इंटरनेट मानक

Disclosure to Promote the Right To Information
Whereas the Parliament of India has set out to provide a practical regime of right to
information for citizens to secure access to information under the control of public authorities,
in order to promote transparency and accountability in the working of every public authority,
and whereas the attached publication of the Bureau of Indian Standards is of particular interest
to the public, particularly disadvantaged communities and those engaged in the pursuit of
education and knowledge, the attached public safety standard is made available to promote the
timely dissemination of this information in an accurate manner to the public.

“जान1 का अ+धकार, जी1 का अ+धकार” “प0रा1 को छोड न' 5 तरफ”
Mazdoor Kisan Shakti Sangathan Jawaharlal Nehru
“The Right to Information, The Right to Live” “Step Out From the Old to the New”

IS 7920-3 (1996): Statistical vocabulary and symbols, Part
3: Design of experiments [MSD 3: Statistical Methods for
Quality and Reliability]

“!ान $ एक न' भारत का +नम-ण”
Satyanarayan Gangaram Pitroda
“Invent a New India Using Knowledge”

“!ान एक ऐसा खजाना > जो कभी च0राया नहB जा सकता ह”

“Knowledge is such a treasure which cannot be stolen”

. OBIS 1996 BUREAU OF INDIAN STANDARDS MANAK BHAVAN. IS7920(Part3):1996 ( Reaffirmed 2002 ) m?-s?m n tiilwcblY vr=m& dk 5Kb5 m3 JniPihTm (w@m) Indian Standard STATISTICALVOCABULARY ANDSYMBOLS PART 3 DESIGN OF EXPERIMENTS ( Second Revision ) ICS 01. 01.120. 9 BAHADUR SHAH ZAFAR MARG NEW DELI-i1 110002 September 1996 Price Group 8 .040.080.

Background information on the design of_experiments is given in Annex A. It was also felt that the definitions of some of the terms covered in IS 7920 : 1985 ‘Statistical vocabulary’ need modification. The examples accompanying the definitions of certain terms are generally intended to illustrate simple applications of those definitions and to provide the experienced person with a reference to illustrate the concepts to less experienced practitioners. considerable assistance has been derived from the definitions given in IS0 3534-3 : 1989 ‘Statistics . An informative list of published Indian Standards on Statistical Methods is given at Annex B. this revision of the vocabulary on statistical terms has been taken up. In line with the three parts of IS0 3534.Statistical Methods for Quality and Reliability Sectional Committee. an alphabetical indd to the terminology has been provided at Annex D. after the draft finalized by the Statistical Methods for Quality and Reliability Sectional Committee had been approved by the Management and Systems Division Council. has also published International Standards on statistical vocabulary as the three parts of IS0 3534. For ease of reference. many other standards have been published in this field using additional terms. IS 7920. / The composition of the Committee responsible for the formulation of this standard is given at Annex C. Since the publication of the revised standard in 1985. was first published in 1976 and revised in 1985. Keeping in view the above. the terminology has been divided in the following three parts of IS 7920: Part 1 Probability and general statistical terms Part 2 Statistical quality control Part 3 Design of experiments In preparing this part. the Technical Committee on Application of Statistical Methods (ISOAC 69). The Indian Standard covering the terminology of statistical terms used in quality control and sampling. . At the international level. MSD 3 FOREWORD This Indian Standard (Second Revision) was adopted by the Bureau of Indian Standards.Vocabulary and symbols Part 3: Design of experiments’.

The levels may be limited to certain to be answered. presence or The unit may be a patient in a hospital. mented tray. have discontinuous or multimodal relationships ing valid and relevant conclusions from the experiment. or which is not 1 ‘I . The selection of an appropriate design for any experiment is a that cause abrupt changes within the range of the function of many considerations such as the type of questions experiment. a group of absence of a catalyst. The terms %bs$nce. relatively simple statistical interpretation of the results. although it is possible to provide the most efficient and economical methods of reach. and the selection of for assuming model relationships. such as the variety of a material. such as temperature.6 Experimental Error 2.5). ment are sometimes called ‘principal factors’.2 Factor (see 2. Extrapolation beyond the range of these A planned arrangement in which an experimental levels is usually inappropriate without a firm basis programme is to be conducted. NOTE-‘ Version’ is a general term applied both to quan- b) Types of designs. The more restrictive term ‘level’ is frequently used to express more precisely the c) Methods of analysis. the degree of generality to be attached to the selected fmed values (known or unknown) or they conclusions. in the experiment. Interpolation the levels (versions) of one or more factors or factor within the range may depend on the number of combinations to be included in the experiment.3 Level. a section of a compart- Those factors which are to be studied in the experi.5 Experimental Unit (test results) and of which different levels Entity to which a treatment is applied or assigned (versions) are included in the experiment.4 Treatment includes the randomization procedure for allocating treat. A properly designed experiment will permit dependent on this selection. It is usually NOTE . or they may be qualitative. 2. animals. The ‘arrangement’ 2.. An assignable cause which may affect the responses 2.. and the type of equipment. levels and the spacing of these levels.1 Design of Experiment. 2 GENERAL TERMS Responses observed at the various levels of a factor provide information for determining the effect of 2. Experiment Design the factor within the range of levels of the experi- ment. cause like. 140°C and 160 C. Version (of a Factor) It is that part of the total variability in the responses A given value.’ ments to experimental units. which may not be possible otherwise. and titative and qualitative factors.The purpose of designing an experiment is to reasonable to interpolate. IS 7920( Part 3):1996 Indian Standard STATISTICAL VOCABULARY AND SYMBOLS PART 3 DESIGN OF EXPERIMENTS ( Second Revision ) 1 SCOPE Example: This standard (Part 3) gives definitions of the terms lbo versions of a catalyst may be presence and used in the field of design of experiments. the homogeneity of the experimental units and the cost of performing the range to be studied. a plot of land. The method of analysis is experiment. factors and blocks. 120 C. Four levels of a heat treatment may be are classified into the following logical groups: LOOC. the magnitude of the effect for which a high may represent purely random selection over the probability of detection (power) is desired. NOTE--Factors may be quantitative. a) General terms. Levels or combination of levels (versions) of each of the factors assigned to an experimental unit 2. a production batch. Example : speed of execution and voltage applied. quantitative characteristic. a specification of procedure or a (test results) which is not due to any assignable specific setting of a factor. etc.

Duplication usually invol- ves a fresh experimental unit. ly measured in an experiment as a pooled variance of sets of replication may require deliberately different. Average Effect basis for the assumption. number of units studied in a day. The analysis of the experiment The execution of a treatment more than once under results has to take into account the effects of blocks. this is generally referred to as NOTE . when present. causes. minimize this risk. Experimental error isusual. this hypothesis is often a reasonable assumption. refers to complete set are called ‘incomplete blocks’. contributing to the experimental error should be included in tions and the experimental operations are carefully control. similar conditions. such as a newsample. environmental condi.Blocks are usually selected to allow for assignable ‘partial replication’. conditions. a single unit is involved. the occurrence of experimental error is may be limited to repetition under essentially the same con- inevitable in practical experimentation. Therefore. in addition to those introduced as principal factors to be studied which it may be difficult. This variation intro. A group of relatively homogenous experimental units. Randomization is a process to grouping the experimental units into blocks. similar method for assuring that each unit has an equal times they are selected in order to broaden the interpretation chance of being selected for each treatment. an independent resetting of the levels ments are dealtwith in pairs. to keep constant for all of the experimental units in the com. the pairs are of the factors being studied on that unit. though duplicate observations for the same treatment. the to the experimental units. In order to do this effectively. NOES NOTE . When duplicate considered as ‘blocks’. but some. even its estimate. In a a single element of an experiment. the principal factors. replication led. When the versions are relatively close. 2. similar. such as different facilities or locations. However. replication is required to increase the precision of they are repeated. Each of the repetitions is called a ‘replicate’.8 Block Factors 1 An essential element in the design of experiment is to Those assignable causes which form the basis for provide estimates of effects free from biases due to un- detected assignable causes. proximity to underground water or thickness 2. having common conditions. repetition of the same treatment to different experimental units in Blocks which accommodate a complete set of treat- ment combinations are called ‘complete blocks’. these should be identified as duplicate observations Example rather than as ‘duplicates’ to reflect the narrower degree of duplication.10 Duplication plete experiment. In some experiments. It is 2 Randomization provides independence in the usually assumed that the block factors do not interact with observations.. wh~en paired comparison experiment. as contrasted to replication. In other situations. the responses at each level (version) of a factor 2 . For some experiments. such as exposure to the wind. of the results by including a wider range of conditions. a short time duces a degree of uncertainty into conclusions that ate interval or a common-batch of materials. if the versions differ considerably or if there is no priority 2. and therefore has to be taken into account in reaching conclusions. NOTES 2. When a subset of the treatments within an experi- ment is repeated. etc. ditions. Recent usage has broadened the definition of The term ‘block’ originated in agricultural experi. operators. Theoperational procedure for assignment NOTE! . Those which accommodate only a portion of the NOTE-Duplication. such as the same facility or location. 2 Experimental error is the unexplained or residual part of the total variation. observations are made on the same sample or unchanged settings.11 Randomization of the arable layer. drawn from the results. the assumption should be verified so that an appropriate method of analysis may be chosen for A term describing ameasure for the comparison of the experimental data. duplication to more than once rather than restricting it to ments in which a field was subdivided into sections twice.Generally the versions of the block factors are ‘at random’ involves the~use of Random Numbers or some imposed by the available experimental conditions. The effect of these assignable causes may be minimized within blocks. where two treat.Since experimental error is almost invariably 1 It is a common characteristic of experiments that. For example. For other experiments requiring more general results.IS 7920 ( Part 3 ) : 1996 associated with any deliberate variation in the 2.12 Main Effect. the same block. their results vary from trial to trial.9 Replication experimental conditions and which introduces a degree of uncertainty in the conclusions that are The repetition of a complete set of all the treat- drawn from the experimental results. blocks’are Qe procedure used to allot treatments at random based on batches of raw material. the replication process. all elements though the experimental materials. ments to be compared in an experiment. or even impossible.7 Block when factors which produce no effect (average or differential) are included in the experiment. longer time intervals or different batches of materials. a ‘pseudo-replication’ occurs 2. or.

