CHAPTER 11.

GOODNESS OF FIT AND CONTINGENCY TABLES

The chi-square distribution was discussed in Chapter 4. We now turn to some

applications of this distribution. As previously discussed, chi-square is a continuous distribution,
however, its application is not limited to continuous data. In fact it is the most important
distribution used for the evaluation of discrete or categorical data, for example, the classification
of experimental units as dead or alive, sick or healthy, white, green or blue, scores of 1 to 5, etc.

11.1 Goodness of fit for discrete distributions

Goodness of fit involves a comparison of the frequency observed in the sample with the
expected frequency based on some theoretical model. If the differences between the observed
and the expected frequencies are so great that they are unlikely to be due to chance alone, we
conclude that the sample is not taken from the population that was used to calculate the expected
frequencies. Suppose we have k observed frequencies, O1, O2, ..., Ok and their corresponding
expected frequencies, E1, E2, ..., Ek, then the expression

k
χ = Σ ( 0i − E i ) 2 / E i
2

i=1

is approximately χ2 distributed with k-1 degrees of freedom. The closer the agreement between
the observed and expected values, the smaller will be the value of χ2, and a value of zero
indicates perfect agreement. Calculated values of χ2 exceeding those in Appendix Table A-5
indicate that, at the corresponding probability level, there is significant disagreement between
observed and expected values.

Chi-square, defined in terms of observed and expected frequencies, becomes a discrete

variable and can only assume certain, but not all, non-negative values. As the number of classes
(k) increases, χ2 as defined above approaches a continuous variable. When frequencies for only
two classes are involved, a correction must be made for the non-continuity, and the observed χ2
is adjusted as,

2
( 0i − E i − 1 / 2 )
2

adj. χ = Σ
2
with df = 1
Ei
i=1

In adjusted χ2 the absolute value of each of the differences between observed and expected
frequencies is reduced by 1/2 before being squared. The correction 1/2 is applicable only in the
case of one degree of freedom (k=2). For more degrees of freedom the corrections are more
complicated and are not generally used, but for 1 df the adjusted value of χ2 is always
appropriate.
An example in the use of adjusted χ2

It is hypothesized that blood groups are inherited in a simple Mendelian manner so that a
cross of parents both of whom have AB blood should give 3/4 type AB or AA children and 1/4
type BB children (theoretical model). Suppose that among 400 children from such parents, 292
are of type AB or AA. Does the observation conform to the theoretical hypothesis?

Blood Type: AB or AA BB
____________________________________________________

observed 292 108

frequency

expected 300 100

frequency
____________________________________________________

( 292 − 300 − .5) 2 ( 108 − 100 −.5) 2

adj. χ =
2
+
300 100
= 0.1875 + 0.5625 = 0.75

Referring to Appendix Table A-5 with 1 df, we find that the observed χ2 is not
significant (p < 0.10).

Thus the observed frequencies of the sample support the hypothetical ratio.

χ2 with 3 or more classes.

In a cross between ivory and red snapdragons the following counts were observed in the
F2 generation.

Phenotype Number of plants

Red 20
Pink 55
Ivory 25

100

On the basis of these data, can segregation be assumed to occur in the simple Mendelian ratio of
1:2:1?
 __________________________________________
Color Red Pink Ivory
__________________________________________

Observed 20 55 25
frequency

Expected 25 50 25
frequency
__________________________________________

( 20 − 25) 2 (55 − 50) 2 ( 25 − 25) 2

χ2 = + +
25 50 25
= 1.0 + 0.5 + 0.0 = 1.5

Note since this χ2 has 2 df, the adjustment for noncontinuity is not necessary. Again, the
calculated χ2 is much smaller than the tabular χ2 at the 10% level. Therefore we conclude that
the observed color frequencies could conform with Mendelian ratio.

11.2 Goodness of fit for continuous distributions

It is often of interest to know whether a given set of data approximates a continuous

distribution such as the normal or chi-square.

To illustrate the procedure, we will use the % sucrose data presented in Table 2-1,
summarized in Table 2-2, and graphed in Figure 2-1. We will test to see if these data can be
considered to be normally distributed. To compute the expected frequencies for each class
interval, we need to determine the probability associated with each interval. This procedure is
summarized in Table 11-1 and the computational steps follow the table.
Table 11-1. Observed and expected frequencies to test the goodness of fit of percent sucrose
values to a normal distribution.

