This question landed here after two migration requests from "Medical Sciences" and "Operations Research"

I'm reading several prospective studies published in top journals where a baseline modifiable exposure was associated with a final outcome. Just as an example might be this study.

The Authors aimed to evaluate whether oral hygiene behaviour can alleviate cardiovascular risk. For what I understand, the exposure persistency (oral hygiene behaviour) was not controlled during the study period, because any repeated assessment was performed during the follow-up. Basically, they associated the final outcome (10 years later) to the baseline exposure.

  1. Is that methodologically correct?
  2. In a prospective cohort study, is there any need to control the exposure during the study period?

1 Answer 1


Observational studies, even prospective ones like you cite, necessarily have limitations. The question is how much those get in the way of interpreting the results successfully.

It does seem that the study in question only used baseline covariate values, including those related to oral health, in their analysis. Instead of asking whether that was "methodologically correct," it might be better to ask how that should inform your interpretation of the results.

Note that some of the strongest associations of cardiovascular disease development were with number of dental caries and number of missing teeth at baseline. Those aren't "modifiable exposures" in the way that frequency of tooth brushing or dental visits per year might be--they either stay the same or get worse. Notably, the initial apparent association between baseline periodontal disease (which might be modifiable over time) and cardiovascular disease development disappeared with adjustment for other clinical characteristics.

With respect to caries and missing teeth, in an ideal world one might have followed their numbers over time and included them in the model. But how would you have modeled them? In survival analysis, it's the instantaneous value of a covariate that is used to estimate hazard ratios. Is it the number of teeth you have missing today that helps determine your risk of developing cardiovascular disease? Or is it a life-long history represented reasonably well by the number of missing teeth that you had when you entered the study? How best to include time-varying covariates in a survival model depends a lot on an understanding of the underlying subject matter. I'm not sure that any specific methodological rules can help with that.

There is, however, an open question about interpretation. The authors of that study and of an associated commentary argue that interventions in dental health might thus improve cardiovascular health. The danger in that interpretation is that baseline dental health might just serve as a proxy for individuals' tendencies to watch out for their health overall. If that's the case, dental-health interventions alone might not have any effect on cardiovascular health. But that's a matter of interpreting the results in a causal framework, not of methodology per se.

The issues you raise are very important, particularly in studies when time-varying exposures to therapies are inherently involved. For a comprehensive introduction to these issues, study for example the Causal Inference book by Hernán and Robins.


Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.