assuming little response pattern shows curvature rather than a experimental error. the other factors. Note that the response of Contrast coefficients . ly more convenient to use integers rather than etc. Sometimes. AS If the levels are equally spaced. and planning the design so that they are con- parameter or the estimate of this parameter founded with block effects or other pre-selected principal factor or differential effects. is that changes in version of. effects (interactions) with block effects. block the factor is not important.14 Confounding in the experiment. in using version 2. and it diminishes or invalidates A term describing a measure of dependence of the the effectiveness of an experiment. In such a case. Since an ‘interaction’ indicates a ‘differen- 2. Contrast 2 .17) factor on each of the several levels (versions) of are examples of confounding effects or differential other factor(s).19) and aliases (2. ment designs. The effects of these factors should not simply be described in terms of averages over all levels (versions) of the A linear_function of the observations or parameters other factors (‘main effects’) involved.21.16) and fac- of the responses for each level (version) of the torial experiment with partial confounding (3. level (version) of one factor on the level(s) of other factor(s) by providing the differential comparison The confounded factorial design (3. it is implied that the effect of variation within the factor is dependent upon the levels (versions) of effects. Y2 . differ in sign and/or magnitude. NOTE . . for which the sum of the coefficients is zero. however. IS 7920 ( Part 3 ) : 1996 averaged over all levels (versions) of other factors 2. . NOTE . fractions for contrast coefficients.1 0 +1 factor A at version 2 of factor B would show no for question 1 effect. the ex- effect of A tremes ofA in the experiment. This is accomplished by deliberately pre- selecting certain effects or differential effects.Y. Large effects of the factor may result at various levels (versions) of other factors. See frac- tional factorial design (3. ‘. . this does not necessarily mean that actions) with the effects of other factor(s). each effect is confounded with one or more other cient magnitude. the first question it might be logical to ask 2 20 20 20 20 is whether the@ is an overall linear trend. but do not affect the responses when A2 should fall on the connecting AI and A3 or. . A second-question might be whether there is evidence that the An interpretation of these results.be-compared to A2 (If there is no curvature. simple linear trend. Also. . while keeping the other more see 2. If the results for version 2 of factor B had Contrast 1 -At +A3 been 8. the linear function alY1 Example : + aaY2 + . as being of In a model. Combining indistinguishably the main effect of a NOTE . This Average 15 21 24 could be done by comparing AI and A3.18). be equal to their average). changes in version of factor other words. B affect the responses when using version 1 of factor A in an opposite direction to when version 3 of Response AI A2 A3 factor A is used.When an interaction is determined lo be of suffi. . AZ.It should be noted that even though a ‘maineffect’ factor or a differential effect between factors(inter- is indicated to be small. It is frequent- factors (ABC) is called ‘three-factor interaction’. of factor B. there would be no interaction since the differences between the results of the two Contrast coefficients -1n +1 -112 versions of Sat each version ofA (or the 3 versions for question 2 of A at each version of B) would be constant. not all ai’s are equal to zero.. effect of B Example 1: Factor B Version 1 2 3 A factor is applied at three levels and the results are Version 1 10 22 28 20 represented bv Al. Here the average of A1 and A3 factor A affect the responses when using version 1 could.+ anYn is a contrast if and only if. but may factors(s) or interaction(s).1/2A1 +A2 -1i2A3 An interaction involving two factors @B) is called This example illustrates a regression type study of a ‘two-factor interaction’ one involving three equally spaced continuous variables. 3 . the term ‘main effect’ may describe the little interest.Confounding is an important technique which The process of averaging in these cases would tend to make permits the effective use of specified blocks in some experi- the ‘main effect’ appear smallersee 213. 20 and 26. important effects free from such complications. etc.13 Interaction design during the running of an experiment or from incom- plete planning of the design. Factor A Average Z ai = 0. confounding results from inadvertent changes to a 2. but separately for each such level (version).15 Contrast tial effect’. With observations Yr.

when multiplied in corresponding pairs. The form of 2 ai. the response for is negligible. N assemblies. but the usual combinations) the size of the experiment can be reduced to meaning of this term is the algebraic sum of the responses l/4. At.1 0 +1 2. Example 2: NOTE .20 Response Variable ailai2 0 0 +1 The variable that shows the observed results of an X ait ai = 1 Therefore. that for the previous problem. At A2 A3 NOTE .1.16 Orthogonal Contrasts factorials having more than 2 levels. A2 and As use the various factors and rows represent the (N) treatments. 3 There exist generalizations of the above to fractional 2. s symbols and strength t. Example 1: whether the selection is in the controlof the experimenter or not. selected. +2. each treatment combination will have a set of coefficients associated with it. uses a new manufacturing tech. the other can be estimated.This~is sometimes referred to as an ‘independent ai1 Contrast 1 -1 0 + 1 variable’.17 Orthogonal Array (. not orthogonal experimental treatment. 1 In a 2” fractional factorial design.The columns of an orthogonal array form the nique while the other two. ai Contrast 2 0 -1 +1 2. NOTE .18 Aliases technique seem to differ from A2 and A3 which are using the old one? Contrast A1 &th the average of Two or more effects (main or interaction) in a A2 and A5 Second. do the two suppliers using the fractional factorial experiment.1/8.19 Predictor Variable coefficients of the two sets satisfy the condition that. Sometimes the term ‘contrast’is used only 2 With a large number of factors (and factorial treatment to refer to the pattern of the coefficients. An identity I is equated to the defining contrast or definmg Contrast coefficients 0 -1 +I contrasts) and using the convention that A2 = BI = C 21. such as a factor level in an experiment. ai = 0 Therefore. Two contrasts are orthogonal if the contrast 2. one of which.21 Model ai Contrast 2 -1 +2 -1 An equation which is intended to provide a func- +1 0 -1 tional description of the observations/sources of ail ai2 information in terms of parameters. the aliases can be deter- Response AI A2 A3 mined once the ‘defining contrast’ (in the case of a half replicate) or ‘defining contrasts’ (for a fraction smaller than Contrast coefficients -2 +1 +1 half) are stated. See examples 1 Here A and BCD are aliases. experiment. . orthogonal the model can be expressed as: 4 ‘I .19 Contrast 2 -A2 +A3 I = ABCD NOTE . customary one.1). about Contrast 1 -2A1 +A2 +A3 which all information may be sacrificed for the experiment. a factor might lead to a different pair of questions. t-plet of s symbols occurs equally often. The defining contrast is that effect (or for question 1 effects). . if in txNsub-arrayof thekxNarray. for question 2 the multiplication of the letters on both sides of the equation shows the aliases In the example of 3. A = A*BCD = BCD arbitrarily provided that pi = 0 condition is met.IS 7920 ( Part 3 ) : 1996 the coefficients for contrast 2 would appear as 2. Questions and AB = A*B*CD = CD of logical interest from an experiment may be expressed as contrastswith carefully selected coefficients. An array of order kxN with entries from a set of s Example 2: symbols is an orthogonal array with k constraints Another example dealing with discrete versions of (or factors). If we assume that one of them and 2 above. First.I. usualiy thought to be of no consequence. does vendor AI with the new 2. or in general to ld to form a 2”’ fractionalfactorial multiplied by the appropriate coefficients.The coefficients for a contrast may be selected so that. though the NOTES interpretation of the results will differ.This is sometimes referred to as the ‘dependent variable’. Asindicated in the examples. AI A2 A3 ai1 Contrast 1 . the A variable the levels (versions) of which are Sum of those products is equal to zero. every ordered Let us suppose there are three sources of supply. a linear combi- customary technique differ? Contrast A2 and AS nation of which is estimable but they can not be The pattern of contrast coefficients is similar to estimated separately.

