Research
Methods 1: Statistics Problem-Sheet 5: Chi-Square:
1. A sample of 100 voters are asked which of four
candidates they would vote for in an election. The number supporting each
candidate is given below:
Higgins Reardon White Charlton
41 19 24 16
Do the data suggest that all candidates are equally
popular? [Chi-Square = 14.96, with 3 d.f.: p<0.05].
2. Children of three ages are asked to indicate their
preference for three photographs of adults. Do the data suggest that there is a
significant relationship between age and photograph preference? What is wrong
with this study? [Chi-Square = 29.6, with 4 d.f.: p<0.05].
Photograph:
A B C
Age of child: 5-6
years: 18 22 20
7-8 years: 2 28 40
9-10 years: 20 10 40
3. A study of conformity using the Asch paradigm involved
two conditions: one where one confederate supported the true judgement, and
another where no confederate gave the correct
response
Support: No Support:
Conform:
18 40
Not Conform: 32 10
Is there a significant difference between the
"support" and "no support" conditions in the frequency with
which individuals are likely to conform? [Chi-Square = 19.87, with 1 d.f.:
p<0.05. OR: Chi-Square = 18.1. See the comment at the end of this handout].
4. A researcher is interested in sex differences in
social play in cats. She observes sixteen kittens (eight male and eight
females), and records the sex of the participants in 500 play-bouts, taking a note of the sex of the
animal initiating a bout and the sex of the animal who responds to the
initiation. Are there sex-differences in the kittens' play? Perform a
Chi-Square test on these data. Is there anything wrong with doing this test?
[Chi-Square = 3.97, with 1 d.f.: not significant. OR: Chi-Square = 3.58. See the comment at the end of this handout].
respondent
male female
initiator of bout: male: 200 150
female: 100 50
5. We want to test whether short people differ with
respect to their leadership qualities (Genghis Khan, Adolf Hitler and Napoleon
were all stature-deprived, and how many midget MP's are there?) The following
table shows the frequencies with which 43 short people and 52 tall people were
categorised as "leaders", "followers" or as "unclassifiable".
Is there a relationship between height and leadership qualities? [Chi-Square =
10.71, with 2 d.f.: p<0.01].
6. The Government decide to find out how many people use
various kinds of transport to commute from Brighton to London each day. They go
to Victoria train station, and ask a random sample of 80 individuals what type
of transport they use. Here are the raw data, coded as follows: "1"
if the person uses a car; "2" if they use a train; "3" if
they walk; and "4" if they fly. The null hypothesis is that all four
forms of transport are used equally frequently by the general commuting public.
(a) Draw a frequency histogram of these data.
(b) Perform a Chi-Square Goodness of Fit test on these
data.
1,4,2,3,4,4,3,4,4,2 2,2,1,4,3,3,2,2,2,2
3,2,4,4,3,2,1,1,2,3 4,4,3,3,2,1,3,2,4,3
2,1,2,2,2,1,2,3,2,4 3,2,2,3,4,2,3,4,2,2
2,1,2,1,2,2,2,3,4,3 2,2,1,3,2,2,2,1,2,2
What problems are there with this
study?
Chi-Square
in the 1 d.f. case, and the use of Yates' Correction:
If you have only 1 d.f., as in the case of a 2x2
contingency table, some textbooks suggest that you apply what's known as
"Yates' Correction" to the Chi-Square formula. When the d.f. are very
small, (and you can't get much smaller than 1!), the Chi-Square sampling distribution
becomes increasingly distorted. As a consequence, the obtained value of
Chi-Square tends to overestimate the "true" discrepancy between the
observed and expected frequencies. Yates' correction is a simple bodge that
makes the formula produce a lower value of Chi-Square than it otherwise would
do. This makes the Chi-Square test more "conservative": i.e., it
makes it harder to get a statistically significant result. Here is the
"corrected" formula:
In English, this means you do the following:
(a) Get the absolute value (i.e., ignore the sign) of
each observed frequency minus its accompanying expected frequency. (The
vertical lines mean "ignore the sign of").
(b) From each of these absolute values, subtract 0.5.
(c) Divide each value in step (b) by the associated
expected frequency.
(d) Add together all of the results of step (c).
(e) Finally, look up the value of Chi-Square in the usual
way.
Statisticians are divided about the need for Yates'
correction: some say you should use it, others say it doesn't matter, and some
say it makes things worse! Consequently, the answers to the problems on this
sheet have been calculated both ways. The normal-type answers are what you
would get if you did not use Yates' correction, and the italic-type answers are
what you would get if you did. Either is acceptable, as long as you make it
clear whether or not you used Yates' correction. Remember - if you do use it,
it is only to be used when you have
just one degree of freedom.
