjects or items are automatically eliminated by the program.
The program will also produce item characteristic curves for each item and report the
point-biserial correlation and the biserial correlation of each item with the total test score.
The Rasch method of calibrating item difficulty and subject ability has several
desirable properties. One can demonstrate that the item difficulties estimated are
independent of the abilities of subjects on which the estimates are based. For example,
should you arbitrarily divide a large group of subjects into two groups, those who have
total scores in the top half of the class and those who have scores in the bottom half of the
class, then complete the Rasch analysis for each group, you will typically obtain item
difficulties from each analysis that vary little from each other. This "sample-free"
property of the item difficulties does not, of course, hold for item difficulties estimated by
classical methods, i.e. proportion of a sample passing an item. Ability estimates of
individuals are similarly "item-free". A subset of items selected from a pool of Rasch
calibrated items may be used to obtain the same ability estimates of an individual that
would be obtained utilizing the entire set of items (within errors of estimation). This
aspect of ability estimation makes the Rasch scaled items ideal for "tailored" testing
wherein a subject is sequentially given a small set of items which are optimized to give
maximum precision about the estimated ability of the subject.
Estimating Parameters in the Rasch Model: Prox. Method
Item difficulties and subject abilities in the Rasch model are typically expressed in
base e logarithm values. Typical values for either difficulties or abilities range between -
3.0 and 3.0 somewhat analogous to the normally distributed z scores. We will work
through a sample to demonstrate the calculations typically employed to estimate the item
difficulties of a short test of 11 items administered to 127 individuals (See Applied
Psychometrics by R.L. Thorndike, 1982, pages 98-100). In estimating the parameters, we
will assume the test items involved the student in generating a response (not multiple
choice or true false) so that the probability of getting the item correct by chance is zero.
We will also assume that the items all have equal slopes, that is, that the change in
probability of getting an item correct for a given change in student ability is equal for all
items. By making these assumptions we need only solve for the difficulty of the item.
The first task in estimating our parameters is to construct a matrix of item failures
for subjects in each total score group. A total score group is the group of subjects that
have the same score on the test (where the total score is simply the total number of items
correctly answered). Our matrix will have the total test score as columns and individual
items as rows. Each element of the matrix will represent the number of students with the
same total test score that failed a particular item. Our sample matrix is
TOTAL TEST SCORE TOTAL
1 2 3 4 5 6 7 8 9 10 FAILED
ITEM
1 10 10 10 7 7 4 2 1 0 0 51
2 10 14 14 12 17 12 5 1 0 0 85
3 10 14 11 11 7 6 3 0 0 0 62
4 1 1 0 1 0 0 0 0 0 0 3
5 10 8 9 6 6 3 1 1 0 0 44
6 10 14 14 15 21 21 12 6 2 1 116
7 10 14 11 13 19 22 8 5 0 1 103
8 10 14 8 8 12 7 1 0 1 0 61
9 10 14 14 14 20 18 11 4 0 1 106
10 10 14 14 14 19 20 9 9 1 2 112
11 9 10 4 4 5 2 0 0 0 0 34
No.in
Grp. 10 14 14 15 22 23 13 9 2 5 127
We begin our estimation of the difficulty of each item by calculating the odds of
any subject failing an item. Since the far right column above is the total number of
subjects out of 127 that failed the items in each row, the odds of failing an item are
no. failing
odds = ____________________
no. subjects - no. failing
If we divided the numerator and denominator of the above ratio by the number of
subjects we would obtain for any item i, the odds
P
i
odds = ___________
1.0 - P
i
Next, we obtain the natural logarithm of the odds of failure for each item. The mean and
variance of these log odds are then obtained. Now we calculate the deviation of each
item's log odds from the mean log odds of all items. To obtain the PROX. estimate of the
item difficulty we multiply the deviation log odds by a constant Y. The constant Y is
obtained by
1 + V / 2.89
Y
2
= _____________
1 - UV / 8.35
where V is the variance of the log odds of items and
W is the variance of the log odds of abilities.
Clearly, we must first obtain the variance of log odds for abilities before we can
complete our PROX. estimates for items. To do this we must obtain the odds of subjects
in each total score group obtaining their total score out of the total number of possible
items. For subjects in each total score group the odds are
No. items passed
odds = ____________________________
No. items - No. items passed
For example, for subjects that have a total score of 1, the odds of getting such a score are
1 / (11 - 1) = 1 / 10 = .1.
Note that if we divide the above numerator and denominator by the number of test items,
the formula for the odds may be expressed as
P
j
odds = _________
1 - P
j
We obtain the logarithm of the score odds for each subject, and like we did for items,
obtain the mean and variance of the log odds for all subjects. The variance of subject's
log odds is denoted as U in the "expansion" factor Y above. A similar expansion factor
will be used to obtain Prox. estimates of ability and is calculated using
1 + U / 2.89
X
2
= _______________
1 - UV / 8.35
The Prox. values for items is now obtained by multiplying the expansion factor Y (square
root of the Y
2
value above) times the deviation log odds for each item. The Prox. values
for abilities is obtained by multiplying the corresponding expansion value X times the log
odds for each score group. The calculations are summarized below:
ITEM FAILED PASSED ODDS LOG ODDS DEVIATION PROX.
1 51 76 .67 -0.3989 -0.61 -0.87
2 85 42 2.02 0.7050 0.49 0.70
3 62 65 .95 -0.0473 -0.26 -0.37
4 3 124 .02 -3.7217 -3.93 -5.62
5 44 83 .53 -0.6347 -0.84 -1.20
6 116 11 10.55 2.3557 2.15 3.08
7 103 24 4.29 1.4567 1.25 1.79
8 61 66 0.92 -0.0788 -0.29 -0.41
9 106 21 5.05 1.6189 1.41 2.02
10 112 15 7.47 2.0104 1.80 2.58
11 34 93 .37 -1.0062 -1.22 -1.75
MEAN LOG ODDS DIFFICULTY = 0.21
VARIANCE LOG ODDS DIFFICULTY = 2.709
TOTAL
SCORE PASSED FAILED ODDS LOG ODDS PROX. ABILITY
1 1 10 .10 -2.30 -3.93
2 2 9 .22 -1.50 -2.56
3 3 8 .38 -0.98 -1.71
4 4 7 . -0.56 -0.94
5 5 6 .83 -0.18 -0.31
6 6 5 1.20 0.18 0.31
7 7 4 1.75 0.56 0.94
8 8 3 2.67 0.98 1.71
9 9 2 4.50 1.50 2.56
10 10 1 10.00 2.30 3.93
MEAN LOG ODDS ABILITY = -0.28
VARIANCE LOG ODDS ABILITY = 1.038
Y EXPANSION FACTOR = 1.4315
X EXPANSION FACTOR = 1.709
Theoretically, the number of subjects in total score group j that pass item i are
estimates of the item difficulty di and the ability bj of subjects as given by
b
j
- d
i
= log[ p
ij
/ ( n
j
- p
ij
) ]
where p
ij
is the proportion of subjects in score group j that pass item i and nj is the
number of subjects in score group j. The Prox. estimates of difficulty and ability may be
improved to yield closer estimates to the pij values through use of the Newton-Rhapson
iterations of the maximum-likelihood fit to those observed values. This solution is based
on the theory that
(b
j
- d
i
)
e
p
ij
= _______________
(b
j
- d
i
)
1 + e
It is possible, using this procedure, that values do not converge to a solution. The Rasch
program included in the statistics package will complete a maximum of 25 iterations
before stopping if the solution does not converge by that time.
If the Rasch model fits the data observed for a given item, the success and failure
of each score group on an item should be adequately reproduced using the estimated
parameters of the model. A chi-squared statistic may be obtained for each item by
summing, across the score groups, the sum of two products: the odds of success times the
number of failures and the odds of failure times the number of successes. This chi-
squared value has degrees of freedom N - n where N is the total number of subjects and k
is the total number of score groups. It should be noted that subjects with scores of 0 or all
items correct are eliminated from the analysis since log odds cannot be obtained for these
score groups. In addition, items which are failed or passed by all subjects cannot be
scaled and are eliminated from the analysis.
We will demonstrate the Rasch one parameter logistic scaling procedure with the
file labeled "TESTITEMS.txt" that contains 51 subjects measured with 8 items. The
form for specifying the analysis is shown below:
Figure 74 Rasch One Parameter Logistic Procedure Form.
In the above, we have specified all of the options. We will not, however, reproduce all of
the results but simply show samples of some of the output. When the OK button is
clicked, the following is obtained:
Figure 75 Log Item Difficulty - Rasch One Parameter Logistic Analysis.
Figure 76 Log Ability Scores - Rasch One Parameter Logistic Analysis.
Figure 77 Sample Item Information Function Plot - Rasch Analysis.
Rasch One-Parameter Logistic Test Scaling (Item Response Theory)
Written by William G. Miller
case 1 eliminated. Total score was 8
Case 2 Total Score = 7 Item scores 1 1 1 1 1 1 1 0
.
.
.
Row 1 for item 1 eliminated.
Row 7 for item 8 eliminated.
Total number of score groups = 7
Matrix of Item Failures in Score Groups
Score Group 1 2 3 4 5 6 7 Total
ITEM

2 6 0 0 0 0 0 0 6
3 6 6 0 0 0 0 0 12
4 6 6 6 0 0 0 0 18
5 6 6 6 6 0 0 0 24
6 6 6 6 6 6 0 0 30
7 6 6 6 6 6 6 0 36
Total 6 6 6 6 6 6 6 42
Item Log Odds Deviation Squared Deviation
1 -1.79 -1.79 3.21
2 -0.92 -0.92 0.84
3 -0.29 -0.29 0.08
4 0.29 0.29 0.08
5 0.92 0.92 0.84
6 1.79 1.79 3.21
Score Frequency Log Odds Freq.x Log Freq.x Log Odds Squared
1 6 -1.95 -11.68 22.72
2 6 -1.10 -6.59 7.24
3 6 -0.51 -3.06 1.57
4 6 0.00 0.00 0.00
5 6 0.51 3.06 1.57
6 6 1.10 6.59 7.24
7 6 1.95 11.68 22.72

Prox values and Standard Errors

Item Scale Value Standard Error
2 -2.546 0.627
3 -1.302 0.485
4 -0.409 0.443
5 0.409 0.443
6 1.302 0.485
7 2.546 0.627
Y expansion factor = 1.4212
Score Scale Value Standard Error
1 -2.726 1.498
2 -1.539 1.144
3 -0.716 1.023
4 0.000 0.991
5 0.716 1.023
6 1.539 1.144
7 2.726 1.498
X expansion factor = 1.4010
Maximum Likelihood Iteration Number 0
.
.
.
Maximum Likelihood Estimates
Item Log Difficulty
2 -2.42
3 -0.93
4 0.38
5 1.64
6 2.80
7 4.07
Score Log Ability
1 -1.99
2 -0.51
3 0.66
4 1.62
5 2.44
6 3.18
7 3.98
Goodness of Fit Test for Each Item
Item Chi-Squared Degrees of Probability
No. Value Freedom of Larger Value
1 10.59 35 1.0000
2 13.30 35 0.9997
3 14.00 35 0.9994
4 13.56 35 0.9996
5 12.97 35 0.9997
6 10.99 35 1.0000
Item Data Summary
ITEM PT.BIS.R. BIS.R. SLOPE PASSED FAILED RASCH DIFF
2 1.000 0.000 0.00 42.00 6 -2.423
3 0.612 0.950 3.03 36.00 12 -0.931
4 0.791 1.000 30.00 30.00 18 0.377
5 0.866 1.000 30.00 24.00 24 1.641
6 0.866 1.000 30.00 18.00 30 2.802
7 0.791 1.000 30.00 12.00 36 4.069
In the above output one sees that cases with all items correct or all items incorrect are
eliminated. In addition, items that have not been missed by any one or have been missed
by everyone are eliminated from the analysis. The initial solution to the estimation of log
difficulty or log ability is obtained by an approximation method. An iterative maximum
likelihood method is then applied to these initial estimates until stable estimates are
obtained for the model. The results are used to plot the item and ability log values as
well as the item information functions. An "ideal" test would have items that had equally
spaced item difficulties with item information functions that showed discrimination
across the ability levels with steep peeks at the log value. This would yield test results
like an interval scale of measurement where each possible total score was measured in
equally spaced intervals. Equal item discrimination slopes are assumed in the Rasch
model but there are several items in this analysis that are quite different from the
remaining ones. The chi-square tests of the item fit to the model indicate a poor fit for
this method of analysis.
Spearman-Brown Reliability Prophecy
The Spearman-Brown "Prophecy" formula has been a corner-stone of many
instructional text books in measurement theory. Based on "Classical True-Score" theory,
it provides an estimate of what the reliability of a test of a different length would be
based on the initial test's reliability estimate. It assumes the average difficulty and inter-
item covariances of the extended (or trimmed) test are the same as the original test. If
these assumptions are valid, it is a reasonable estimator. Shown below is the
specification form which appears when you elect this Measurement option from the
Analyses menu:
Three Parameter Logistic Scaling
Subject responses to test items may be a function of three parameters of the
individual items: (1) the difficulty of the item, (2) the degree to which the item
discriminates among individuals of differing ability, and (3) the probability of obtaining
the correct answer by chance alone. These three parameters may be estimated by this
procedure.
Kuder-Richardson #21 Reliability
The Kuder-Richardson formula #20 was developed from Classical Test Theory
(true-score theory). A shorter form of the estimate can be made using only the mean,
standard deviation and number of test items if one can assume that the inter-item
covariances are equal. Below is the form which appears when this procedure is selected
from the Measurement option of the Analyses menu:
Figure 78 Kuder-Richardson Formula 21 Reliability Calculation.
Note that we have entered the maximum score (total number of items), the test mean, and
the test standard deviation. When you click the Compute button, the estimate is shown in
the labeled box.
Reliability Adjustment for Score Variability
Researchers will sometimes use a test that has been standardized on a large,
heterogenous population of subjects. Such tests typically report rather high internal-
consistency reliability estimates (e.g. Cronbach's estimate.) But what is the reliability if
one administers the test to a much more homogeneous population? For example, assume
a high school counselor administers a "College Aptitude Test" that reports a reliability of
0.95 with a standard deviation of 15 (variance of 225) and a mean of 20.0 for the national
norm. What reliability would the counselor expect to obtain for her sample of students
that obtain a mean of 22.8 and a standard deviation of 10.2 (variance of 104.04)? This
procedure will help provide the estimate. Shown below is the specification form and our
sample values entered. When the compute button is clicked, the results shown are
obtained.
Figure 79 Reliability Changes With Changes in Variance.
Weighted Composite Test Reliability
The reliability for a combination of tests, each of which has its own estimate of
reliability and a weight assigned to it, may be computed. This composite will typically
be greater than any one test by itself due to the likelihood that the subtests are correlated
positively among themselves. Since teachers typically assign course grades based on a
combination of individual tests administered over the time period of a course, this
reliability estimate in built into the Grading System. See the description and examples in
that section.
Figure 80 Estimation of Reliability for Composite Tests.
Composite Test Reliability
File Analyzed: C:\Projects\Delphi\OS4\comprel.OS4
Correlations Among Tests with 10 cases.
Variables
VAR1 VAR2 VAR3
VAR1 1.000 0.939 0.152
VAR2 0.939 1.000 0.139
VAR3 0.152 0.139 1.000
Means with 10 valid cases.
Variables VAR1 VAR2 VAR3
5.500 5.500 5.500
Variances with 10 valid cases.
9.167 9.167 9.167
Standard Deviations with 10 valid cases.
3.028 3.028 3.028
Test Weights with 10 valid cases.
1.000 1.000 2.000
Test Reliabilities with 10 valid cases.
0.900 0.700 0.800
Composite reliability = 0.871
Successive Interval Scaling
Successive Interval Scaling was developed as an approximation of Thurstones
Paired Comparisons method for estimation of scale values and dispersion of scale values
for items designed to measure attitudes. Typically, five to nine categories are used by
judges to indicate the degree to which an item expresses an attitude (if a subject agrees
with the item) between very negative to very positive. Once scale values are estimated,
the items responded to by subjects are scored by obtaining the median scale value of
those items to which the subject agrees.
To obtain Successive interval scale values, select that option under the
Measurement group in the Analyses menu on the main form. The specifications form
below will appear. Select those items (variables) you wish to scale. The data analyzed
consists of rows representing judges and columns representing the scale value chosen for
an item by a judge.
Figure 81 Successive Interval Scaling.
When you click the OK button on the box above, the results will appear on the
printout form. An example of results are presented below.
SUCCESSIVE INTERVAL SCALING RESULTS
0- 1 1- 2 2- 3 3- 4 4- 5 5- 6 6- 7
VAR1
Frequency 0 0 0 0 4 4 4
Proportion 0.000 0.000 0.000 0.000 0.333 0.333 0.333
Cum. Prop. 0.000 0.000 0.000 0.000 0.333 0.667 1.000
Normal z - - - - -0.431 0.431 -
VAR2
Frequency 0 0 1 3 4 4 0
Proportion 0.000 0.000 0.083 0.250 0.333 0.333 0.000
Cum. Prop. 0.000 0.000 0.083 0.333 0.667 1.000 1.000
Normal z - - -1.383 -0.431 0.431 - -
VAR3
Frequency 0 0 4 3 4 1 0
Proportion 0.000 0.000 0.333 0.250 0.333 0.083 0.000
Cum. Prop. 0.000 0.000 0.333 0.583 0.917 1.000 1.000
Normal z - - -0.431 0.210 1.383 - -
VAR4
Frequency 0 3 4 5 0 0 0
Proportion 0.000 0.250 0.333 0.417 0.000 0.000 0.000
Cum. Prop. 0.000 0.250 0.583 1.000 1.000 1.000 1.000
Normal z - -0.674 0.210 - - - -
VAR5
Frequency 5 4 3 0 0 0 0
Proportion 0.417 0.333 0.250 0.000 0.000 0.000 0.000
Cum. Prop. 0.417 0.750 1.000 1.000 1.000 1.000 1.000
Normal z -0.210 0.674 - - - - -
VAR6
Frequency 1 2 2 2 2 2 1
Proportion 0.083 0.167 0.167 0.167 0.167 0.167 0.083
Cum. Prop. 0.083 0.250 0.417 0.583 0.750 0.917 1.000
Normal z -1.383 -0.674 -0.210 0.210 0.674 1.383 -
INTERVAL WIDTHS
2- 1 3- 2 4- 3 5- 4 6- 5
VAR1 - - - - 0.861
VAR2 - - 0.952 0.861 -
VAR3 - - 0.641 1.173 -
VAR4 - 0.885 - - -
VAR5 0.885 - - - -
VAR6 0.709 0.464 0.421 0.464 0.709
Mean Width 0.80 0.67 0.67 0.83 0.78
No. Items 2 2 3 3 2
Std. Dev.s 0.02 0.09 0.07 0.13 0.01
Cum. Means 0.80 1.47 2.14 2.98 3.76
ESTIMATES OF SCALE VALUES AND THEIR DISPERSIONS
Item No. Ratings Scale Value Discriminal Dispersion
VAR1 12 3.368 1.224
VAR2 12 2.559 0.822
VAR3 12 1.919 0.811
VAR4 12 1.303 1.192
VAR5 12 0.199 1.192
VAR6 12 1.807 0.759
Z scores Estimated from Scale values
0- 1 1- 2 2- 3 3- 4 4- 5 5- 6 6- 7
VAR1 -3.368 -2.571 -1.897 -1.225 -0.392 0.392
VAR2 -2.559 -1.762 -1.088 -0.416 0.416 1.201
VAR3 -1.919 -1.122 -0.448 0.224 1.057 1.841
VAR4 -1.303 -0.506 0.169 0.840 1.673 2.458
VAR5 -0.199 0.598 1.272 1.943 2.776 3.000
VAR6 -1.807 -1.010 -0.336 0.336 1.168 1.953
Cumulative Theoretical Proportions
0- 1 1- 2 2- 3 3- 4 4- 5 5- 6 6- 7
VAR1 0.000 0.005 0.029 0.110 0.347 0.653 1.000
VAR2 0.005 0.039 0.138 0.339 0.661 0.885 1.000
VAR3 0.028 0.131 0.327 0.589 0.855 0.967 1.000
VAR4 0.096 0.306 0.567 0.800 0.953 0.993 1.000
VAR5 0.421 0.725 0.898 0.974 0.997 0.999 1.000
VAR6 0.035 0.156 0.369 0.631 0.879 0.975 1.000
Average Discrepency Between Theoretical and Observed Cumulative Proportions = 0.050
Maximum discrepency = 0.200 found in item VAR4
Guttman Scalogram Analysis
Guttman scales are those measurement instruments composed of items which,
ideally, form a hierarchy in which the total score of a subject can indicate the actual
response (correct or incorrect) of each item. Items are arranged in order of the proportion
of subjects passing the item and subjects are grouped and sequenced by their total scores.
If the items measure consistently, a triangular pattern should emerge. A coefficient of
reproducibility is obtained which may be interpreted in a manner similar to test
reliability.
Dichotomously scored (0 and 1) items representing the responses of subjects in
your data grid rows are the variables (grid columns) analyzed. Select the items to analyze
in the same manner as you would for the Classical Item Analysis or the Rasch analysis.
When you click the OK button, you will immediately be presented with the results on the
output form. An example is shown below.
Figure 82 Guttman Scaling of Affective Items.
GUTTMAN SCALOGRAM ANALYSIS
Cornell Method
No. of Cases := 101. No. of items := 10
RESPONSE MATRIX
Subject Row Item Number
Label Sum Item 10 Item 9 Item 1 Item 3 Item 5 Item 2
0 1 0 1 0 1 0 1 0 1 0 1
1 10 0 1 0 1 0 1 0 1 0 1 0 1