b) The residual random variablesare mutually independent.12) with replication in two blocks treatments are randomly allocated to the (a randomized block design). effects (also called residual NOTES error) 1 No block factors are involved in a completely randomized NOTE-The assumptions about the residual errorcom.2 Randomized Block Design c) All the residual random vaiiables have the samevariance. to improve the process should be obtained along with a product. Each block contains as many experimental Two quantitative factors (Xl) at two levels and (X2) units as there are number of treatments. representing assignable causes/factors) 3 TYPES OF DESIGNS f Z(Ranclom variables rep- 3. as is experimental units within each block and frequently the case. IS 7920 ( Part 3 ) : 1996 Observed value = Z(Parameters or terms may require considerable replication so as to improve the efficiency of comparison of the effect of the change in levels. The rows and columns represent leveli of variable-and the response variable. the rows might be days and the 5 . and this of the square. Change ofversions (levels) experiment. not of primary interest in the this knowledge at minimum cost. and that designed experiments using relatively small NOTE . ponent are as under: 2 Completely randomized design is advocated only under the assumption that all the experimental units are more or a) The expected value of each residual random variable is less homogenous. two blocking factors. the fitted randomization for each block. each treat- at three levels. A design in which the treatments are assigned at senting non-assignable random to the full set of experimental units.1 Completely Randomized Design resenting assignable effects) + Random variable repre. D 4 A B C E NOTE-The principal theses of EVOP are that knowledge The treatments are shown by the-Latin letters. 3. B. The experiment (see 3. by ‘balancing out’ their contributions. with two blocking factors where the n Bb = effect of blocks.22 Response Surface I in such a manner that each level (version) appears exactly once in each row and exactly once in each It is a functional relationship between the predictor column. If it is assumed.23 Evolutionary Operation (EVOP) 3 C DAB A sequential form of experimentation conducted in production facilities during regular production. Y = observation corresponding to a treatment Block Bo = constant =_ BI = linear effect of factor Xl B2 = linear effect of factor X.Latin square designs are generally used to shifts in factor levels (within production tolerances) can yield eliminate two block effects. C and D are assigned at BbXb + e random to the experimental units in each of three where blocks. zero.Asequential form of experimentation is often used in conjunction with the mapping of response surfaces in Example (n = 4) Factor 2 (columns) which the responses of the earlier phases are used to help predict where to-select additional treatment combinations Factor 1 1 A B C D for study so as to optimize results efficiently. the units within each block Example: being more homogenous than units in different blocks. equation of the model can be written as: Example : Y =Bo+BlXl +B~XZ+B~~X$+BIZXIX~+ Four treatments A. are to be studied in a 2x3 factorial ment appearing precisely once in each block. and levels (versions) of a factor are allocated to n2 e = random error experimental units arranged in n rows or n columns 2. (rows) 2BCDA 2. grouped into blocks. For example.! Bz = curvature effect of factor X2 3. The of the factors forany EVOP experiment is usually quitesmall blocks are customarily identified with the rows and columns in order to avoid making out-of-tolerance products. design. A design in which the experimental units are d) Each of residual random variable is normallydistributed. that the response pattern can replicated in several blocks with a separate be approximated by a polynomial model.3 Latin Square Design B12 = Interaction (Differential) effect of the two factors (linear) A design. NOTE.

Greek letters a) Each block contains k ( < t) distinct treat- ments NOTE. when the assumptions t = 7. there are: 1(2x2).3 are also pertinent here. Of these. k.Latin letters Factor 4 . 1(3x3). . 4(4x4). for minimizing the amount of the ex.4 Graeco-Latin Square Design principal factor appears the same number of times ruin the experiment and that the following relations hold true: A design with three blocking factors where the n levels (versions) of a factor are allocated to . r. given Example (n = 4) Factor 2 (columns) 5 integers (t.b = l. or among themselves. the following conditions: size 6. if a treatment p is Graeco-Latin square’. i) any two treatments are either lst. together in the same number 1 of blocksfrom the If this assumption is not valid. ran.k = 4. mth associates. coluirms and greek letters binations (t. r (k-l) = 1(t-1) and b 1 t experimental units arranged in n rows and n For randomization. Randomization can be achieved by assigning at ran. A generalization of the Graeco-Latin square for . Graeco. and the effect of the Erample : factor is confounded with such interactions.The design implies that every version of the 3. modified by the extension to four factors. b. balanced incomplete block designs.7’ Partially Balanced Incomplete Block Dksi& 2 (PBIBD) 3 An incomplete block design with t treatments and 4 b blocks is called a PBIBD with m ( 12 ) associate classes if: Factor 3 . Block 1 2 3 6 This is equivalent-to a fractional factorial with the 2 3 4 7 assumption of no interactions. the relation being more than four factors is known as the ‘hyper symmetric. Versions of the perimentation.* bk = tr.b. 6 ‘I . A) is possible for constructing represent levels of three blocking factors. 2nd. 3. 161280 (5x5) Latin squares. Factor 1 1 (rows) 3. d) satisfying the above 3 condi- .5 -Incomplete Block Design cipal factor and of each of the block factors has to-be the same. it is not necessary that a BIBD will exist. Sometimes other principal factors principal factor are used in the block positions so that there may be three principal factors without any block factors. that is.The comments in the note in 3.IS 7920 ( Part 3 ) : 1996 columnsoperators. . k. A design in which the experimental units are dom the versions of the principal factor to the letters. grouped into blocks in which there are insufficient domly selecting a Latin square from the listings or by the procedures described in statistical tables and assigning the number of experimental units available within a versions of the block factors at random to the rows and block to run a complete set of treatments columns of the square [There are: 1 (2-x2). Since appears exactly once in each row and exactly once each letter in the above equations represents an in each column and also appears with greek letter integer. 56(5x5) ‘standard’ Latin squares 3. arrange the blocks and versions columns in such a manner that each level (version) within each block independently at random. 576 (‘replicate’)~of the experiment.6 Balanced Incomplete Block Design (BIBD) in which the first row and first column are in alphabetical order. However. The rows. The design is particularly useful. . 7 1 2 5 NOTE . Some fractional 3 4 5 1 factorial designs ~form LaWsquares and it may be 4 5 6 2 more desirable to approach the problem from the 5 ~6 1 3 fractional factorial viewpoint to understand the 6 7 1 4 assumption-being made concerning interactions. A Graeco-latin square dots not exist for squares of satisfying. (4x4). it is clear that only a restricted set of com- exactly once. the measure of b blocks. residual error will be increased. and from which the other Latin squares can be derived An incomplete block design in which each block by permuting the rows and columns]. tions. r. generally used to eliminate three block Cl There exists a relation among treatments effects.The number ofversions(n) of the prin. contains the same number k of different versions from the t versions of a single principal factor A basic assumption is that these block factors do not interact (cause differential effects) with the arranged so that every pair of versions occurs principal factor under study. 12 (3x3). the I~ associate of a then so is a to #. b) Each treatment appears in r blocks Latin squares are.A= 2 are valid.

