Outbreak investigations must always be case-control studies?

Submitted by floX on Wed, 10/14/2009 - 09:57

Dear All,
Have just had a tricky discussion with a trainee who has shaken my belief and would be grateful for suggestions.
I've got a dataset from a workplace where there was norovirus illness associated with a sandwich company. We sent questionnaires to everyone at the workplace. These assessed exposure (ate sandwiches on one of two days involved; which sandwiches) and illness (GI symptoms).
In my thinking, the workforce constituted a retrospective cohort as they all had the opportunity to be exposed to the sandwiches, and our analysis starts with assessment of exposure rather than selecting cases and controls. Therefore I analysed using the very helpful cstable command as a cohort.
The trainee is arguing that it should be a case-control study. His arguments appear to be

  1. it is retrospective
  2. it is "cross-sectional" data
  3. the OR would be higher therefore give a better estimate of association.

Of course we could select controls randomly from our cohort (as in the computer skills module) but this is not the point.

Any comments from the panel?
Many thanks
best wishes
Chris

‹ X-mas Riddle 2010 "confidence intervals for cohort studies in outbreak settings are useless" ›

Comments

Esther's view

Submitted by floX on Wed, 10/14/2009 - 10:01.

In my humble opinion this is definitely a cohort analysis.
(*all of the following represents my humble opinion & should be interpreted with caution)

The difference between a (traditional) case-control and a cohort study is not due to the prospective/retrospective nature (as your trainee suggests), but due to the sampling of the ill to non-ill. If you choose all or a random sample of persons within the company, then you can calculate a measure of risk (e.g. all persons ill/total number of persons). In a (traditional) case-control study you over-sample cases, so you cannot get this measure of risk. Here, you are not over-sampling cases.

This is not really a cross-sectional study, as you have a measure of time, through recall of food consumption prior to onset of illness (if they got ill). Of course, it is perhaps prone to bias (as it's retrospective). (where did your trainee study? I've come across other people who had problems with retrospective cohort studies - perhaps it's linked to teaching at a given university?)

If you analyse this study as a case-control study, your odds ratio will overestimate the risk ratio. While you can do this - it really depends on what you want. A risk ratio is much easier to interpret. If you need to present this to the public, or the sandwich company a very high OR might be misleading. Your trainee says a higher OR gives a better measure of association. It does not. It just gives a higher number. Should chicken sandwiches suddenly be more dangerous in a case-control study rather than in a cohort study? Or do I misunderstand??

I don't know if this helps.
Good luck & have fun with the outbreak!
Cheers,
Esther

Florian's comment

Submitted by floX on Wed, 10/14/2009 - 10:05.

Hi Chris,
I agree with Esther on all points and would like to add:

  • If you start from disease status (outcome) and measure exposure, you do a case control study. Fix outcome, measure exposure = case control
  • If you start from exposure and measure outcome, you do a cohort study.
    fix exposure, measure outcome=cohort study.
  • time sequence of ascertainment of data in relation to time of onset of symptoms is irrelevant to cohort vs. case-control. It is relevant to validity of data (recall bias etc), though.
  • OR are bigger than RR and might point to the likely vehicle of infection. However, that should then be tested independently, ideally non-statistically (=food sampling, lab). As Esther said, what you want is the probability of getting disease after eating food X compared to non-eating. That probability is better approximated by the RR, not the OR.
  • outbreak investigations are special and do not really fit into "classic" epi-studies. n's are much smaller, time is very short and you usually do have a cause-effect model (food > ill), unlike a cross-sectional study with very limited cause-effect "weight" that only looks for associations.

Cheers,
Florian