modify glmmTMB.cpp

• enum valid_family is the list of distributions that glmmTMB knows about. Give your distribution an unused index (in a range with other similar distributions) and add it to the list.

...
truncated_genpois_family =404,
truncated_compois_family =405,
// new
zo_truncated_poisson_family = 410,
...

• add a case to the giant switch statement that handles the conditional likelihood (search for // Observation likelihood at the beginning): in this case we can most easily do this by modifying the code for the zero-truncated Poisson family (case truncated_poisson_family):
• mu(i) is the value of the location parameter for the current observation
• phi(i) is the value of the dispersion parameter for the current observation. This value always uses a log link, so it will be a value on [0,Inf). You should decide whether this value needs to be transformed or combined with mu(i) to form the traditional scale or dispersion parameter for your distribution. There are lots of examples in glmmTMB.cpp
• use logspace addition/subtraction if possible (logspace_add and logspace_sub functions in TMB: see here for more information). This isn't necessary but will make your computations more stable.
• if you want to be able to simulate data, add a SIMULATE{} condition that samples a random deviate from your distribution

case zo_truncated_poisson_family:
   log_nzprob = logspace_sub(Type(0), -mu(i));  // log(1-exp(-mu(i)));
   // now subtract the prob(X==1)
   log_nzprob = logspace_sub(log_nzprob, log(mu(i)) - mu(i));
   // log-Poisson likelihood minus the 'missing mass'
   tmp_loglik = dpois(yobs(i), mu(i), true) - log_nzprob;
   // this is a utility for use in ther zero-inflated case
   tmp_loglik = zt_lik_nearzero(yobs(i), tmp_loglik);
   SIMULATE{
    // conveniently, this built-in function already allows truncation
    //  at different points
    yobs(i) = glmmtmb::rtruncated_poisson(1, asDouble(mu(i)));
   }
 break;

modifying family.R

We might be able to get away with specifying family= as a list, but it's better to implement it as a new function.

#' @rdname nbinom2
#' @export
zo_truncated_poisson <- function(link="log") {
    r <- list(family="zo_truncated_poisson",
              variance=function(lambda) {
                  stop("haven't implemented variance function")
                  ## should figure this out ...
                  ## (lambda+lambda^2)/(1-exp(-lambda)) - lambda^2/((1-exp(-lambda))^2)
              })
    return(make_family(r,link))
}

As you can see, I haven't yet worked out the variance of a zero-one-truncated Poisson. This will only cause problems if/when a user wants to estimate Pearson residuals.

Ideally a $dev.resids() component should also be added, to return the deviance residuals (i.e., \(2 (\log L(y_i) - \log L_{\textrm{sat}}(y_i))\), where \(L_{\textrm{sat}}\) is the log-likelihood of \(y_i\) under the saturated model; see the $dev.resids components of families built into base R for examples.

For some families, the variance and deviance-residuals function require extra information such as a dispersion parameter. For the nbinom1 and nbinom2 families, glmmTMB does some additional stuff to store the value of the dispersion parameter in the environment of the variance/deviance residual functions (which share an environment), and to retrieve the dispersion parameter from the environment (search for ".Theta" in the R code for the package).

You should also document your new family, probably in the ?glmmTMB::family_glmmTMB page. This material is located in R/family.R, above the nbinom2 family function.

modifying glmmTMB.R

There may not be any other R code that needs to be updated, depending on the details of the family you are adding. Again, it's best to try to work by analogy with the closest family to the one you're adding. In this case, the only occurrence of truncated_poisson in glmmTMB.R is in the definition of which families have no dispersion parameter:

.noDispersionFamilies <- c("binomial", "poisson", "truncated_poisson",
                           "zo_truncated_poisson")

updating NAMESPACE

You need to make sure that your new family function is exported (listed in the NAMESPACE file). The easiest way to do this is by running devtools::document().

updating enum.R

The R file that keeps the C++ and R code in sync with respect to which families are available and which numeric code corresponds to which family is enum.R. Do not edit this file by hand: instead, run make enum-update

Reinstall

Re-install the package from source (R CMD INSTALL or install.package(..., repos = NULL))

Test

library(glmmTMB)
set.seed(101)
dd <- data.frame(y = rpois(500, exp(1)))
table(dd$y)
##  0   1   2   3   4   5   6   7   8   9 
## 34  91 117 116  68  45  17   7   3   2 
dd <- dd[dd$y>1,,drop=FALSE]
table(dd$y)
##   2   3   4   5   6   7   8   9 
## 117 116  68  45  17   7   3   2 
glmmTMB(y ~ 1, family = "zo_truncated_poisson", data = dd)

This appears to give the right answer (i.e. the estimated value of the intercept (log-link), 1.015, is close to the true value of 1).

Formula:          y ~ 1
Data: dd
      AIC       BIC    logLik  df.resid 
1184.2750 1188.2020 -591.1375       374 

Number of obs: 375

Fixed Effects:

Conditional model:
(Intercept)  
      1.015  

If you are adding the material for long-term use you should also add some tests to tests/testthat/test-families.R

C++ code

• add to the valid_covStruct enum in glmmTMB.cpp
• modify termwise_nll. In the case of homdiag we can re-use the existing diag_covstruct code (since everything is vectorized so should work equally well with a length-1 or length-\(p\) vector of (log) standard deviations)

R code

• modify parFun to specify the number of parameters
• modify documentation of glmmTMB()
• run make enum-update