.. k.2. . They are arranged Three versions of factor A are tested in two so that not ail pairs of versions occur together in the same replicate runs. . . . IS 7920 ( Part 3 ) : 1996 ii) each treatment has ni~l”hassociate.2 .. .j. Bi Ei t=6. k = 1. . . fr = bk r=k=4. 6 i where . i. the same number of the b blocks.. t. compared with greater precision than others.:. 11. . This type of arrangement is common in industrial experimen- version l). tation as well as in agriculture (whence the name is derived). moulds into which the alloy might be poured.. one series of treatments requires a larger ex- 11 = 4 blocks and n2 = 4 versions (No.9 Split-Plot Design nlAl+n2&+. . The design implies that every version of the principal factor appears the Replicate I Replicate II same number of times r in the experiment.2. 2. r. t = 7 = 6. &.nl. . . ni+nz+. Within each version ofA.r=4. 3. For instance. columns.k=4. b=6.2. 9 The integers. . NOTE ... the number has levels smaller than the number of treatments. some versions can therefore be two versions of factor B are studied.3.+ nmAm = r(k-1) &+ = nj . and the other In this design every version occurs r = 4 times and factors can be varied readily within the runs or areas. . are connected NOTE.. . m)... j?) of i’ associate.nl=l.s are equal. piL is independent of the pair Example : (a .m.5 and perimental unit while another series can be compared with 6) that appears together with version 1 in 12 = 2 smaller amounts.nz= 4. is the A design in which the group of experimental units j Kronecker (plot) to which the same version of a principal factor is assigned is subdivided (split) so that one delta taking the value 1 if i = j and 0. Al and A2 are the amounts of alloy than the comparison of different types of same whatever the starting version may be. b) which number of treatments and the other block factor are mutually z*associates. the comparison of different typesof furnaces used to prepare~analloy would needgreater blocks. n.. not all d. m. 12=2 NOTE . replicates setve the role of blqcks Versions of the to the first-stage principal factor (A) and each plot assigned principal factor to eve of the three versions of A serves the role of blocks for the additional second-stage principal factor B (within plot Block factor) studied within k Thus.. The types of . different measures of residual 4 1 5 2 error are obtained for the within-plot and the between-plot 5 2 6 3 effects.. b. j.1=2. In a split-plot design.In the example. It is possible to extend this design further in order to 6 3 4 1 introduce a third-stage factor. ..rZ1= 4. . These parameters nl. Plot A1 1 Example: . version 4) that appears together with version 1 in Frequently. if we start with any version (for example. of the blocking factors~say. the number ni being A type of design with two block factors called as independent of the treatment rows and columns with the number of levkls of one chosen. 4x7 Youden Square Design d) &Y two -treatments that are mutually i”h associates appear together in li blocks (i = 1. .8 Youden Square Design i = 1. The number columns when treated as blocks form a BIBD. we find nl = 1 version (for example. .and& i. equals the iii) given a pair of treatments (a .k= 1.It can be seen that tows form a Randomized block by the following relations: design and columns a BIBD with parameters. of treatments that are simultaneously The treatments areallocated in such a manner that j” associate of a and @ associate of every treatment appears once in each row and the p is&.&. zi n2. the experimental error for the 2 5 3 6 within-plot factor B should be smaller than that for the full experiment (if there indeed is some effect of varying the first 3 6 1 4 factor). m&led the parameters of a PBIBD.+ nm = t-l 3. or more additional principal factors may be studied otherwise n& = ng& = nkpfj within each version of that factor. n2. . 6. This type of design is frequent- ly used where large runs or areas are obtainable from a factor the levels of which are not easily changed.PBIBD is an incomplete block design in which Example: each block contains the same numberk of different versions from the I versions of the principal factor.

The product of these 2 In some applications. proportions of each of theq components (Xi) in the or a subset of these combinations. tot-s) while the speed variations constitute the between-plot factor (or first-stage factor). practical considerations sometimes necessitate its use. are arranged in techniques. the appropriate arrangements (after randomization) might be as shown below: B3 Bt I32 Bq BI B4 B2 B3 BI B3 B2 B4 A2 AI A2 A3 A2 A1 A1 A3 A3 NOTE -The design sacrifices precision on the main effects (average effects) of A and B in order to provide higher precision on the interactions (differential effects). shaped under varying pressures and smoothed strips across plots in each replication. In others. a timeconsuming and expensive task. When a factorial experiment includes factors all having the same number of levels. Thus an experiment with three factors all run at formed with two or more factors. Another example is a large machine the Design speed of which can be changed only by replacing the gear A split-plot design in which the versions of the train. Two Way Split-Plot and the types of moulds as theversions of the second stage (within-plot) factor. each being studied two levels would be referred to as a 23 factorial 8 ‘I . These latter constitute the within-plot factors (or second-stage fk- ways. mixture shall satisfy the conditions 0 5 Xi I 1 and The most commonly used designs for the selected arrange- ment of the factorial treatment combinations are the com- 3 Xi = 1.10 Split-Block Design. an experi- mixture designs is appropriate for experimenting with the ment based on 3 levels of factorA. In industrial experimentation. tive restraints. treatments. for example. such as paints. The describe the experiment design in which these combinations.IS 7928 ( Part 3 ) : 1996 furnaces are regarded as the vetsions of the first-stage factor 3. foods. Thus. it is using different polishing agents with relative ease of shifting considered as a split-plot design in two different from one level (version) of these factors to another. will be studied. NOTES A factorial experiment is usually described symboli- 1 In some fields of application the experiment mixtures are cally as the product of the number of levels described by the terms ‘formulation’ or ‘blend’. gasoline. The material second-stage factor. the description 3.12 Factorial Experiment (general) is usually given in terms of the number of levels raised to the power equal to the number of factors. the randomized block design and i=l the balanced incomplete block design. Example: For a 3 x 4 design. 3. which will generally be more accurately determined than in either randomized blocks or the ordinary split-plot design. but others Also are terms of these proportions. used. and each experimental point is defin&l in pletely randomized design. The use of (versions) of each factor. For example. there may be opera. factor A may be different procedures of bleaching_ by chlorine peroxide and factor B those of rising by different amounts of hydrogen peroxide in the cooling process. 2 levels of factor formulations of manufactured products. so that infre- quent changes to this factor are desired. B and 4 levels of factor C would be referred to as a 3x2x4 factorial experiment. An experiment in which all possible treatments n.11 Miiure Experiment at two or more levels (versions). components shall-be mixed and the response is a NOTE . in the textile industry. so that at least one component cannot attain 0 or 100 per cent (‘reduced domain’). are examined so that interactions (different effects) as well as main An experiment in which two or more ingredients or effects can be estimated. the proportions of the components of the mixture may vary between 0 and 100 per cent of the numbers indicates the number of factorial mixture (‘mmplete domain’). instead of being randomized manufactured at each speed can be heat-treated by several independently within each plot.The term is descriptive of obtaining the various property of the resulting mixture that does not factors in all possible combinations. but in itself does not depend upon the amount of the mixture. rubber and textiles.

+ 1. -1.. For example.1 ).1. +I ). listed in the same order as the factors.[a-(1)] = [“b-a]- of various treatments. Example illustrating contrasts: is the contrast of the contrasts for studying the effects of factor A at the ‘high’ level ofB and at the In a 2” factorial experiment with factors A and B low level of B or the contrast of the contrasts for A: [a-(l)] + [ab -b] studying the effects of factor B at the ‘high’ level of A and at the ‘low’ level of A. In the above example the code would the factor. (+ l. 221. one can denote the first. contrasts when levels are equally spaced.121 studied. (+l.+ 1. A popular code for of unit deviations from the centre level. ( -1.212. (+ LO. + 1 ). AIBICZ. AzBzCl.AW High Low High Low High Low High Factor B Low Low High High Low Low High High Factor C Low Low Low Low High High High High Code (1) a b ‘ab C ac bc abc This type of identification has advantages for is the contrast for studying the effect of factor B at defining blocks. 100.I.1. are used. B and C: Level Factor A J. Factorial experiments. -1. A factorial experiment in which n factors are 111. (+l. 121. -1 . O.12). +1 ). the ‘low’ level ofA plus the contrast for studying the effect of factor B at the ‘high’ level of A. confounding and forming aliases. these terms being of the form B: [b-(l)] + [ab -a] (a 2 l). for example. At Br Cl. product’of two terms.020. three factors . the -1~coefficient might be treatments for a factorial experiment are: assigned to 3 mm and the + 1 to 7 mm with a) Use a letter to indicate the factor and a 0 being assigned to the non-included 5 mm numerical subscript the level (version) of level.4. with two levels.13 2n Factorial Experiment 000. ( +1. A1 B2C1. NOTE-The 2” factorial experiment is a special ease of the b) Describe the levels in terms of the number general factorial experiment (see 3. AzBzC2. Sometimes only the subscripts. 211.232 Unequal spacing of the levels or weighted Alternatively. 101.132.1 is used when the capital 9 . +l). emphasis for the various levels can also be second and the third level by 0. 120$01. 222. 122. B appear as and C in a 2 x 3 x 2 factorial experiment. if a unit of deviation between these Some commonly used notations for describing the levels is 4 mm. +l).1). regardless of the form of analysis used essentially involve forming contrasts AB:rL-Flj. each at two levels (versions). AzBsCl. particularly in analysing 111. 0. (. 231. such as: This descriptive coding has many advantages. + 1.021. IS 7920 ( Part 3 ) : 1996 experiment (n being equal to 3) and has 8 factorial ing the levels are usually given in terms of treatments. 112. in. AIBSCL AzBsC2 ( .. Al&Cl. the coefficients describ- Example illustrating the Note above: A 23 factorial experiment with factors A. In the case of an even number indicate a factor at a high level by the corresponding small of levels where there is no actual treatment letterand theletter isomitted when the factorisat a lowlevel.011.010. +1. representing treatments in a 2” factorial experiment is to cluding sign. -1). A2B1 Cl. 131.. AIBzCZ.where . 0. 12 combinations would be: ( -1. When all the factors are at low level the treatment code is (1). ( + 1.-l ).I). 110. half-unit deviations. The (-1.. (b 2 l). is the contrast for studying the effect of factor A at the ‘low’ level ofB plus the contrast for studying the E ac h contrast can be derived from ‘symbolic effect of factor A at the ‘high’ level of B.I). 1 and 2 reflected in the coefficients. respectively to obtain the following combinations: 3. + 1). at the centre level. A&CL ( -1. .

