I'm doing multiple regression through forced method by using 6 predictors. One of them was categorical in nature with more than 2 categories (3 to be precise). I've used 2 dummy variables to clarify my analysis. I've heard from somewhere that while using such scenario, we should use either higher or lower category dummy variable. So while considering higher categorical dummy variable, my lower level of categorical dummy variable shows significant trend however when I do the same with lower categorical dummy variable as reference category, my higher categorical dummy variable shows the significant relationship. But my problem is that there value of Beta and significance values is different than before. Is that supposed to happen?
In this situation it can often be helpful to consider other parameterisations of the categorical predictor. Two main alternatives:
- effect coding (-1, 0, 1 and -1, 1, 0)
- cell means coding (no intercept and dummy code all three categories)
With either coding it is often sensible to centre continuous predictors. The former models deviations from a weighted grand mean and the latter models the adjusted means of each category directly (assuming centred covariates).
Effect coding allows you to mimic an ANOVA/ANCOVA style parameterisation (though software often fits these as a form of verparameterized model). If your software allows you to fit ANOVA/ANCOVA models then it may handle the reparameterization for you and allow you to test the effect of the categorical variable as a factor and conduct pairwise tests or other contrasts.
Dear Claudia.
Thanks for your answer. So while changing the reference category or as you are baseline category, beta in response variable (DV) does change?
- Badges
- Science topic
- Similar topics
- Psychology