Yi Oi Est.Z Ei Contri-
Mid- End- Obs. Y− Y Cum. Int. Exp. bution
Class point point freq. S prob. prob. freq. to χ2
1 4.8 5.6 1 -2.54 0.0055 0.0055 0.6 0.27
2 63 7.1 4 -1.96 0.0250 0.0195 2.0 2.00
3 7.8 8.6 4 -1.38 0.0838 0.0588 5.9 0.61
4 9.3 10.1 13 -0.81 0.2090 0.1252 12.5 0.02
5 10.8 11.6 10 -0.23 0.4090 0.2000 20.0 5.00
6 12.3 13.1 24 0.35 0.6368 0.2278 22.8 0.06
7 13.8 14.6 23 0.92 0.8186 0.1818 18.2 1.27
8 15.3 16.1 17 1.50 0.9332 0.1146 11.5 2.63
9 16.8 17.6 4 2.08 0.9812 0.0480 4.8 0.13

Y = 12.2, S = 2.6, χ2 = 11.99 with 9 - 3 = 6 df

1. Columns 1 through 4 are recorded from the frequency table, Table 2-2.
2. Standardize the class interval end points, (Y - Y )/S, where Y = 12.2 and S = 2.6 as
previously calculated.

3. Determine the cumulative probability for each standardized value from Appendix Table
A-4. For example,

p(Z < -2.54) = p(Z > 2.54) = 0.0055

p(Z < 0.35) = 1 p p(Z > 0.35) = 1 - 0.3632 = 0.6368
p(Z < 2.08) = 1 p p(Z > 2.08) = 1 - 0.0188 = 0.9812

4. Calculate the probability for each class interval, e.g., for class 2, the interval probability =
0.0250 - 0.0055 = 0.0195.

5. The expected frequency for each class is calculated by multiplying the interval
probability by the sample size, e.g., for class 2, the expected frequency = (0.0195) • (100)
= 1.95 ~ 2.00.

6. The contribution of each class to the overall 2 is equal to

(Observed frequency - expected frequency)2 = (O - E)2

expected frequency E

7. The calculated χ2 is the sum of each class contribution

χ2 = 0.27 + 2.00 + ... + 0.13 = 11.99

8. The degrees of freedom for χ2 depends on the number of parameters that must be
estimated for computing the expected frequencies. In this case, we have 9 classes,
therefore there are 8 df for classes. This is further reduced by a degree of freedom for
mean and a degree of freedom for standard deviation. Thus, there are 6 df for the
calculated χ2 .

For Appendix Table A-5, χ2 0.05,6 = 12.592. Although the calculated χ2 is not
significant at the 5% level, it is nearly so. Therefore we cannot be too sure that the data
of Figure 2-1 are normally distributed. However, we can conclude that the data are near
enough to being normally distributed to have no effect on the AOV procedures we are
using in the evaluation of this variable.

11.3 Contingency Tables

A closely related application of a χ2 distribution is the test of independence, also known

as Pearson's test for association. This test is very similar to the test of goodness of fit and some
people prefer to treat them as the same test with minor variation. In this section, we are
concerned with the hypothesis of statistical independence between two variables, each of which
is classified into a number of categories or attributes.

Suppose a group of persons or set of objects is classified according to two criteria of

classification, one criterion being entered in rows and the other in column. This two-way table is
called a contingency table. If there are j rows and k columns, the table is known as a j x k table.
From such a table we are interested in determining whether a relationship exists between the two
criteria of classification or if they are independent. For a j x k table there are (j-1)(k-1) degrees
of freedom.

A 2x2 contingency table

The following data show the effect of a certain type of fumigation on fruit spoilage.

Spoiled Unspoiled Totals

Unfumigated 8 16 24
Fumigated 2 14 16

Totals 10 30 40

Does the amount of fruit depend upon whether it has been fumigated?