First-year
Research Methods: Worked Solutions to Problem Sheet 5:
Question
1:
A Chi-Squared Goodness-of-Fit test is appropriate
here. The null hypothesis is that there is no preference for any of the
candidates: if this is so, we would expect roughly equal numbers of voters to
support each candidate. Our expected frequencies are therefore 100/4 = 25 per
candidate.
O |
41 |
19 |
24 |
16 |
E |
25 |
25 |
25 |
25 |
(O-E) |
16 |
-6 |
-1 |
-9 |
(O-E)2 |
256 |
36 |
1 |
81 |
(O-E)2 --------- E |
10.24 |
1.44 |
0.04 |
3.24 |
Adding together the last row
gives us our value of c2 :
(O - E)2
å -----------------
= 10.24+ 1.44 + 0.04 + 3.24 = 14.96,
with 4 - 1 = 3 degrees of freedom.
E
The critical value of Chi-Square for a 0.05 significance
level and 3 d.f. is 7.82. Our obtained Chi-Square value is bigger than this,
and so we conclude that our obtained value is unlikely to have occurred merely
by chance. In fact, our obtained value is bigger than the critical Chi-Square
value for the 0.01 significance level (13.28). In other words, it is possible
that our obtained Chi-Square value is due merely to chance, but highly
unlikely: a Chi-Square value as large as ours will occur by chance only about
once in a hundred trials. It seems more reasonable to conclude that our results
are not de to chance, and that the data do indeed suggest that voters do not
prefer the four candidates equally.
Question
2:
|
photograph: |
|
|||
age of child: |
A: |
B: |
C: |
row totals: |
|
5-6 years |
18 |
22 |
20 |
60 |
|
7-8 years |
2 |
28 |
40 |
70 |
|
9-10 years |
20 |
10 |
40 |
70 |
|
column totals: |
40 |
60 |
100 |
200 |
|
(a) Work out the row, column and grand totals (as
shown in the shaded parts of the table, above).
(b) Work out the expected frequencies, using the formula:
(row total * column total)
E = --------------------------------------
grand
total
For each cell of the above table, this gives us:
O: |
18 |
22 |
20 |
2 |
28 |
40 |
20 |
10 |
40 |
E: |
12 |
18 |
30 |
14 |
21 |
35 |
14 |
21 |
35 |
Next, work out (O - E):
(O-E): |
6 |
4 |
-10 |
-12 |
7 |
5 |
6 |
11 |
5 |
Square each of these, to get
(O - E)2 :
(O - E)2: |
36 |
16 |
100 |
144 |
49 |
25 |
36 |
121 |
25 |
Divide each of the above
numbers by E, to get (O - E)2
/ E:
(O - E)2 ---------- E |
3 |
0.89 |
3.33 |
10.29 |
2.33 |
0.71 |
2.57 |
5.76 |
0.71 |
Chi-squared is the sum of
these:
c2 = 29.60.
d.f. = (rows - 1) * (columns - 1) = 2 * 2 = 4.
The critical value of Chi-Square in the table for a 0.001
significance level and 4 d.f. is 18.46. Our obtained Chi-Square value is bigger
than this: therefore we have a Chi-Square value which is so large that it would
occur by chance only about once in a thousand times. It seems more reasonable
to accept the alternative hypothesis, that there is a significant relationship
between age of child and photograph preference.
Question
3:
Here we have a 2x2 contingency table. Chi-Square is
the appropriate test to use, but since we have 1 d.f., we will modify the
formula to include "Yates' correction for continuity".
|
support |
no support |
row totals: |
conform: |
18 |
40 |
58 |
not conform: |
32 |
10 |
42 |
column totals: |
50 |
50 |
100 |
(a) Calculate the row, column and grand totals.
(b) Calculate the expected frequency for each cell of the
table, by multiplying together the appropriate row and column totals and then
dividing by the grand total.
(c) Subtract each expected frequency from its associated
observed frequency; but then apply Yates' correction, by subtracting 0.5 from
the absolute value of each O-E value. (The vertical bars in the formula mean
"ignore any minus signs").
