0

If I standardize the coefficients (scale() command in R) of a beta regression with a logit link, how do I interpret them? I would say:

Of how many standard deviations (SD) the log odds of the y changes for a one SD increase in the covariates.

But is this correct?

Karolis Koncevičius
  • 4,282
  • 7
  • 30
  • 47
Jackk
  • 131
  • 5

1 Answers1

1

Standardizing the coefficients $\beta_1, \dots, \beta_k$ with scale() does not make sense. This would subtract the mean and divide by the standard deviation across the coefficients: $\tilde \beta_j = (\beta_j - \bar \beta)/\mathrm{SD}(\beta_1, \dots, \beta_k)$.

Multiplying $\beta_j$ with the standard deviation of the corresponding regressor $\mathrm{SD}(x_j)$ would have the same effect as standardizing the regressor $x_j$ (e.g., with scale()). This makes sense if you want to capture the marginal effect of a one standard deviation change in $x_j$ as opposed to a one unit change in $x_j$.

Additionally, you might find this answer about interpretation of betareg coef useful.

Achim Zeileis
  • 13,510
  • 1
  • 29
  • 53
  • To clarify, does this mean that multiplying the beta coefficient by the standard deviation of the predictor gives you a standardized beta that is comparable with standardized coefficients from other types of regression models, e.g. GLMs? – qdread Sep 06 '18 at 13:51
  • 1
    Yes, this behaves the same way as other models based on linear predictors. – Achim Zeileis Sep 06 '18 at 23:54