Anda di halaman 1dari 11

# Chapter 3: Three-Way Contingency Tables 1

## Chapter 3: Three-Way Contingency Tables

Example. Florida court cases with multiple murders,
197687 (Radelet & Pierce, Florida Law Review, 1991,
pp. 134).
Victims Defendants Death Penalty Percentage
Race Race Yes No Yes
White White 53 414 11.3
Black 11 37 22.9
Black White 0 16 0.0
Black 4 139 2.8
Total White 53 430 11.0
Black 15 176 7.9
We treat


 

 

## victims race as a control variable.

Chapter 3: Three-Way Contingency Tables 2

53 414 0 16
and are called partial tables.
11 37 4 139


## is controlled (held constant) in the partial table. Odds

ratios in the partial tables are called conditional odds ratios:
  
    
 
   
   

   
  
   
 !
   

## For black victims ,

 
Controlling for victims race, the odds of receiving the death
penalty were lower for white defendants than black
defendants.

## Combining the partial tables yields the - marginal table,

53 430
, which has
15 176

   "   " 
  #
 

  

## Ignoring victims race, odds of death penalty were higher for

white defendants than black defendants.
Chapter 3: Three-Way Contingency Tables 3

## Simpsons paradox occurs when all partial tables show

reverse association from that in the marginal table.
Cause? Note that
For each defendants race, death penalty much more
likely when victim is white.
Very strong assoc. between defendants and victims race
(whites kill whites, blacks kill blacks).
Moral: It can be very dangerous to collapse contingency
tables.
Chapter 3: Three-Way Contingency Tables 4

  
Def. In a  


association if
 
 
 
    
  
  

Example.
  

## smoker (yes,no), lung cancer (yes,no).

 
 

Homogen. Nonhomogen.
 

## age Assoc. Assoc. Indep.

 "  "
1.  45 
  "
2. 4565  
  "
3. 65  

  
 
Note. When conditional odds ratios are equal, then
 

    
 
so are the conditional and odds ratios.
Chapter 3: Three-Way Contingency Tables 5

## Testing Conditional Independence

 
In a   table, want to test
   
 
  
   "
H

## Let  be the sample count in row and column

of partial

 
   
   
table :       .
 
     

In partial table , we reduce data to a single count   by
conditioning on row and column marginal totals. Then, under
H ,

   
 
   
 

   
        
     


   
We expect
    
 
     


when
  
 
when 
  
 
  when  
Chapter 3: Three-Way Contingency Tables 6

## Summarize differences across strata using

Cochran-Mantel-Haenszel statistic 
   
   
 


CMH
 

Under H , CMH is approximately chi-squared with d.f. for
large samples. P-value is right-tail prob.
  

CMH test most appropriate when  has same direction
in each partial table.
Example. Death penalty data.
H : death penalty and def. race cond. indep., given victims
race.
  
  
 
  "
Recall    and 
  

to
each cell). So both odds ratios have assoc. in the same
direction (both less than 1).

    !

  

CMH df P-value

## Fairly strong evidence of an association.

How strong is the association?
Chapter 3: Three-Way Contingency Tables 7

## When the association seems similar in each partial table, we

summarize it by the Mantel-Haenszel estimator
    




  


 
 


     
MH

Example.
 


    
    
    
 

MH   
 

## Controlling for victims race, estimated odds of death penalty

for white defendants is 41% of estimated odds of death
penalty for black defendants.
     
From SAS, a 95% CI for MH is   . We can
 
conclude that MH  .
   
Note asymmetry of CI around  MH 


.
Chapter 3: Three-Way Contingency Tables 8

## Note. Under the assumption that

  
 

  
 


(homogeneous association),

 MH estimates the common odds ratio.

## SAS also provides the Breslow-Day statistic for testing

  
    
 
hypothesis H 
 

of
homogeneity of odds ratios. Its large sample dist. is
 
chi-squared with d.f.
  

  

## Example. . Breslow-Day stat. , df ,

   
   
 
 . It is plausible that
 

P-value .
CMH test extends to   tables. (See Section 7.3


## and SAS PROC FREQ), and there are analogous tests

using models.
For small samples, there is an exact test of conditional
independence (need special SAS PROC STATXACT)
(Section 7.3).
Chapter 3: Three-Way Contingency Tables 9

## CMH Analysis in SAS

data cmh;
input victim \$ defend \$ penalty \$ count @@;
cards;
w w yes 53 w w no 414
w b yes 11 w b no 37
b w yes 0 b w no 16
b b yes 4 b b no 139
;
proc freq order=data;
weight count;
tables victim*defend*penalty / cmh;
run;

## TABLE 1 OF DEFEND BY PENALTY

CONTROLLING FOR VICTIM=w

DEFEND PENALTY

Frequency|
Percent |
Row Pct |
Col Pct |yes |no | Total
---------+--------+--------+
w | 53 | 414 | 467
| 10.29 | 80.39 | 90.68
| 11.35 | 88.65 |
| 82.81 | 91.80 |
---------+--------+--------+
b | 11 | 37 | 48
| 2.14 | 7.18 | 9.32
| 22.92 | 77.08 |
| 17.19 | 8.20 |
---------+--------+--------+
Total 64 451 515
12.43 87.57 100.00
Chapter 3: Three-Way Contingency Tables 10

## TABLE 2 OF DEFEND BY PENALTY

CONTROLLING FOR VICTIM=b

DEFEND PENALTY

Frequency|
Percent |
Row Pct |
Col Pct |yes |no | Total
---------+--------+--------+
w | 0 | 16 | 16
| 0.00 | 10.06 | 10.06
| 0.00 | 100.00 |
| 0.00 | 10.32 |
---------+--------+--------+
b | 4 | 139 | 143
| 2.52 | 87.42 | 89.94
| 2.80 | 97.20 |
| 100.00 | 89.68 |
---------+--------+--------+
Total 4 155 159
2.52 97.48 100.00
Chapter 3: Three-Way Contingency Tables 11

## SUMMARY STATISTICS FOR DEFEND BY PENALTY

CONTROLLING FOR VICTIM

## Statistic Alternative Hypothesis DF Value Prob

------------------------------------------------------
3 General Association 1 5.796 0.016

## Estimates of the Common Relative Risk (Row1/Row2)

95%
Type of Study Method Value Confidence Bounds
----------------------------------------------------------
Case-Control Mantel-Haenszel 0.412 0.200 0.848
(Odds Ratio)