Random-effects regression with within-person predictors

#1
Hi,

I generated a random-effects logistic regression model and included within-person predictors by person-mean-centering.

Not all individuals in my dataset have multiple data points, i.e. the individuals gave interviews, and while a large number of them did multiple interviews, many only did one interview.

My research question is whether interviewer characteristics affect how interviewers measure respondent characteristics. To answer the question, I analyze the within-person effects of interviewer characteristics.

Now, my question is:

Who (i.e. which interview respondents) are not included in this analysis?

1. Obviously, respondents who only did one interview are not included, since there are literally no within-person comparisons to make.
2. Next, Williams (2018) (see link below) states that "for all subjects where the dependent variable [i.e. measured respondent characteristics) is a constant ... the case is dropped from the statistical analysis." This make sense.

3. However, what about the independent variable being constant, i.e. what about about respondents who were interviewed multiple times by interviewers whose characteristics are constant across all interviews? My intuition tells me that these individuals would also be excluded from the analysis, but I cannot fully wrap my head around it and I couldn't find any literature on that.

Any input would be greatly appreciated!


Link to William (2018): https://www.google.com/url?sa=t&rct...VsRandom.pdf&usg=AOvVaw2Uq3PAcGrUIML-wmaaA8tL
 
#3
If the predictor is a constant over time how can you predict with it. Don't you have to have some variation to run the regression?
Sometimes interviewer characteristics (e.g. gender) vary across a respondent's interviews, sometimes they don't vary across a respondent's interviews.

I assume that those respondents who were interviewed only by men or only by women would also be dropped from the analysis. But I'm not positive, so I wanted to check my intuition by posting here. In other words, the within-respondent results are only based on respondents who have variation in both the IV and the DV across interviews, correct?
 

noetsi

Fortran must die
#4
I only know of within and between effects in the context of multilevel models. But I don't see how you can run any regression on something that does not vary. Indeed things that vary little are often dropped out because they attenuate slopes (I think of all the slopes in the model).
 
#5
I only know of within and between effects in the context of multilevel models. But I don't see how you can run any regression on something that does not vary. Indeed things that vary little are often dropped out because they attenuate slopes (I think of all the slopes in the model).
My apologies for not providing more detail.

In a nutshell, I generated an interviewer random-effects binary logistic regression model and estimated within-respondent effects (using mean-centering) to analyze whether interviewer characteristics (e.g. gender) influence how they record certain respondent characteristics. Here, the random-effects results are based on all interviews, including interviews of respondents who were only interviewed once (and thus cannot, by definition, be included in within-responded analyses).

However, regarding the within-person effects: some respondents were always interviewed by women, so -- as you suggest -- they drop out of the analysis. Likewise, respondents whose characteristics were recorded the same way across interviews also drop out of the within-respondent analyses. In other words, only when both the IV (interviewer characteristics) and the DV (recorded respondent characteristics) vary across a respondent's interviews are they included in the within-person results.