Why must we use random effects?

[Karin Jensen]

We are analysing data which has baseline measurements, exposure (three levels) and an outcome measurement. We interested in the effects (if any) of the exposure levels. We are regressing the outcome measurements on the exposure groups and the baseline measurements, plus adjustments for age and sex.

We have been told that we should use random effects modelling because we have more than one measurement on each individual. Is this right? If not, what could we say to defend what we have done?

Thankyou,
Karin

[Philip Jones]

Karin, If each observation only has one baseline measurement and one outcome measurement, you do not need (or want) a random effects model. If you had multiple outcome measurements per person/unit, then if you did not explicitly model the intra-person clustering, your standard errors would be incorrect, since they assume independence. From your description, you mentioned "an outcome measurement" which implies only one per observation.

[Richard Williams]

Karin, it is not clear how your data are set up. Do you have, say, the same people measured at 10 different times, and each time you want to compare their current status with their original status? Or maybe current with immediate previous? Or what? If you have 10 observations of reach person, are you currently just treating them as though they were 10 different people?

I am guessing the same people were measured at multiple times, and if so Philip is right, your standard errors will be wrong if you just treat those records as if they were all from different people.

Anyway, I suspect you are getting good advice, but if we knew more about your data and what you have already been doing we might be able to better answer your question.

[Clyde Schechter]

One clarification to Philip's reply. If you have only a single baseline and a single follow-up measurement per person, then you do not need to do random effects modeling (though you could if you wanted to). But if you choose not to do random effects modeling, you must reshape your data to a single observation per person, containing both measurements, and then include the baseline measurement as a covariate in your model, with the follow-up measurement as the dependent variable.

[Karin Jensen]

Thanks for all your replies. Sorry that I was not clearer in my initial message.

Clyde that is *exactly* what we are doing. How could I reply to justify this to a referee who has mentioned random effects modelling?

[Philip Jones]

Why don't you refer them to the long history of ANCOVA (which is just a special case of regression)? ANCOVA has been used for decades to control for imbalances in continuous variables (prototypically one variable, but since it is just regression, easily extendible to multiple continuous and/or categorical variables).

[Philip Jones]

We are analysing data which has baseline measurements, exposure (three levels) and an outcome measurement. [SNIP] We have been told that we should use random effects modelling because we have more than one measurement on each individual. [SNIP]

Upon re-reading your original post, why don't you just clarify that the Reviewers' assertion (that is there is more than one measurement of the outcome variable) is just incorrect? You had previously stated, and now confirmed, that there is indeed only one measurement for the outcome variable. This should suffice!

[Richard Williams]

What difference would it make whether or not you used random effects? Would you learn anything different by using random effects?

[Karin Jensen]

Thanks for all your replies. I think I can make a sensible response now.