confounded in other replicates are corresponding coefficients in A and B.19 Fractional Factorial Design Example: A factorial experiment with only an adequately In a 23 factorial experiment with only room for 4 chosen fraction of the treatments required for the treatments per block.’ -U -b+ub A factorial experiment with several replicates in at Note that the coefficient of each treatment com.15 Randomized Block Factorial Design bc ubc abc UC A factorial experiment run in a randomized block The estimate of BC can be obtained only from design in which each block includes a complete set replicate 1 and that of ABC only from replicate 2. the following arrangement is selected so term contrasts the effect ofA at the high and the that theABCinteraction is confounded in replicate low levels of B (or the effect of B at the high and 1 and the BC interaction in replicate 2: low levels ofA). 3. AB and AC 3. These ‘defining contrasts”cannot bc estimated and thus are sacrificed. free from confounding. Where only one effect is to be confounded with Thus. confounded in all replicates. All factor level combinations are included in the experiment. 3. There are generalized rules for more complex B : (a + 1)(LJ . blocks. C. confounding.17 Factorial Experiment with Partial B : -u+b+ub Confounding AB: -::. can be sacrificed through confounding with blocks NOTE .UC . Similarly B represents the blocks of 4 (see 3.ub + c . A factorial experiment in which only a fraction of the treatments are run in each block and where the 3. A : . The AB replicates.16 Confounded Factorial Design are obtainable using both replicates and therefore will have greater precision.bc + ubc] sometimes called ‘fractional replication’. B. that is.14 Completely Randomized Factorial Design Block 1 Block 2 Block 1 Block 2 mm A factorial experiment (including all replications) a (1) b run in a completely randomized design. 43. B. the ABC interaction complete factorial experiment. be determined once the effect(s) to be confounded is (are) defined. The remaining estimates ofA. This proper. ab b UC C Fractional factorial designs are often used very effectively in screening tests to determine which bc ubc factor or factors are effective. generally interaction(s) of high Block 1 Block 2 order. thus eliminating any confounding with blocks for them.1) will show the balance of the plus and minus signs in each fiese expressions are usually written iti a standard block. :: b C UC C bat ub 3.actions. the treatments with a positive sign are assigned to one block and those with a negative sign to A : (a -1) (b + 1) the other. as in thii example. a ‘main In a 23 factorial experiment requiring the use of effect’ or ‘average effect’. of factorial treatments. etc) AB: (a -1) (b . This procedure is [ABC : .(1) + a . or negligible. that is.18 Factorial Experiments with Total selection of the treatments assigned to each block Confounding is arranged so that one (or more) prescribed A factorial experiment with several replic+s in effect(s) is (are) confounded with the blppk which some interactions or main effects are effect(s). or as part of a sequen- 10 . In this case.b + ab 3. ty is general in Yfactorial experiments. After nor- -Example: malization.1) situations. an ‘interaction’or ‘differen- Replicate 1 Replicate 2 tial effect’. By ‘adequately chosen’ is meant selection U (1) according to specified rules which include consideration of effects to be confounded and aliased. A check on all the other effects (A. while the other effects remain free from . the A term represents the effect of A averaged over the two levels of B.IS 7920 ( Part 3 ) : -1996 letter (A.The treatments to be assigned to each block can it is not. order.(1) + a + b . least one of which some main effects or bination in AB (+l or -1) is the product of the inter. B) is included in the cOntrast and + 1 when NOTE .16) and carried out with 2 average effect of B over both levels of A.The fraction selected is obtained by choosing one without loss of any other effect if the blocks include: or several ‘defining contrasts’ which are considered of minor importance.

nij = number of times the level combination founded and inestimable. NOTE . because only half of the full factorial experiments is run. A : (a .1 )(d . and follows: N = Total number of experimental units. This example. of columns.1) (b + 1) (c + 1) (d + 1) is not equal to Example: BCD: (a+l) (b-l) (c-l) (d-l) Two half-replicates of a 24 factorial experiment (see 3.13. + A comparable. ni = number of times the level i occurs inbne column. ab abd but another approach to these situations is through ac acd the use of the composite design defined in 3. it will be found ttt’at (i. had the complete set of fac- may be made to a fractional.cd - biased estimates of main effects or of misjudging (1) : BCD the relative importance of various factors.21.bd .l)(c .1 )(b . biased estimates will result. When there is a large number of factor level combinations AB:abcd+ab+cd+(l)-ac-ad-bc- resulting from a large number of factors to be bd :CD tested. From the + sign fractional replicate in the above nj = number of times the level j occurs in the example. Because only those elements of the ABCD interaction having where the same sign are run. nij .1) tlij = N Those factorial combinations in the second column are those for every combination of (i. When this assumption is not valid. the factorial combinations in the first column are those with the a ‘+’ (plus) sign in the Mathematically. j) occurs in any two columns.13) for the code interpretation with defining when all 16 combinations are included instead of contrast : ABCD only 8. but there are risks of getting A : abed + ab + ac + ad . Note that. IS 7920 ( Part 3 ) : 1996 tial series of tests. but more difficult. each contrast representstwo effects. In addition.(minus) sign. and the comments thereon. the estimates of the A and BCD or AB and fact that. (1) : A necessary and sufficient condition for a design to Either of these half-replicates can be used as a be an orthogonal design is that each level of one tfractional replicate’. The usefulness of these designs stems from the them. That is not likely to occur. approach is abed abc available when there are more than two versions. we would compute the factorial effects as other column.In the example. 11 . Effects represented by the same contrast are named binations with one experiment. in general. it is often impracticable to test all the com. In such cases resort ‘aliases’. factor occurs with each level of the other factor with proportional frequency. this condition may be rewritten as development of the symbolic product of theAED ‘defining follows: contrast’ as illustrated in the example of 3. higher order interactions are BC effects would no longer be identical. have been limited to the 2” factorial experiments.20 Orthogonal Design bc bd ZJ A design in which all the effects can be estimated cd independently of one another. (no ALICD contrast measure is obtainable) so that the&CD interaction is completely con. X tZj ABCD : ((I . j) level and every ~pair with a . that is partial replica. ad bed 3. torial treatments been run instead of only half of tion.bc .

those obtained from a 2” factorial (or its fraction). and useful. . as compared with 9. 0.21 Composite Design An experiment to examine the effect of two or more A design developed specifically for fitting second factors in which the same level (version) of a factor order response surfaces to study curvature. . The degree of difficulty is dependent on the nature 3.12 factorial experiment (general)] the (2n + 1) additional combinations for a ‘central If two vendors are to be compared by evaluating two composite design’ are: shipments from each. full 23 l/2 replicate 1 2 3 4 5 6 7 factorial of 24 factorial experiment experiment 1 -1 -1 1 -1 1 1 -1 (1) (1) 2 1 -1 -1 -1 -1 1 1 a ad 3 -1 1 -1 -1 1 -1 1 b bd 4 1 1 1 -1 -1 -1 -1 ab ab 5 -1 -1 1 1 -1 -1 1 C cd 6 1 -1 -1 1 1 -1 -1 ac ac 7 -1 1 -1 1 -1 1 -1 bc bc 8 1 1 1 1 1 1 1 abc abed 23 Factorial contrast name l/2 replicate 24 (or24-1) fractional factorial contrast name NOTE-O~rthogonaldesignSincludeawidevarietyofspecialdesignssuch asalatinsquaredesign (3. .19) and so forth. Example: Example: If the coded levels of each factor are . Experiment 1 3.(O. 0 ). + a). and 25 units respectively.6.. Dierarchical of the non-orthogonality. 0). *. . However. non. adjusted by the mean of the array. 9.IS 7920 ( Part 3 ) : 1996 Example Table-of orthogonal design derived for l/2 replicate of a 24 factorial experiment is shown below: Treatment Array No.lS be evaluated independently. If shipment factor of vendor A are nested within that 12 ‘I . is equal to zero. simple since each main effect and interaction may NOTE . although additional duplicate runs orthogonal designs. lead to more difficult. The reduction in experiment size results in confounding. The total number of treatments to be tested is The differences between the two versions of the (2” + 2n +l) for a 2” factorial experiment. Statistical analysis of the results from experiments 2”-k’fraction is taken.14) a fractional factorial design (3. statistical interpretation.3). k a.O. and thereby sacrificing. ( + a. cannot be used with all levels (versions) of theother structed by adding further selected treatments to factors. Treatments for No.27 and 81 accidental (such as by the loss of data due to missing in the 3” factorial experiment.1 and + 1 in the 2” factorial experiment [see notation (b) under WY-x-j comment on 3. . It is also possible. A and that of vendor B or similarly for the second (0. shipment. which are already defined or derived. the number of treatments to using orthogonal designs is generally relatively be tested is (T-k + 2n + 1). to construct an orthogonal design by using appropriate tables of orthogonal arrays in which the sum of products of elements in any pair of arrays.acompletely randomized factorial design (3. there ordinarily is no direct relationship between the first shipment of vendor (090. .3 and 4 fheexperiment requires.O..22 Nested Experiment. . or information about interactions. which may be planned or of the centre point are usual. . . See the first note in 4.0).For n = 2. sometimes impossible. con. 0.. all tests or gross errors).

