Matched pairs and conditional logistic regression

Question

Hello,

I am trying to compare measurements of lymph node sizes between two groups.  One group is African American Women (n=55) which were matched 1:2 with Caucasian women (n=110) by age.  Each patient also has a corresponding lymph node measurement with both a left and right.  Although a simple t-test could be performed, I believe a conditional logistic regression would be the best to allow comparison between matched patients.  I have my data coded such that each matched group of 3 (1 AA and 2 Caucasian women) has a number and then each patient's race is coded either AA or C.  Should I include both group number and race in the model effects and then lymph node size under the Y in the fit model?  Additionally, how can I control for the left and right measurements on the same patient?

Any help is greatly appreciated.

Sincerely,

Lars

ljgrimm · Answer

Ah, you are certainly right.  I apologize for not being clear in my initial query.

* Our primary hypothesis is between the average of AA and C.  
* There should be no relationship between left and right on an AA vs C level, but there should be on the patient specific level.  For example, patient 1's left and right lymph nodes are likely similar in size.  
* We matched the women 1:2 because there were only a limited number of AA women but an overwhelming number of potential C women.  So to increase statistical power we did a 1:2 ratio insted of just a 1:1.

I hope that clears things up.

stephen_pearson · Answer

It depends on the exact research question you are trying to answer:

* How does your hypothesis relate to left and right?
* Why are the women matched 1:2?
* Is it the average difference between lymph nodes or the average of the two C versus the AA. You could concievably carryout anywhere from 1 (overall average) to 4 pair wise by Left/Right C1/C2 comparisons for each pair.

You need to clearly decide this before trying to analyze the data or you run the risk of data snooping (deciding the test after looking at the data can lead to picking a test which is most likely to give a result consistent with your hypothesis rather than being a test of your hypothesis).