Date: 2016-10-07 02:24 Sender: Ben HansenI tested out a fix that works by adjusting the value of `bb`. The new lines are between the first and last in the block below, they being adjacent line s of the current .vcov.avar1 definition.
```
p <- ncol(x)
## Next 2 lines added post-.vcov.MM, addressing #6471.
## This assumes initial S-estimate solved sum( loss )/(n-p) == bb,
## not mean( loss ) == bb as assumed in .vcov.avar1
adj <- (n-p)/n
bb <- bb * adj
r.s <- r / scale # final scaled residuals
```
I tried this change in a very small simulation study. In small samples (n=50, p=2) it improved coverage, but the improvement was extremely modest: a nominally 5% test that previously had 10% type 1 error rate came down to a type 1 error rate of 9%. |