Lecture #8

advertisement
Action Research
Data Manipulation and Crosstabs
INFO 515
Glenn Booker
INFO 515
Lecture #8
1
Parametric vs. Nonparametric
Statistical tests fall into two broad
categories – parametric & nonparametric
 Parametric methods




INFO 515
Require data at higher levels of measurement interval and/or ratio scales
Are more mathematically powerful than
nonparametric statistics
But often require more assumptions about the
data, such as having a normal distribution, or
equal variances
Lecture #8
2
Parametric vs. Nonparametric

Nonparametric methods




Use nominal or ordinal scale data
Still allows us to test for a relationship,
and its strength and direction (direction
only if ordinal)
Often has easier prerequisites for being
tested (e.g. no distribution limits)
Ratio or interval scale data may be
recoded to become nominal or ordinal
data, and hence be used with
nonparametric tests
INFO 515
Lecture #8
3
Significance and Association

… are useful for inferring population
values from samples (inferential statistics)



Significance establishes whether chance
can be ruled out as the most likely explanation
of differences
Association shows the nature, strength, and/or
direction of the relationship between two (or
among three or more) variables
Need to show significance before
association is meaningful
INFO 515
Lecture #8
4
Common Tests of Significance

We’ve been introduced to three common
tests of significance:




z test (large samples of ratio or interval data)
t test (small samples of ratio or interval data)
F test (ANOVA)
Shortly we’ll explore a fourth one

Pearson’s chi-square 2 (used for nominal or
ordinal scale data)
{ is the Greek letter chi, pronounced ‘kye’, rhymes with ‘rye’}
INFO 515
Lecture #8
5
Common Measures of Association

Association measures often range in value
from -1 to +1 (but not always!)


Absence of association between variables
generally means a result of 0
Examples



Pearson’s r (for interval or ratio scale data)
Yule’s Q (ordinal data in a 2x2 table)
Gamma (ordinal – more than 2x2 table)
{A “2x2” table has 2 rows and 2 columns of data.}
INFO 515
Lecture #8
6
Common Measures of Association

Notice these are all for nominal scale data





INFO 515
Phi (, ‘fee’) (nominal data in a 2x2 table)
Contingency Coefficient (nominal – table larger
than 2x2)
Cramer’s V (nominal - larger than 2x2)
Lambda (l) - nominal data
Eta () – nominal data
Lecture #8
7
Significance and Association
Tests of significance and measures of
association are often used together
 But you can have statistical significance
without having association

INFO 515
Lecture #8
8
Significance and Association Examples
Ratio data: You might use F to determine
if there is a significant relationship, then
use ‘r’ from a regression to measure its
strength
 Ordinal data: You might run a chisquare to determine statistical significance
in the frequencies of two variables, and
then run a Yule’s Q to show the
relationship between the variables

INFO 515
Lecture #8
9
Crosstabs

Brief digression to introduce crosstabs
before discussing non-parametric methods


INFO 515
Crosstabs are a table, often used to display
data, sorted by two nominal or ordinal
variables at once, to study the relationship
between variables that have a small number
of possible answers each
Generally contains basic descriptive statistics,
such as frequency counts and percentages
Lecture #8
10
Crosstabs

Used to check the distribution of data, and
as a foundation for more complex tests


Look for gaps or sparse data (little or no
contribution to the data set)
Rule of thumb - put independent variable
in the columns and dependent variable in
the rows
INFO 515
Lecture #8
11
Percentages

Can show both column and row
percentages in crosstabs, rather than just
frequency counts (or show both counts
and percentages)


Make sure percentages add to 100%!
Raw frequency counts of variables don’t
always provide an accurate picture

INFO 515
Unequal numbers of subjects in groups (N)
might make the numbers appear skewed
Lecture #8
12
Crosstabs Example

Open data set “GSS91 political.sav”



