The effect of deviding a variable into different sized groups in logistic regression

Question

My dependent is highly skewed to the left with range of 0-162 (a score of a questionnaire). I'm including it into a logistic regression model with 2 other cofactors. The total N is around 500. I have looked different models using the dependants as dichotomous (median) and 4 class (quartiles). When using median I get statistically signifant (p and 95% CI) OR regarding the higher scoring group but when using quartiles, none of the groups ORs are significant. I also get significant OR with highest group when using division of 50/25/25% (this makes sense in a clinical setting).

Could someone clarify for me possible reasons why I don't get signicant OR when using quartiles but do get when using median or combining the two lowest socring groups in logistig regression model?

Thank you!

Answer 1

A technical reason you are not getting consistent results is hard to provide without investigating your data and implementation. In any case, it seems that this is actually a question of interpretation. Simply put, a significant p-value in this case indicates that your term is providing a meaningful addition to the model. So, there is information contained in the dichotomous coding of your data that is useful. That information may be lost or obscured when you break it into quartiles. Arguably, this is supported by the fact that combining the two lower quartiles provides significance. This could be an issue for you or it may be irrelevant, depending on your research question.