6 10 0 1 0 1 0 1 0 1 0 1 0 1

20 10 0 1 0 1 0 1 0 1 0 1 0 1

46 10 0 1 0 1 0 1 0 1 0 1 0 1

68 10 0 1 0 1 0 1 0 1 0 1 0 1

77 10 0 1 0 1 0 1 0 1 0 1 0 1

50 9 0 1 0 1 0 1 1 0 0 1 0 1

39 9 1 0 0 1 0 1 0 1 0 1 0 1
etc.
TOTALS 53 48 52 49 51 50 51 50 50 51 48 53
ERRORS 3 22 19 9 5 20 13 10 10 10 10 13
Subject Row Item Number
Label Sum Item 8 Item 6 Item 4 Item 7
0 1 0 1 0 1 0 1
1 10 0 1 0 1 0 1 0 1

6 10 0 1 0 1 0 1 0 1
etc.

65 0 1 0 1 0 1 0 1 0

10 0 1 0 1 0 1 0 1 0

89 0 1 0 1 0 1 0 1 0

TOTALS 46 55 44 57 44 57 41 60
ERRORS 11 11 17 3 12 11 11 15
Coefficient of Reproducibility := 0.767
Differential Item Functioning
Anyone developing tests today should be sensitive to the fact that some test items
may present a bias for one or more subgroups in the population to which the test is
administered. For example, because of societal value systems, boys and girls may be
exposed to quite different learning experiences during their youth. A word test in
mathematics may unintentionally give an advantage to one gender group over another
simply by the examples used in the item. To identify possible bias in an item, one can
examine the differential item functioning of each item for the sub-groups to which the
test is administered. The Mantel-Haenszel test statistic may be applied to test the
difference on the item characteristic curve for the difference between a "focus" group and
a "reference" group. We will demonstrate using a data set in which 40 items have been
administered to 1000 subjects in one group and 1000 subjects in another group. The
groups are simply coded 1 and 2 for the reference and focus groups. Since there may be
very few (or no) subjects that get a specific total score, we will group the total scores
obtained by subjects into groups of 4 so that we are comparing subjects in the groups that
have obtained total item scores of 0 to 3, 4 to 7, , 40 to 43. As you will see, even this
grouping is too small for several score groups and we should probably change the score
range for the lowest and highest scores to a larger range of scores in another run.
When you elect to do this analysis, the specification form below appears:
Figure 83 Differential Item Functioning.
On the above form you specify the items to be analyzed and also the variable defining the
reference and focus group codes. You may then specify the options desired by clicking
the corresponding buttons for the desired options. You also enter the number of score
groups to be used in grouping the subject's total scores. When this is specified, you then
enter the lowest and highest score for each of those score groups. When you have
specified the low and hi score for the first group, click the right arrow on the "slider" bar
to move to the next group. You will see that the lowest score has automatically been set
to one higher than the previous group's highest score to save you time in entering data.
You do not, of course, have to use the same size for the range of each score group. Using
too large a range of scores may cut down the sensitivity of the test to differences between
the groups. Fairly large samples of subjects is necessary for a reasonable analysis. Once
you have completed the specifications, click the Compute button and you will see the
following results are obtained (we elected to print the descriptive statistics, correlations
and item plots):
Mantel-Haenszel DIF Analysis adapted by Bill Miller from
EZDIF written by Niels G. Waller
Total Means with 2000 valid cases.
Variables VAR 1 VAR 2 VAR 3 VAR 4 VAR 5
0.688 0.064 0.585 0.297 0.451
0.806 0.217 0.827 0.960 0.568
0.350 0.291 0.725 0.069 0.524
0.350 0.943 0.545 0.017 0.985
0.778 0.820 0.315 0.203 0.982
0.834 0.700 0.397 0.305 0.223
0.526 0.585 0.431 0.846 0.115
0.150 0.817 0.909 0.793 0.329
Total Variances with 2000 valid cases.
0.215 0.059 0.243 0.209 0.248
0.156 0.170 0.143 0.038 0.245
0.228 0.206 0.199 0.064 0.250
0.228 0.054 0.248 0.017 0.015
0.173 0.148 0.216 0.162 0.018
0.139 0.210 0.239 0.212 0.173
0.249 0.243 0.245 0.130 0.102
0.128 0.150 0.083 0.164 0.221
Total Standard Deviations with 2000 valid cases.
0.463 0.244 0.493 0.457 0.498
0.395 0.412 0.379 0.196 0.495
0.477 0.454 0.447 0.253 0.500
0.477 0.233 0.498 0.129 0.124
0.416 0.384 0.465 0.403 0.135
0.372 0.459 0.489 0.461 0.416
0.499 0.493 0.495 0.361 0.319
0.357 0.387 0.288 0.405 0.470
Total Score: Mean = 21.318, Variance = 66.227, Std.Dev. = 8.138
Reference group size = 1000, Focus group size = 1000
Correlations Among Items with 2000 cases.
Variables
VAR 1 VAR 2 VAR 3 VAR 4 VAR 5
VAR 1 1.000 0.162 0.389 0.308 0.406
VAR 2 0.162 1.000 0.190 0.275 0.259
VAR 3 0.389 0.190 1.000 0.368 0.382
VAR 4 0.308 0.275 0.368 1.000 0.423
Etc.
Variables
VAR 1 0.213 0.234 0.203 0.230 0.273
VAR 2 0.229 0.107 0.075 0.123 0.211
VAR 3 0.209 0.233 0.206 0.274 0.255
VAR 4 0.236 0.180 0.156 0.221 0.284
VAR 5 0.253 0.241 0.196 0.248 0.289
VAR 6 0.149 0.338 0.254 0.282 0.227
VAR 7 0.305 0.183 0.158 0.197 0.290
VAR 8 0.163 0.271 0.259 0.278 0.222
VAR 9 0.086 0.167 0.228 0.236 0.121
Etc.
VAR 37 0.155 1.000 0.250 0.276 0.204
VAR 38 0.118 0.250 1.000 0.242 0.181
VAR 39 0.180 0.276 0.242 1.000 0.262
VAR 40 0.288 0.204 0.181 0.262 1.000
Item-Total Correlations with 2000 valid cases.
0.527 0.352 0.556 0.514 0.563
0.507 0.509 0.488 0.302 0.579
0.566 0.502 0.556 0.352 0.586
0.564 0.371 0.582 0.171 0.200
0.532 0.511 0.574 0.511 0.235
0.507 0.570 0.591 0.569 0.507
0.580 0.584 0.590 0.501 0.411
0.465 0.482 0.415 0.513 0.556
Conditioning Levels
Lower Upper
0 3
4 7
8 11
12 15
16 19
20 23
24 27
28 31
32 35
36 39
40 43
Figure 84 Frequency Plot of Subject Item Scores - Differential Item Functioning.
etc.
etc. for all items. Note the difference for the item plot shown above! Next, the output
reflects multiple passes to "fit" the data for the M-H test:
COMPUTING M-H CHI-SQUARE, PASS # 1
Cases in Reference Group
Score Level Counts by Item
Variables
0- 3 8 8 8 8 8
4- 7 38 38 38 38 38
8- 11 65 65 65 65 65
12- 15 108 108 108 108 108
16- 19 153 153 153 153 153
20- 23 175 175 175 175 175
24- 27 154 154 154 154 154
28- 31 167 167 167 167 167
32- 35 94 94 94 94 94
36- 39 38 38 38 38 38
40- 43 0 0 0 0 0
Variables
0- 3 8 8 8 8 8
4- 7 38 38 38 38 38
8- 11 65 65 65 65 65
12- 15 108 108 108 108 108
16- 19 153 153 153 153 153
20- 23 175 175 175 175 175
24- 27 154 154 154 154 154
28- 31 167 167 167 167 167
32- 35 94 94 94 94 94
36- 39 38 38 38 38 38
40- 43 0 0 0 0 0
Variables
0- 3 8 8 8 8 8
4- 7 38 38 38 38 38
8- 11 65 65 65 65 65
12- 15 108 108 108 108 108
16- 19 153 153 153 153 153
20- 23 175 175 175 175 175
24- 27 154 154 154 154 154
28- 31 167 167 167 167 167
32- 35 94 94 94 94 94
36- 39 38 38 38 38 38
40- 43 0 0 0 0 0
Variables
0- 3 8 8 8 8 8
4- 7 38 38 38 38 38
8- 11 65 65 65 65 65
12- 15 108 108 108 108 108
16- 19 153 153 153 153 153
20- 23 175 175 175 175 175
24- 27 154 154 154 154 154
28- 31 167 167 167 167 167
32- 35 94 94 94 94 94
36- 39 38 38 38 38 38
40- 43 0 0 0 0 0
Variables
0- 3 8 8 8 8 8
4- 7 38 38 38 38 38
8- 11 65 65 65 65 65
12- 15 108 108 108 108 108
16- 19 153 153 153 153 153
20- 23 175 175 175 175 175
24- 27 154 154 154 154 154
28- 31 167 167 167 167 167
32- 35 94 94 94 94 94
36- 39 38 38 38 38 38
40- 43 0 0 0 0 0
Variables
0- 3 8 8 8 8 8
4- 7 38 38 38 38 38
8- 11 65 65 65 65 65
12- 15 108 108 108 108 108
16- 19 153 153 153 153 153
20- 23 175 175 175 175 175
24- 27 154 154 154 154 154
28- 31 167 167 167 167 167
32- 35 94 94 94 94 94
36- 39 38 38 38 38 38
40- 43 0 0 0 0 0
Variables
0- 3 8 8 8 8 8
4- 7 38 38 38 38 38
8- 11 65 65 65 65 65
12- 15 108 108 108 108 108
16- 19 153 153 153 153 153
20- 23 175 175 175 175 175
24- 27 154 154 154 154 154
28- 31 167 167 167 167 167
32- 35 94 94 94 94 94
36- 39 38 38 38 38 38
40- 43 0 0 0 0 0
Variables
0- 3 8 8 8 8 8
4- 7 38 38 38 38 38
8- 11 65 65 65 65 65
12- 15 108 108 108 108 108
16- 19 153 153 153 153 153
20- 23 175 175 175 175 175
24- 27 154 154 154 154 154
28- 31 167 167 167 167 167
32- 35 94 94 94 94 94
36- 39 38 38 38 38 38
40- 43 0 0 0 0 0
Cases in Focus Group
Variables
0- 3 7 7 7 7 7
4- 7 47 47 47 47 47
8- 11 94 94 94 94 94
12- 15 139 139 139 139 139
16- 19 177 177 177 177 177
20- 23 174 174 174 174 174
24- 27 141 141 141 141 141
28- 31 126 126 126 126 126
32- 35 68 68 68 68 68
36- 39 27 27 27 27 27
40- 43 0 0 0 0 0
Variables
0- 3 7 7 7 7 7
4- 7 47 47 47 47 47
8- 11 94 94 94 94 94
12- 15 139 139 139 139 139
16- 19 177 177 177 177 177
20- 23 174 174 174 174 174
24- 27 141 141 141 141 141
28- 31 126 126 126 126 126
32- 35 68 68 68 68 68
36- 39 27 27 27 27 27
