# Learning error of mlogit

I am using the mlogit R package.

For a subset of my dataset, mlogit is learned ok. For another subset of my dataset, I get this error.

Error in solve.default(H, g[!fixed]) : system is computationally singular: reciprocal condition number = 1.12239e-16

When I looked for this error in contexts outside of mlogit, one suggested remedy it to use qr.solve() instead of solve.default().

1. How do I modify mlogit to use qr.solve()?
2. What in my data can possibly be causing this problem?
3. Can this error result from a certain category being fully predictable from some attribute?
4. How do I check?
• Mlogit is regression model. The most likely problem is that your regressors are linearly related. Using different methods will not change the problem. Supply more details about your data. – mpiktas Apr 20 '12 at 6:43
• This error can also come if you specify an alternative-specific variable as being generic. Did you remember to | your formula, as in mlogit(DV ~ Generic | Alt.Spec). Maybe you already checked this. – gregmacfarlane Jun 27 '12 at 1:22