Frequency matching multiple variables

[Gabriel Bouz]

Hello,
I am conducting a database analysis involving an exposed cohort and an unexposed cohort. Essentially before running multivariate analysis, I would like to establish a 2:1 cohort match. However, instead of ensuring no significant difference between just age and gender, I would like to ensure no significant differences in other variables/comorbidities between the two cohorts. I’ve seen posts regarding matching 2:1 based on age and/gender but I was wondering if I could expand that to multiple variables. My total sample size is in the thousands. Thank you for your help.

[Matt Warkentin]

Hi Gabriel,

To answer your main question -- yes, you can definitely frequency match on many more variables than just age and sex.

My question to you, why are you frequency matching at all in a cohort study? If you are conducting secondary data analysis of cohort data, and you have all the data, frequency matching provides no statistical or logistical advantage. If you think there are variables that you need to adjust for, why not simply adjust for them in multivariable regression? It sounds like you have the sample size for including a number of regressors for adjustment. Even if you decided to frequency match, you would still need to adjust for these variables in any regression analysis. Not to mention, isn't exposure already determined based on whomever in the cohort naturally exposed themselves? Frequency matching is more commonly used in designing case-control studies.

[Gabriel Bouz]

Thanks Matt for your reply. The reason why I would like to frequency match is because I'd like both the exposed and unexposed cohorts to have statistically insignificant differences in comorbidities (for example, diabetes, age, gender, etc) so I can more strongly assess the impact of the exposure on the development of complications. Right now, I have two separate data files for each cohort. But I haven't figured out a code to find a sample of observations, ensuring a 2:1 control to case ratio, with no significant differences in comorbidities.