Another column ‘expected mean square E(M. .) (second column). 13 . The question could be framed in terms of Monday A technique of estimation of a parameter which morningversus Friday afternoon production. See of factor B. be crossed as in factorial experiments and other See the note in 4. In this arrangement. For example. ment factor of vendor B are nested within this other version of the vendor factor.F. the arrangement in 3.S. degrees of freedom (analogous to the . 3. parameters of the model or estimating variance Example: components. independent. if that makes a more meaningful question. See nested experiment versions that can be compared across other factors (X22). classification and could be ease of correlated errors. rather than in terms of two component from the model. . regression analysts and contrast analysis are all based on the method of least squares and provide different computational Vendor A B and interpretative advantages stemming from certain balanceswithin theexperimental arrangementswhich permit Day Monday 1 1 convenient groupings of the data. Factor A An analysis of variance table usually contains columns for . the expected mean squares version Cs or C4 and C7 or C8 might be eliminated. arranged as a factorial experiment. source of variation (first column). shipments 1 and 3 of the above example 4 METHODS OF ANALYSIS might represent Monday morning production and 4.S. When the levels (versions) In the example for a fully nested experiment (3.It is not unusual to find that experiments consist It is sometimes possible to redefine the factor into of both factorial and nested segments.23. the would be estimated with only half the precision of differences between the two versions of the ship.1 Method of Least Squares shipments 2 and 4 Friday afternoon production. show the composition of the ‘components of so that factor C is studied in only versions 1 and 3 variance’assignable to the appropriate sources. denominator n-l in the definition of sample Factor C Cl c2 c3 c4 cs C6 Cl CI variance s2) (D. run within only a subset of the versions of the first or succeeding factors. Friday 2 2 L I 4.S. similarly.4). unrelated shipments.23 Fully Nested Experiment A technique which subdivides the total variation of A nested experiment in which the second factor is a set of data into meaningful component parts nested within each level (version) of the first factor associated with specific sources of variation for the and each succeeding factor is nested within each purpose of testing some hypothesis on the level of the previous factor. although the method may be generalized to the rather than nested. NOTE . mean square (the sum of squares divided by 3. which minimizes Ze2. This would now represent a crossed [that is each level (version) of a factor is NOTE -The experimental errors associated with the individual observations ordinarily are assumed to be used with all levels (versions) of the other factors].4.25 Partially Nested Experiment NOTE-Generally.) may often be added to serve as a guide showing Example: which mean squares under the model are to be compared in an F-test.23). factors nested within the crossed combinations.24 Staggered Nested Experiment the degrees of freedom) (M. fS 7920 ( Part 3 ) : 1996 version of the vendor factor and.2 Analysis of Variance (ANOVA) 3.) (fourth Anested experimentin which the nested factors are column). nested experiments are used to evaluate A nested experiment in which several factors may studies in terms of components of variance rather than in terms of differences in response levels or prediction models. are selected at random. The usual analysis of variance.) (third column). sum of squares (S. the variability of C random effects model (4. ’ t_ version . where e is the random error has a common thread.

. it is inappropriate to compute components of variance.2 .=n-1 s. t.. NOTE . in ST=&+&+& conjunction with the F ratio.-Y..Basic assumptions are that the effects due to all NOTE . Averages and interactions are usually looked at by as: summarizing in 2-way (or k-way) tables.=iX. 14 .) b i i v.. i= 1.. are assumed to be selected at random from the distribution of versions to be studied for those The least square estimates of .u. factors.d) 4.2 . zero mean and have equal variances (homoscedasticity) throughout all subdivisions of the data.-Y.j-Y..j+Y.3). etc).=zm(Yij-Y... The assumption of a normal distribution is required only for this test of significance and confidence The model associated with the observations is given intervals.2.Components of variance are meaningful only for the sources of variation are additive and that the experimen.With fiied levels. the primary interest is usually .=t-1 s. = (r .=CZYij/rt in obtaining components of variance estimates and it is inap- i j propriate to compute estimates of the effects of the selected h factor levels.) WV I Block vg=r-1 1 Treatment v.)2= 2 factors are fued. When the assumption Xj=p +CZi+&+eij.r times possible to use transformations (for example logarithms.. it is some- j= 1.)2 i j In the ANOVA table. Ml is the.. and I 4. =Y .=rqqj-YL)2 1 I 2 v. it is assumed that selected (fured) A model in which the levels (versions) of all factors levels are designated. NOTE .5 Miied Effects Model = Y. -Y.4 Iiandom Effects Model / e ij is the experimental error. ai is the effect of the ith treatment.. r).j= 1.. obtained by: NOTE-With random levels..j+Y./t i A model in which the levels (versions) of some *2 = rz(Yij-X. but for other factors they are u i j [(f - l)(r.. For this example.2. significance for the effects of these sources of variation and/or to obtain estimates of the variances attributable to F (M. these sources. /si is the effect of thejth block...i/r 4.1) S. ( (D.. = Yi. =zY.xXZi(Yij-Yi.. sine inverse.-statistic. The technique.. is used to provide a teat of VT= vt + VJj+ v.1) (r . with = Zbj XC?Zi’ = 0: eij -N(O..’ Then the following ANOVA table is computed: Analysis of Variance (ANOVA)Table sum 01 -7 ittzzIF scl. This example as- sumes a jked effits model (see 4. l)] SC selected at random.t.IS 7920 ( Part 3 ) : 1996 Example: In a randomized block design the observation obtained from the ? oft treatments in the/* of r blocks is denotedbyxi (i= 1. the random level factors and their interactions with ‘fuced- tal errors are independently and normally distributed with effect’ factors..3 Fled Effects Model A model in which the levels (versions) of all factors are fixed rather than random selections where over the-range of versions to be studied for those p is the general mean.F. ai.. of normal distributions of error cannot be made. /$ and C? are factors.

S. Sa:By Sw4 Subtraction Total NOTE .) Replicates 8 sR=~(~j-~.)2 SRI8 Residual ’ 12 SE SE/I2 Lack of fit 4 SL = SE . replicate is run and the assumed model is selected NOTE .2x3~ sx. by the method of least squares). for example. Example : tion (for example. Regression Analysis Table for Example Addenda for Replicated Experiment Source Degrees of Sum of Squares Mean Freedom (S.S.S.6 Regression Analysis experiments.ZX.S. The ‘residual’ sum of squares might then be partitioned into 2 elements associated with ‘replicates’ and ‘lack of fit’ with 8 and 4 degrees of freedom respectively.=B. IS 7920 ( Part 3 ) : 1996 4.% Regression of Y on X2 (B) B2 = . three quantitative factors tical significance against an appropriate null studied in a 23 factorial.iY.) ZXijY.: sx2 wi CX. 1 Sx2=B2Zxziy.) Square (D. lack of balance x1 is the coded level of factor A.) Square (Da F) (M. A designed experiment which is orthogonally and then testing the resulting predictions for statis.SR SU4 15 ‘I .If the 2’ experiment were replicated within the same block the degrees offreedom for the ‘total’ would become 15 and the ‘residual’ would become 12. It is also useful for where designated experiments with missing data since the balance required for ordinary use of the analysis~of variance is not x0 is equal to 1 required for regression analysis. For balanced e is the random unexplained error.X) Source Regression Degrees of Sum of Squares Mean Coefficient Freedom (S. in which only a single hypothesis model. subsequent terms) of the hypothesis tests as well as losing x3 is the coded level of factor C. the two techniques are simplyvariations of the method of least squares and produce comparable resulta. Regression Analysis Table for Example (notation: x = X . 3i Residual . Regression of Y on XI (A) B1 = z xL i 1 SX.Regression analysis plays a role similar to the iiS analysis of variance and is particularly pertinent when the levels-of the factors are continuous and emphasis is more on Y = bg. The process of estimating the parameters of a model by optimizing the value of an objective func.) (M.iyi Regression of Yon X3 (C) B3 = zx 1 Sx3=B. + blxl + b2x2 + bJx3 + e the model than on the~hypothesis tests. increases the order-dependency (common elements are included in the first correlated term and not included in x2 is the coded level of factor B. and other advantages of balanced experiments.: % 1 XX. 4 . balanced. However.F.