40- 43 0 0 0 0 0
Variables
0- 3 7 7 7 7 7
4- 7 47 47 47 47 47
8- 11 94 94 94 94 94
12- 15 139 139 139 139 139
16- 19 177 177 177 177 177
20- 23 174 174 174 174 174
24- 27 141 141 141 141 141
28- 31 126 126 126 126 126
32- 35 68 68 68 68 68
36- 39 27 27 27 27 27
40- 43 0 0 0 0 0
Variables
0- 3 7 7 7 7 7
4- 7 47 47 47 47 47
8- 11 94 94 94 94 94
12- 15 139 139 139 139 139
16- 19 177 177 177 177 177
20- 23 174 174 174 174 174
24- 27 141 141 141 141 141
28- 31 126 126 126 126 126
32- 35 68 68 68 68 68
36- 39 27 27 27 27 27
40- 43 0 0 0 0 0
Variables
0- 3 7 7 7 7 7
4- 7 47 47 47 47 47
8- 11 94 94 94 94 94
12- 15 139 139 139 139 139
16- 19 177 177 177 177 177
20- 23 174 174 174 174 174
24- 27 141 141 141 141 141
28- 31 126 126 126 126 126
32- 35 68 68 68 68 68
36- 39 27 27 27 27 27
40- 43 0 0 0 0 0
Variables
0- 3 7 7 7 7 7
4- 7 47 47 47 47 47
8- 11 94 94 94 94 94
12- 15 139 139 139 139 139
16- 19 177 177 177 177 177
20- 23 174 174 174 174 174
24- 27 141 141 141 141 141
28- 31 126 126 126 126 126
32- 35 68 68 68 68 68
36- 39 27 27 27 27 27
40- 43 0 0 0 0 0
Variables
0- 3 7 7 7 7 7
4- 7 47 47 47 47 47
8- 11 94 94 94 94 94
12- 15 139 139 139 139 139
16- 19 177 177 177 177 177
20- 23 174 174 174 174 174
24- 27 141 141 141 141 141
28- 31 126 126 126 126 126
32- 35 68 68 68 68 68
36- 39 27 27 27 27 27
40- 43 0 0 0 0 0
Variables
0- 3 7 7 7 7 7
4- 7 47 47 47 47 47
8- 11 94 94 94 94 94
12- 15 139 139 139 139 139
16- 19 177 177 177 177 177
20- 23 174 174 174 174 174
24- 27 141 141 141 141 141
28- 31 126 126 126 126 126
32- 35 68 68 68 68 68
36- 39 27 27 27 27 27
40- 43 0 0 0 0 0
Insufficient data found in level: 0 - 3
CODES ITEM SIG. ALPHA CHI2 P-VALUE MH D-DIF S.E. MH D-DIF
C R 1 *** 8.927 276.392 0.000 -5.144 0.338
C R 2 *** 10.450 68.346 0.000 -5.514 0.775
C R 3 *** 7.547 280.027 0.000 -4.750 0.305
C R 4 *** 10.227 298.341 0.000 -5.464 0.350
C R 5 *** 12.765 393.257 0.000 -5.985 0.339
B 6 *** 0.571 15.476 0.000 1.316 0.331
A 7 * 0.714 6.216 0.013 0.791 0.310
A 8 * 0.705 5.694 0.017 0.822 0.335
B 9 ** 0.493 6.712 0.010 1.664 0.621
B 10 *** 0.621 17.349 0.000 1.121 0.267
A 11 * 0.775 4.511 0.034 0.599 0.275
A 12 *** 0.687 9.833 0.002 0.883 0.277
B 13 *** 0.647 11.904 0.001 1.024 0.294
B 14 ** 0.568 7.160 0.007 1.331 0.482
B 15 *** 0.600 19.747 0.000 1.199 0.267
B 16 *** 0.601 18.326 0.000 1.198 0.278
A 17 0.830 0.486 0.486 0.438 0.538
A 18 *** 0.709 8.989 0.003 0.807 0.264
A 19 0.582 1.856 0.173 1.270 0.834
A 20 1.991 1.769 0.183 -1.618 1.072
A 21 * 0.725 5.783 0.016 0.754 0.308
A 22 * 0.743 4.023 0.045 0.697 0.337
B 23 *** 0.572 20.804 0.000 1.315 0.286
A 24 * 0.723 5.362 0.021 0.764 0.321
A 25 0.555 1.782 0.182 1.385 0.915
B 26 *** 0.540 16.456 0.000 1.447 0.353
A 27 *** 0.687 9.240 0.002 0.884 0.287
A 28 ** 0.735 6.822 0.009 0.723 0.271
A 29 *** 0.681 9.458 0.002 0.904 0.289
A 30 * 0.756 4.342 0.037 0.658 0.306
A 31 *** 0.724 8.016 0.005 0.758 0.263
A 32 * 0.745 6.513 0.011 0.693 0.266
A 33 ** 0.738 6.907 0.009 0.715 0.267
A 34 0.944 0.089 0.766 0.135 0.360
A 35 0.769 2.381 0.123 0.618 0.383
A 36 0.819 1.530 0.216 0.469 0.357
A 37 * 0.709 5.817 0.016 0.809 0.326
A 38 * 0.665 4.552 0.033 0.960 0.431
A 39 0.779 3.305 0.069 0.588 0.312
B 40 *** 0.644 13.215 0.000 1.034 0.280
No. of items purged in pass 1 = 5
Item Numbers:
1
2
3
4
5
COMPUTING M-H CHI-SQUARE, PASS # 2
Cases in Reference Group
Variables
0- 3 8 9 9 9 9
4- 7 44 49 47 47 47
8- 11 76 92 77 91 81
12- 15 150 163 148 158 159
16- 19 183 191 185 184 176
20- 23 188 178 187 173 180
24- 27 176 177 173 170 172
28- 31 129 103 128 123 130
32- 35 46 38 46 45 46
36- 39 0 0 0 0 0
40- 43 0 0 0 0 0
Variables
0- 3 9 9 9 9 9
4- 7 49 49 49 49 49
8- 11 92 92 92 92 92
12- 15 163 163 163 163 163
16- 19 193 193 193 193 193
20- 23 180 180 180 180 180
24- 27 184 184 184 184 184
28- 31 106 106 106 106 106
32- 35 24 24 24 24 24
36- 39 0 0 0 0 0
40- 43 0 0 0 0 0
Variables
0- 3 9 9 9 9 9
4- 7 49 49 49 49 49
etc.
CODES ITEM SIG. ALPHA CHI2 P-VALUE MH D-DIF S.E. MH D-DIF
C R 1 *** 11.643 331.803 0.000 -5.769 0.354
C R 2 *** 13.737 91.260 0.000 -6.157 0.776
C R 3 *** 10.273 346.784 0.000 -5.474 0.324
C R 4 *** 13.828 358.692 0.000 -6.173 0.373
C R 5 *** 16.713 460.515 0.000 -6.618 0.356
A 6 0.823 1.758 0.185 0.457 0.327
A 7 1.061 0.144 0.704 -0.139 0.310
A 8 0.992 0.000 1.000 0.019 0.331
A 9 0.682 1.770 0.183 0.901 0.619
A 10 0.887 1.033 0.309 0.283 0.264
A 11 1.147 1.261 0.261 -0.321 0.272
A 12 0.958 0.100 0.752 0.102 0.272
A 13 0.932 0.262 0.609 0.166 0.291
A 14 0.909 0.134 0.714 0.225 0.480
A 15 0.875 1.318 0.251 0.314 0.261
A 16 0.887 0.949 0.330 0.281 0.273
A 17 1.187 0.417 0.518 -0.403 0.532
A 18 1.017 0.009 1.000 -0.039 0.260
A 19 0.809 0.168 0.682 0.498 0.844
B 20 2.529 3.536 0.060 -2.180 1.075
A 21 1.031 0.028 0.866 -0.072 0.308
A 22 1.073 0.175 0.675 -0.166 0.337
A 23 0.861 1.390 0.238 0.352 0.284
A 24 1.105 0.448 0.503 -0.235 0.319
A 25 0.737 0.357 0.550 0.717 0.911
A 26 0.797 2.143 0.143 0.534 0.348
A 27 1.002 0.002 1.000 -0.004 0.284
A 28 1.095 0.540 0.463 -0.212 0.268
A 29 1.031 0.039 0.843 -0.073 0.281
A 30 1.124 0.707 0.400 -0.274 0.303
A 31 1.041 0.094 0.759 -0.095 0.261
A 32 1.073 0.320 0.572 -0.165 0.265
A 33 1.099 0.619 0.431 -0.222 0.264
A 34 * 1.375 4.152 0.042 -0.748 0.355
A 35 1.160 0.686 0.408 -0.348 0.384
A 36 1.276 2.370 0.124 -0.572 0.354
A 37 0.982 0.004 1.000 0.042 0.323
A 38 0.923 0.121 0.728 0.188 0.428
A 39 1.090 0.354 0.552 -0.203 0.308
A 40 0.953 0.126 0.722 0.114 0.274
One should probably combine the first two score groups (0-3 and 4-7) into one
group and the last three groups into one group so that sufficient sample size is available
for the comparisons of the two groups. This would, of course, reduce the number of
groups from 11 in our original specifications to 8 score groups. The chi-square statistic
identifies items you will want to give specific attention. Examine the data plots for those
items. Differences found may suggest bias in those items. Only examination of the
actual content can help in this decision. Even though two groups may differ in there item
response patterns does not provide sufficient grounds to establish bias - perhaps it simply
identifies a true difference in educational achievement due to other factors.
Polytomous DIF Analysis
The purpose of the differential item functioning programs are to identify test or
attitude items that "perform" differently for two groups - a target group and a reference
group. Two procedures are provided and selected on the basis of whether the items are
dichotomous (0 and 1 scoring) or consist of multiple categories (e.g. Likert responses
ranging from 1 to 5.) The latter case is where the Polytomous DIF Analysis is selected.
When you initiate this procedure you will see the dialogue box shown below:
Figure 85 Form for Differential Item Function of Polytomous Items.
The results from an analysis of three items with five categories that have been
collapsed into three category levels is shown below. A sample of 500 subject's attitude
scores were observed.
Polytomous Item DIF Analysis adapted by Bill Miller from
Procedures for extending item bias detection techniques
by Catherine Welch and H.D. Hoover, 1993
Applied Measurement in Education 6(1), pages 1-19.
Conditioning Levels
Lower Upper
0 1
2 3
4 5
For Item 1:
Observed Category Frequencies
Item Group Level Category Number
1 2 3 4 5
1 Ref. 1 46 51 39 64 48
1 Focal 1 40 41 38 46 42
1 Total 1 86 92 77 110 90
1 Ref. 2 2 0 0 0 0
1 Focal 2 1 0 0 0 0
1 Total 2 3 0 0 0 0
1 Ref. 3 12 8 1 0 0
1 Focal 3 15 6 0 0 0
1 Total 3 27 14 1 0 0
t-test values for Reference and Focus Means for each level
Mean Reference = 3.069 SD = 24.396 N = 248
Mean Focal = 3.043 SD = 21.740 N = 207
Level 1 t = -0.011 with deg. freedom = 453
Mean Focal = 1.000 SD = 1.000 N = 1
Level 2 t = 0.000 with deg. freedom = 0
Mean Focal = 1.286 SD = 4.088 N = 21
Composite z statistic = -0.076. Prob. > |z| = 0.530
Weighted Composite z statistic = -0.248. Prob. > |z| = 0.598
Generalized Mantel-Haenszel = 0.102 with D.F. = 1 and Prob. > Chi-Sqr. = 0.749
For Item 2:
1 2 3 4 5
2 Ref. 1 56 46 47 48 51
2 Focal 1 37 38 49 35 48
2 Total 1 93 84 96 83 99
2 Ref. 2 2 0 0 0 0
2 Focal 2 1 0 0 0 0
2 Total 2 3 0 0 0 0
2 Ref. 3 12 8 1 0 0
2 Focal 3 9 11 1 0 0
2 Total 3 21 19 2 0 0
Mean Focal = 3.092 SD = 22.466 N = 207
Mean Focal = 1.000 SD = 1.000 N = 1
Mean Focal = 1.619 SD = 5.094 N = 21
Composite z statistic = 0.075. Prob. > |z| = 0.470
Weighted Composite z statistic = 0.673. Prob. > |z| = 0.250
Generalized Mantel-Haenszel = 1.017 with D.F. = 1 and Prob. > Chi-Sqr. = 0.313

