I am using a linear regression model to determine the effect of diet score at baseline as predictor of weight at follow up where:

lm(weight_followup ~ Diet_Score*weight_base)

When I center my predictor variables I get the following:

enter image description here

However, the main effect changes when I do the same un centered: enter image description here

From clinical knowledge, the second set of results (uncentered) makes more sense, as Diet Score (aka quality) weight at follow up goes down.

Not sure what to use, as some posts suggest that if using continous predictors where 0 is not in the data range than predictors should centered.

All suggestions welcome. Its a large dataframe of N = 3000.

  • 3
    $\begingroup$ When you have a significant interaction effect, as you have here, the interpretation of what you call a "main effect" is tricky. It represents the difference from the intercept value (which also differs between your models) corresponding to the value of that predictor when the other predictor is at a value of 0 on its scale in that model. Regardless of which model you choose, the predicted weight at follow-up (and its standard error) would be the same for any combination of DASH score and baseline weight. $\endgroup$ – EdM Oct 9 '19 at 17:00
  • $\begingroup$ The two models are identical, as demonstrated by the statistics posted above and beneath each table. The differences in output are because the software is testing different sets of hypotheses about coefficients. $\endgroup$ – whuber Oct 9 '19 at 19:48

Browse other questions tagged or ask your own question.