Not sure what this warning message means

lynnerperry

A student tried this code at the end of notebook 4 (create your own code for class competition) and received the warning message. We were wondering what it means.

Thanks,

LPerry

my_model <- lm(default_rate ~ (region * hbcu * net_tuition * enrollment) + SAT_avg + (pct_fed_loan * pct_PELL ),data = train)

Warning message in predict.lm(my_model, newdata = test):
“prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases”

dashiell-young-saver

Here's a helpful thread on this! https://www.statology.org/prediction-from-rank-deficient-fit-may-be-misleading/

When you use the * instead of the + when listing the predictors, it creates additional interaction terms. I'm thinking that maybe there are so many terms in their model, it actually outpaces the number of data values, leading to this warning. If they add a new cell and do summary(my_model) you can see how many predictors are in the model. If it's a LOT of predictors, that's what may be causing the warning.

dylarm

dashiell-young-saver To add on to this, it doesn't seem so much to me that it's strictly the number of terms is exceeding the number of values (this one has 52 coefficients while the winning model last year had something in the 60s), but also what those terms are. When I ran summary on it, there were something like 14 coefficients that couldn't be determined due to singularities. And trying out different combinations, it seems the interaction that usually causes this is region*hbcu, likely because there are no HBCUs at least two of the regions.

dashiell-young-saver

@dylarm You're totally right! Good catch there. @lynnerperry - does that make sense? We'd be happy to provide more explanation if would be helpful.