For item 3:
1 2 3 4 5
3 Ref. 1 35 38 52 68 55
3 Focal 1 42 41 37 42 45
3 Total 1 77 79 89 110 100
3 Ref. 2 2 0 0 0 0
3 Focal 2 1 0 0 0 0
3 Total 2 3 0 0 0 0
3 Ref. 3 8 10 3 0 0
3 Focal 3 7 10 4 0 0
3 Total 3 15 20 7 0 0
Mean Focal = 3.034 SD = 21.784 N = 207
Mean Focal = 1.000 SD = 1.000 N = 1
Mean Focal = 1.857 SD = 5.102 N = 21
Composite z statistic = -0.023. Prob. > |z| = 0.509
Weighted Composite z statistic = -1.026. Prob. > |z| = 0.848
Generalized Mantel-Haenszel = 3.248 with D.F. = 1 and Prob. > ChiSqr. = 0.071
Course Grades System
The grade book system is designed for teachers. The teacher can enter
information for each student in a course and their raw scores obtained on each test
administered during the term. Procedures are built into the grade book system to
automatically convert raw test scores into standard (z) scores, ranks and letter grades.
When the Grade Book System is first started, you will see the following screen:
Figure 86 The Grading System Initial Form.
When the teacher clicks on the button to add a test, the following dialog box
appears in which information about the test is recorded:
Figure 87 Grading System Dialog Form for Test Information.
When a test is added, the grid is expanded to include columns for raw scores, z
scores, ranks and letter grade. Clicking the Add Student button adds a new row for
student information. Students may be added or deleted at any time. Simply place the
cursor in the row where a student is to be deleted or added. One can also alphabetize the
grid on the student last names. If students are added or deleted, the user can recalculate
the z scores, ranks and grades if desired. Shown below is a grid in which fictitious data
has been entered.
Figure 88 Example Test Data Entered in the Grading System Form.
When creating a new grade book, one clicks on the Enter a Title for the Grade
Book and clicks the Do Above button below that option. The following dialogue
appears in which to make the entry:
Figure 89 Dialog Form for the Grading System to Create a Grade Book.
The teacher also enters information about the grade book by checking the Enter
Grading Procedures and clicking the Do Above. Shown below is an example of how the
Grading Procedures dialogue box might be completed. These choices and values as
stored as part of the grade book file.
Figure 90 Specifying Grading Procedures in the Grading System.
When raw scores have been entered for a test, the teacher can then elect to convert
those scores to z scores, ranks and grades. When the Do Above button is clicked, the
conversions are completed. A message box will first pop up to notify the teacher of the
mean and standard deviation that was computed for the test as shown below:
Figure 91 Test Results Message from the Grading System.
The converted scores then appear in the grid as shown in the snapshot below:
Figure 92 Example Results of the Grading System Score Conversions.
The teacher can also request a bar graph of the distribution of raw scores, z
scores, or grades. Shown below is a graph of the z scores for the sample data above.
Figure 93 A Plot of Test Scores from the Grading System.
The distribution is also printed as shown below:
A variety of listings is available to the teacher. For example, he or she may want
a list of the students scores on a given test. To obtain the list, place the cursor in the grid
on the column of the raw test scores of the test summary desired, then select the option
for the test summary and click the Do Above button. Individual student reports can also
be generated in a similar manner for distribution to the students.
General Unfolding Model Analysis
The Unfolding Model suggests that the probability of a subjects response to an
attitude item being positive or negative is a function of the distance that the subject's
attitude parameter (theta) is from the items parameter (delta). Instead of a monotonic
increase in probability of responding, say positively, as subjects attitude increase toward
the positive direction, the probability decreases in both directions away from the point
where the item and subject attitude parameters are most similar. This is a departure from
the latent trait theory underlying methods such as the Rasch model and the Thurstone
scaling models (Paired Comparisons, Successive Interval, etc.)
Shown below is an example of a GUM analysis. The data that was analyzed is
found in a file labeled gumcap.txt. The data represents judgments as to the degree of
positive attitude reflected in each item (recorded in the grid columns) as judged by a
group of judges (represented by grid rows.) Following the selection of the variables for
your GUM analysis, you are presented with the following dialogue box in which you
specify the parameters for your analysis. Suggested values are already entered for many
of the requested values.
Figure 94 Generalized Unfolding Model Specification Form.
Descriptive statistics for the items are presented first. These are then followed by
the estimated parameters and bar graphs of the distribution of those parameters. A log
file may also be viewed and printed.
GUMJML (Adapted from the original FORTRAN program written by James S. Roberts)
No. Cases = 240, No. Items = 12
CORRELATION MATRIX
Correlations
ITEM2 ITEM9 ITEM10 ITEM12 ITEM13 ITEM14
ITEM2 1.000 0.273 -0.426 0.574 0.500 0.577
ITEM9 0.273 1.000 -0.230 0.366 0.290 0.315
ITEM10 -0.426 -0.230 1.000 -0.562 -0.452 -0.527
ITEM12 0.574 0.366 -0.562 1.000 0.590 0.655
ITEM13 0.500 0.290 -0.452 0.590 1.000 0.579
ITEM14 0.577 0.315 -0.527 0.655 0.579 1.000
ITEM15 0.467 0.235 -0.392 0.533 0.429 0.517
ITEM16 0.534 0.310 -0.481 0.587 0.542 0.605
ITEM17 -0.552 -0.309 0.541 -0.609 -0.533 -0.583
ITEM18 0.134 0.437 -0.107 0.252 0.139 0.311
ITEM20 -0.423 -0.308 0.425 -0.518 -0.432 -0.463
ITEM24 -0.526 -0.355 0.622 -0.567 -0.489 -0.567
Correlations
ITEM15 ITEM16 ITEM17 ITEM18 ITEM20 ITEM24
ITEM2 0.467 0.534 -0.552 0.134 -0.423 -0.526
ITEM9 0.235 0.310 -0.309 0.437 -0.308 -0.355
ITEM10 -0.392 -0.481 0.541 -0.107 0.425 0.622
ITEM12 0.533 0.587 -0.609 0.252 -0.518 -0.567
ITEM13 0.429 0.542 -0.533 0.139 -0.432 -0.489
ITEM14 0.517 0.605 -0.583 0.311 -0.463 -0.567
ITEM15 1.000 0.442 -0.483 0.196 -0.367 -0.452
ITEM16 0.442 1.000 -0.527 0.210 -0.421 -0.612
ITEM17 -0.483 -0.527 1.000 -0.253 0.504 0.576
ITEM18 0.196 0.210 -0.253 1.000 -0.146 -0.199
ITEM20 -0.367 -0.421 0.504 -0.146 1.000 0.536
ITEM24 -0.452 -0.612 0.576 -0.199 0.536 1.000
Means
Variables ITEM2 ITEM9 ITEM10 ITEM12 ITEM13 ITEM14
1.267 1.729 3.754 0.908 1.354 1.392
1.592 1.508 3.192 2.017 3.029 2.792
Standard Deviations
1.496 1.522 1.590 1.360 1.496 1.496
1.647 1.489 1.591 1.579 1.564 1.679
Grid search of theta values on final iteration
FINAL RESULTS
6 passes completed.
Maximum change in any item parameter between last two iterations = 0.008412
Constant tau # 0 = 0.00000, Std. Error = 0.00000
Constant tau # 1 = -2.18278, Std. Error = 0.05318
ITEM DELTA STD.ERROR CLASS TAU STD.ERROR
ITEM2 -1.32894 0.06446
0 0.00000 0.00000
1 -2.18278 0.05318
2 -1.97099 0.05283
3 -1.60056 0.05444
4 -0.96555 0.05789
5 -0.74944 0.06515
.
.
.
ITEM24 2.47016 0.05496
0 0.00000 0.00000
1 -2.18278 0.05318
2 -1.97099 0.05283
3 -1.60056 0.05444
4 -0.96555 0.05789
5 -0.74944 0.06515
SUBJECT THETA STD.ERROR
1 2.06814 0.41482
2 3.16067 0.34780
.
.
.
239 2.40370 0.46135
240 1.51229 0.27371
RESPONSE FREQUENCIES
RESPONSE VALUES
0 1 2 3 4 5
ITEM2 104 55 33 23 11 14
ITEM9 66 53 51 37 16 17
ITEM10 19 12 15 32 47 115
ITEM12 138 43 30 11 8 10
ITEM13 91 61 45 14 13 16
ITEM14 93 53 41 26 14 13
ITEM15 90 44 40 25 22 19
ITEM16 83 53 43 32 18 11
ITEM17 26 15 23 59 57 60
ITEM18 52 52 45 41 31 19
ITEM20 24 24 25 63 56 48
ITEM24 34 27 35 50 47 47
Figure 95 GUM Theta Plot.
Figure 96 GUM Item Delta Values Plot.
Item Banking System
The Item Banking System is designed to provide a mechanism for entering and
saving multiple choice, true and false, completion and essay types of test questions. A
provision is made to save codes that help identify the major and minor topical areas
covered by each item as well as difficulty parameters for each item. The system can then
be used to generate a test based on a blueprint for that test. The test parameters can
include item selection based on a random sample from the item bank, item topics and
range of item difficulty.
To use the system, select the Measurement group under the Analysis option and
move to the Item Bank System group. There are three options under the Item Banking
sub-menu. The first is used to add or edit items of an item bank. It provides the option
for creating a new bank of items or editing and adding items to an existing bank.
If entering or revising itembank items, you will see the dialogue box below.
Figure 97 Form for Creating or Editing an Item Bank
Click on the Open New Bank button to start a new item bank or click on the Open Old
Bank button to modify an existing item bank. If you are starting a new item bank, DO
NOT add any extension to the file name (e.g. txt) because an extension of .BNK will
automatically be added to the bank name. The name you choose will become the name
of several other files that have additional extensions (e.g. ITM, MAJ, MIN, etc.) Once
you have created or re-opened an existing item bank, you can click on the Add New Item
button to begin the entry of a new item. The panel on the right side of the box is used to
enter information about the new item. You will need to click on a button for the type of
item, click on the None or enter the name of any bitmap which is to be shown with the
item, enter or select a major and minor integer code, indicate the number of choices (if a
multiple choice item format is used), and the correct choice. If you are creating a new
item bank, there will not be any major or minor topic codes in the drop-down
combination boxes. Simply type an integer for a major or minor code in the appropriate
combination box and press the enter key. That code will be added to the drop-down list
and available the next time you open the item bank. Once codes have been added to the
combination box lists, you can simply click once on a item in the list to indicate your
code choice. Should you accidentally add the same code more than once, select the last
duplicate from the list and when it is in the text box, double click it. A dialogue box will
appear which asks you if you would like to delete it.
When you are ready to enter the actual stem and foils (choices) for an item, click
on the Enter Info For This Item button. A series of screens will appear in which you
type the stem first, then each of the foils. DO NOT press the return key after each entry
but simply click the continue button. If you press the enter key while entering text for
your stem or foils, extra lines will appear when the test is typed or displayed on the
screen. When you have completed entering the information, stem and foils for an item,
click the Save This Item button before continuing to the next item or exiting the
program.
When an existing item bank is opened, the left panel will show the total number
of items in the item bank and the current item will be displayed (item 1 at first.) You can
scroll through the existing items (and their foils) to examine each one or select one for
revision. If you select an item to revise and click the Revise button, the panel on the right
will show the current information for the item. You can then click the changes desired in
the right panel and click the Enter Info for this item button to re-enter new stem and
foil information.
Figure 98 Sample Item Bank Opened for Editing
The display below is an example of an entry of a foil (choice) for a multiple choice item.
Figure 99 Entry Form for an Item Stem in the Item Bank System
Once you have created an item bank, you may then develop specifications for
administration of a test which contains items from the item bank. From the main menu,
select the Specify A Test option. You will then see the dialogue box shown below.
You may choose to administer all items in the bank, a random selection of items, items
selected for specific major and minor category codes or by hand-picking the items to
include in the test. You need to specify whether the tests are to be printed or
administered on the screen. You must also indicate the number of students to be tested
and enter identification numbers or names for each student. Once you have completed
the specifications for administering the test, click on the OK button. Your specifications
will be saved.
Figure 100 Item Bank Test Specification Form
The dialogue box below shows an example of entering a student Identification for
a subsequent test administration.
Figure 101 Entry of a Student ID in the Item Bank System Test Specification
To create tests or administer them on the screen (which ever you specified in the
previous section) you go to the Measurement group of the Statistics options and move to
the Item Banking group and click the Generate a Test option. You will then see the
following Form.
Figure 102 Item Bank System Test Administration Form
Click on the Open Item Bank to select the item bank for which test specifications have
been created. You can then click the Proceed with the test button to administer the test
(or print it if that method of administration had been specified in the test specifications.)
If a test is administered on the screen, each item is displayed and the student is
prompted to enter the CAPITAL letter of the correct choice (for multiple choice items) or
enter words, phrases, etc. on separate paper if short answer or essay types of questions are
asked. For MC and TF items, entering the letter of the correct choice in the Your Answer
box will result in immediate scoring of the question and calculating the current total
score.
Figure 103 Sample On-Screen Test Administration
When a subject completes the test on the screen or a test has finished printing, you
are returned to the test administration dialogue box. If the test were administered on the
screen and automatically scored, the total score will be displayed. Click the All Done
button to return to the main form.
Figure 104 Results Shown for an On-Screen Test
Hoyt and Cronbach Reliability
Test reliability may be assessed by means of an analysis of variance in which the
variance of the items within subjects and the variance between subjects are the major
components of the estimates. Four estimates are obtained - unadjusted and adjusted item
and test reliabilities. As an example, we will analyze a data file with eight test items.
The subject responses are coded 0 and 1 for wrong and right answers. There are 17
subjects in the file labeled ITEMREL.txt. The form for specifying the analysis is shown
below:
Figure 105 Reliability Analysis Using Repeated Measures Analysis of Variance
When the Compute button is clicked, the following output is obtained:
-----------------------------------------------------------
SOURCE DF SS MS F Prob. > F
-----------------------------------------------------------
SUBJECTS 16 12.882 0.805
WITHIN SUBJECTS 119 21.000 0.176
TREATMENTS 7 9.882 1.412 14.222 0.000
RESIDUAL 112 11.118 0.099
-----------------------------------------------------------
TOTAL 135 33.882 0.251
-----------------------------------------------------------
TREATMENT (COLUMN) MEANS AND STANDARD DEVIATIONS
VARIABLE MEAN STD.DEV.
VAR1 0.941 0.243
VAR2 0.824 0.393
VAR3 0.706 0.470
VAR4 0.588 0.507
VAR5 0.471 0.514
VAR6 0.353 0.493
VAR7 0.235 0.437
VAR8 0.118 0.332
RELIABILITY ESTIMATES
TYPE OF ESTIMATE VALUE
Unadjusted total reliability 0.781
Unadjusted item reliabiity 0.308
Adjusted total (Cronbach) 0.877
Adjusted item reliability 0.471
Three Parameter Logistic Scaling
Subject responses to test items may be a function of three parameters of the
individual items: (1) the difficulty of the item, (2) the degree to which the item
discriminates among individuals of differing ability, and (3) the probability of obtaining
the correct answer by chance alone. Each of these parameters may be estimated by this
procedure when two of them are known. Typically, one has estimates of the chance
probability of the item (parameter 3) and an initial estimate of slope (e.g. 1.0 as is
assumed in the Rasch model.) These estimates are entered in a file and used to obtain the
third parameter estimates in an iterative manner.
Statistical Process Control
Statistical Process Control (SPC) is a collection of statistical procedures applied
to industrial and manufacturing processes. The goal of SPC is typically to minimize the
variability of parts produced or services provided, reduce waste and achieve closer
tolerances. A number of companies strive for a "six sigma" control meaning they hope to
achieve the majority of their parts within tolerance such that the number out of tolerance
is less than the area under the normal curve beyond six standard deviations (a few parts
per million.)
XBAR Chart
Charts are the main method of communicating the degree to which processes are
under control. Samples are frequently taken of the process or product during
manufacturing and the mean measurement obtained for each. By plotting these means
the quality control person can judge whether or not the process is in control or needs
adjustment. As the process continues, the data collected consists of a lot number (the
sample number) and the measurement for quality of each work piece. A sample file
labeled BoltSize.txt may be loaded into OpenStat and analyzed by several of the SPC
procedures. When you elect the XBAR Chart procedure, you see the following dialog:
Figure 106 X Bar Chart Form
On the above form you can accept the default upper and lower control limits or enter your
own. You can also enter target specification limits for a product. To select variables,
simply click the variable representing the sample or lot number first, then click the
variable representing the criterion measurement. When you click the Compute button
you should see:
X Bar Chart Results
Group Size Mean Std.Dev.
_____ ____ _________ __________
1 5 19.88 0.37
2 5 19.90 0.29
3 5 20.16 0.27
4 5 20.08 0.29
5 5 19.88 0.49
6 5 19.90 0.39
7 5 20.02 0.47
8 5 19.98 0.43
Grand Mean = 19.97, Std.Dev. = 0.359, Standard Error of Mean = 0.16
Lower Control Limit = 19.493, Upper Control Limit = 20.457
Figure 107 X Bar Control Chart for BoltSize.txt file
You can see by the above that the means are well within the control limits.
Range Chart
As tools wear the products produced may begin to vary more and more widely
around the values specified for them. The mean of a sample may still be close to the
specified value but the range of values observed may increase. The result is that more
and more parts produced may be under or over the specified value. Therefore quality
assurance personnel examine not only the mean (XBAR chart) but also the range of
values in their sample lots. Again, examine the boltsize.txt file with the option Statistics /
Statistical Process Control / Control Charts / Range Chart. Shown below is the
specification form and the results:
Figure 108 Specifications for an SPC Range Chart
Group Size Mean Range Std.Dev.
_____ ____ _________ _______ ________
1 5 19.88 0.90 0.37
2 5 19.90 0.70 0.29
3 5 20.16 0.60 0.27
4 5 20.08 0.70 0.29
5 5 19.88 1.20 0.49
6 5 19.90 0.90 0.39
7 5 20.02 1.10 0.47
8 5 19.98 1.00 0.43
Mean Range = 0.89
Figure 109 SPC Range Chart for BoltSize.txt
The S Chart
The sample standard deviation, like the range, is also an indicator of how much
values vary in a sample. While the range reflects the difference between largest and
smallest values in a sample, the standard deviation reflects the square root of the average
squared distance around the mean of the values. We desire to reduce this variability in
our processes so as to produce products as similar to one another as is possible. The S
control chart plot the standard deviations of our sample lots and allows us to see the
impact of adjustments and improvements in our manufacturing processes.
Examine the boltsize.txt data with the S Control Chart. Shown below is the
specification form for the analysis and the results obtained:
Figure 110 Specification Form for the SPC Sigma Chart
X Bar Chart Results
Group Size Mean Std.Dev.
_____ ____ _________ ________
1 5 19.88 0.37
2 5 19.90 0.29
3 5 20.16 0.27
4 5 20.08 0.29
5 5 19.88 0.49
6 5 19.90 0.39
7 5 20.02 0.47
8 5 19.98 0.43
Mean Sigma = 0.37
Figure 111 S Control Chart for the BoltSize.txt Data.
CUSUM Chart
The cumulative sum chart, unlike the previously discussed SPC charts (Shewart
charts) reflects the results of all of the samples rather than single sample values. It plots
the cumulative sum of deviations from the mean or nominal specified value. If a process
is going out of control, the sum will progressively go more positive or negative across the
samples. If there are M samples, the cumulative sum S is given as
M _
S = ( X
i
-
o
)
i=1
_
Where X
i
is the observed sample mean and
o
is the nominal value or (overall mean.)
It is often desirable to draw some boundaries to indicate when a process is out of control.
By convention we use a standardized difference to specify this value. For example with
the boltsize.txt data, we might specify that we wish to be sensitive to a difference of 0.02
from the mean. To standardize this value we obtain
0.02
= -----------

x
or using our sample values as estimates obtain
0.02 0.02
= ------ = ------- = 0.0557
S
x
0.359
A V Mask is then drawn starting at a distance d from the last cumulative sum value
with an angle back toward the first sample deviation. In order to calculate the distance
d we need to know the probabilities of a Type I and Type II error, that is, the probability
of incorrectly concluding that a shift to out-of-control has taken place and the probability
of failing to detect an out-of-control condition. If these values are specified then we can
obtain the distance d as
2 1 -
d = (----) ln (---------)

2

When you run the CUSUM procedure you will note that the alpha and beta error rates
have been set to default values of 0.05 and 0.20. This would imply that an error of the
first type (concluding out-of-control when in fact it is not) is a more expense error than
concluding that the process is in control when in fact it is not. Depending on the cost of
shut-down and correction of the process versus scraping of parts out of tolerance, you
may wish to adjust these default values.
The angle of the V mask is obtained by

