Do you have missing observations in your data? If so, I suspect that the number of valid observations differ across the models depending on how many terms are in the models...

A plausible answer is in your error message. The models should have the same numbers of cases (at each level). As Salvatore S. Mangiafico says, one cause of this is missing values and how you have dealt with them, but because you decided not to share any output or the code you input for your models we are just guessing.

Salvatore S. Mangiafico Daniel Wright thank you the replies! I do have some missing data. and here are the codes for model2 and model4

M2

There are some "tricky" things to handle missing values. A very simplistic solution is that you can replace the missing values with the mean/median/modes of the specific variable. Another option would be to impute values to predict what "could" have been the value using another models for example using a random forest model from the missForest package. (https://stat.ethz.ch/education/semesters/ss2012/ams/paper/missForest_1.2.pdf). But, I always feel the results should be more descriptively addressed than conclusive. Hence, you use the data to predict the data(?), which is somehow circle reasoning resulting in a bias. The question is how many missing values do you have per variable if there are a "lot" then imputation also does not work. But if you have only1414-1407=7 missing values on 1414 samples you can just as well leave them out or set them to any random value as change of the model output will be neglectable.

When you're comparing M2 and M4, you'd have to keep only the 1407 observations that are complete for both of them. This should be easy to do in R, I think with the complete.cases function.

Or use a method mentioned by Wim Kaijser.