O: |
18 |
40 |
32 |
10 |
E: |
29 |
29 |
21 |
21 |
Next, work out (O - E):
(|O-E|- 0.5): |
10.5 |
10.5 |
10.5 |
10.5 |
Square each of these, to get
(O - E)2 :
(|O-E|- 0.5)2: |
110.25 |
110.25 |
110.25 |
110.25 |
Divide each of the above
numbers by E, to get (O - E)2
/ E:
(O - E)2 ----------- E |
3.80 |
3.80 |
5.25 |
5.25 |
Chi-squared is the sum of
these:
c2 = 18.10.
d.f. = (rows - 1) * (columns - 1) = 1 * 1 = 1.
Our obtained value of Chi-Squared is bigger than the
critical value of Chi-Squared for a 0.001 significance level. In other words,
there is less than a one in a thousand chance of obtaining a Chi-Square value
as big as our obtained one, merely by chance. Therefore we can conclude that
there is a significant difference between the "support" and "no
support" conditions, in terms of the frequency with which individuals
conformed.
Question
4:
|
respondent: |
|
||
initiator: |
male |
female |
row totals: |
|
male |
200 |
150 |
350 |
|
female |
100 |
50 |
150 |
|
column totals: |
300 |
200 |
500 |
|
(a) Calculate the marginal totals and expected
frequencies.
The expected frequencies are obtained by multiplying the
row total by the column total, and then dividing by the grand total. For
example, the expected frequency for male-male play is (350*300)/500 = 210; the
expected frequency for male-female play is (150*200)/500 = 60; and so on.
Expected frequencies are shown in brackets below:
|
respondent: |
|
|||
initiator: |
male |
female |
row totals: |
||
male |
200 (210) |
150 (140) |
350 |
||
female |
100 (90) |
50 (60) |
150 |
||
column totals: |
300 |
200 |
500 |
||
NB: the expected frequencies should add up to the same
Grand Total (500 in this case) as the observed frequencies!
(b) Since we have a 2x2 table (and
hence 1 d.f.) we will work out Chi-Square incorporating Yates' correction for
continuity. For each cell in the table, work out:
This gives us the following numbers:
|
respondent: |
|
||
initiator: |
male |
female |
||
male |
0.4298 |
0.6446 |
||
female |
1.0028 |
1.5042 |
||
(c) Adding these together gives us our value of
Chi-Square: 3.581, with 1 d.f. This
is smaller than the critical value of Chi-Square for 1 d.f. at the 0,.05
significance level, which is 3.841. We would therefore conclude that the
frequencies with which male and female animals play together in opposite-sex
and liked-sexed pairs do not differ from those that we would expect by chance.
In practice, there is a problem with
this analysis which invalidates the Chi-Square test: we have 500 observations
from only 16 animals, which means that each animal must have contributed more
than observation to the total. This means that the observations are not
independent - thus violating an important assumption for the use of the
Chi-Square test.
Question 5:
Expected frequencies are in
brackets:
|
height: |
|
|||
|
short |
tall |
row totals: |
||
leader: |
12 (19.92) |
32 (24.08) |
44 |
||
follower: |
22 (16.29) |
14 (19.71) |
36 |
||
unclassifiable: |
9 (6.79) |
6 (8.21) |
15 |
||
column totals: |
43 |
52 |
95 |
||
Chi-Square =
3.146 + 2.602 + 1.998 + 1.652 + 0.720 + 0.595 = 10.712, with 2 d.f.
10.712 is bigger than the tabulated
value of Chi-Square at the 0.01 significance level. We would conclude that
there seems to be a relationship between height and leadership qualities. Note
that we can only say that there is a relationship between our two variables,
not that once causes the other. There could be all kinds of explanations for
such a relationship.
Question 6:
Our 80 subjects fall into four
transport categories as follows:
type of transport: |
|||
car (1) |
train (2) |
walk (3) |
fly (4) |
11 |
36 |
18 |
15 |
Our
null hypothesis is that each form of transport is used equally frequently.
Therefore our expected frequencies are (80/4) = 20, 20, 20 and 20. The
appropriate analysis is a Chi-Squared Goodness-of-Fit test.
Take each observed frequency;
subtract 20; square the result; and then divide by 20. This gives 4.05, 12.80,
0.20 and 1.25. Adding these values together gives us our value of Chi-Square:
18.30, with 3 d.f., p<0.001. In other words, our obtained frequencies are
markedly different from those that we would expect to obtain by chance.
Strictly speaking, this is all we can say. However, "eyeballing" the
data suggests that more people travel by train than we would expect, and fewer
by car (well, these are fictional data!)
The problem with this study lies not
in the statistical analysis, but in the way that the data were obtained:
standing in a train station is likely to bias the sample in favour of people
travelling by train.