= tan
-1
(----)
2k
where k is a scaling factor typically obtained as k = 2
x
The specification form for the CUSUM chart is shown below for the data file
labeled boltsize.txt. We have specified our desire to detect shifts of 0.02 in the process
and are using the 0.05 and 0.20 probabilities for the two types of errors.
Figure 112 CUMSUM Control Chart Specification Form
CUMSUM Chart Results
Group Size Mean Std.Dev. Cum.Dev. of
mean from Target
_____ ____ ________ ________ ___________
1 5 19.88 0.37 -0.10
2 5 19.90 0.29 -0.18
3 5 20.16 0.27 0.00
4 5 20.08 0.29 0.10
5 5 19.88 0.49 -0.00
6 5 19.90 0.39 -0.08
7 5 20.02 0.47 -0.04
8 5 19.98 0.43 -0.04
Mean of group deviations = -0.005
Mean of all observations = 19.975
Std. Dev. of Observations = 0.359
Standard Error of Mean = 0.057
Target Specification = 19.980
Figure 113 SPC CUMSUM Chart for BoltSize.txt Data.
P Chart
In some quality control processes the measure is a binomial variable indicating
the presence or absence of a defect in the product. In an automated production
environment, there may be continuous measurement of the product and a tagging of the
product which is non-conforming to specifications. Due to variation in materials, tool
wear, personnel operations, etc. one may expect that a certain proportion of the products
will have defects. The p Chart plots the proportion of defects in samples of the same size
and indicates by means of upper and lower control limits, those samples which may
indicate a problem in the process.
To demonstrate the p Chart we will utilize a file labeled pchart.txt. Load the file
and select the Analyses / Statistical Process Control / p Chart option. The specification
form is shown below along with the results obtained after clicking the Compute Button:
Figure 114 SPC p Control Chart Specification Form.
Target proportion = 0.0100
Sample size for each observation = 1000
Average proportion observed = 0.0116
Defects p Control Chart Results
Sample No. Proportion
__________ __________
1 0.012
2 0.015
3 0.008
4 0.010
5 0.004
6 0.007
7 0.016
8 0.009
9 0.014
10 0.010
11 0.005
12 0.006
13 0.017
14 0.012
15 0.022
16 0.008
17 0.010
18 0.005
19 0.013
20 0.011
21 0.020
22 0.018
23 0.024
24 0.015
25 0.009
26 0.012
27 0.007
28 0.013
29 0.009
30 0.006
Target proportion = 0.0100
Sample size for each observation = 1000
Average proportion observed = 0.0116
Figure 115 SPC p Chart
Defect (Nonconformity) c Chart
The previous section discusses the proportion of defects in samples (p Chart.)
This section examines another defect process in which there is a count of defects in a
sample lot. In this chart it is assumed that the occurrence of defects are independent, that
is, the occurrence of a defect in one lot is unrelated to the occurrence in another lot. It is
expected that the count of defects is quite small compared to the total number of parts
potentially defective. For example, in the production of light bulbs, it is expected that in
a sample of 1000 bulbs, only a few would be defective. The underlying assumed
distribution model for the count chart is the Poisson distribution where the mean and
variance of the counts are equal. Illustrated below is an example of processing a file
labeled Defects.OS4.
Figure 116 Defects Conformity Chart (c Chart) Form
Clicking the Compute Button results in the following:
Defects c Control Chart Results
Sample Number of
Noncomformities
______ _______________
1 12.00
2 15.00
3 8.00
4 10.00
5 4.00
6 7.00
7 16.00
8 9.00
9 14.00
10 10.00
11 5.00
12 6.00
13 17.00
14 12.00
15 22.00
16 8.00
17 10.00
18 5.00
19 13.00
20 11.00
21 20.00
22 18.00
23 24.00
24 15.00
25 9.00
26 12.00
27 7.00
28 13.00
29 9.00
30 6.00
Total Nonconformities = 347.00
No. of samples = 30
Poisson mean and variance = 11.567
Figure 117 SPC c Chart for Defects.
Defects Per Unit u Chart
Like the count of defects c Chart described in the previous section, the u Chart
describes the number of defects per unit. It is assumed that the number of units observed
is the same for all samples. We will use the file labeled Defects.txt as our example. In
this set of data, 30 observations of defects for 1000 units each are recorded. The
assumption is that defects are distributed as a Poisson distribution with the mean given as
_ c
u = --------- where c is the count of defects and n is the number of units observed.
n
and ____ ___
/_ / _
_ / u _ / u
UCL = u + sigma / ------ and LCL = u - sigma / --------
n n
The specification form and results for the computation following the click of the
Compute button are shown below:
Figure 118 Defects Per Unit SPC Form.
When you click the Compute button you obtain:
Defects u Control Chart Results
Sample No Defects Defects Per Unit
______ __________ ________________
1 12.00 0.01
2 15.00 0.01
3 8.00 0.01
4 10.00 0.00
5 4.00 0.01
6 7.00 0.02
7 16.00 0.01
8 9.00 0.01
9 14.00 0.01
10 10.00 0.01
11 5.00 0.01
12 6.00 0.02
13 17.00 0.01
14 12.00 0.02
15 22.00 0.01
16 8.00 0.01
17 10.00 0.01
18 5.00 0.01
19 13.00 0.01
20 11.00 0.02
21 20.00 0.02
22 18.00 0.02
23 24.00 0.01
24 15.00 0.01
25 9.00 0.01
26 12.00 0.01
27 7.00 0.01
28 13.00 0.01
29 9.00 0.01
30 6.00 0.00
Total Nonconformities = 347.00
No. of samples = 30
Def. / unit mean = 0.012 and variance = 0.003
Figure 119 Defects Per Unit SPC Chart for the file Defects.txt.
Financial Procedures
Personal and business financial planning typically attempts to increase the value
or worth of the planner. In many societies one must borrow money to purchase major
items such as homes and automobiles. In addition, many may discover they have some
cash they can invest. A small entrepreneur may have borrowed funds for capital
investment in his or her company and has a need to determine how to price a product or
service to maintain a positive cash flow. While these activities are not specifically
"statistics" in the sense of testing hypotheses or estimating parameters, they are often
performed by the same individuals responsible for completing statistical analyses (for
example, financial officers.)
Loan Amortization
Assume you wish to buy a car for $20,000.00 over 5 years with 12 equal
payments per year. You would like an amortization table that shows you how much
interest you are paying each month (and total.) In addition, you would like to know at
each month how much you still owe and how much you have paid on the amount
borrowed. You select the Loan Amortization option and complete the dialogue box
shown below:
Figure 120 Loan Amortization Specification Form.
The results obtained are:
Payment Schedule Program by W. G. Miller
----------------------------------------------------------------------------
Name of Borrower : I.M. Costly
Amount borrowed = $ 20000.00 at 5.00 percent over 5.0 years.
----------------------------------------------------------------------------
PAYMENT PAYMENT INTEREST BALANCE TOTAL TOTAL
NUMBER AMOUNT PAYED REMAINING INTEREST PAID
----------------------------------------------------------------------------
9/13/2004 377.42 83.33 19705.91 83.33 377.42
9/13/2004 377.42 82.11 19410.59 165.44 754.85
9/13/2004 377.42 80.88 19114.04 246.32 1132.27
9/13/2004 377.42 79.64 18816.26 325.96 1509.70
9/13/2004 377.42 78.40 18517.24 404.36 1887.12
9/13/2004 377.42 77.16 18216.97 481.52 2264.55
9/13/2004 377.42 75.90 17915.45 557.42 2641.97
9/13/2004 377.42 74.65 17612.67 632.07 3019.40
9/13/2004 377.42 73.39 17308.63 705.45 3396.82
9/13/2004 377.42 72.12 17003.33 777.57 3774.25
9/13/2004 377.42 70.85 16696.75 848.42 4151.67
9/13/2004 377.42 69.57 16388.89 917.99 4529.10
----------------------------------------------------------------------------
----------------------------------------------------------------------------
----------------------------------------------------------------------------
9/13/2005 377.42 68.29 16079.76 986.28 4906.52
9/13/2005 377.42 67.00 15769.33 1053.28 5283.95
9/13/2005 377.42 65.71 15457.61 1118.98 5661.37
9/13/2005 377.42 64.41 15144.59 1183.39 6038.79
9/13/2005 377.42 63.10 14830.27 1246.49 6416.22
9/13/2005 377.42 61.79 14514.64 1308.28 6793.64
9/13/2005 377.42 60.48 14197.69 1368.76 7171.07
9/13/2005 377.42 59.16 13879.43 1427.92 7548.49
9/13/2005 377.42 57.83 13559.83 1485.75 7925.92
9/13/2005 377.42 56.50 13238.91 1542.25 8303.34
9/13/2005 377.42 55.16 12916.64 1597.41 8680.77
9/13/2005 377.42 53.82 12593.04 1651.23 9058.19
----------------------------------------------------------------------------
----------------------------------------------------------------------------
----------------------------------------------------------------------------
9/13/2006 377.42 52.47 12268.09 1703.70 9435.62
9/13/2006 377.42 51.12 11941.78 1754.82 9813.04
9/13/2006 377.42 49.76 11614.11 1804.58 10190.47
9/13/2006 377.42 48.39 11285.08 1852.97 10567.89
9/13/2006 377.42 47.02 10954.67 1899.99 10945.32
9/13/2006 377.42 45.64 10622.89 1945.63 11322.74
9/13/2006 377.42 44.26 10289.73 1989.90 11700.16
9/13/2006 377.42 42.87 9955.18 2032.77 12077.59
9/13/2006 377.42 41.48 9619.24 2074.25 12455.01
9/13/2006 377.42 40.08 9281.89 2114.33 12832.44
9/13/2006 377.42 38.67 8943.14 2153.00 13209.86
9/13/2006 377.42 37.26 8602.98 2190.27 13587.29
----------------------------------------------------------------------------
----------------------------------------------------------------------------
----------------------------------------------------------------------------
9/13/2007 377.42 35.85 8261.40 2226.11 13964.71
9/13/2007 377.42 34.42 7918.40 2260.54 14342.14
9/13/2007 377.42 32.99 7573.97 2293.53 14719.56
9/13/2007 377.42 31.56 7228.10 2325.09 15096.99
9/13/2007 377.42 30.12 6880.79 2355.20 15474.41
9/13/2007 377.42 28.67 6532.04 2383.87 15851.84
9/13/2007 377.42 27.22 6181.83 2411.09 16229.26
9/13/2007 377.42 25.76 5830.16 2436.85 16606.69
9/13/2007 377.42 24.29 5477.03 2461.14 16984.11
9/13/2007 377.42 22.82 5122.43 2483.96 17361.53
9/13/2007 377.42 21.34 4766.35 2505.31 17738.96
9/13/2007 377.42 19.86 4408.78 2525.17 18116.38
----------------------------------------------------------------------------
----------------------------------------------------------------------------
----------------------------------------------------------------------------
9/13/2008 377.42 18.37 4049.73 2543.54 18493.81
9/13/2008 377.42 16.87 3689.18 2560.41 18871.23
9/13/2008 377.42 15.37 3327.12 2575.78 19248.66
9/13/2008 377.42 13.86 2963.56 2589.64 19626.08
9/13/2008 377.42 12.35 2598.48 2601.99 20003.51
9/13/2008 377.42 10.83 2231.89 2612.82 20380.93
9/13/2008 377.42 9.30 1863.76 2622.12 20758.36
9/13/2008 377.42 7.77 1494.10 2629.88 21135.78
9/13/2008 377.42 6.23 1122.90 2636.11 21513.21
9/13/2008 377.42 4.68 750.16 2640.79 21890.63
9/13/2008 377.42 3.13 375.86 2643.91 22268.06
9/13/2008 377.42 1.57 -0.00 2645.48 22645.48
----------------------------------------------------------------------------
Future Value
FutureValue returns the future value of an investment of PresentValue where Payment is
invested for NPeriods at the rate of Rate per period. The PaymentTime parameter
indicates whether the investment is an ordinary annuity or an annuity due (enter
EndOfPeriod if payments are at the end of each period, StartOfPeriod if they are at the
beginning).
EXAMPLE: What is the value 20 years from now of a savings account if you invest
$100.00 per month in an account that pays 5.25% annual interest, assuming that
the interest is compounded monthly?
Rate Per Period= 0.0525 / 12 = 0.004375
Number of Periods = 20 years times 12 months = 240 periods
Payment = -100.00 (NOTE: payments are always negative)
Present Value = 0.0
ANSWER: Future Value = 42311.1776128932 or
approximately $42,311.18
Shown below is the dialogue box you would complete for the above example:
Figure 121 Future Value Form.
Payment
Payment calculates the fully amortized payment of borrowing PresentValue at Rate
percent per period over NPeriods. It assumes that interest is paid at the end of each
period.
FutureValue is the value that the investment will reach at some point. PaymentTime
indicates whether the cash flows occur at the beginning or end of the period.
Example: How much would you have to pay into a savings account monthly in order for
that savings account to be worth $50,000.00 after 20 years, assuming that the savings
account pays 5.25% annual percentage rate (APR) compounded monthly?
Rate = 0.0525 / 12 = 0.004375
Number of Periods = 20 * 12 = 240
Present Value = 0.0
Future Value = 50000.00
Time of Payment = After the period
Answer: Payment = -118.172083172565
NOTE: Payments are always indicated as negative
Shown below is the dialogue box you would complete for the above example:
Figure 122 Payment Calculation Form.
Period Payment
Period Payment gives the part of the payment that is principal. The
InterestPayment function gives the part of the payment that is interest. The Rate
parameter is the interest rate. Period is the number of periods into the loan for which the
principal is desired, and NPeriods is the number of periods of the loan. FutureValue is the
value the investment will reach at some point. Payment Time indicates whether the cash
flows occur at the beginning or end of the period.
EXAMPLE: What is the principle payment on payment 1 of an auto loan of $20,000.00
borrowed at 7.5% Annual Percentage Rate (APR) over four years with monthly
payments?
Rate = 0.075 / 12 = 0.00625
Present Value = 20000.00
Period = 1
Number of periods = 4 * 12 = 48
Future Value = 0.0
ANSWER: (Note: payments are always negative values)
The first payment on the principle is about $72.02
Figure 123 Period Payment Form.
Interest Payment
Interest Payment calculates the portion of a loan payment that reflects the interest.
Rate represents the fixed periodic interest rate. Period identifies the payment period.
NPeriods is the number of periods of the loan. PresentValue represents the amount
borrowed (the principal). FutureValue is the future value of the investment.
PaymentTime indicates whether the cash flows occur at the beginning or end of the
period.
EXAMPLE: What is the interest payment on payment 1 of an auto loan of $20,000.00
borrowed at 7.5% Annual Percentage Rate (APR) over four years with monthly
payments?
Rate per period = 0.075 / 12 = 0.00625. Present Value = 20000.00.
Period = 1, Number of periods = 4 * 12 = 48. Future Value = 0.0.
ANSWER: -125.00 (Note: payments are always negative values)
The first payment contains $125.00 of interest charges.
The dialogue box for the calculations is shown below:
Figure 124 Interest Paid On A Loan Form.
Interest Rate
InterestRate calculates the interest rate required in order for an investment of
PresentValue, with periodic payments of Payment, to be worth FutureValue within
NPeriods compounding periods. If NPeriods represents years, an annual interest rate
results; if NPeriods represents months, a monthly interest rate results, and so on. The
PaymentTime parameter indicates whether the cash flows occur at the beginning or end
of the period.
EXAMPLE: What is the Annual Percentage Rate (APR) on a four-year auto loan for a
$20,000.00 car when the monthly payment is $483.58?
Number of periods = 4 years * 12 months = 48.
Payment = -483.58 (NOTE: payments are always negative.)
Present Value = 20000, Future Value = 0.0, Payment Time = End of Period.
ANSWER: Rate = 0.0625 * 12 or about 7.5% APR.
The dialogue box for this procedure is shown below with the example entered:
Figure 125 Interest Rate Calculation Form.
Number of Periods
NumberOfPeriods computes the number of payment periods required for an
investment of PresentValue to reach a value of FutureValue, while making regular
payments of Payment and accruing interest at the rate of Rate per compounding
period. PaymentTime indicates whether the cash flows occur at the beginning or end of
the period.
EXAMPLE: How long would you have to make payments on a car loan for a $20,000.00
car assuming that the payment is to be no more than $450.00, and that the annual
percentage rate (APR) on the loan is 7.5% ?
Rate = 0.075 / 12 = 0.00625
Payment = -$450.00 (NOTE- payments are always negative)
Present Value = $20000.00
Future Value = 0.0
ANSWER: Number of Payments = 52.2301263064951 or
about 53 payments.
Figure 126 Number of Payment Periods Form.
Present Value of an Investment
PresentValue calculates the present value of an investment where Payment is
received for NPeriods and is discounted at the rate of Rate per period. FutureValue is the
value the investment may reach at some point. PaymentTime indicates whether the cash
flows occur at the beginning or end of the period.
EXAMPLE: What was the amount borrowed in a 7.5% Annual Percentage Rate (APR)
four-year auto loan when the payment is $500.00 per month?
Rate per period = 0.075 / 12 = 0.00625
Number of periods = 4 years * 12 months per year = 48
Payment = -500.00
Future Value = 0.0
Payment Time = End of Period
ANSWER: Present Value = 20679.1855679664
or about $20679.19 was borrowed.
The dialogue box for this procedure with the example entered is shown below:
Figure 127 Present Value Form.
Double Declining Balance
Double Declining Balance determines accelerated depreciation values for an
asset, given the initial cost, life expectancy, end value, and depreciation period.
EXAMPLE:
What is the depreciation value for a computer with a life expectancy of three years if it
initially cost $2,000.00 with no expected value at the end of the three years?
Initial Cost = 2000.00
Life Expectancy = 3 years
End Value = 0.0
Depreciation Period = 3 years
ANSWER: Approximately $148.15
The answer is obtained using the dialogue box shown below:
Figure 128 Double Declining Value Form.
Straight Line Depreciation
SLN Depreciation calculates the straight-line depreciation allowance for an asset
over one period of its life. The function divides the Cost minus the Salvage by the
number of years of useful Life of the asset. Cost is the amount initially paid for the asset.
Salvage is the value of the asset at the end of its useful life.
To compute accelerated depreciation (allowing higher depreciation values in the
first years of the assets life), use the SYD Depreciation (accelerated) function.
EXAMPLE:
What is the depreciation value that one may claim for a computer purchased for $2000.00
