How do I get confidence intervals without inverting a singular Hessian matrix in R?

I'm a student working on an epidemiology model in R, using maximum likelihood methods. I created my negative log likelihood function. It's sort of gross looking, but here it is:

NLLdiff = function(v1, CV1, v2, CV2, st1 = (czI01 - czV01), st2 = (czI02 - czV02), st01 = czI01, st02 = czI02, tt1 = czT01, tt2 = czT02) { 
    prob1 = (1 + v1 * CV1 * tt1)^(-1/CV1)
    prob2 = ( 1 + v2 * CV2 * tt2)^(-1/CV2) 
    -(sum(dbinom(st1, st01, prob1, log = T)) + sum(dbinom(st2, st02, prob2, log = T)))
 }

The reason the first line looks so awful is because most of the data it takes is input there. czI01, for example, is already declared. I did this simply so that my later calls to the function don't all have to have awful vectors in them.

I then optimized for CV1, CV2, v1 and v2 using mle2 (library bbmle). That's also a bit gross looking, and looks like:

ml.cz.diff = mle2 (NLLdiff, start=list(v1 = vguess, CV1 = cguess, v2 = vguess, CV2 = cguess), method="L-BFGS-B", lower = 0.0001)

Now, everything works fine up until here. ml.cz.diff gives me values that I can turn into a plot that reasonably fits my data. I also have several different models, and can get AICc values to compare them. However, when I try to get confidence intervals around v1, CV1, v2 and CV2 I have problems. Basically, I get a negative bound on CV1, which is impossible as it actually represents a square number in the biological model as well as some warnings.

Is there a better way to get confidence intervals? Or, really, a way to get confidence intervals that make sense here?

What I see happening is that, by coincidence, my hessian matrix is singular for some values in the optimization space. But, since I'm optimizing over 4 variables and don't have overly extensive programming knowledge, I can't come up with a good method of optimization that doesn't rely on the hessian. I have googled the problem - it suggested that my model's bad, but I'm reconstructing some work done before which suggests that my model's really not awful (the plots I make using the ml.cz.diff look like the plots of the original work). I have also read the relevant parts of the manual as well as Bolker's book Ecological Models in R. I have also tried different optimization methods, which resulted in a longer run time but the same errors. The "SANN" method didn't finish running within an hour, so I didn't wait around to see the result.

In a nutshell: my confidence intervals are bad. Is there a relatively straightforward way to fix them in R?

My vectors are:

czT01 = c(5, 5, 5, 5, 5, 5, 5, 25, 25, 25, 25, 25, 25, 25, 50, 50, 50, 50, 50, 50, 50)
czT02 = c(5, 5, 5, 5, 5, 10, 10, 10, 10, 10, 25开发者_StackOverflow社区, 25, 25, 25, 25, 50, 50, 50, 50, 50, 75, 75, 75, 75, 75)
czI01 = c(25, 24, 22, 22, 26, 23, 25, 25, 25, 23, 25, 18, 21, 24, 22, 23, 25, 23, 25, 25, 25)
czI02 = c(13, 16, 5, 18, 16, 13, 17, 22, 13, 15, 15, 22, 12, 12, 13, 13, 11, 19, 21, 13, 21, 18, 16, 15, 11)
czV01 = c(1, 4, 5, 5, 2, 3, 4, 11, 8, 1, 11, 12, 10, 16, 5, 15, 18, 12, 23, 13, 22)
czV02 = c(0, 3, 1, 5, 1, 6, 3, 4, 7, 12, 2, 8, 8, 5, 3, 6, 4, 6, 11, 5, 11, 1, 13, 9, 7)

and I get my guesses by:

v = -log((c(czI01, czI02) - c(czV01, czV02))/c(czI01, czI02))/c(czT01, czT02)
vguess = mean(v)
cguess = var(v)/vguess^2

It's also possible that I'm doing something else completely wrong, but my results seem reasonable so I haven't caught it.

You could change the parameterization so that the constraints are always satisfied. Rewrite the likelihood as a a function of ln(CV1) and ln(CV2), that way you can be sure that CV1 and CV2 remain strictly positive.

NLLdiff_2 = function(v1, lnCV1, v2, lnCV2, st1 = (czI01 - czV01), st2 = (czI02 - czV02), st01 = czI01, st02 = czI02, tt1 = czT01, tt2 = czT02) { 
prob1 = (1 + v1 * exp(lnCV1) * tt1)^(-1/exp(lnCV1))
prob2 = ( 1 + v2 * exp(lnCV2) * tt2)^(-1/exp(lnCV2)) 
-(sum(dbinom(st1, st01, prob1, log = T)) + sum(dbinom(st2, st02, prob2, log = T)))
 }