In any contingency table, we set up the hypothesis that the two criteria of classification
are independent. The marginal totals are accepted as part of the hypothesis. For a population in
which the distribution in the classes is shown by the marginal totals and the classes are
independent, we are asking what proportion of a large series of samples similar to the one under
consideration will deviate as much or more from the theoretical as the one observed. On the
basis of the marginal total the expected entry in the upper left-hand corner would be (24) •
(10)/40 = 6. After this entry or any other has been calculated, all the remaining entries can be
obtained by subtraction from the marginal totals. Since only one value need be calculated, for a
2 x 2 table, there is only one degree of freedom. This checks with the (j - 1) (k - 1) degrees of
freedom for a jxk table since in this case j = k = 2. The expected (theoretical) values are given
below:

Spoiled Unspoiled Totals

Unfumigated 6 18 24
Fumigated 4 12 16
Totals 10 30 40

Therefore,

( 8 − 6 − 0.5) 2 ( 16 − 18 − 0.5) 2
χ2 = +
6 18
( 4 − 2 − 0.5) 2
( 12 − 14 − 0.5) 2
adjusted + +
4 12
2 2 2
(15.) (15
.) (15
.) . )2
(15
= + + +
6 18 4 12
= 0.27 + 0.13 + 0.56 + 0.19
= 1.25

Note the use of the correction for non-continuity for 1 df. From Appendix Table A-5, χ2 0.05,1
= 3.84. Since χ2 = 1.25 < 3.84, there is no reason to reject the null hypothesis of independence.
It appears, therefore, that fumigation has had no significant effect in reducing spoilage.

A 2x3 contingency table

The following are tabulated data on 82 strains of oats divided into 2 groups according to the
presence or absence of awns, and into 3 groups according to yield. Do these data permit the
conclusion that more of the awned strains occur in the highest yielding classes than do awnless
strains?

Yield Class (weight in grams)

151-200 201-250 251-325
Awned 6 7 21
Awnless 18 21 9
The χ2 is calculated as follows:

Yield Class (expected values in parentheses)

151-200 201-250 251-325 Total
Awned 6(10) 7(11.6) 21(12.4) 34
Awnless 18(14) 21(16.4) 9(17.6) 48
Totals 24 28 30 82

( 6 − 10) 2 (18 − 14 ) 2 ( 7 − 11.6) 2 ( 21 − 16.4 ) 2

χ =
2
+ + +
10 14 11.6 16.4
( 21 − 12.4 ) 2
( 9 − 17.6) 2
+ +
12.4 17.6
16 16 2116 . 2116
. 70.56 70.56
= + + + + +
10 14 11.6 16.4 12.4 17.6

= 1.60 + 1.14 + 1.82 + 1.29 + 5.69 + 4.01

= 15.55

Since χ2 = 15.55 > χ 20.05, 2 = 5.99 , the observed values are not distributed as expected on the basis
of the marginal totals, and the awned strains do occur in the highest yielding classes more
frequently.
 SUMMARY

1. The chi-square distribution can be used to test the goodness of fit of data to a
hypothesized model.

( 0i − E i ) 2
χ2 = Σ
Ei

Where Oi is the observed frequency and Ei is the expected frequency from the
hypothesized model. Assume observations are classified into k frequency classes, then

a) χ2 has k-1 df if no parameter of the hypothesized model needs to be

estimated from the data.

b) χ2 has k-1-n df in n parameters of the model are to be estimated from the

data.

2. The chi-square distribution can also be used to test the hypothesis of statistical
independence between two variables, each of which is classified into a number of
categories or attributes. The two-way table of the classified frequencies is called a
contingency table.

( 0ij − E ij ) 2
χ2 = Σ
E ij

where Oij is the observed frequency in the ith row and the jth column, with expected
frequency Eij. If there are j rows and k columns, the χ2 has (j - 1) (k - 1) df.

3. In case the calculated 2 has 1 df, the above formula needs to be modified as

( 0i − E i − 0.5) 2
adj. χ 2 = Σ
Ei
EXERCISES

1. Suppose 50 years rainfall data are summarized in the table below. Test whether the data
approximate a normal distribution.

