Chi-square and ANOVA Analysis Guide

Chi-square and Analysis
of Variance
harishram@hotmail.com
KDUE73YTIL
Introduction to Statistics
This file is meant for personal use by harishram@hotmail.com only.

Sharing or publishing the contents in part or full is liable for legal action.
Agenda
● Chi-square test
○ Goodness of Fit
○ Independence of Attributes
● Analysis
harishram@hotmail.com of variance
KDUE73YTIL
○ One way ANOVA
■ Total Variation
■ Variation within treatment
■ Variation between treatment
○ Post-Hoc Test for ANOVA
● Appendix
KDUE73YTIL
Tests for Categorical Data

Tests for categorical data
● The collected data is at times best represented by categories
● These categories are summarized by their frequency of occurrence. It may be of

KDUE73YTIL
interest whether this frequency is equal to the expectation/claim
● It may also be of interest to know whether the categories are statistically independent

Chi-square tests
● These interests are tested by non-parametric ways
● Tests based on the chi-square distribution are used

KDUE73YTIL
● The chi-square tests are used to test:
○ The goodness of fit
○ The independence of two attributes
● Chi-square tests are also used to test for population variance

Chi-square distribution
● Chi-square distribution is sum of squared of standard normals

Let X1, X2, …, Xn be n standard normal variates,
then Y =
harishram@hotmail.com X21 + X22 + … + X2n,
KDUE73YTIL
Y follows χ2 distribution with n degrees of freedom.
● The mean of the distribution is n and

its variance is 2n
● The distribution is positively skewed

χ2 Test for Goodness of Fit
KDUE73YTIL

Chi-square test for goodness of fit
At an emporium, the manager is interested in knowing age group which visits the mall
during the day. He defines categories as - children (age < 13), teenagers (13 ≤ age < 20),
adults (20 ≤ age < 55) and senior citizens (55 ≤ age). Moreover, he wishes to plan his
inventory of goods accordingly.
KDUE73YTIL
He claims that out of all the people who visited 5% are children, 38% are teenagers, 2%
are senior citizens are remaining are adults.
Can the owner verify the managers claim?

● The hypothesis to test whether the data fits the a specified distribution
H0: There is no difference Against H1: There is difference

KDUE73YTIL between observed between observed
frequencies and expected frequencies and expected
frequencies frequencies
● Failing to reject H0, implies that there is no difference between observed frequencies
and expected frequencies

● The test statistic is given by

observed frequency
KDUE73YTIL
total number of observations
estimated frequency
● Under H0, the test statistics follows χ2 distribution with k-p-1 d.f
where k: number of class frequencies
This file is meant for personal Ref: Test statistic for goodness
p: number of parameter estimated for use by harishram@hotmail.com only.
fitting of fit (A.1)
Decision Rule:
Reject H0 if χ2calc ≥ χ2k-p-1,α

KDUE73YTIL
or
Reject H0 if p-value ≤ α
Where, α is the level of significance (l.o.s.)

Test for goodness of fit
Question:
At an emporium, the manager is interested in knowing age group which visits the mall
during the day. He defines categories as - children, teenagers, adults and senior citizens.
KDUE73YTIL
He plans to have his inventory of goods accordingly. He claims that out of all the people
who visited 5% are children, 38% are teenagers, 2% are senior citizens are remaining are
adults.
From a sample of 180 people it was seen that 25 were children, 50 were teenagers, 90
were adults and 15 were senior citizens
Test the manager’s claim at 95% confidence level.

Solution:
We can tabulate the given data as follows:

KDUE73YTIL
Manager Claimed The frequency expected from The frequency observed
Frequency 180 customers (ei) from 180 customers (Oi)
Children 5% 0.05 x 180 = 9 25
Teenagers 38% 0.38 x 180 = 68.4 ≅ 68 50
Adults 55% 0.55 x 180 = 99 90
Senior Citizens 2% 0.02 x 180 = 3.6 ≅ 4 15

Solution:
To test, H0: The managers claim is correct Against H1: The managers claim is
false
KDUE73YTIL
The test statistic is

Solution:
Here there are 4 class frequencies, i.e k = 4. Since no parameter was calculated p = 0
KDUE73YTIL
From the statistical table for χ2 distribution, χ2k-p-1,α = χ23,0.05 = 7.815
The test statistic χ2calc = 64.27
Since χ2calc > χ2k-p-1,α , reject H0.