Use Analyze / Descriptive Statistics /
Crosstabs...
Set the Row(s) as “region”, and the
Column(s) as “relig”
Note the default scope of an SPSS
crosstab is to show frequency Counts,
with row and column totals
INFO 515
Lecture #8
13
Crosstabs Example
REGION OF INTERVIEW * RS RELIGIOUS PREFERENCE Crosstabulation
Count
RS RELIGIOUS PREFERENCE
REGION OF
INTERVIEW
Total
INFO 515
NEW ENGLAND
MIDDLE ATLANTIC
E. NOR. CENTRAL
W. NOR. CENTRAL
SOUTH ATLANTIC
E. SOU. CENTRAL
W. SOU. CENTRAL
MOUNTAIN
PACIFIC
PROTES
TANT
34
88
156
92
217
104
92
57
115
955
CATHOLIC
49
86
77
24
53
4
24
15
49
381
Lecture #8
JEWISH
0
8
4
0
6
0
1
1
12
32
NONE
6
13
17
3
14
3
7
5
33
101
OTHER
1
6
1
3
7
1
4
1
5
29
Total
90
201
255
122
297
112
128
79
214
1498
14
Crosstabs Example

Repeat the same example with
percentages selected under the “Cells…”
button to get detailed data in each cell




Percent within that region (Row)
Percent within that religious preference
(Column)
Percent of total data set (divide by Total N)
Gets a bit messy to show this much!
INFO 515
Lecture #8
15
Crosstabs
Example
REGION OF INTERVIEW * RS RELIGIOUS PREFERENCE Crosstabulation
RS RELIGIOUS PREFERENCE
REGION OF
INTERVIEW
NEW ENGLAND
MIDDLE ATLANTIC
E. NOR. CENTRAL
W. NOR. CENTRAL
SOUTH ATLANTIC
E. SOU. CENTRAL
W. SOU. CENTRAL
MOUNTAIN
PACIFIC
Total
INFO 515
Lecture #8
Count
% within REGION OF
INTERVIEW
% within RS RELIGIOUS
PREFERENCE
% of Total
Count
% within REGION OF
INTERVIEW
% within RS RELIGIOUS
PREFERENCE
% of Total
Count
% within REGION OF
INTERVIEW
% within RS RELIGIOUS
PREFERENCE
% of Total
Count
% within REGION OF
INTERVIEW
% within RS RELIGIOUS
PREFERENCE
% of Total
Count
% within REGION OF
INTERVIEW
% within RS RELIGIOUS
PREFERENCE
% of Total
Count
% within REGION OF
INTERVIEW
% within RS RELIGIOUS
PREFERENCE
% of Total
Count
% within REGION OF
INTERVIEW
% within RS RELIGIOUS
PREFERENCE
% of Total
Count
% within REGION OF
INTERVIEW
% within RS RELIGIOUS
PREFERENCE
% of Total
Count
% within REGION OF
INTERVIEW
% within RS RELIGIOUS
PREFERENCE
% of Total
Count
% within REGION OF
INTERVIEW
% within RS RELIGIOUS
PREFERENCE
% of Total
PROTES
TANT
34
CATHOLIC
49
JEWISH
0
37.8%
54.4%
3.6%
12.9%
2.3%
88
NONE
6
OTHER
1
Total
90
.0%
6.7%
1.1%
100.0%
.0%
5.9%
3.4%
6.0%
3.3%
86
.0%
8
.4%
13
.1%
6
6.0%
201
43.8%
42.8%
4.0%
6.5%
3.0%
100.0%
9.2%
22.6%
25.0%
12.9%
20.7%
13.4%
5.9%
156
5.7%
77
.5%
4
.9%
17
.4%
1
13.4%
255
61.2%
30.2%
1.6%
6.7%
.4%
100.0%
16.3%
20.2%
12.5%
16.8%
3.4%
17.0%
10.4%
92
5.1%
24
.3%
0
1.1%
3
.1%
3
17.0%
122
75.4%
19.7%
.0%
2.5%
2.5%
100.0%
9.6%
6.3%
.0%
3.0%
10.3%
8.1%
6.1%
217
1.6%
53
.0%
6
.2%
14
.2%
7
8.1%
297
73.1%
17.8%
2.0%
4.7%
2.4%
100.0%
22.7%
13.9%
18.8%
13.9%
24.1%
19.8%
14.5%
104
3.5%
4
.4%
0
.9%
3
.5%
1
19.8%
112
92.9%
3.6%
.0%
2.7%
.9%
100.0%
10.9%
1.0%
.0%
3.0%
3.4%
7.5%
6.9%
92
.3%
24
.0%
1
.2%
7
.1%
4
7.5%
128
71.9%
18.8%
.8%
5.5%
3.1%
100.0%
9.6%
6.3%
3.1%
6.9%
13.8%
8.5%
6.1%
57
1.6%
15
.1%
1
.5%
5
.3%
1
8.5%
79
72.2%
19.0%
1.3%
6.3%
1.3%
100.0%
6.0%
3.9%
3.1%
5.0%
3.4%
5.3%
3.8%
115
1.0%
49
.1%
12
.3%
33
.1%
5
5.3%
214
53.7%
22.9%
5.6%
15.4%
2.3%
100.0%
12.0%
12.9%
37.5%
32.7%
17.2%
14.3%
7.7%
955
3.3%
381
.8%
32
2.2%
101
.3%
29
14.3%
1498
63.8%
25.4%
2.1%
6.7%
1.9%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
100.0%
63.8%
25.4%
2.1%
6.7%
1.9%
100.0%
16
Recoding
An interval or ratio scaled variable, like
age or salary, may have too many distinct
values to use in a crosstab
 Recoding lets you combine values into a
single new variable -- also called
collapsing the codes
 Also helpful for creating histogram
variables (e.g. ranges of age or income)

