0

I am using the glmer() from the lme4 package. In the model, I have 6 different variables and multiple samples from the same patient. Five of the variables are scaled the same way (2^(difference patients and technical control). The last variable was on a different scale(non-parametric, values between 10^1 and 10^6). I log-transformed this data and corrected them for the technical controls the same way as variable 1-5. Should I now also transform and scale my matrix before running the model? or is that not necessary when doing a glmer()? 

model <- glmer(
  resp ~ variable1 + variable2 + variable3 +
    variable4 + variable5 + variable6 + (1 | pat),
  family = binomial(), data = datmat)
Dimitris Rizopoulos
  • 17,519
  • 2
  • 16
  • 37
esseattle
  • 3
  • 1

1 Answers1

2

There are two main reasons to standardise predictors: optimisation performance and interpretability of results. When you glmer estimates the model parameters without problems (no warning messages; estimates and their standard errors reasonable) then you don't need to standardise your predictors.

Depending on issues such as sample size, collinerarity of predictors, including non-orthogonal polynomial terms, complexity of the hierarchical structure and distribution, glmer's internal optimisation may find it difficult to estimate model parameters. In this case, standardisation can help. That said, I have yet to encounter a situation where glmer struggled. It seems to be a particularly well-crafted bit of model fitting.

Which leaves interpretability of predictors: parameter estimates of standardised predictors can be directly compared: the larger the absolute value, the more "important" (rather: influential) they are for the response. When transforming, consider Andrew Gelman's advice to use 2 times the standard deviation (rather than one) when you also have categorical predictors (https://statmodeling.stat.columbia.edu/2006/06/21/standardizing_r/ ) or re-code dummy categories to -1/1 (rather than 0/1: http://andrewgelman.com/2009/06/09/standardization/).

This question has been asked in various guises, e.g. Why shouldn't I standardize my predictors when putting them into a regression model?

Carsten
  • 341
  • 2
  • 3