7.815
The managers claim is false, his claim is different than what was observed from the data
Python solution:
KDUE73YTIL
As p-value < 0.05, we reject H0.

● If the expected frequencies ei ≥ 5 and the total frequencies are large (≥ 50)
the test can be used
KDUE73YTIL
● If ei < 5, the class is merged with the neighbouring class for observed and
expected frequencies until the it becomes ≥ 5
● It is not applicable for testing the goodness of fit of a straight line or any
curve (exponential curve, second degree curve)

χ2 Test for Independence of Attributes
KDUE73YTIL

Chi-square test for independence of attributes
● The hypothesis to test independence of attributes
H0: The attributes are independent against H1: The attributes are
dependent
KDUE73YTIL
● Failing to reject H0, implies that the attributes are independent
● Decision rule: Reject H0 at α l.o.s if χ2(r-1)(s-1) ≥ χ2(r-1)(s-1);α

or
Reject H0 if p-value ≤ α
Where, α is the level oforsignificance
Sharing (l.o.s.)
publishing the contents in part or full is liable for legal action.
Chi-square test for independence of attributes
observed frequency in ith row

harishram@hotmail.com and jth column
KDUE73YTIL
total number of observations
estimated frequency in ith

row and jth column
● Under H0, the test statistics follows χ2 distribution with (r-1)(c-1) d.f
where r levels for a category and c levels for another category
This file is meant for personal use by harishram@hotmail.com only. Ref: Test statistic for
independence of attributes (A.2)
Test for independence of attributes
Question:
A study was conducted to test the effect of the malaria parasite - plasmodium falciparum -
on heterozygous and homozygous humans. The vaccine was given to a cohort of 252
KDUE73YTIL
humans. Test whether the heterozygous humans are better protected than homozygous.
Infected with plasmodium Not infected with plasmodium

falciparum falciparum
Heterozygous humans 93 51
Homozygous humans 68 40
Solution:
Let X: The zygote type - Homozygous or Heterozygous

Y: Whether infected or not with malaria parasite
KDUE73YTIL
Here X and Y are two attributes.
To test, H0: The attributes are independent Against H1: The attributes are
dependent
Here there are 2 rows and 2 columns.
Let us computed the expected frequency.

Solution:
In order to compute the Infected with Not infected with

expected frequency, first
KDUE73YTIL plasmodium plasmodium Total
compute the total of the falciparum falciparum
each column and row.
Heterozygous
93 51 144
humans
Homozygous
68 40 108
humans
Total 161 91 252

Solution:
Infected with
The expected plasmodium
Not infected with
Total
frequencies are
harishram@hotmail.com plasmodium falciparum
KDUE73YTIL falciparum
computed as
Heterozygous
144
humans
Homozygous
108
humans
Total 161 91 252

Solution:
Infected with
Similarly compute the plasmodium
Not infected with
Total
expected frequencies
harishram@hotmail.com plasmodium falciparum
KDUE73YTIL falciparum
for other classes
Heterozygous
144
humans
Homozygous
108
humans
Total 161 91 252

Solution:
The observed frequency (Oij)

The expected frequency (eij)
KDUE73YTIL

Solution:
The test statistic is computed as

KDUE73YTIL

Solution:
Here there are 2 levels of one attribute and 2 levels of another.

Thus the degrees of freedom are (r-1)(c-1) = (2-1)(2-1) = 1
KDUE73YTIL
From the statistical table for χ2 distribution, χ2(r-1)(s-1),α = χ21,0.05 = 3.841
The test statistic χ2calc = 0.070
Since χ2calc < χ2k-p-1,α , we fail to reject H0.

