scan1() with model="binary" giving lots of NAs

[kbroman]

scan() with model="binary" is giving lots of NA results for the LOD scores (at least in the case of DO mouse data)...much more than I would have expected. Here's an example plot for one chromosome; note the gaps.

[kbroman]

This happens whenever one of the coefficient's estimates goes to +/- infinity. glm() gives the warming "glm.fit: fitted probabilities numerically 0 or 1 occurred". But I should at least be able to get LOD scores if not coefficient estimates.

[kbroman]

Rather simplistic solution: don't allow the "linear predictor" to be too large. If > 100 (< -100), set to 100 (or -100).

[kbroman]

Still having a few problems (with Attie DO data using coat color), though clearly much less than before.

Posted some data and code in a gist.

• fit1() gives NA for the LOD score; in other cases I was getting negative LOD scores.
• glm() doesn't converge within 25 iterations (the default), and if we increase the number of iterations, it gives a log likelihood that is less than that for the null hypothesis (with just an intercept).
• My crude R function goes to -Inf for the log likelihood after 15 iterations.