INFO 515
Lecture #8
17
Recoding Example

Use Transform / Recode /
Into Different Variables…




INFO 515
Move “age” from the dropdown list for the
Numeric Variable
Define the new Output Variable to have Name
“agegroup” and Label “Age Group”
Click “Change” button to use “agegroup”
Click on “Old and New Values” button
Lecture #8
18
Recoding Example



For the Old Value, enter Range of 18 to 30
Assign this to a New Value of 1
Click on “Add”
Repeat to define ages 31-50 as agegroup
New Value 2, 51-75 as 3, and 76-200 as 4
 Click “Continue” and now a new variable
exists as defined

INFO 515
Lecture #8
19
Recoding
Example
INFO 515
Lecture #8
20
Recoding Example

Now generate a crosstab with “agegroup”
as columns, and “region” as the rows
REGION OF INTERVIEW * Age Group Crosstabulation
Count
1.00
REGION OF
INTERVIEW
Total
INFO 515
NEW ENGLAND
MIDDLE ATLANTIC
E. NOR. CENTRAL
W. NOR. CENTRAL
SOUTH ATLANTIC
E. SOU. CENTRAL
W. SOU. CENTRAL
MOUNTAIN
PACIFIC
25
36
56
29
66
15
38
22
48
335
Age Group
2.00
3.00
40
17
89
66
115
71
41
37
115
95
57
30
55
27
24
24
106
48
642
415
Lecture #8
4.00
8
12
13
15
21
10
8
9
12
108
Total
90
203
255
122
297
112
128
79
214
1500
21
Second Recoding Example
Prof. Yonker had a previous INFO515 class
surveyed for their height (in inches) and
desired salaries ($/yr)
 Rather than analyze ratio data with few
frequencies larger than one, she recoded:



INFO 515
Heights into: Dwarves for people below
average height, and Giants for those above
Desired salaries were recoded into Cheap and
Expensive, again below and above average
Lecture #8
22
Second Recoding Example

The resulting crosstab was like this:
New Salary * New Height Crosstabulation
New Salary
Cheap
Expensiv
Total
INFO 515
Count
% within New Height
Count
% within New Height
Count
% within New Height
Lecture #8
New Height
Dwarves
Giants
9
7
69.2%
53.8%
4
6
30.8%
46.2%
13
13
100.0%
100.0%
Total
16
61.5%
10
38.5%
26
100.0%
23
Pearson Chi Square Test

The Chi Square test measures how much
observed (actual) frequencies (fo) differ
from “expected” frequencies (fe)