3.841
The attributes are independent.

Python solution:
KDUE73YTIL
As p-value > 0.05, we fail to reject H0.

Independence of attributes
Question:
A psychologist wants study whether the happiness quotient of children in the house is
related to the
KDUE73YTIL family income. He collects data of 1300 children is there enough evidence to
claim that they are related.
Low income Moderate income High income
Happy 245 354 243
Unsatisfied 98 220 140

KDUE73YTIL Tests based on Chi-squared distribution for categorical data
are one tailed tests.

Analysis of Variance
KDUE73YTIL

Question
Ryan is a production manager at an industry manufacturing alloy wires. They have 4
machines - A, B, C and D.
KDUE73YTIL
Ryan wants to study whether all the machines have equal efficiency based on the tensile
strength of the alloy wire.
Is it possible to test his claim?

Solution
The trivial solution is conducting multiple t tests. However performing multiple t tests has an
effect on the type I error.
KDUE73YTIL
As the number of t-tests increases the probability of at least one type I error increases.
However, it is possible to test Ryans claim by using one way analysis of variance (one way
ANOVA) where the probability of type I error does not change

Multiple t tests and type I error
● For a true null hypothesis, the probability of not obtaining a significant result is 0.95 if
the α = 0.05
KDUE73YTIL
● Say you conduct the t-test twice, the probability of not obtaining one or more
significant result is 0.95 x 0.95 = 0.9025
● Thus the probability of at least one type error is 1-0.9025 = 0.0975 (for two t-tests)

Multiple t tests and type I error
Probability of not obtaining one or more significant Probability of at least

Number of t tests
result one type I error
KDUE73YTIL
3 t tests 0.95 x 0.95 x 0.95 = 0.857 0.143
4 t tests 0.95 x 0.95 x 0.95 x 0.95 = 0.815 0.185
5 t tests 0.95 x 0.95 x 0.95 x 0.95 x 0.95 = 0.774 0.226
As the number of tests increase the probability of at least one type error also increases

ANOVA - History
● ANOVA was first introduced by Prof R. A. Fisher in 1920’s
● He developed
KDUE73YTIL
ANOVA while dealing with agronomic data

One way ANOVA
● A t-test is used when two unpaired data are compared
● To test the equality of population means for two or more unrelated samples ANOVA
KDUE73YTIL
technique is used
● Each group is considered to be a treatment
● It is based on the F-distribution

F distribution
● Let X be χ2m distribution and

let Y be χ2n
KDUE73YTIL
● Then the ratio follows F

distribution with (m,n) df

F distribution
As n1 and n2 become large the F

harishram@hotmail.com distribution becomes symmetric
KDUE73YTIL

One way ANOVA - assumption
● The samples should be independent
● Each sample
harishram@hotmail.com should be from normally distributed population
KDUE73YTIL
● The population variance of the samples should be equal (homoscedastic)

One way ANOVA
● The null hypothesis to be tested is
H : The averages of all treatments

0
H1: At the least one treatment has
KDUE73YTIL are
same. a different average
Against i.e. µ = µ = … = µ
1 2 n
● Failing to reject H0, implies that all treatments have the same average

One way ANOVA
● Suppose Ryan collects data for tensile strength of wires produced by each machine
● It is said
harishram@hotmail.com there are 4 treatments (t = 4) A B C D
KDUE73YTIL
68.7 62.7 55.9 80.7

● Each treatment has 5 observations (ni = 5)
where i = 1, 2, …, t 75.4 68.5 56.1 70.3
70.9 63.1 57.3 80.9

● Total number of observations is given by N
79.1 62.2 59.2 85.4
78.2 60.3 50.1 82.3

One way ANOVA
● Let µi (i=1, 2, …, t) denote the average strength due to each machine
● For our
KDUE73YTIL
example, t = 4 A B C D
68.7 62.7 55.9 80.7
● The test hypothesis can be written as 75.4 68.5 56.1 70.3
70.9 63.1 57.3 80.9

H0: µ1 = µ2 = µ3 = µ4 Against H1: At least µi is different
79.1 62.2 59.2 85.4
78.2 60.3 50.1 82.3