and expected to have a useful life of three years with no residual value?
ANSWER: Approximately $666.67
Shown below is the dialogue box and example:
Figure 129 Straight Line Depreciation Form.
Accelerated Depreciation
SYD Depreciation (for "sum-of-years-digits depreciation") calculates depreciation
amounts for an asset using an accelerated depreciation method. This allows for higher
depreciation in the earlier years of an asset's life.
Cost is the initial cost of the asset. Salvage is the value of the asset at the end of
its life expectancy. Life is the length of the asset's life expectancy. Period is the period
for which to calculate the depreciation.
EXAMPLE:
What is the depreciation value for the first period that one may claim for a computer
purchased for $2000.00 and expected to have a useful life of three years with no residual
value?
ANSWER: $1000.00 for the first year. If computed for year 2 the answer is
approximately $666.67 and if done for year 3 yields $333.33.
Below is the dialogue box for this procedure with the problem entered for year 1:
Figure 130 Accelerated Depreciation Schedule Form.
Internal Rate of Return
This procedure determines the internal rate of return on an investment. It
references the main grid that contains cash flow information and uses the supplied
internal rate of return estimate to calculate results.
Before using this function, open a file containing expected cash flow amounts over a
period of time. It is assumed that the amounts are received at regular intervals. Negative
amounts are interpreted as cash outflows, and positive amounts as inflows. The first
amount must be a negative number, to reflect the initial investment. The following
amounts can all be the same for each time period, or they can be different (including a
mixture of negatives, positives, or zeros).
Specify the estimated rate of return as the Guess parameter. Pass the array of expected
cash flow amounts as the CashFlows parameter.
Figure 131 Internal Rate of Return Form.
The above was obtained using a file labeled CashFlow.txt. The file contains the
following information:
Receipts
-2000.00
100.00
100.00
200.00
100.00
-50.00
75.00
200.00
250.00
35.00
100.00
400.00
Net Present Value
The Net Present Value procedure calculates the current value of an array of
estimated cash flow values, discounted at the given interest rate of Rate. Payment Time
indicates whether the cash flows occur at the beginning or end of the period. Net Present
Value helps determine how much an investment is currently worth, based on expected
earnings, although its accuracy depends on the accuracy of the cash flows in the array.
Before using this function, open a file containing expected cash flow amounts
over a period of time. It is assumed that the amounts are received at regular intervals.
Negative amounts are interpreted as cash outflows, and positive amounts as inflows. The
first amount must be a negative number, to reflect the initial investment. The following
amounts can all be the same for each time period, or they can be different (including a
mixture of negatives, positives, or zeros).
Specify the estimated rate of return as the Guess parameter. Pass the array of expected
cash flow amounts as the CashFlows parameter.
Figure 132 Net Present Value Form.
Simulation Procedures
The Simulations menu of OpenStat contains procedures for creating simulated
data and analyses to assist you in understanding various statistics and statistical tests.
Correlation
In learning the fundamental ideas about the Pearson Product-Moment correlation,
it is useful to have a graphic image of the plot between two, approximately normally
distributed variables. It is even more helpful if that image shows the straight line of "best
fit" for the "least-squares" solution of the relationship between the two variables and if
the marginal distributions are shown along with the means, standard deviations and
correlation of the variables.
This procedure provides a specification form for specification of population data.
You enter the desired means and standard deviations for two variables and the desired
population correlation as well as the number of data points. When you click the Do It
button, the plot of the data points, correlation and marginal distributions is obtained.
Shown below is the printed data and the plot for 100 points with population values of 100
for the means, 5 for the standard deviations and 0.95 for the correlation:
POPULATION PARAMETERS FOR THE SIMULATION
Mean X = 100.000, Std. Dev. X = 5.000
Mean Y = 100.000, Std. Dev. Y = 5.000
Product-Moment Correlation = 0.950
Regression line slope = 0.950, constant = 5.000
SAMPLE STATISTICS FOR 100 OBSERVATIONS FROM THE POPULATION
Mean X = 98.991, Std. Dev. X = 5.400
Mean Y = 98.905, Std. Dev. Y = 5.404
Product-Moment Correlation = 0.960
Regression line slope = 0.961, constant = 3.793
Figure 133 Simulated Correlation Between Two Variables.
Generating Multivariate Normal Distribution Data
To become familiar with various multivariate statistical tests, one needs to have
data to analyze. With this procedure you can generate one or more distributions of
variable values from defined populations. As the user, you specify the population
correlations among the variables as well as the means and standard deviations of the
variables. The procedure then generates a sample of values from the specified population
for whatever sample size you request. Shown below is the specification form for this
procedure and the results placed in the data grid. Be sure you have closed all files before
using this procedure. You may save the generated file and then use it to demonstrate
various statistical analysis procedures in your OpenStat package.
Figure 134 Simulation of a Multivariate Distribution.
POPULATION CORRELATION MATRIX, MEANS AND STANDARD DEVIATIONS
VAR. 1 VAR. 2 VAR. 3 VAR. 4 VAR. 5
VAR. 1 1 .8 .3 -.4 .5
VAR. 2 .8 1 .2 -.01 .3
VAR. 3 .3 .2 1 -.6 .7
VAR. 4 -.4 -.01 -.6 1 -.03
VAR. 5 .5 .3 .7 -.03 1
MEANS: 20 30 40 50 60
STD.DEV.s 3 4 5 6 7
POPULATION COVARIANCE MATRIX
9.000 9.600 4.500 -7.200 10.500
9.600 16.000 4.000 -0.240 8.400
4.500 4.000 25.000 -18.000 24.500
-7.200 -0.240 -18.000 36.000 -1.260
10.500 8.400 24.500 -1.260 49.000
Determinant of your population matrix = -268372.9843
Figure 135 Simulated Multivariate Data in the Grid .
If you now use the Correlation procedure to analyze the simulated data you obtain the
sample statistics:
CORRELATION MATRIX
Correlations
VAR1 VAR2 VAR3 VAR4 VAR5
VAR1 1.000 0.529 0.352 -0.273 0.437
VAR2 0.529 1.000 0.329 -0.148 0.564
VAR3 0.352 0.329 1.000 -0.606 0.562
VAR4 -0.273 -0.148 -0.606 1.000 -0.086
VAR5 0.437 0.564 0.562 -0.086 1.000
Means
Variables VAR1 VAR2 VAR3 VAR4 VAR5
19.734 30.464 40.070 50.942 59.603
Standard Deviations
Variables VAR1 VAR2 VAR3 VAR4 VAR5
2.802 5.015 5.142 6.807 9.203
Estimating Sample Sizes Needed for Type I and Type II Error Control
The beginning researcher soon learns about the relationship between a dependent
variables standard deviation and the errors of inference for samples of various size.
Unfortunately, many beginning researchers only attend to the probability of rejecting a
hypothesis about their variable due to random sampling variability that occurs from
sample to sample drawn from a population. They should, of course, be equally
concerned about accepting a null hypothesis when, in fact, in the population sampled they
should reject the null hypothesis. These two types of error are often known as Type I and
Type II errors and each may have different costs and consequences. The cost, for
example, in human lives by accepting the null hypothesis that two drug treatments are
equally effective when, in fact, one is more effective may be a greater cost than the
accidental rejection of the null hypothesis for two equally effective treatments.
One can estimate the sample size needed to control for BOTH Type I and Type II
errors for many statistical analyses. Books and articles abound for estimating these sizes
depending on the statistical test to be used. Frequently however, one can obtain a
reasonable estimate if the statistic to be employed can be approximately transformed into
a z-test between two means. For example, in completing an Analysis of Variance, the F
ratio of variability between groups to the variability within groups is used as the test
statistic. However, if one considers two of the groups to be contrasted, a z-test could be
used for inferring whether or not there means differ beyond that expected due to random
sampling variability. Thus the sample size may be determined for all of the groups by
simply specifying the Type I and Type II error controls desired for any two groups and
using the standard deviation estimated by the square root of the pooled within group
variance. Shown below is the specifications form for this procedure:
Figure 136 Form for Sample Size of a z-Test.
When the Continue button is clicked, the results below are obtained:
Figure 137 Sample Size Estimate for a z-Test.
Notice a very small size is needed (N=4) for a difference of 5 in our example. Now if
you select a much smaller size for the alternative mean such as 101, you would get:
Figure 138 Sample Size for a z-Test (Smaller Difference Between Means.)
Notice that now we would require a much larger sample (N = 289) in order to control
both type I and type II error at the 0.05 level.
Generating Power Curves for z-Tests
The student of statistics learns that "power" of a statistical test involves the
sensitivity of the test for accepting the null hypothesis when, in fact, one should. It is
obtained as 1.0 - beta where beta is the probability of making a Type II error (accepting
the null hypothesis due to random sampling variability when one should have rejected.)
This power is a function of the alpha rate accepted by the researcher (probability of a
Type I error) as well as the difference between the null and alternative hypothesized
statistic and the standard deviation of the statistic (which is, itself, a function of sample
size.)
This procedure plots the power curves for various levels of Type I error given the
standard deviation of the statistic. Shown below is the specification form for the plot and
the results obtained.
Figure 139 Form for Simulation of Power Curves.
Figure 140 Power Curves for a Mean of 100 and N = 50 with Levels of Alpha
Power of the z-test for Alternate Hypotheses
Alpha Levels: 0.01 0.05 0.10 0.20
X = 100.00 0.010 0.050 0.100 0.200
X = 100.04 0.011 0.055 0.109 0.214
X = 100.07 0.013 0.061 0.119 0.229
X = 100.11 0.015 0.067 0.129 0.245
X = 100.14 0.017 0.074 0.140 0.261
X = 100.18 0.019 0.082 0.151 0.277
X = 100.21 0.021 0.089 0.163 0.294
.
.
.
X = 102.62 0.915 0.980 0.992 0.998
X = 102.65 0.923 0.982 0.993 0.998
X = 102.69 0.930 0.984 0.994 0.998
X = 102.72 0.936 0.986 0.995 0.999
X = 102.76 0.942 0.988 0.996 0.999
X = 102.79 0.948 0.989 0.996 0.999
Theoretical Distributions
Various sample statistics have unique sampling distributions. A researcher often
wishes to know the "critical value" for rejecting a hypothesis with a specified probability.
For example, what is the critical value for rejecting a null hypothesis at the 95%
confidence limit when utilizing a chi-square distribution with 10 degrees of freedom?
This procedure lets the user plot a z, t, chi-square or F statistic sampling
distribution and identify the critical value for a given probability. Shown below is the
specification form and an example plot:
Figure 141 Simulated Distribution Cumulative Distribution Frequency.
Figure 142 Probability Distribution of a Simulated F Statistic.
Generate Sequential Values
It is sometimes convenient to be able to create a variable that simply consists of a
sequence of values for N cases. Sometimes it is desired that the sequence start at a
specific value for case 1 and subsequent cases have values incremented by some value X.
This procedure provides that capability as shown in the specification form below and a
snap-shot of the values generated in the data grid.
Figure 143 Generation of Sequential Values.
The values generated in the Main Form grid are shown below:
1
6
11
16
21
26
31
36
41
46
51
56
61
66
71
76
81
86
91
96
Random Number Generation
Your OpenStat package contains this procedure which permits you to generate
random values from several theoretical distributions such as a flat distribution, normal
distribution (Gaussian), etc. You may want to create such data in your data grid simply
to plot such values for educational purposes or to experiment with statistical procedures
in which the assumption of normality is violated. In any event, one can have some fun!
The specification form below is seen when this procedure is selected. A snap-shot of the
values entered in the data grid is also presented.
Figure 144 Simulation of Scores from a Theoretical Distribution.
Figure 145 Form for the Degrees of Freedom in Simulation Chi-Square Values
Figure 146 Generated Chi-Square Values in the Grid.
Simulate 2-Way Analysis of Variance
In teaching or learning statistics, it is valuable to have the ability to create sample
problems. This procedure permits you to create data that fits an analysis of variance
model with effects for rows and for columns as well as the interaction of rows and
columns. The data generated can then be analyzed with the Analysis of Variance
procedures. Of course, students might first try to do the calculations themselves by hand!
The dialogue box for this Simulation option is shown below:
Figure 147 Simulation Form for a Two Way Analysis of Variance.
RESULTS FOR SIMULATED TWO-WAY ANOVA DATA
No. of Treatment Levels for Factor A = 2
No. of Treatment Levels for Factor B = 3
No. of observations per cell = 20
Population Mean Desired = 100.000
Population Standard Deviation Desired = 15.000
FACTOR A TREATMENT EFFECTS
Level Effect
1 0.000
2 0.000
FACTOR B TREATMENT EFFECTS
Level Effect
1 0.000
2 0.000
3 0.000
TREATMENT INTERACTION EFFECTS
A Level B Level Effect
1 1 5.000
1 2 -5.000
1 3 -0.000
2 1 -5.000
2 2 5.000
2 3 -0.000
Figure 148 Simulated Grid Data for a Two Way Analysis of Variance.
The analysis of the above generated data with the ANOVA procedure yields:
Two Way Analysis of Variance
Variable analyzed: YScore
Factor A (rows) variable: ATreat. (Fixed Levels)
Factor B (columns) variable: BTreat. (Fixed Levels)
SOURCE D.F. SS MS F PROB.> F Omega Squared
Among Rows 1 395.220 395.220 2.065 0.153 0.007
Among Columns 2 675.566 337.783 1.765 0.176 0.011
Interaction 2 4700.517 2350.258 12.278 0.000 0.155
Within Groups 114 21821.096 191.413
Total 119 27592.398 231.869
Omega squared for combined effects = 0.173
Note: Denominator of F ratio is MSErr
GROUP Row Col. N MEAN VARIANCE STD.DEV.
Cell 1 1 20 108.432 192.908 13.889
Cell 1 2 20 87.823 202.714 14.238
Cell 1 3 20 101.097 287.111 16.944
Cell 2 1 20 90.136 123.634 11.119
Cell 2 2 20 100.111 197.112 14.040
Cell 2 3 20 96.217 145.000 12.042
Row 1 60 99.117 293.848 17.142
Row 2 60 95.488 167.121 12.928
Col 1 40 99.284 240.052 15.494
Col 2 40 93.967 233.502 15.281
Col 3 40 98.657 216.621 14.718
TOTAL 120 97.303 231.869 15.227
TESTS FOR HOMOGENEITY OF VARIANCE
---------------------------------------------------------------------
Hartley Fmax test statistic = 2.32 with deg.s freem: 6 and 19.
Cochran C statistic = 0.25 with deg.s freem: 6 and 19.
Bartlett Chi-square statistic = 9.24 with 5 D.F. Prob. larger value = 0.100
---------------------------------------------------------------------
Figure 149 Plot of Cell Means for the Simulated Two Way ANOVA Data.
Project Evaluation Generator
When you select the evaluation generator procedure you will see the dialogue box
shown below:
Figure 150 Form for Specifying a Simulated Evaluation Report.
You complete the information for the project name, authors, company and date. Click on
the buttons for the type of evaluation and the model for the evaluation. Check the targets
and resource requirement boxes that will be a part of your proposal. Once you have
made those selections, a Rich Text File (compatable to a variety of word processing
programs) will be created. You can edit this file to enter specific plans and information
for your evaluation. The generator simply provides a standard template that you can use
to create evaluation proposals. Some of the topic headings generated automatically are
shown below:
Project Description
Scope of the Evaluation
Alternative Outcomes and Consequences
Methodology of the Evaluation
Quantitative Methodology Utilized
Qualitative Methodology Employed
Models Employed in the Evaluation
CIPP Model of Evaluation
Scriven's Evaluation Model
Exception Reporting Evaluation Model
Causal Evaluation Model
Anecdotal Reporting Evaluation Model
Chaos Theory Evaluation Model
Resource Requirements of the Evaluation
Personnel Resource Requirements
Equipment Resource Requirements
Space Requirements
Time Requirment Estimates
Computer Software Requirements
Evaluation Cost Estimates
Targets of the Evaluation
Focus on Personnel
Focus on the Evaluation of Materials Utilized
Focus on Processes Employed
Focus on Products Produced
Focus on Plant Facilities
Focus on Environmental Aspects
Types of Evaluation Included
Diagnostic Evaluation Types
The following diagnostic measurement instruments will be used in this project:
Formative Evaluation Types
Formative evaluation instruments are used to modify a subject's or object's
behavior as a consequence of its use. In this project, the following formative instruments
will be used:
Summative Evaluation Types
A summative evaluation instrument is used to assess the current status of a subject
or object. In this study, the following summative instruments will be employed:
Plot a z Distribution
Use this procedure to create a normal z score distribution with a tail representing a
Type I (alpha) area of rejection for a null hypothesis. When you click on this simulation
choice, you are first asked to indicate the probability of the Type I error as shown below.
The curve is then created.
Figure 151 Form for Specifying Alpha for a Simulated z Score Distribution.
Figure 152 Simulated z-Score Distribution.

