I would like to estimate the association between exposure to an environmental contaminant and all-cause mortality. Unfortunately the dataset I can get access to is incomplete - exposure data are only available for approximately 5% of the entire cohort, and these 5% are reasonably random (essentially based on accessibility in a freezer - but not linked to collection date etc).

In an ideal world, I would like to conduct a survival analysis as I could do with other data - but I don't know whether this is really appropriate. Would it be more appropriate to use logistic regression here?

Edit: they are reasonable random as samples were treated after collection and processed in a way that was not linked to collection date/site.


If the reasons for missingness invalidate a survival regression model, they certainly invalidate a logistic regression as well.

If missingness still gives you a sample that is a simple-random-sample (SRS), then you can analyze those data as-is. The only issue might be small sample size: then collect more data. Weighting or imputation methods would only make a difference if you "knew something" about the remaining samples in the freezer.

It does concern me somewhat that "availability in a freezer" is listed as reason you don't think you might have a non-representative sample. If you "sampled" the average expiry of milk at the grocery store by picking the first cartons from the shelf, the estimate would be biased because of a FIFO system: the older milks are setup to be bought first. In general, anything preferentially sampled because of convenience, time, or space tends to be non-representative (meaning some methods must be applied to calibrate estimates toward a target population).

You can confirm this by actually performing an SRS of specimens from the freezer and ascertaining the collection date and also the exposure to determine if there is an association or otherwise.

  • $\begingroup$ Thanks - the way they were stored ensures that their position in the freezer is not linked to collection site/date. They also don't differ from the rest of the sample in main factors (age, sex, collection date). $\endgroup$
    – Gux
    Mar 18 '19 at 19:52

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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