INFO 515
Is a nonparametric test, a.k.a. the Goodness of
Fit statistic
Does not require assumptions about the shape
of the population distribution
Does not require variables be measured on an
interval or ratio scale
Lecture #8
24
Chi Square Concept

Chi Square test is like the ANOVA test



ANOVA proved whether there was a difference
among several means – proved that the means
are different from each other in some way
Chi square is trying to prove whether the
frequency distribution is different from a
random one – is there a significant difference
among frequencies?
Allows us to test for a relationship (but not
the strength or direction if there is one)
INFO 515
Lecture #8
25
Chi Square Null Hypothesis

Null hypothesis is that the frequencies in
cells are independent of each other (there
is no relationship among them)


Each case is independent of every other case;
that is, the value of the variable for one
individual does not influence the value for
another individual
Chi Square works better for small sample
sizes (< hundreds of samples)

INFO 515
WARNING: Almost any really large table will
have a significant chi square
Lecture #8
26
Assumptions for Chi Square

A random sample is the “expected” basis
for comparison


No zero values are allowed for the
observed frequency, fo


Each case can fall into only one cell
And no expected frequencies, fe, less than one
At least 80% of expected frequencies, fe,
should be greater than or equal to
five (≥5)
INFO 515
Lecture #8
27
Expected Frequency

The expected frequency for a cell is based
on the fraction of things which would fall
into it randomly, given the same general
row and column count proportions as the
actual data set


fe = (row total) * (column total) / N
So if 90 people live in New England, and 335
are in Age Group 1 from a total sample of
1500, then we would expect
fe = 90*335/1500 = 20.1 people in that cell
See slide 21
INFO 515
Lecture #8
28
Expected Frequency
So the general formula for the expected
frequency of a given cell is:
fe = (actual row total)*
(actual column total)/N
 Notice that this is NOT using the average
expected frequency for every cell

fe = N / [(# of rows)*(# of columns)]
INFO 515
Lecture #8
29
Calculating Chi Square

The Chi square value for each cell is the
observed frequency minus the expected
one, squared, divided by the expected
frequency
Chi square per cell = (fo-fe)2/fe


Sum this for all cells in the crosstab
For the cell on slide 28, the actual
frequency was 25, so Chi square for that
cell is = (25-20.1)2/20.1 = 1.195
Note: Chi square is always positive
INFO 515
Lecture #8
30
Calculating Chi Square

Page 36/37 of the Action Research
handout has an example of chi square
calculation, where
fo is the observed (actual) frequency
fe is the expected frequency

E.g. fe for the first cell is 20*30/60 = 10.0
Chi square for each cell is (fo-fe)2/fe
 Sum chi square for all cells in the table

No comments about fe fi fo fum! Is that clear?!?!
INFO 515
Lecture #8
31
Interpreting Chi Square

When the total Chi square is larger than
the critical value, reject the null
hypothesis


See Action Research handout page 42/43 for
critical Chi square (2) values
Look up critical value using the ‘df’ value,
which is based on the number of rows and
columns in the crosstab:
df = (#rows - 1)(#columns - 1)

INFO 515
For the example on slide 21,
df = (9-1)(4-1) = 8*3 = 24
Lecture #8
32
Interpreting Chi Square

Or you can be lazy and use the old
standby:

if the significance is less than 0.050, reject the
null hypothesis
if the significance is less than 0.050, reject the
null hypothesis
if the significance is less than 0.050, reject the
null hypothesis
if the significance is less than 0.050, reject the
null hypothesis
INFO 515
Lecture #8
33
Chi Square Example
Open data set “GSS91 political.sav”
 Use Analyze / Descriptive Statistics /
Crosstabs...
 Set the Row(s) as “region”, and the
Column(s) as “agegroup”
 Click on “Statistics…” and select the
“Chi-square” test