9) provide mechanisms that let the ex- being studied (a person or a product) must be able perimenter consider beforehand how to reduce the to behave in a normal fashion following the experi.Usually the concomitant variable cannot be using the F-statistic for the mean square of that accounted for in the design of the experiment and its undesirable effect on the results has then to be taken into source and the appropriate error term. The com- conditions surrounding the -experiments and for bination of design and methods of analysis (see 4) arranging the experiments so as to provide the reflects how the design is effective. Arrange- is over. count for the contribution of the_concomitant vari- able. The ‘replicates’ term ordinarily would be Y. For example.1) is inde. deals with the inter-relationships between multiple factors of interest to the experimenter.3) and “split-plots” relatively small changes per step because the unit (see 3. _)” . if those parts of the experiment using low under conditions of isolation where all other fac- dosage of a drug were conducted in the morning and tors are held “constant” or at some “ideal” level.IS 7920 ( Part 3 ) : 1996 The statistical significance of each source is tested NOTE . One- While “design of experiments” (see 2. “incomplete blocks” in an existing process where there is a restriction to (see 3. To ac- which would-be included in the ‘residual’ term.i = pi.7 Analysis of Covariance obtain B by pooling the within-version value for EXzij and ZXcij Xj with A technique for estimating and testing the effects of treatments when one or more concomitant .Xc. but they can tion of the data collected. ANNEX A (Foreword) DESIGN OF EXPERIMENTS Design of experiments is essentially a strategy for methods should be considered because they help in experimentation that accounts for environmental the understanding of design differences. the ‘regression’ terms may differ in the amount of some chemical constituent would be tested against the ‘residual’ term. ’ “randomization” (see 2. it is necessary to limit blear manner.11) and “blocking” deal.). Variability exists. stances to gain insight into that factor. be adjusted to account for the X. values. In planning an experiment. which can be measured. model” equipment requiring further work to relate to “production” status or they may be run in routine The area of “factorial experimentation” (see 3. an analysis of variance may the potential contribution of model inadequacy produce biased results if not adjusted for. but not ad- two replicates situation. For the account in the analysii.2).)Yij c (xi j _ x. “Latin squares” (see 3. effects of unwanted variability and how to get more ment. For the present in each unit. A regression analysis could be used to 4.5). ments.12) type environments.. Because -of the con- represents a measure of experimental error free of comitant variable X. experimentation takes place ments into “blocks” (see 3. For taken into consideration. would the usually are not representative of what happens to environmental factor of time of day be confounded that factor in the “real” world where there is with the levels of dosage? Topics such as simultaneous variation of many things. and it must be biases introduced by the environment. those with high dosage in the afternoon. Studies of some factors example. factor-at-a-time studies may be useful in some in- pendent in a sense from the analysis and interpreta. and the ‘regression’ terms would also The model for a singk factor analysis of variance be tested against ‘replicates’. answer to the questions of interest in an efficient. The Y values could then variables influence the response variable but they themselves are not influenced by the treat. The experiments may be run on “laboratory meaningful answers. be tested against the ‘replicates’ (‘experimental error’) term to determine whether the model is Example: inadequate. Experimentation may take place in a laboratory with issues of how to minimize the unwanted effects where there is a high degree of freedom to change of these “noise” elements that are usually so the levels of the factors of interest because the test numerous they could not be eliminated even if it specimens are not to be used after the experiment were economical or realistic to do so. frequently used analysis also be misleading if that factor behaves differently 16 ‘I .=i(Xcij-X. the model might become Yij = pi + B (Xc i j) . the ‘lack of fit’ term would justed. the experimental units single replicate situation. In other cases. .

(see 2. Title IS No. 3 Special control charts 397 Method for statistical quality 5420 Guide on precision of test (Part 2) : 1973 methods : Part 2 Inter- (Part 4) : 1987 control during production : Part laboratory testing 4 Master control systems 1548 Manual on basic principles of lot 6200 Statistical tests of significance : (Part 1) : 1981 sampling : Part 1 Itemized lot (Part 1) : 1995 Part 1 t. For work near the optimum points. ofthe factors shows greatest promise d) show the pattern and arrangement of of a real change. or a ment is frequently used to create a “model” failure maystem from unknown interaction effects. Title 397 Method for statistical quality 2500 Sampling inspection procedures (Part 1) : 1972 control during production : Part (Part 3) : 1995 : Part 3 Attribute sampling plans 1 Control charts for variables AS0 2859-2 : indexed by limiting quality (LQ) 397 Method for statistical quality 1985 for isolated lot inspection (Part 2) : 1985 control during production: Part (Under Print) 2 Control charts for attributes 4905 : 1968 Methods for random sampling and count of defects 5420 Guide on precision of test 397 Method for statistical quality methods : Part 1 Principles and (Part 1) : 1969 (Part 3) : 1980 control during production : Part applications . selection of factors and levels. IS 7920 ( Part 3 ) : 1996 in the presence. is of immense value. screening to determine if there is any apparent Good experiment design should: interest in the factor.2l)adding supplemen- tary points to the two-level factorial experiment.21) of how the factors relate to selected Factorial experiments may be at two versions or levels.22) serves as a levels of each factor. of selecting specific portion of a complete factorial C) reflect whether an experiment series or a experiment. factors of potential interest or to optimize some mits a step forward comes from the synergism effects. For finding out one-shot experiment is desirable. ings in carrying out the experiment and. Frequently the “breakthrough” that per. which. They may also include three furnish required information or more levels or versions to allow for estimation a) with minimum effort. A “response surface” (see 2. The size of the experiment is an obvious consideration in experiment efficien- b) lead to pre-experiment determination of whether the questions of interest can be cy. Normal & F-tests sampling 6200 Statistical tests of significance : (Part 2) : 1977 Part 2 x2 Test 2500 Sampling inspection proce- (Part 1) : 1992 dures: Part 1 Attribute sampling 6200 Statistical tests of significance : /Is0 2859-l : plans indexed by acceptable (Part 3) : 1984 Part 3 Tests for normality 1989 quality level (AQL) for lot.13). absence or at other levels of other Experimentation is generally carried out to find factors. 6200 Statistical tests of significance : by-lot inspection (Part 4) : 1983 Part 4 Nonparametrictests 2500 Sampling inspectionprocedures 7200 Presentation of statistical data : (Part 2) : 1965 : Part 2 Inspection by variables (Part 1) : 1989 Part 1 Tabulation and sum- for percent defective marization 17 . which limits interpretation to map these models and may be useful in predic- linear relationships but may be sufficient for tion and location of the next phase of experiments.19). “screening” experiments using experiment points to avoid misunderstand- small fractional replications can be very effective. a means clearly answered in the experiment. ANNEX B (Foreword) LIST OF INDIAN STANDARDS ON STATISTICAL METHODS IS No. and “fractional replication” (see 3. of “curvilinear” effects. if any. curvature effects may be studied by the creation of e) encourage the use of prior knowledge and experience in describing assumptions and “composite” designs (see 3. For optimization. the data from the experi- revealed in a study of “interactions” (see 2.

Estimation of mean. N. G. MSD 3 Chairman Representing Prof S. -IO427 Designs for industrial ex- relation (Part 2) : 1986 Part 2 perimentation : 7600 : 1975 Analysis of variance Orthogonal arrays 7920 Statistical vocabulary and 10645 : 1983 Method -for estimation of (Part 1) : 1994 symbols : Part 1 Probability and process capability general statistical terms 7920 Statistical vocabulary and 12347 : 1988 Analysis of means . AGGARWAL University of Delhi. W. Ministry of . Quality and Reliability. Calcutta SHRIU. V. BISWAS(Alternate) DR AsttoK KUMAR Defend Research’and Development Organization. Calcu~i:~ PROFARVINDSEX-H(Altemare) SHRIM. Title IS No. New Delhi ~DR 0. LAHIRI Indian Jute Industries Research A-. KRISHNAN Bharat Heavy Electricals Limited. Delhi PROF S. P. DAS Tata Engineering and Locomotive Company Limited. BHOWMIK(Ahmare) DR B. New Delhi StiRI M. MOHAN Indian Statistical Institute.IS 7920(Part3):1996 IS No. DATEY Quality Management Institute. NAIK Bajaj Auto Limited. DAS Indian Association for Productivity. DU?TA (Alremate) PROFS. Defence. DATE (Akmute) (Continued on Page 19) 18 ‘I . New Delhi SHRIM. KATHURIA Indian Agricultural Statistics Research Institute. KUMAR(Alremare) SHRIBHAGWANDAS Central Statistical Organization. M. Calcutta Members SHRI S. P. P. 13131 : 1991 Statistical tolerance interval - periments Methods for determination 8900 : 197s Criteria for the rejection of 14277 : 1995 Statistical interpretation of test outlying observations results .91 KUMAR@[ternare) PROF M. (Part 1) : 1982 experimentation : Part 1 sentation of data Standard designs 7300 : 1995 Methods for regression and car. A graphi- (Part 2) : 1994 symbols : Part 2 Statistical cal procedure quality control 12348 : 1988 Use of probability papers 7920 Statistical vocabulary and (Part 3) : 1995 symbols : Part 3 Design of ex. Jamshedpur SHRIA. L. New Delhi SI-IRISuniz. 9300 Statistical models for industrial standard deviation and (Part 1) : 1979 applications : Part 1 Discrete regression coefficient - models Confidence interval (under 9. BHADE Tata Iron and Steel Company Limited. Hyderabad SHRI JAVEDAFAQUE (A&mute) DR A. MUKERJEE University af Calcutta. AGARWAL POWERGRID Corporation of India Limited.sociation. Pune SHRI S. Calcutta .100 Statistical models for industrial print) (Part 2) : 1989 applications : Part 2 SP 28 : 1994 Handbook on statistical quality Continuous models control ANNEX C ( Foreword ) COMMITTEE COMPOSITION Statistical Methods for Quality and Reliability Sectional Committee. New Delhi DR PRA~NE~HU(ABemare) SHRIA. Jamshedpur SHRIG. K. DR DEBABRATARAY (Alternate) SI1RIS. R. Title 7200 Presentation of statistical data : 10427 Designs for industrial (Part 2) : 1975 Part 2 Diagrammatic repre.

