I don't know if this is still relevant for you, but probably for someone
else.
I ran into the same problem an found two dirty workarounds that can help
in many cases (but not all). Both can also be combined:
1. repeat the call to "segmented" with different seeds and different
initial estimates of the breakpoints (psi). Wrap this with "try" to
catch
the error:
<code>
res= NULL
while (class(res)[1]!="segmented")
{
some_seed= runif(1) #you may also systematically vary this
some_psi = runif(2) #you may also systematically vary this.
#Better tailor this to the distribution of x values in your_data
model_lm = lm(y~x, data=your_data)
res = try(segmented(obj=model_lm, seg.Z=~x, model=TRUE,
psi=list(x=some_psi),
control=seg.control(seed=some_seed)), silent=TRUE)
}
</code>
2. insert dummy points between your actual data to reduce the chance of
"only 1 datum in an interval". When you supply these dummy data with
low
weights, their effect on the outcome should be low:
<code>
model_lm = lm(y~x, data=data_with_dummyentries,
weights=data_with_dummyentries$weight)
res=segmented(obj=model_lm, seg.Z=~x, model=TRUE, psi=list(x=c(10,20)))
</code>
Hope it helps,
Melwin