I'm running a binomial GLM in R. The data for the model comes from survey responses. The response variable is 'change in wellbeing' and the predictor variables are derived from several other questions from the survey.
Many of the predictor variables are categorical, with several levels. The number of responses in each level can be very different, e.g., we have a variable 'employment status' with the following levels and numbers of respondents in each level:
Retired - 650
Employed (full time) - 150
Employed (part time) - 200
Student - 3
Unemployed - 10
There are similar patterns for predictor variables such as 'ethnicity' and 'disability', with one or two of the levels containing most of the respondents, and a few other factor levels with very low numbers of respondents in them.
As far as I'm aware, it's still okay to run models when some of the predictor variables look like this. The problem being that the statistical power for those levels with very low sample sizes will be very poor (e.g., from the above example, it's unlikely we'd be able to tell if being unemployed is a strong predictor of wellbeing). However, I'm wondering if there may be other problems or consideration to make when data looks like this.
My questions are:
- Is it still okay to run models when the data look like this? Or may some factor levels need grouping?
- Will data like this cause any computational problems when running a model?
- Are there any other considerations to make when data looks like this?