One way ANOVA
● In one way ANOVA, the entire population variance is split into two component
○ Variation within treatment

KDUE73YTIL
○ Variation between treatment
● Total variation = Within treatment variation + Between treatment variation

Total variation
● It is the total sum of squares (TSS)
● Let xij be the observations in the ith treatment and jth row
KDUE73YTIL
● is the grand mean, i.e. the mean of all observations
● The total variation is given by
Summation over all Summation over all

treatments observation in the treatment
Within treatment variation
● It is the treatment sum of squares (TrSS)
● Let xi. be the observations in the ith treatment with ni in observation in each treatment
and
KDUE73YTIL
is the mean over ith treatment
● The treatment variation is given by

This file is meant for personal use by harishram@hotmail.com
observation in the treatment only.
treatments
Between treatment variation
● It is the error sum of squares (ESS)
● Let xi. be the observations in the ith treatment and is the mean over jth row
KDUE73YTIL
● The error sum of squares is given by

treatments observation in the treatment
Error sum of squares
During problem solving, the error sum of squares is obtained as:

KDUE73YTIL

One way ANOVA

Mean Treatment Sum of
Squares
KDUE73YTIL
Mean Error Sum of Squares
● Under H0, the test statistic follows F-distribution with (dfTr, dfe) degrees of freedom
One way ANOVA
Decision Rule: If Fcal ≥ F(t-1,N-t),α or p-value ≤ α, then we reject H0 at α% level of significance
KDUE73YTIL
F(t-1,N-t),α
One way ANOVA
To ease the entire computational process, an ANOVA table is prepared as follows:
Source
harishram@hotmail.comof Degrees of Mean Sum of
KDUE73YTIL Sum of Squares F-ratio
variation freedom Squares
Treatment t-1 TrSS s2t
Error N-t ESS s2e
Total N-1 TSS

One way ANOVA - procedure
1. State the hypothesis to be tested
2. Compute the sum of squares

KDUE73YTIL a. The total sum of squares, TSS = ∑tj=1∑nii=1 (xij - x̄..)2
b. The treatment sum of squares TrSS = ∑tj=1∑nii=1 ni( xij - x̄i.)2
c. The Error sum of squares, ESS = TSS - TrSS
3. Compute mean sum of squares

a. s2t = Mean group sum of squares (MTrSS) = TrSS/(t-1)
b. s2e = Mean error sum of squares (MESS) = ESS/(N-t)
One way ANOVA - procedure
4. Compute the F-ratio
KDUE73YTIL
4. Prepare the ANOVA table
5. Write the decision and conclusion accordingly

One way ANOVA
Question:
Ryan is a production manager at an industry A B C D

manufacturing alloy seals. They have 4 machines
KDUE73YTIL 68.7 62.7 55.9 80.7
- A, B, C and D. Ryan wants to study whether all
75.4 68.5 56.1 70.3
the machines have equal efficiency.
70.9 63.1 57.3 80.9
Ryan collects data of tensile strength (in N/m2)
79.1 62.2 59.2 85.4
from all the 4 machines as given.
78.2 60.3 50.1 82.3
Test at 5% level of significance.
One way ANOVA
Solution:
Ryan is a production manager at an industry manufacturing alloy seals. They have 4

machines -
KDUE73YTIL A, B, C and D
Let µ1 be the average tensile strength due to machine A

µ2 be the average tensile strength due to machine B
µ3 be the average tensile strength due to machine C
µ4 be the average tensile strength due to machine D
To test, H0: µ1 = µ2 = µ3 = µ4 Against H1: At least one µi is different (i=1, 2, 3, 4)

One way ANOVA
The plot shows the

difference between the
KDUE73YTIL
average efficiency for

each machine, which
indicates the rejection
of H0.

