I have a problem with running my logistic regression. When I run my analysis, I get really strange values and I cannot find anywhere how I can fix it. I already changed my reference category and that led to less strange values but they are still there. Also, this only happens to two of my eight predictors. These two predictors have multiple levels/categories.
Can someone explain to me what's wrong and how I can fix it?
Hi Zoë,
The problem you're running into in your logistic regression is that you have several factor levels which only have a few observations. If we examine the country of birth variable (in welk land bent u geboren?) we find: Suriname (N = 2), Nederlandse Antillen of Aruba (N = 1), Marokko (N = 4), ... and so on. When you fit a logistic regression model, the precision of your coefficients (size of the standard error, i.e. S.E.) depends upon how many observations you have across each of these factor levels. And in particular, whether you have sufficient events and non-events across these levels. For N = 1 in "Nederlandse Antillen of Aruba", you can never have both an event and a non-event if you only have a single observation. This will lead to estimation problems in your regression analysis, which is reflected by the large standard errors you observe in your output. Sample sizes such as N = 4 are also too small for proper estimation of logistic regression coefficients.
To fix this issue, you could consider grouping all small sample observations (i.e., everything except for Nederland) into one category called "Other". But this also depends on your research question. I don't have enough information on what you're exactly analysing to give you advice on whether this makes sense for your particular application (e.g. it is unclear what you exactly aim to investigate and we also do not know what your dependent variable is). But if the event rate is still too low after grouping the factor levels, the problem will remain unresolved. In that case, it is wiser to omit this variable from your regression analysis.
Hope this helps!
Hello Michael,
Thank you for your answer!
I'm looking to identify which factors predict dropout from treatment (dependent variable). Prior to my logistic regression, I did a Chi-square test to compare dropouts and completers. I am afraid that when I group categories, my Chi-square test for these variables will not be significant anymore.
You are using categorical independent variables. How could there be a logistic relationship???
- Badges
- Science method