Hello,

I am doing some multiple regressions for my master thesis using bootstrapping. I am getting warning messages sounding like this:

"For models with dependent variable [dependent variable], the following variables are constants or have missing correlations in split file $bootstrap_split=[different numbers] : [independent variable]. They will be deleted from the analysis.

Therefore some of the results are based on a reduced number of bootstrapping samples.

Is this reduction something I have to accept or can I do something about it?

Best wishes,

Katharina

Bootstrapping involves doing random sampling from the estimation data, so for variables without much variation, it is certainly possible that some of these samples wouldn't be usable. This is more likely if the full dataset is very small.

You may be able to compensate for this by specifying a larger number of samples (the default is 1000). If the problem is due to having a very small sample for estimation, the bootstrap estimates may not be very accurate anyway.

Hello Jon,

thank you for your response. My first idea was aswell to define a higher number of samples to make sure that every result is based on at least 1000 bootstrap samples.

What would you consider as a very small sample? I have around 90 participants for a regression with 3 predictors. However due to missing values the sample reduces to 50 to 60 for most of the analyses.

That sample size is reasonable for bootstrapping, and you can check the replication issue by upping the number from 1000 to, say, 5000 to see if it makes any real difference.

I would be more worried about the high proportion of missing values unless these are NA cases. You might want to consider multiple imputation to address that.