Notice we’re still using the Crosstab command!
INFO 515
Lecture #8
34
Chi Square Example
Chi-Square Tests
Pears on Chi-Square
Likelihood Ratio
Linear-by-Linear
Ass ociation
N of Valid Cases
Value
43.260a
43.557
1.062
24
24
Asymp. Sig.
(2-s ided)
.009
.009
1
.303
df
1500
a. 0 cells (.0%) have expected count les s than 5. The
minimum expected count is 5.69.
INFO 515
Lecture #8
35
Chi Square Example
Note that we correctly predicted the ‘df’
value of 24
 SPSS is ready to warn you if too many
cells expected a count below five, or had
expected counts below one
 The significance is below 0.050, indicating
we reject the null hypothesis
 The total Chi square for all cells is 43.260

INFO 515
Lecture #8
36
Chi Square Example
The critical Chi square value can be looked
up on page 42/43 of Yonker
 For df = 24, and significance level 0.050,
we get a critical Chi square of 36.415



Since the actual Chi square (43.260) is greater
than the critical value (36.415), reject the null
hypothesis
Chi square often shows significance falsely
for large sample sizes (hence the earlier
warning)
INFO 515
Lecture #8
37
Chi Square Example

What are the other tests? They don’t apply
here...



The Likelihood Ratio test is specifically for loglinear models
The Linear-by-Linear Association test is a
function of Pearson’s ‘r’, so it only applies to
interval or ratio scale variables
Notice that SPSS doesn’t realize those
tests don’t apply, and blindly presents
results for them…
INFO 515
Lecture #8
38
One-variable Chi square Test
To check only one variable’s distribution,
there is another way to run Chi square
 Null hypothesis is that the variable is
evenly distributed across all of its
categories
 Hence all expected frequencies are equal
for each category, unless you specify
otherwise


INFO 515
Expected range can also be specified
Lecture #8
39
Other Chi square Example

Use Analyze / Nonparametric Tests /
Chi-square…

NOT using the Crosstab command here
Add “region” to the Test Variable List
 Now df is the number of categories in
the variable, minus one



df = (# categories) - 1
Significance is interpreted the same
INFO 515
Lecture #8
40
Other Chi square Example
Test Statistics
Chi-Squarea
df
Asymp. Sig.
REGION OF
INTERVIEW
290.352
8
.000
a. 0 cells (.0%) have expected frequencies les s than
5. The minimum expected cell frequency is 166.7.
INFO 515
Lecture #8
41
Other Chi square Example
So in this case, the “region” variable has
nine categories, for a df of 9-1 = 8
 Critical Chi square for df = 8 is 15.507, so
the actual value of 290 shows these data
are not evenly distributed across regions
 Significance below 0.050 still, in keeping
with our fine long established tradition,
rejects the null hypothesis

INFO 515
Lecture #8
42
Whodunit?
The chi-square value by itself doesn’t tell
us which of the cells are major
contributors to the statistical significance
 We compute the standardized residual to
address that issue
 This hints at which cells contribute a lot to
the total chi square

INFO 515
Lecture #8
43
Residuals

The Residual is the Observed value minus
the Estimated value for some data point

Residual = fo - fe
If this variable is evenly distributed, the
Residuals should have a normal
distribution
 Plots of residuals are sometimes used to
check data normalcy (i.e. how normal is
this data’s distribution?)

INFO 515
Lecture #8
44
Standardized Residual
The Standardized Residual is the Residual
divided by the standard deviation of the
residuals
 When the absolute value of the
Standardized Residual for a cell is greater
than 2, you may conclude that it is a
major contributor to the overall chi-square
value


INFO 515
Analogous to the original t test, looking for
|t| > 2
Lecture #8
45
Standardized Residual
Extreme values of Standardized Residual
(e.g. minimum, maximum) can also help
identify extreme data points
 The meaning of residual is the same for
regression analysis, BTW, where residuals
are an optional output

INFO 515
Lecture #8
46
Standardized Residual Example
In the crosstab region-agegroup example
 Click “Cells…” and select Standardized
Residuals
 In this case, the worst cell is the
combination W. Nor. Central region Age Group 4, which produced a
standardized residual of 2.1

