Question about analyzing matched case-control data

Question about analyzing matched case-control data

by Kristen Krysko -

Hello, 

When one matches on variables in a case control study then performs conditional logistic regression based on the pairs, I wanted to confirm whether taking the pairs into account in the analysis controls for the matched factors, or whether these should also be included as covariates for adjustment? 

For example, if one matches on age, sex and neighborhood in a case control study with an indicator variable for the pairs, then if you did conditional logistic regression, would you just include the "pair" variable as the group, or would you also have to include age, sex and neighbourhood as covariates to adjust for? i.e. does including the pair variable already adjust for the matching factors? 

i.e. clogit outcome exposure, group(pair)    

OR 

clogit outcome exposure age sex neighborhood, group(pair)  


Thank you, 

Kristen 

Re: Question about analyzing matched case-control data

by June Chan -

Hi Kristen, 

Based on what you describe, I would say it should be the first line, and including the "pair" variable handles that "adjustment":

clogit outcome exposure, group(pair)    


I also checked with one of the biostats faculty who agrees.

That being said, I'm mtg with Dr. Van Blarigan tmrw to confirm/clarify things on this topic, and will post/udpdate as needed. 


thanks!

June



Re: Question about analyzing matched case-control data

by Kristen Krysko -

Thank you very much.  It would be great if you could update on this once you speak with Dr. Van Blarigan as it will be helpful for future studies. 

Thank you,

Kristen

Re: Question about analyzing matched case-control data

by June Chan -

Hi Kristen and all, 

We had a good discussion today in class around the topics raised above in this trail. To answer Kristen's explicit question, there are no changes to the response above. Dr. Van Blarigan agrees with the answer posted.

In class, we talked about some general points about matching and "adjusting for matching in the analysis". 

In general, if you match, it should be accounted for in the analysis. This can happen in different ways. 

One can do an unconditional logistic regression analysis and include the matching factor variables into the regression statement so that the model "adjusts" for the matching factors. Typically, the matching factors are also then commented on in the footnotes of the table with the unconditional logistic regression (e.g., Table 2 in the Poynter paper).

One can also do conditional logistic regression, which inherently takes into consideration the matching factors by modeling each pair as their own strata and allowing each pair to have their own intercept (the stata command is shown in Kristen's line of code). By having the "group(pair)" term in the clogit statement, you are telling the model to stratify the analysis by the pair variable in your dataset (if you have 50 matched pairs, the pair variable would range from 1-50).  You do not need to additionally add terms for the matching factors to the model statement (hence why in Table 3, where they did conditional logistic regression, age and sex and clinic are not mentioned in the footnote... they have been addressed via the matching and conditional logistic model).

However, if you think that you did not match finely enough to address confounding, you might wish to add finer strata of the matching factor into the model, even with the conditional logistic model.  What was discussed in class was - say you matched on age in 5 yrs groups... you may still want to add age as a continuous variable in the model, to adjust for residual confounding. The example in table 3 is that you see "ethnicity" was still added to the model. While not stated explicitly, we believe this is bc they matched on "jewish/non-jewish" but in Table 1, there is still an association for subtypes of Jewish and risk of CRC... so presumably they wished to address this by including "ethnicity" in the conditional logistic model of Table 3... and while not stated in the footnote, we assume this was for finer strata of ethnicity (eg., jewish subtypes) than just jewish/non-jewish. 

Hope that helps!  

Good Luck!

JMC

Re: Question about analyzing matched case-control data

by Kristen Krysko -

Hi Dr. Chan, 

That's very helpful. Thank you!

Kristen