Midpoint
Class (inches) Frequency Endpoint
1 8 3 9.5
2 11 9 12.5
3 14 14 15.5
4 17 10 18.5
5 20 7 21.5
6 23 0 24.5
7 26 4 27.5
8 29 2 30.5
9 32 1 33.5

Y = 16.5 and S = 5.4

2. Of 64 offspring of a certain cross of guinea pigs, 34 are red, 10 are black and 20 are
white. According to the genetic model, these numbers should be in the ratio of 9:3:4.
Are the data consistent with the model?

3. In an experiment involving the crossing of two hybrids of a species of flower the results
shown below are observed. Are these results consistent with the expected population
9:3:3:1?

Magenta Flower Magenta Flower Red Flower Red Flower

Green Stigma Red Stigma Green Stigma Red Stigma
120 49 36 12

4. It was hypothesized that the F2 generation of a particular barley cultivar would show a 9
hooded to 3 long-awned to 4 short-awned ratio. The observed data showed 348 hooded to
115 long-awned to 157 short-awned individuals. Test the hypothesis.

5.Suppose in a particular orange growing area of California that on the average 25% of the
oranges are graded as best grade, 40% are placed in the above average grade, 25% are
graded as average and 10% are graded as poor. A random sample of 500 bushels from
one orchard in the region yields 100 bushels of best grade, 300 bushels of above average
 grade, 50 bushels of average grade and 50 bushels of poor grade oranges. Is the quality
of oranges from this orchard representative of this area?
(χ2 = 167.5)

6. A study was conducted to establish whether there is a difference in susceptibility to

mildew between two types of pasture grass. The following data were obtained:

Type Mildewed Not Mildewed Total

A 107 289 396
B 291 81 372
Total 398 370 768

Test the hypothesis that there is no difference between the two grasses in their
susceptibility to mildew. (χ2 = 199.40)

7. Two barley fields containing two different varieties of barley were examined for rust.
The results are presented below.

Variety Rust No Rust Total

1 470 30 500
2 436 64 500
Total 906 94 1000

Is there evidence of heterogeneity in response to rust? (χ2 = 12.787)

8. During an epidemic of cholera the following data on the effectiveness of inoculation as a

means of preventing the disease were obtained.

Not Attacked Attacked

Inoculated 192 4
Not inoculated 113 34

Do these data indicate the effectiveness of the inoculation on the basis of the 1% level of
significance? (Yes, since adj. χ2 = 35.79)
9. Twenty-two animals are suffering from a disease, the severity of which is about the same
in each case. In order to test the therapeutic value of a serum, it is administered to 10 of
the animals; 12 remain uninoculated as a control. The results are shown below.

Recovered Died
Inoculated 7 3
Not inoculated 3 9
 Has inoculation been effective? (5% level) (No, since adj. χ2 = 2.82)

10. It is suspected that different combinations of temperature and humidity affect the number
of defective articles produced in a certain workroom. Do the following data confirm this
suspicion? (5% level) (Yes, since adj. χ2 = 5.95)

Humidity
Low High
Low 10 5
Temperature
High 4 16

11. A plant breeder wants to know if the sterility of rice is a genetic problem. Samples were
taken from a large field study of 400 plots and the sterility of each plot was rated as
follows:

Genotypes
Sterility A B C D
No problem 20 15 12 10
Moderate 70 60 80 50
Severe 10 25 8 40

Test the hypothesis that the severity of sterility is independent of genetic make-up or
genotype. (χ2 = 43.807)

12. An alfalfa breeder conducted a study on inheritance of resistance to anthracnose. The

following data were obtained:

Growth Habit Resistant Not Resistant Total

Standard 55 45 100
Intermediate
standard 70 30 100
Intermediate
alpha 82 18 100
Alpha 88 12 100
Total 298 102 400

Can he conclude that resistance to crown rot and growth habit are independent?
( χ2 = 33.636)
13. A pastry chef wishes to determine whether the proportion of unsatisfactory bear claws is
affected by oven temperature. He has baked batches of them at 350 , 400 , and 450 , and
has obtained the following results:

350 400 450 Total

Number satisfactory 132 128 111 371
Number unsatisfactory 14 17 35 66
Totals 146 145 146 437

If the chef wishes to test the null hypothesis of identical proportions at an = 0.05
significance level, what conclusion should he reach? (χ2 = 13.712)