INFO 515
Lecture #8
47
Standardized Residual Example
REGION OF INTERVIEW * Age Group Crosstabulation
1.00
REGION OF
INTERVIEW
NEW ENGLAND
MIDDLE ATLANTIC
E. NOR. CENTRAL
W. NOR. CENTRAL
SOUTH ATLANTIC
E. SOU. CENTRAL
W. SOU. CENTRAL
MOUNTAIN
PACIFIC
Total
INFO 515
Count
Std. Res idual
Count
Std. Res idual
Count
Std. Res idual
Count
Std. Res idual
Count
Std. Res idual
Count
Std. Res idual
Count
Std. Res idual
Count
Std. Res idual
Count
Std. Res idual
Count
25
1.1
36
-1.4
56
-.1
29
.3
66
.0
15
-2.0
38
1.8
22
1.0
48
.0
335
Lecture #8
Age Group
2.00
3.00
40
17
.2
-1.6
89
66
.2
1.3
115
71
.6
.1
41
37
-1.6
.6
115
95
-1.1
1.4
57
30
1.3
-.2
55
27
.0
-1.4
24
24
-1.7
.5
106
48
1.5
-1.5
642
415
4.00
8
.6
12
-.7
13
-1.3
15
2.1
21
-.1
10
.7
8
-.4
9
1.4
12
-.9
108
Total
90
203
255
122
297
112
128
79
214
1500
48
Crosstab Statistics for 2x2 Table

2x2 tables appear so often that many
tests have been developed specifically
for them




INFO 515
Equality of proportions
McNemar Chi-square
Yates Correction
Fisher Exact Test
Lecture #8
49
Crosstab Statistics for 2x2 Table

Equality of proportions tests prove
whether the proportion of one variable is
the same as for two different values of
another variable


e.g. Do homeowners vote as often as renters?
McNemar Chi-square tests for frequencies
in a 2x2 table where samples are
dependent (such as pre-test and
post-test results)
INFO 515
Lecture #8
50
Crosstab Statistics for 2x2 Table

Yates Correction for Continuity
chi-square is refined for small observed
frequencies



fe = ( |fo-fe| - 0.5)/fe
Corrections are too conservative; don’t use!
Fisher Exact Test – assumes row/column
frequencies remain fixed, and computes
all possible tables; gives significance value
like Chi square
INFO 515
Lecture #8
51
Nominal Measures of Association

Are used to test if each measure is zero
(null hypothesis) using different scales




Phi
Cramer’s V
Contingency Coefficient
All three are zero iff Chi square is zero

INFO 515
“iff” is mathspeak for ‘if and only if’
Lecture #8
52
Nominal Measures of Association

The usual Significance criterion is used
for all three


If significance < 0.050, reject the null
hypothesis, hence the association is significant
Notice that direction is meaningless for
nominal variables, so only the strength
of an association can be determined
INFO 515
Lecture #8
53
Phi
For a 2x2 table, Phi and Cramer’s V are
equal to Pearson’s r
 Phi (φ) can be > 1, making it an unusual
measure of association


Phi = sqrt[ (Chi square) / N]
Phi = 0 means no association
 Phi near or over 1 means strong
association

INFO 515
Lecture #8
54
Cramer’s V
Cramer’s V ≤ 1
 V = sqrt[ Chi Square / (N*(k – 1) ]
where k is the smaller of the number of
columns or rows
 Is a better measure for tables larger than
2x2 instead of the Contingency Coefficient

INFO 515
Lecture #8
55
Contingency Coefficient
a.k.a. C or Pearson’s C or Pearson’s
Contingency Coefficient
 Most widely used measure based on
chi-square
 Requires only nominal data
 C has a value of 0 when there is no
association

INFO 515
Lecture #8
56
Contingency Coefficient
The max possible value
of C is the square root
of (the number of
columns minus 1,
divided by the number
of columns)
Cmax = sqrt( (#column
- 1) / #column)
 C = sqrt[ Chi Square /
(Chi Square + N) ]

Maximum Contingency Coefficient
1
0.95
Cmax
0.9
0.85
0.8
0.75
0.7
0
INFO 515
Lecture #8
2
4
6
8
10
12
14
Number of Columns
57
Download