1
$\begingroup$

I collected data over multiple locations and years on crop yield and want to regress yield as a function of rainfall and heat-stress which are in two different units. Suppose my dataframe has 5 columns: year, location, yield, rainfall and temperature. These are my steps:

 dat[,4:5] <- scale(dat[, 4:5], center = T, scale  T)
 model <- lmer(yield ~ rain + temp + (1|location) + (1|year), data = dat)

After I get this model and I want to use the model for prediction. Suppose I collect new data from different years or locations called dat1 which has 4 columns: year, location, rainfall and temperature.

My confusion is since the fitted model takes in standardised rainfall and temperature, how do I standardise these two variables in the new data dat1? Do I simple do:

dat1[,3:4] <- scale(dat1[,3:4], center = T, scale = T)
predict(model, newdata = dat1)

Or do I have to standarise the new data using mean and standard deviation of the original data dat?

$\endgroup$
  • 1
    $\begingroup$ standarise the new data using mean and standard deviation of the original data $\endgroup$ – Bryan Krause Jun 21 '18 at 22:32
  • $\begingroup$ Standardization should not be necessary at all, keep the original variables! that makes interpretation easier. And, if for any reason you choose to standardize (what is your reason?), then follow advice from @Bryan Krause $\endgroup$ – kjetil b halvorsen Jun 22 '18 at 9:29
  • $\begingroup$ I standardised because all of my variables (I have more variables than what I showed before) are in different units and I keep getting warning message that my predictors are on a different scale and consider rescaling. $\endgroup$ – 89_Simple Jun 22 '18 at 9:31
3
$\begingroup$

Just because there is a warning message about predictors being on a different scale does not mean that you need to standardize them. Standardization can lead to difficulties with interpretation.

Generally, you only need to rescale the offending variable(s) by multiplying or dividing it by something appropriate, and since it is only a warning, you don't really need to do anything, although there are some situations where variables on vastly differing scales can cause numerical instability during model fitting.

You are correct that if you standardize the variables that are used as inputs for your model, you will have to standardize those in your test/prediction dataset too.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.