Plot a Chi Square Distribution
You can create a theoretical chi-square distribution with an area specified as a
rejection region for a Type I error. Since a chi-square statistic with one degree of
freedom is a squared normal z score, one can simulate a chi-square with k degrees of
freedom by the simple addition of k squared z scores. When you elect this simulation
option, you are asked to indicate the probability of the Type I error and degrees of
freedom. You then obtain the curve as shown below.
Figure 153 Form for Specifying Type I Error in Simulating Chi-Square Values.
Figure 154 Simulated Chi-Square Distribution.
Plot an F Distribution
The theoretical F statistic is the ratio of two chi-square statistics, each divided by
their degrees of freedom. To simulate this statistic, one can first generate the numerator
and denominator chi-square statistics using randomly selected squared z scores. The
number of such squared scores added together is the degree of freedom for each. This is
then divided into the sum and the ratio obtained to yield the F statistic. When you elect
this simulation option you are asked for the probability of the Type I error (area of
rejection) as well as the numerator and denominator degrees of freedom. Once these are
entered, the distribution curve is generated. These forms are shown below:
Figure 155 Type I Error Rate for Simulation of F Statistics.
Figure 156 Numerator Degrees of Freedom for Simulated F Values.
Figure 157 Denominator Degrees of Freedom for Simulated F Values.
Figure 158 Simulated F Distribution.
Resampling and Bootstrap Demo
A number of new methods have been developed over the past 30 years to provide
estimates of statistics and their confidence intervals when the shape of the sampling
distribution may not be assumed to be known or to minimize the effect of "outliers". For
example, the "jack-knife" method repeatedly removes one observation at a time and
recalculates the statistic. The average or median of these repeated sample statistics may
provide a better estimate of the parameter and the distribution itself a better estimate of
the statistics confidence interval. The number of generated statistics is equal to the
number of permutations of the original sample.
The "re-sampling" method randomly selects N-1 values from a sample and
obtains the statistic. This process is repeated a large number of times (often 10,000 times
or more) to create a distribution of the statistic. The mean or median of these sample
statistics can be used as an estimate of the population parameter and the distribution of
the values used to identify the upper and lower values of the confidence interval.
The "bootstrap" method first creates a pseudo-population of values by repeatedly
adding the sample to this population a large number of times. More than 100,000
observations may be in the pseudo-population. Once the population is created, a large
number of samples that are of the same size as the original sample are drawn at random
from the population. The mean or median of the sample statistics may be used to
estimate the population parameter and the distribution itself used to estimate confidence
interval boundaries.
This procedure lets you "experiment" with the methods of re-sampling and
bootstrapping to estimate the mean and confidence interval of a sample. To use the
program you first load a file of observed sample data in the main grid and then select this
option from the Simulation menu. On the specification dialogue there is a "slider" by
which you can elect the number of samples to generate. You select the variable to
analyze from the list of variables in your data grid (file.) You may change the confidence
interval from the default 95% (0.95) if desired. You may also elect to plot the
distribution of generated samples and print the results on the output form.
When you click the OK button, the results are created directly on the specification
form. The mean, standard error of the mean and confidence interval are first obtained by
the traditional Neyman-Pearson method (assumes a normal distribution.) Next, the re-
sampling method is used to create a large number of samples and the mean of the sample
means is reported. The distribution of sample means is sorted into ascending order and
the means in the upper and lower portion of this distribution that correspond to the alpha /
2 regions are identified. These are the estimated values of the confidence interval.
Finally, the bootstrap method is applied. A population the size of the value selected on
the slider is first created. Samples are repeated drawn at random from the created
population to create a distribution of sample means (the number drawn is the size selected
on the slider.) The mean of this distribution of means gives an estimate of the population
parameter. The confidence interval is obtained in the traditional way that assumes a
normal distribution.
Examining the plot of the generated samples is instructive. If the original sample
is from a population of highly skewed values, the generated plot will tend to reflect this.
The re-sampling confidence interval may better reflect the "true" population distribution
better than one that assumes a normal distribution.
A variety of methods for re-sampling and bootstrapping estimates have been
developed for other statistics such as the median, the product-moment correlation, the
analysis of variance, etc. A free program may be obtained from David C. Howell of the
Department of Psychology, University of Vermont.. His email address is
David.Howell@uvm.edu. A good introduction to the topic area is provided at
http://seamonkey.ed.asu.edu/~alex/teaching/WBI/resampling.html .
Figure 159 Bootstrap Demonstration for the Weight Variable from the cansas.txt
File.
Figure 160 Distribution of Resampling Estimates of the Mean.
Figure 161 Distribution of Means Estimated by the Bootstrap Method.
Results for Resampling and Bootstrap Demo
File = C:\Program Files\WGM Consulting\OpenStat\cansas.txt with 20 cases.
Selected Confidence Interval = 0.950 (z = 1.960)
Traditional statistics:
MEAN STD. DEV. SE MEAN CI LOW CI HIGH
178.600 24.691 5.521 167.779 189.421
Resampling statistics:
MEAN SE MEAN CI LOW CI HIGH
178.343 5.753 167.158 189.895
Bootstrap statistics:
MEAN SE MEAN CI LOW CI HIGH
178.620 5.169 168.489 188.752
Matrix Manipulation
MatMan was written to provide a platform for performing common matrix and
vector operations. It is designed to be helpful for the student learning matrix algebra and
statistics as well as the researcher needing a tool for matrix manipulation.
When you first start the MatMan program, you will see the main program form
below. This form displays four "grids" in which matrices, row or column vectors or
scalars (single values) may be entered and saved. If a grid of data has already been
saved, it can be retrieved into any one of the four grids. Once you have entered data into
a grid, a number of operations can be performed depending on the type of data entered
(matrix, vector or scalar.) Before performing an operation, you select the grid of data to
analyze by clicking on the grid with the left mouse button. If the data in the selected grid
is a matrix (file extension of .MAT) you can select any one of the matrix operations by
clicking on the Matrix Operations "drop-down" menu at the top of the form. If the data is
a row or column vector, select an operation option from the Vector Operations menu. If
the data is a single value, select an operation from the Scalar Operations menu.
Figure 162 The Matrix Manipulation Procedure Main Form.
The heart of all operations you perform involve values entered into the cells of a
grid. These values may represent values in a matrix, a column vector, a row vector or a
scalar. Each grid is like a spreadsheet. Typically, you select the first row and column
cell by clicking on that cell with the left mouse key when the mouse cursor is positioned
over that cell. To select a particular grid, click the left mouse button when the mouse
cursor is positioned over any cell of that grid. You will then see that the grid number
currently selected is displayed in a small text box in the upper left side of the form
(directly below the menus.)
Keyboard Input
You can input data into a grid directly from the keyboard to create a file. The file
may be a matrix, row vector, column vector or a scalar. Simply click on one of the four
grids to receive your keystrokes. Note that the selected grid number will be displayed in
a small box above and to the left of the grids. Next, click on the Files menu and move
your cursor down to the Keyboard entry option. You will see that this option is expanded
for you to indicate the type of data to be entered. Click on the type of data to be entered
from the keyboard. If you selected a matrix, you will be prompted for the number of
rows and columns of the matrix. For a vector, you will be prompted for the type (column
or row) and the number of elements. Once the type of data to be entered and the number
of elements are known, the program will "move" to the pre-selected grid and be waiting
for your data entry. Click on the first cell (Row 1 and Column 1) and type your (first)
value. Press the tab key to move to the next element in a row or, if at the end of a row,
the first element in the next row. When you have entered the last value, instead of
pressing the tab key, press the return key. You will be prompted to save the data. Of
course, you can also go to the Files menu and click on the Save option. This second
method is particularly useful if you are entering a very large data matrix and wish to
complete it in several sessions.
Clearing a Grid
Individual grids are quickly reset to a blank grid with four rows and four columns
by simply moving the mouse cursor over a cell of the grid and clicking the RIGHT mouse
button. CAUTION! Be sure the data already in the grid has been saved if you do not
want to lose it!
Using the Combination Boxes
In the upper right portion of the MatMan main form, there are four "Combo
Boxes". These boxes each contain a drop-down list of file names. The top box labeled
"Matrix" contains the list of files containing matrices that have been created in the current
disk directory and end with an extension of .MAT. The next two combo boxes contain
similar lists of column or row vectors that have been created and are in the current disk
directory. The last contains name of scalar files that have been saved in the current
directory. These combo boxes provide documentation as to the names of current files
already in use. In addition, they provide a "short-cut" method of opening a file and
loading it into a selected grid.
Five types of files are loaded when you first start the execution of the MatMan
program. The program will search for files in the current directory that have file
extensions of .MAT, .CVE, .RVE, .SCA and .OPT. The first four types of files are
simply identified and their names placed into the corresponding combination boxes of
matrices, column vectors, row vectors and scalars. The last, options, is a file which
contains only two integers: a 1 if the script should NOT contain File Open operations
when it is generated or a 0 and a 1 if the script should NOT contain File Save operations
when a script is generated or a 0. Since File Open and File Save operations are not
actually executed when a script or script line is executed, they are in a script only for
documentation purposes and may be left out.
Operations and Operands
At the bottom of the form (under the grids) are four "text" boxes which are
labeled Operation, Operand1, Operand2 and Operand3. Each time you perform an
operation by use of one of the menu options, you will see an abbreviation of that
operation in the Operation box. Typically there will be at least one or two operands
related to that operation. The first operand is typically the name of the data file
occupying the current grid and the second operand the name of the file containing the
results of the operation. Some operations involve two grids, for example, adding two
matrices. In these cases, the name of the two grid files involved will be in operands1 and
operands2 boxes while the third operand box will contain the file for the results.
You will also notice that each operation or operand is prefixed by a number
followed by a dash. In the case of the operation, this indicates the grid number from
which the operation was begun. The numbers which prefix the operand labels indicate
the grid in which the corresponding files were loaded or saved. The operation and
operands are separated by a colon (:). When you execute a script line by double clicking
an operation in the script list, the files are typically loaded into corresponding grid
numbers and the operation performed.
Menus
The operations which may be performed on or with matrices, vectors and scalars
are all listed as options under major menu headings shown across the top of the main
form. For example, the File menu, when selected, provides a number of options for
loading a grid with file data, saving a file of data from a grid, etc. Click on a menu
heading and explore the options available before you begin to use MatMan. In nearly all
cases, when you select a menu option you will be prompted to enter additional
information. If you select an option by mistake you can normally cancel the operation.
The Operations Script
Located on the right side of the main form is a rectangle which may contain
operations and operands performed in using MatMan. This list of operations and their
corresponding operands is known collectively as a "Script". If you were to perform a
group of operations, for example, to complete a multiple regression analysis, you may
want to save the script for reference or repeated analysis of another set of data. You can
also edit the scripts that are created to remove operations you did not intend, change the
file names referenced, etc. Scripts may also be printed.
Each time an operation is performed on grid data, an entry is made in a "Script"
list shown in the right-hand portion of the form. The operation may have one to three
"operands" listed with it. For example, the operation of finding the eigenvalues and
eigenvectors of a matrix will have an operation of SVDInverse followed by the name of
the matrix being inverted, the name of the eigenvalues matrix and the name of the
eigenvectors matrix. Each part of the script entry is preceded by a grid number followed
by a hyphen (-). A colon separates the parts of the entry (:). Once a series of operations
have been performed the script that is produced can be saved. Saved scripts can be
loaded at a later time and re-executed as a group or each entry executed one at a time.
Scripts can also be edited and re-saved. Shown below is an example script for obtaining
multiple regression coefficients.
CURRENT SCRIPT LISTING:
FileOpen:1-newcansas
1-ColAugment:newcansas:1-X
1-FileSave:1-X.MAT
1-MatTranspose:1-X:2-XT
2-FileSave:2-XT.MAT
2-PreMatxPostMat:2-XT:1-X:3-XTX
3-FileSave:3-XTX.MAT
3-SVDInverse:3-XTX.MAT:1-XTXINV
1-FileSave:1-XTXINV.MAT
FileOpen:1-XT.MAT
FileOpen:2-Y.CVE
1-PreMatxPostVec:1-XT.MAT:2-Y.CVE:3-XTY
3-FileSave:3-XTY.CVE
FileOpen:1-XTXINV.MAT
1-PreMatxPostVec:1-XTXINV.MAT:3-XTY:4-BETAS
4-FileSave:4-Bweights.CVE
Occasionally you may want to edit a script you have created or loaded. For
example, you may see a number of Load File or Save File operations in a script. Since
these are entered only for documentation and cannot actually be executed by clicking on
them, they can be removed from the script. The result is a more compact and succinct
script of operations performed. You may also want to change the name of files accessed
for some operations or the name of files saved following an operation so that the same
operations may be performed on a new set of data. To begin editing a script, move the
mouse cursor to the Script menu and click on the Edit option. A new form appears which
