Hello,
I previously asked how to perform a binary logistic regression with two outcomes (response to treatment: yes=1 or no=0) measured at two different time points. I got a helpful response but now I wonder:
1) Would it be statistically wrong or just a different way of analyzing the data to perform the binary logistic regression for the two time points separately?
2) How would this affect the results? Would the results answer different questions? Would the separated or the simultaneous/combined analysis provide higher statistical power?
The basic research question is: Which predictor variables are good predictors for a response at time point 1, and which are good predictors for a response at time point 2?
Binary logistic regression for repeated measures data - separate vs combined analysis of datasets for two time points
-
Biochemist
- Posts: 37
- Joined: Mon Feb 19, 2024 9:44 am
Re: Binary logistic regression for repeated measures data - separate vs combined analysis of datasets for two time point
Hey @Biochemist,
a light answer.
A good starting point for identifying basic predictors might be to run a separate analysis, but for a more complete understanding a combined analysis with interaction terms would be recommended (especially if you believe the effect of the predictors may change in time).
Obviously, when choosing the approach it is also necessary to consider the sample size and the possible rarity of the outcome.
So your possible "Separate Analysis vs. Combined Analysis" wouldn't seem statistically wrong, but it's a different approach that has its tradeoffs.
Analyzing each time point separately may be valid for identifying specific predictors for each time point, but may not capture the impact of the predictors on response variation between the two points.
Therefore, if the separate analysis identifies independent predictors for the response at each time point, it may be less powerful if the sample size is limited (especially for rare outcomes).
Combined analysis identifies predictors that influence response at both time points and those that influence change over time (interaction effects), and may be more powerful, especially if there is a correlation between responses at the two time points.
Repeating myself
Since you want to identify specific predictors for each time point, separate analyzes might be a good first step.
However, if you want to get a more complete view, consider a combined analysis with interaction terms that will allow you to see whether the effect of a predictor on response differs between time points.
Cheers,
Maurizio
a light answer.
A good starting point for identifying basic predictors might be to run a separate analysis, but for a more complete understanding a combined analysis with interaction terms would be recommended (especially if you believe the effect of the predictors may change in time).
Obviously, when choosing the approach it is also necessary to consider the sample size and the possible rarity of the outcome.
So your possible "Separate Analysis vs. Combined Analysis" wouldn't seem statistically wrong, but it's a different approach that has its tradeoffs.
Analyzing each time point separately may be valid for identifying specific predictors for each time point, but may not capture the impact of the predictors on response variation between the two points.
Therefore, if the separate analysis identifies independent predictors for the response at each time point, it may be less powerful if the sample size is limited (especially for rare outcomes).
Combined analysis identifies predictors that influence response at both time points and those that influence change over time (interaction effects), and may be more powerful, especially if there is a correlation between responses at the two time points.
Repeating myself
Since you want to identify specific predictors for each time point, separate analyzes might be a good first step.
However, if you want to get a more complete view, consider a combined analysis with interaction terms that will allow you to see whether the effect of a predictor on response differs between time points.
Cheers,
Maurizio
-
Biochemist
- Posts: 37
- Joined: Mon Feb 19, 2024 9:44 am
Re: Binary logistic regression for repeated measures data - separate vs combined analysis of datasets for two time point
Thank you, Maurizio, for your elaborate answer. I now understand the benefits of the combined (and more complex) analysis. Good to know that the separate analysis is not statistically wrong.
I think I will go with the approach you suggested and start with separate analyses first and then - after getting a first picture of the predictors for the individual time points - continue with the combined analysis in a second step.
I think I will go with the approach you suggested and start with separate analyses first and then - after getting a first picture of the predictors for the individual time points - continue with the combined analysis in a second step.