BIS Basic Statistical Methods Subcommittee. New Delhi Panel for Basic Methods including Terminology. K. Madras SHRIP. V.LAHIRI Indian Jute Industries’ Research Association.SAMPATH (RETD) Management and Statistics Centre. New Delhi SHRID. R.R. C. PAHUJA Steel Authority of India Limited. DRDO. Quality and~Reliability~Calcutta 19 . Pune PROF A. New Delhi . K. MSD 3 : 1 Convener PROQ S. New Delhi SHRI G. SEN DCM Limited. L. TALWAR. New Delhi MOtlbOS DR A~HOKKUMAR Defer& Science Centre. M. Ministty of Defence. NANKANA In Petsonal capacity (B-109 Malviya Nagar. MOHAN Indian Statistical Institute. New Delhi SHRlRAhiANISUBRAMANIAM SRF Limited. New Delhi DRA. NANKANA In Personal capacity (B-109 Malviya Nags. Madras SHRt c. DRDO. NA~K Bajaj Auto Limited. W. New Delhi SHR~GOPALBANDHU (Altemute) SHR~C. New Delhi) SHRIK. New Delhi SHWkKSAHA (Altema~) SHRID. ‘ DR D. MUKHWEE University of Calcutta. New Delhi DR S. R. P. DATEY Quality Management Institute. New Delhi SHIU M. DATEY Quality Management Institute. RANGANATHAN Lucas-TVS Limited. RAY Indian Associition for Productivity. Calcutta PROQ S. Ministry of Defense. N. SEN DCM Limited. Calcutta Members PROQ M. Delhi DR ASHOK KUMAR Defence Science Centre. Director General. MDS 3 : l/F-l Convener SHRI G. Madras SHRJS. New Delhi) SHRl KK PAHUJA Steel Authority of India Limited. AGGARWAL University of Delhi. SEDDEY Army Statistical Organization. SFSHAGIRI(Al&mate) AIR~MDER.W.SUBBARAhWA H (Altema&z) SHRI A. R. IS7920(Part3):1996 (Continuedfrom Page 18) PROP k N. BIS (&officio Msnbs) Director (MSD) Member Secrq SHRNATI BINDUM~A Deputy Director (MSD). MOHAN Indian Statistical Institute.

22 Nested 3.6 Interaction 2.4 D Mixed effects 4.4 of bovariance 4.25 S Staggered nested 3.15 Mired effects model 4.14 Fully 3.10 F V Factor 2.3 Level 2.3 with total confounding 3.18 Fiied effects model 4.21 Staggered 3.20 Partially balanced incomplete block 3.15 Partially balanced incomplete block design 3.16 Model 2.11 Confounded factorial design 3.19 2” 3.6 ’ Hierarchical experiment 3.22 Completely randomized factorial 3.12 T Main 2.24 Snlit block design 3.20 design 3.22 variable 2.11 E Randomized block design 3.1 factorial design 3.2 Randomized block factorial 3.12 Replication 2.16 Partial 3.11 surface 2.20 with partial confounding 3.1 Contrast(s) 2.2 Experiment factorial design 3.1 contrasts 2.5 Design N Balanced incomplete block 3.8 Random effects model 4.5 Orthogonal Latin Square 3.19 Youden Square 3.17 of experiment 2.12 Composite design 3.12 Treatment 2.10 R Youden square 3.25 Fractional factorial 3.17 Version (of a factor) 2.21 Confounding 2.16 Mixture experiment 3.4 Evolutionary Operation (EVOP) 2.5 Orthogonal 2.10 unit 2.8 20 ‘I .1 Nested experiment 3.14 Fixed effects 4.4 Duplication 2.24 Confounded factorial 3.6 Factorial 3.3 Completely randomized M design 3.23 Aliases 218 c Analvsis Graeco-Latin square design 3.9 error 2.1 Regression analysis 4.21 Method of least squares 4.2 Variable Factorial experiment 3.7 -of variance (ANOVA) 4.9 Predictor variable 2.10 Partially nested experiment 3.22 Average effect 212 I B Incomplete block design 3.23 %o-way split plot design 3.3 Random effects 4.19 Two-way split plot 3.35 Y Fractional factorial design 3.9 Fully nested 3.10 Exnerimental Split plot desigi 3.13 Response 2.22 Partially nested 3.14 Main effect 2.20 Mixture 3. 4.4 Incomplete block 3.2 H Regression.10 Randomization 2.5 Balanced incomplete block design (BIBD) 3.6 Completely randomized 3.7 P Randomized block 3.12 Predictor 2.7 L Factors 2.16 Orthogonal 3.23 Composite 3.13 Block 2.19 0 Graeco-Latin square 3.3 array 2.8 C Latin square design 3.6 Two-wav 3.IS 7920 ( Part 3 ) : 19% ANNEX D ( Foreword > INDEX A Fully nested experiment 3.5 Staggered’neated experiment 3.25 Split plot 3.24 Effect Average 2.7 Split block 3.15 design 2.23 Response Hierarchical 3.

type or grade designations. New Delhi-l 10015. Scheme VII M.235 23 15 Western : Manakalaya. This Indian Standard has been developed from Dot : No. India. V. Copyright BIS has the copyright of all its publications. HYDERABAD. Date of Issue Text Affected BUREAU OF INDIAN STANDARDS Headquarters: Manak Bhavan. T. in the course of implementing the standard. MADRAS 600113 235 02 16.337 9120 Northern : SC0 335-336.323 83 75. CHANDIGARH 160022 . SIS.832 78 92 Branches : AHMADABAD. 9 Bahadur Shah’iafar Marg 323 76 17 NEW DELHI 110002 323 38 41 Eastern : l/14 C.832 78 58 MUMBAI 400093 1 832 78 91. E<*NGALORE. marking and quality certification of goods and attending to connected matters in the country. Printed at Dee Kay Printers. GHAZIABAD. BHUBANESHWAR. if the review indicates that changes are needed. I. JAIPUR. Amendments Issued Since Publication Amend No. I. Standards are also reviewed periodically. such as symbols and sizes. New Delhi 110002 Telegrams : Manaksanstha Telephones : 323 0131. Review of Indian Standards Amendments are issued to standards as the need arises on the basis of comments. Sector 34-A.T. Campus. COIMBATORE. 1986 to promote harmonious development of the activities of standardization.337 85 61 CALCUTTA 700054 337 86 26. IV Cross Road. PATNA. of necessary details. Users of Indian Standards should ascertain that they are in possession of the latest amendments or edition by referring to the latest issue of ‘BIS Handbook’ and ‘Standards Monthly Additions’. ‘I . it is taken up for revision. 9 Bahadur Shah Zafar Marg. P. Road.Bureau of Indian Standards BIS is a statutory institution established under the Bureau oflndiun Stundmds Act. THIRUVANANTHAPURAM. a standard-along with amendments is reaffirmed when such review indicates that no changes are needed. Marol. GUWAHATI. No part of these publications may be~reproduced in any form without the prior permission in writing of BIS. MSD 3 ( 31). Andheri (East) 832 92 95.323 94 02 (Common to all offices) Regional Offices : Telephone Central : Manak Bhavan. This does not preclude the free use. Maniktola 337 84 99. LIJCKNOW. I. FARIL’ABAD.235 04 42 1 235 15 19. KANPUR. E9 MIDC. Enquiries relating to copyright be addressed to the Director (Publications).603843 ( 60 20 25 Southern : C. BHOPAL.