Covariate Balancing in post-hoc analysis

data analysis
, ,
1

I am doing post-hoc mixed effect modeling on data from about 16 studies. The gender balance is bad with males making up 0.85 of the dataset. Is the answer to just oversample the female group to make the data less skewed towards male.

I see some discussion about imbalances for RCTs but this work is being done after the fact.

2

This is more general than RCTs. It is not a good idea to molest data, which would bias inference.

1 Like
3

Thank you for the prompt response.

Is there any way of massaging the data to increase the power to detect a gender effect?

4

I am also interested in this. A simple example is that in observational studies to study whether diabetes (or other diseases) is related to myocardial infarction, it is common that the number of people with diabetes accounts for about 0.1 of the total population. But few people apply weighting in this situation.
If you want to condition (if sex = female) and only calculate the results for the female (or other classification with a small population), the sample size may not be enough. If you are concerned about the overall effect, can’t multivariate adjustment (interaction terms if necessarily) handle such data imbalance?

1 Like
5

Imbalance is irrelevant except that it lowers power/precision for sex effects or sex-specific outcome estimates. I’m not sure where the concern from imbalance originated. The data are the data, and since quotas were not used in obtaining the sample, you are just stuck with this.

It is also important to note that imbalance is not even the issue. The issue is the absolute number of females.

The only analytical action that this situation calls for would be if you had a major interest in assessing the magnitude of a sex by X interaction where X represents some other covariate or treatment. You would need to borrow information from males by specifying a Bayesian prior for the interaction effect.

2 Likes
6

I am fairly new and I was not able to effectively answer this inquiry and I got a bit spooked. Thanks for establishing that this is a non issue. Need to get back to plowing through your book.

2 Likes