One way ANOVA
Solution:
The grand mean:

harishram@hotmail.com A B C D
KDUE73YTIL
68.7 62.7 55.9 80.7
The total sum of squares: 75.4 68.5 56.1 70.3
70.9 63.1 57.3 80.9
79.1 62.2 59.2 85.4
78.2 60.3 50.1 82.3

One way ANOVA
Solution: A B C D
68.7 62.7 55.9 80.7

The treatment sum of squares is
KDUE73YTIL 75.4 68.5 56.1 70.3
70.9 63.1 57.3 80.9
79.1 62.2 59.2 85.4
78.2 60.3 50.1 82.3
∑xi 372.3 316.8 278.6 399.6
74.46 63.36 55.72 79.92

One way ANOVA
Solution: A B C D
68.7 62.7 55.9 80.7

The error sum of squares can also be
calculated as
KDUE73YTIL 75.4 64.5 56.1 80.3
70.9 63.1 57.3 80.9
79.1 59.2 55.2 81.4
78.2 60.3 50.1 82.3
74.46 63.36 55.72 79.92
Or can be obtained as This file is meant for personal use by harishram@hotmail.com only.
One way ANOVA
Solution:
Source of Degrees of
harishram@hotmail.com Sum of Squares Mean Sum of Squares F-ratio
KDUE73YTIL variation freedom
Treatment t-1 = 4-1 =3 TrSS = 1778.0655
N-t = 20-4
Error ESS = 296.06
=16
N-1 = 20 -1
Total TSS = 2241.5255
=19

One way ANOVA
Solution:
From the F-table we have F(3,16),0.05 = 3.24

KDUE73YTIL
Since 3.24 < 32.03, we reject H0.
F(3,16),0.05
One way ANOVA
Python solution:
KDUE73YTIL
As p-value < 0.05, we reject H0.

One way ANOVA can be said to check the effect of a nominal variable over a
KDUE73YTIL
numerical variable.

Further analysis
● In the example, Ryan has tested for strength of materials due to 4 machines
● The null hypothesis for ANOVA was rejected

KDUE73YTIL
● Now it is of Ryan’s interest to know which machine(s) has a different outcome
How would he find out?

Further analysis
● If we fail to reject the null hypothesis, it implies that all the treatments have the same
effect
● However,
KDUE73YTIL if the null hypothesis is rejected, it implies that at least one treatment has a
different average
● To know which treatment(s) has/have different outcome
● Can be found out using the post hoc tests

Post-Hoc Tests
KDUE73YTIL

Post-hoc test
● A post hoc is conducted after the null hypothesis of ANOVA is rejected to determine
the different treatments(s)
KDUE73YTIL
● There are various post hoc tests available such as:
○ Tukey’s HSD test (Tukey’s Honest(ly) Significant Difference test)
○ Scheffe test
○ Duncan's Multiple Range test
○ Fisher's’ LSD test (Fisher’s Least Significant Difference test)
○ Bonferroni test
● We will study the Tukey’s HSD test in detail

Post-hoc test
● Consider our example where Ryan wants to find out the which machines had different
result
● Each pair of machines is tested for the statistical difference
KDUE73YTIL
Machine B Machine C
Machine A Machine C Machine B Machine C Machine D
Machine D Machine D
Post-hoc test
Thus the test hypothesis are
H01:
harishram@hotmail.com μmachine_A = μmachine_B Against H11:
KDUE73YTIL
μmachine_A ≠ μmachine_B
H02: μmachine_A = μmachine_C Against H12:
μmachine_A ≠ μmachine_C
H03: μmachine_A= μmachine_D Against H13:
μmachine_A ≠ μmachine_D
H04: μmachine_B = μ or publishing the
Sharing
machine_C H is:liable
contents in part or full
Against 14 μ for legal ≠
machine_A
action.
Post-hoc test
The test statistic is:

Obtained from the
Tukey table
KDUE73YTIL
t: total treatments
f: error degrees of freedom
MSE: Mean error sum of squares (from ANOVA table)
n: number of observations in a group
Post-hoc test
● Consider the absolute difference between two treatments |x̄i - x̄j|
● The decision rule: Reject H0, if the absolute difference ≥Tα

KDUE73YTIL
● The python code:
First create the DataFrame df_machine then use the following function

