I have a matched data set with 10,000 cases and 20,000 controls. Cases are defined as such due to a diagnosis of COPD (Chronic Obstructive Pulmonary Disease - a lung disease caused by smoking). Controls are healthy controls, matched for age and sex and region of residence. Cases and controls are not matched for socioeconomic variables, although I have that data as well. I have no other data for controls, although variables are plenty for cases. Mathching was done in the design phase of the study, and not during the analyses. to sum up, I have 10,000 cases with COPD and each individual has been matched to two age and sex matched controls.

AIMS Estimate the excess risk of developing dementia among cases, as compared to controls.


• I Use R; the "survival" and "RMS" packages.

• I plan to use Cox proportional hazards regression for relative risks and some measure of incidence to estimate absolute risks. I have the following questions:

(1) should I take the matching property into account in the Cox regressions? If that is the case, how do You recommend me doing this? Should I use the cluster()-function? the strata()-function? Or could I just ignore the matching?

(2) When calculating the incidence rate (and incidence rate ratio) do I have to adjust anything; I'm asking because mean age and proportion of women is exactly the same in cases and controls (for obvious reasons). Should I go for Poisson based incidence rates? Any other methods to consider?

I'd be very greateful for any advice on this.

  • 1
    $\begingroup$ Did the matching process involve discarding any patient data that were already collected? If so this is not a good application of matching and you should go back to the original larger cohort and do covariate adjustment without matching in my opinion. $\endgroup$ Nov 23, 2014 at 12:52
  • $\begingroup$ Hi prof Harrell, No it did not involve such a process. We haven't discarded any data actually. Matched controls are only used for one case. Controls have been selected from the general population, via registry-based procedures. I (although I'm not a statician) believe that the matching procedure has not introduced bias. I'm confused about whether and how to use the matching property in the final analyses. $\endgroup$
    – Heala45
    Nov 23, 2014 at 13:08
  • 1
    $\begingroup$ I think you could use a stratified Cox model, in which a distinct underlying hazard function is allowed for each stratum (in this case each matched set). I'm not sure though. $\endgroup$ Nov 24, 2014 at 8:31
  • $\begingroup$ You may be able to ignore the matching and do a regular covariable adjustment. But see doi:10.1002/sim.5879, or citeulike.org/user/harrelfe/article/12464813 to see additional notes. $\endgroup$ Nov 24, 2014 at 12:55


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.