provides options for the editing. The list of operations appears on the left side of the
form and an Options box appears in the upper right portion of the form. To edit a given
script operation, click on the item to be edited and then click one of the option buttons.
One option is to simply delete the item. Another is to edit (modify) the item. When that
option is selected, the item is copied into an "Edit Box" which behaves like a miniature
word processor. You can click on the text of an operation at any point in the edit box,
delete characters following the cursor with the delete key, use the backspace key to
remove characters in front of the cursor, and enter characters at the cursor. When editing
is completed, press the return key to place the edited operation back into the script list
from which it came.
Also on the Edit Form is a "Directory Box" and a "Files Box". Shown in the
directory box is the current directory you are in. The files list box shows the current files
in that directory. You can delete a file from any directory by simply double-clicking the
name of the file in the file list. A box will pop up to verify that you want to delete the
selected file. Click OK to delete the file or click Cancel if you do not want to delete the
file. CAUTION! Be careful NOT to delete an important file like MATMAN.EXE,
MATMAN.HLP or other system files (files with extensions of .exe, .dll, .hlp, .inf, etc.!
Files which ARE safe to delete are those you have created with MatMan. These all end
with an extension of .MAT, .CVE, .RVE ,.SCA or .SCR .
You may quickly repeat the execution of a single operation previously performed
and captured in the script. Simply click on the script item with the left mouse button
when the cursor is positioned over the item to execute. Notice that you will be prompted
for the name of the file or files to be opened and loaded for that operation. You can, of
course, choose a different file name than the one or ones previously used in the script
item. If you wish, you can also re-execute the entire script of operations. Move your
mouse cursor to the Script menu and click on the Execute option. Each operation will be
executed in sequence with prompts for file names appearing before execution each
operation. Note: you will want to manually save the resulting file or files with
appropriate names.
Getting Help on a Topic
You obtain help on a topic by first selecting a menu item, grid or other area of the main
form by placing the mouse over the item for which you want information. Once the area
of interest is selected, press the F1 key on your keyboard. If a topic exists in the help file,
it will be displayed. You can press the F1 key at any point to bring up the help file. A
book is displayed which may be opened by double clicking it. You may also search for a
topic using the help file index of keywords.
Matrix Operations
Once a matrix of data has been entered into a grid you can elect to perform a
number of matrix operations. The figure below illustrates the options under the Matrix
Operations menu.
Figure 163 Matrix Operations Available in MatMan.
References
1. Afifi, A. A. and Azen, S. P. Statistical Analysis. A Computer Oriented Approach.
New York: Academic Press, Inc. 1972.
2. Anderberg, Michael R. Cluster Analysis for Applications. New York: Academic
Press, 1973.
3. Bennett, Spencer and Bowers, David. An Introduction to Multivariate Techniques for
Social and Behavioral Sciences. New York: John Wiley and Sons, Inc., 1977.
4. Besterfield, Dale H. Quality Control, 2
nd
Ed. Englewood Ciffs, N.J.: Prentice-Hall,
Inc., 1986.
5. Bishop, Yvonne M., Fienberg, Stephen E. and Holland, Paul W. Discrete
Multivariate Analysis. Theory and Practice. Cambridge, Mass.: The MIT Press
1975.
6. Blommers, Paul J. and Forsyth, Robert A. Elementary Statistical Methods in
Psychology and Education, 2
nd
Ed. Boston, Mass.: Houghton Mifflin Company,
1977.
7. Borg, Walter R. and Gall, Meridith Damien. Educational Research. An Introduction,
5
th
Ed. New York: Longman, Inc., 1989.
8. Brockwell, Peter J. and Davis, Richard A. Introduction to Time Series and
Forecasting. New York: Springer-Verlag New York Inc., 1996.
9. Bruning, James L. and Kintz, B. L. Computational Handbook of Statistics, 2
nd
Ed.
Glenview, Ill.: Scott, Foresman and Company, 1977.
10. Campbell, Donald T. and Stanley, Julian C. Experimental and Quasi-Experimental
Designs for Research. Chicago, Ill.: Rand McNally College Publishing Company,
1963.
11. Chapman, Douglas G. and Schaufele, Ronald A. Elementary Probability Models and
Statistical Inference. Waltham, Mass.: Ginn-Blaisdell, 1970.
12. Cody, Ronald P. and Smith, Jeffrey K. Applied Statistics and the SAS Programming
Language, 4
th
Ed. Upper Saddle River, N.J.: Prentice Hall, 1997.
13. Cohen, Jacob and Cohen, Patricia. Applied Multiple Regression/ Correlation
Analysis for the Behavioral Sciences. Hillsdale, N.J.: Lawrence Erlbaum Associates,
1975.
14. Cohen, Jacob. Statistical Power Analysis for the Behavioral Sciences, 2
nd
Ed.,
Hillsdale, N.J.: Lawrence Erlbaum Associates, 1988.
15. Comrey, Andrew L. A First Course in Factor Analysis. New York: Academic Press,
Inc., 1973.
16. Cook, Thomas D. and Campbell, Donald T. Quasi-Experimentation. Design and
Analysis Issues for Field Settings. Chicago, Ill.: Rand McNally College Publishing
Company, 1979.
17. Cooley, William W. and Lohnes, Paul R. Multivariate Data Analysis. New York:
John Wiley and Sons, Inc., 1971.
18. Crocker, Linda and Algina, James. Introduction to Classical and Modern Test
Theory. New York: Holt, Rinehart and Winston, 1986.
19. Diekhoff, George M. Basic Statistics for the Social and Behavioral Sciences. Upper
Sadle River, N.J., Prentice Hall, Inc. 1996.
20. Edwards, Allen L. Techniques of Attitude Scale Construction. New York: Appleton-
Century-Crofts, Inc., 1957.
21. Efromovich, Sam. Nonparametric Curve Estimation. Methods, Theory, and
Applications. New York: Springer-Verlag, 1999.
22. Ferrguson, George A. Statistical Analysis in Psychology and Education, 2
nd
Ed.
New York: McGraw-Hill Book Company, 1966.
23. Fienberg, Stephen E. The Analysis of Cross-Classified Categorical Data, 2
nd
Ed.
Cambridge, Mass.: The MIT Press, 1980.
24. Fox, John. Multiple and Generalized Nonparametric Regression. Thousand Oaks,
Cal.: Sage Publications, Inc., 2000.
25. Freund, John E. and Walpole, Ronald E. Mathematical Statistics, 4
th
Ed. Englewood
Cliffs, N.J.: Prentice-Hall, Inc., 1987.
26. Fruchter, Benjamin. Introduction to Factor Analysis. Princeton, N.J.: D. Van
Nostrand Company, Inc., 1954.
27. Gay, L. R. Educational Research. Competencies for Analysis and Application, 4
th
Ed. New York: Macmillan Publishing Company, 1992.
28. Glass, Gene V. and Stanley, Julian C. Statistical Methods in Education and
Psychology. Englewood Cliffs, N.J.: Prentice-Hall, Inc., 1970.
29. Gottman, John M. and Leiblum, Sandra R. How to do Psychotherapy and How to
Evaluate It. A Manual for Beginners. New York: Holt, Rinehart and Winston, Inc.,
1974.
30. Guertin, Wilson H. and Bailey, Jr., John P. Introduction to Modern Factor Analysis.
Ann Arbor, Mich.: Edwards Brothers, Inc., 1970.
31. Gulliksen, Harold. Theory of Mental Tests. New York: John Wiley and Sons, Inc.,
1950.
32. Hambleton, Ronald K. and Swaminathan, Hariharan. Item Response Theory.
Principles and Applications. Boston, Mass.: Kluwer-Nijhoff Publishing, 1985.
33. Hansen, Bertrand L. and Chare, Prabhakar M. Quality Control and Applications.
Englewood Cliffs, N.J.: Prentice-Hall, Inc., 1987.
34. Harman, Harry H. Modern Factor Analysis. Chicago, Ill.: The University of Chicago
Press, 1960.
35. Hays, William L. Statistics for Psychologists. New York: Holt, Rinehart and
Winston, Inc., 1963.
36. Heise, David R. Causal Analysis. New York: John Wiley and Sons, Inc., 1975.
37. Hinkle, Dennis E., Wiersma, William and Jurs, Stephen G. Applied Statistics for the
Behavioral Sciences, 2
nd
Edition. Boston, Mass.: Houghton Mifflin Company, 1988.
38. Huntsberger, David H. and Billingsley, Patrick. Elements of Statistical Inference, 6
th
Ed. Boston, Mass.: Allyn and Bacon, Inc., 1987.
39. Kelly, Louis G. Handbook of Numerical Methods and Applications. Reading, Mass.:
Addison-Wesley Publishing Company, 1967.
40. Kennedy, Jr., William J. and Gentle, James E. Statistical Computing. New York:
Marcel Dekker, Inc., 1980.
41. Kerlinger, Fred N. and Pedhazur, Elazar J. Multiple Regression in Behavioral
Research. New York: Holt, Rinehart and Winston, Inc., 1973.
42. Lieberman, Bernhardt (Editor). Contemporary Problems in Statistics. A book of
Readings for the Behavioral Sciences. New York: Oxford University Press, 1971.
43. Lindgren, B. W. and McElrath, G. W. Introduction to Probability and Statistics, 2
nd
Ed. New York: The Macmillan Company, 1966.
44. Marcoulides, George A. and Schumacker, Randall E. (Editors). Advanced Structural
Equation Modeling. Issues and Techniques. Mahwah, N.J.: Lawrence Erlbaum
Associates, 1996.
45. Masters, Timothy. Practical Neural Network Recipes in C++. San Diego, Calif.:
Morgan Kaufmann, 1993.
46. McNeil, Keith, Newman, Isadore and Kelly, Francis J. Testing Research Hypotheses
with the General Linear Model. Carbondale, Ill.: Southern Illinois University Press,
1996.
47. McNemar, Quinn. Psychological Statistics, 4
th
Ed. New York: John Wiley and Sons,
Inc., 1969.
48. Minium, Edward W. Statistical Reasoning in Psychology and Education, 2
nd
Ed.
New York: John Wiley and Sons, Inc., 1978.
49. Montgomery, Douglas C. Statistical Quality Control. New York: John Wiley and
Sons, Inc., 1985.
50. Mulaik, Stanley A. The Foundations of Factor Analysis. New York: McGraw-Hill
Book Company, 1972.
51. Myers, Jerome L. Fundamentals of Experimental Design. Boston, Mass.: Allyn and
Bacon, Inc., 1966.
52. Nunnally, Jum C. Psychometric Theory. New York: McGraw-Hill Book Company,
Inc., 1967.
53. Olson, Chester L. Essentials of Statistics. Making Sense of Data. Boston, Mass.:
Allyn and Bacon, Inc., 1987.
54. Payne, David A., Editor. Curriculum Evaluation. Commentaries on Purpose,
Process, Product. Lexington, Mass.: D. C. Heath and Company, 1974.
55. Pedhazur, Elazar J. Multiple Regression in Behavioral Research. Explanation and
Prediction. 3
rd
Ed. Fort Worth, Texas: Holt, Rinehart and Winston, Inc., 1997.
56. Press, William H., Flannery, Brian P., Teukolsky, Saul A., and Vetterling, William T.
Numerical Recipes in C. The Art of Scientific Computing. Cambridge University
Press, 1988.
57. Ralston, Anthony and Wilf, Herbert S. Mathematical Methods for Digital
Computers. New York: John Wiley and Sons, Inc., 1966.
58. Rao, C. Radhakrishna. Linear Statistical Inference and Its Applications. New York:
John Wiley and Sons, Inc., 1965.
59. Rao, Valluru and Rao, Hayagriva. C++ Neural Networks and Fuzzy Logic, 2
nd
Ed.
New York: MIS Press, 1995.
60. Rogers, Joey. Object-Oriented Neural Networks in C++. San Diego, Calif.:
Academic Press, 1997.
61. Roscoe, John T. Fundamental Research Statistics for the Behavioral Sciences, 2
nd
Ed. New York: Holt, Rinehart and Winston, Inc., 1975.
62. Rummel, R. J. Applied Factor Analysis. Evanston, Ill.: Northwestern University
Press, 1970.
63. Scheffe', Henry. The Analysis of Variance. New York: John Wiley and Sons, Inc.,
1959.
64. Schumacker, Randall E. and Lomax, Richard G. A Beginner's Guide to Structural
Equation Modeling. Mahwah, N.J.: Lawrence Erlbaum Associates, 1996.
65. Siegel, Sidney. Nonparametric Statistics for the Behavioral Sciences. New York:
McGraw-Hill Book Company, Inc., 1956.
66. Silverman, Eliot N. and Brody, Linda A. Statistics. A Common Sense Approach.
Boston, Mass.: Prindle, Weber and Schmidt, Inc., 1973.
67. SPSS, Inc. SPSS-X User's Guide, 3
rd
Ed. Chicago, Ill.: SPSS, Inc., 1988.
68. Steele, Sara M. Contemporary Approaches to Program Evaluation: Implications for
Evaluating Programs for Disadvantaged Adults. Syracuse, New York: ERIC
Clearinghouse on Adult Education (undated).
69. Stevens, James. Applied Multivariate Statistics for the Social Sciences, 3
rd
Ed.
Mahwah, N.J.: Lawrence Erlbaum Associates, 1996.
70. Stodala, Quentin and Stordahl, Kalmer. Basic Educational Tests and Measurement.
Chicago, Ill.: Science Research Associates, Inc., 1967.
71. Thomson, Godfrey. The Factorial Analysis of Human Ability, 5
th
Ed., Boston,
Mass.: Houghton Mifflin Company, 1951.
72. Thorndike, Robert L. Applied Psychometrics. Boston, Mass.: Houghton Mifflin
Company, 1982.
73. Thorndike, Robert L. (Editor.) Educational Measurement, 2
nd
Edition. One Dupont
Circle, Washington D.C.: American Council on Education, 1971.
74. Veldman, Donald J. Fortran Programming for the Behavioral Sciences. Holt,
Rinehart and Winston, New York, 1967, pages 308-317.
75. Walker, Helen M. and Lev, Joseph. Statistical Inference. New York: Henry Holt and
Company, 1953.
76. Winer, B. J. Statistical Principles in Experimental Design. New York: McGraw-Hill
Book Company, Inc., 1962.
77. Worthen, Blaine R. and Sanders, James R. Educational Evaluation: Theory and
Practice. Belmont, Calif.: Wadsworth Publishing Company, Inc., 1973.
78. Yamane, Taro. Mathematics for Economists. An Elementary Survey. Englewood
Cliffs, N.J.: Prentice-Hall, Inc., 1962.
Index
2-Stage Least-Squares Regression, 102
Accelerated Depreciation, 240
Analysis of Variance, 47
both Type I and Type II error, 24
Box plots, 28
Breakdown, 29
Central tendency and variability, 25
Chi-Square, 133
Classical Item Analysis, 155
Cluster Analysis, 116
Cochran Q Test, 144
Composite Test Reliability, 176
Contingency tables, 42
Correlation, 68
Course Grades, 194
Cox Proportional Hazards Survival
Regression, 84
CUSUM Chart, 218
Defect (Nonconformity) c Chart, 223
Defects Per Unit u Chart, 225
Differential Item Functioning, 181
Discrete and Continuous Variables, 13
Discriminant analysis, 120
Double Declining Balance, 238
Estimating Sample Sizes, 249
Excel Data, 18
Factor Analysis, 109
Files, 16
Fishers Exact Test, 138
Frequencies Analysis, 37
Friedman Two Way ANOVA, 147
Future Value, 230
General Unfolding Model Analysis, 199
Grid, 16
Guttman Scalogram Analysis, 180
Hoyt and Cronbach Reliability, 210
Interest Payment, 233
Interest Rate, 235
Internal Rate of Return, 241
Interval Variables, 12
Item Banking, 204
Item difficulty, 156
Item Discrimination, 155
Item Information Function, 172
Kendalls Coefficient of Concordance,
140
Kendall's Tau and Partial Tau, 152
Kruskal-Wallis One-Way ANOVA, 141
Kuder-Richardson #21 Reliability, 175
Latin Square Designs ANOVA, 64
Loan Amortization, 228
Mann-Whitney U Test, 137
Matrix Manipulation, 273
Multiple Regression and Prediction, 70
Multivariate Analyses, 41
Nested ANOVA, 60
Net Present Value, 243
nominal variable, 12
Non-Linear Regression, 87
Non-parametric Analyses, 41
Number of Periods, 236
Ordinal Variables, 12
P Chart, 221
Parametric Analyses, 41
Path analysis, 126
Payment, 231
Period Payment, 232
Polytomous DIF Analysis, 192
Power Curves, 251
Present Value of an Investment, 237
Printing, 20
Probability of a Binomial Event, 149
Project Evaluation, 261
QQ or PP Plot, 34
Random Number Generation, 256
Range Chart, 214
Ratio Variables, 13
Reliability and validity, 44
Repeated Measures ANOVAs, 54
Resampling and Bootstrap, 268
Runs Test, 150
Sample statistics, 23
Sign Test, 146
simulated data, 18
Simulation, 245
Simultaneous Multiple Regression, 81
Spearman Rank Correlation, 135
Spearman-Brown Reliability Prophecy,
174
Statistical Process Control, 212
Stem and Leaf Plots, 32
Straight Line Depreciation, 239
Successive Interval Scaling, 178
The One Parameter Logistic Model, 164
The S Chart, 216
The t-test, 45
theoretical distributions, 26
Transformations, 27
Univariate Analyses, 41
Variable Labels, 15
Weighted Least-Squares Regression, 93
Wilcoxon Matched-Pairs Signed Ranks
Test, 143
XBAR Chart, 212

Quick Guide

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Quick Guide

Uploaded by

Copyright:

Available Formats

Quick Guide to OpenStat

(B within A)C p(q-1)(r-1)

Within Cell pqr(n-1)

You might also like