When we look for an association or a treatment effect, ideally we need to look at whether the effect is more or less the same across different subgroups by sex, ethnicity, or other characteristics. This is commonly called in a jargon form as ‘subgroup analysis‘.
There could be two types of practice.
What I used to do: doing subgroup analysis in subgroups
For example, in my imaginative study of 500 individuals, if I find an analysis between some type of socialising activity is good for dementia, say relative risk of 0.8. It means participating in this socialising activity provides 20% protection against having dementia. And no, this is just an example and only occurs in my dream).
Now I would like to see whether the relative risk is more or less the same across different types of people. Would it be the same by sex (male or female), by age, by ethnicity, etc.? This question is also important to address diversity and health equality, because a lot of time we find a treatment only benefits a subgroup of people.
If I would like to see whether the association by sex, previously I would just take male individuals (say 240 of 500 individuals were male) in the study and see what the relative risk was, and then take only female individuals (n=260) and see what the relative risk was. If they were, for example, 0.75 and 0.85, I would look at the confidence intervals of them and see whether the difference between 0.75 and 0.85 could be reasonably due to randomness and perhaps not a real meaningful difference.
What I tend to do now: using a effect modification term to do subgroup analysis
Wherever I can, now I no longer separate them into male and female when estimating the relative risk, but just test whether there is an effect modification (i.e. an interaction). For example, in the original analysis (n=500) dementia was predicted by social activity participation, sex, age, and other factors. When I want to see whether there is a sex difference in effect of social activity participation (i.e. subgroup analysis by sex), I would still use the 500 individuals, and dementia was predicted by social activity participation, sex, an effect modification term between social activity participation and sex (i.e. social activity participation X sex), age, and other factors.
From that prediction models I should get a few things: the effect in male individuals, the effect in female individuals, and whether the effect modification was of some significant interest (or worry).
Peace of mind
I now feel more comfortable in this way because the analysis is more similar to the original analysis in a few perspectives. All individuals are contributing to the analysis, and because of this there would be (almost) the same proportions of other factors in the model compared to the original analysis. This way effectively puts a constraint so that the effects (i.e. coefficients) of other factors, such as age, are the same between male and female individuals. If you do them separately, then the effects would not be the same in male and female analysis. Now I can also be more sure the effect difference in male and female individuals, if there is, is not because the 240 male individuals are older, but because they are male per se.
Having said all this, I would not really object either way of practice, and certainly would not frown to those who do things differently. My attitude is when we say subgroup analysis, we are thinking of it as a screening test, and so any finding after that should be taken more seriously. Making a ‘conclusion’ from subgroup analyses is rarely my attention.
But this is probably another topic beyond today.