Post-hoc test
The output is as follows:
KDUE73YTIL
True: reject H0
False: fail to reject H0 (accept H0)
It can been seen that there is statistical difference between pairs of machines (A,B),
(A,C), (B,D), and (C,D).
● For equal number of observations in each treatment, tukey HSD test can be
used
KDUE73YTIL
● However when the data is unequal it is not efficient
● In such a scenario, one may use the Scheffe test

Summary
Parametric Test Non-parametric Test
KDUE73YTIL
Kruskal-
One-way Reject Reject
Wallis H
ANOVA H0 H0
Test
Post-hoc Post-hoc
Test Test
Fail to Fail to
reject H0 reject H0
One - way ANOVA
● Ryan only considered the effect of the machines on the tensile strength
● What if
harishram@hotmail.com he considers the effect of work shifts used?
KDUE73YTIL
● The effect of the quality of material and the effect due to machine can be tested using
two way ANOVA

KDUE73YTIL Thank You


Chi-square and ANOVA Analysis Guide

Uploaded by

Document Information

Original Description:

Original Title

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

Chi-square and ANOVA Analysis Guide

Uploaded by

Copyright:

Available Formats

Chi-square and Analysis

This file is meant for personal use by harishram@hotmail.com only.

○ Post-Hoc Test for ANOVA

Tests for Categorical Data

This file is meant for personal use by harishram@hotmail.com only.

● The collected data is at times best represented by categories

● These categories are summarized by their frequency of occurrence. It may be of

This file is meant for personal use by harishram@hotmail.com only.

● These interests are tested by non-parametric ways

● Tests based on the chi-square distribution are used

○ The goodness of fit

○ The independence of two attributes

● Chi-square tests are also used to test for population variance

This file is meant for personal use by harishram@hotmail.com only.

● Chi-square distribution is sum of squared of standard normals

● The mean of the distribution is n and

● The distribution is positively skewed

This file is meant for personal use by harishram@hotmail.com only.

Can the owner verify the managers claim?

This file is meant for personal use by harishram@hotmail.com only.

H0: There is no difference Against H1: There is difference

This file is meant for personal use by harishram@hotmail.com only.

● The test statistic is given by

Reject H0 if χ2calc ≥ χ2k-p-1,α

Where, α is the level of significance (l.o.s.)

This file is meant for personal use by harishram@hotmail.com only.

Test the manager’s claim at 95% confidence level.

We can tabulate the given data as follows:

Children 5% 0.05 x 180 = 9 25

Teenagers 38% 0.38 x 180 = 68.4 ≅ 68 50

Adults 55% 0.55 x 180 = 99 90

Senior Citizens 2% 0.02 x 180 = 3.6 ≅ 4 15

The test statistic is

This file is meant for personal use by harishram@hotmail.com only.

The test statistic χ2calc = 64.27

Since χ2calc > χ2k-p-1,α , reject H0.

As p-value < 0.05, we reject H0.

This file is meant for personal use by harishram@hotmail.com only.

This file is meant for personal use by harishram@hotmail.com only.

● The hypothesis to test independence of attributes

● Failing to reject H0, implies that the attributes are independent

● Decision rule: Reject H0 at α l.o.s if χ2(r-1)(s-1) ≥ χ2(r-1)(s-1);α

● The test statistic is given by

observed frequency in ith row

total number of observations

estimated frequency in ith

Infected with plasmodium Not infected with plasmodium

Let X: The zygote type - Homozygous or Heterozygous

Here X and Y are two attributes.

Here there are 2 rows and 2 columns.

Let us computed the expected frequency.

In order to compute the Infected with Not infected with

Total 161 91 252

This file is meant for personal use by harishram@hotmail.com only.

Total 161 91 252

This file is meant for personal use by harishram@hotmail.com only.

Total 161 91 252

This file is meant for personal use by harishram@hotmail.com only.

The observed frequency (Oij)

This file is meant for personal use by harishram@hotmail.com only.

The test statistic is computed as

This file is meant for personal use by harishram@hotmail.com only.