thomas-fung/mpcmp

Mean-parametrized Conway-Maxwell Poisson (mpcmp) Regression

Hi there 👋

My name is Thomas and welcome to my Github page!

• 💬 I can speak English, Cantonese and R fluently.
• 🔭 I’m currently working on various projects in modelling count/integer-valued data. Please dm me if you would like to collaborate.
• ⚡ Fun fact: I like baking and I often bake late at night. I also like a good cup of coffee. ☕️
• 📫 How to reach me: For my github repos, filing an an issue in the corresponding github repo. For others, email [email protected]

mpcmp's Issues

In the varying dispersion case, the coef summary table has character instead of double

Based on how the coefficient summary table is constructed, it got converted into characters along the way. One negative side effect is that it doesn't show the Signif. codes.

documentation for nu estimation

Hi Thomas,

In the documentation for glm.cmp, the formulation for the varying dispersion model is given by

Y_i ~ CMP(mu_i, nu_i)

where

E(Y_i) = mu_i = exp(x_i^T beta),

and dispersion parameters are model via

nu_i = exp(s_i^T gamma)

This might not warrant an issue (sorry!!), but in the formulation for for nu_i, what is s_i referring to?

Thank you,

Sean

Error when using gammastart

This issue was submitted us by Sijie Wang.

library(mpcmp)
data(sitophilus)
M.sit <- glm.cmp(formula = ninsect ~ extract,
                 formula_nu = ~extract, 
                 data = sitophilus)
glm.cmp(formula = ninsect ~ extract,
        formula_nu = ~extract, 
        data = sitophilus, 
        gammastart = M.sit$coefficients_gamma)
#> Error in comp_lambdas(mu0, nu0, lambdalb = lambdalb, lambdaub = min(lambdaub, : object 'lambdaold' not found

^{Created on 2022-05-10 by the reprex package (v2.0.1)}

CRAN check issues

Received a message from Prof Ripley that CRAN checks picked up a couple of errors on the devel version on Linux and windows. For some unknown reason, it's rounded one of the coefficients to a different number.

Post hocs

Hi!
First of all, this is a fantastic package! I am a biologist who deals with count data that often is not equidispersed. CMP regression using mpcmp models my data very well. The only thing that would be great to have for me and I guess other users is possibility to do posthoc or contrast comparisons. I am wondering maybe this is in your plans, or do you have a dirty script to do that/ make cmp objects communicate with emmeans or other established R packages?

Is the normalising constant being correctly calculated?

The implementation here seems to include log(j) and not lfactorial(j) or lgamma(j + 1) in the denominator.

I can't generate random number from COM-POISSON distribution if mean is large

I know that mpcmp::dcomp() generates random numbers from COM-Poisson distribution. However, they fail to generate such data if mean is large.

The following example doesn't compute, or takes too much time:

mpcmp::rcomp(1, mu = 500, nu = 1)

How can I generate a COM-Poisson data with large mean efficiently?

Function `comp_lambdas` not available

Although the manual shows the usages of this function, it was not exported to the NAMESAPCE such that it cannot be called.

offset terms in glm.cmp?

Hi Thomas,

Thanks for putting together such a helpful package. I'm interested in including an offset term in my glm.cmp model, but it doesn't seem to get carried through to the final output. For example,

set.seed(1)

#generate variable to use as an offset
cottonbolls$n_samples <- rpois(nrow(cottonbolls), 5)

#fit with offset
M.bolls <- glm.cmp(nc~ 1+stages:def+stages:def2, offset = log(n_samples),data = cottonbolls)
M.bolls$offset

returns a vector of zeros. Is this the correct method to incorporate an offset term into glm.cmp models, or am I going about it wrong?

Best

Sean

mpcmp with identity link

As titled.

glm.cmp would give an error if the model matrix doesn't have full rank

library(tidyverse)
library(mpcmp)
dat <- tribble(~count, ~ code1, ~code2,
               3, "Code 1", "Code 1",
               4, "Code 2", "Code 1",
               2, "Code 2", "Code 1",
               5, "Code 2", "Code 2",
               1, "Code 2", "Code 2")
M1 <- glm(formula=count ~ code1 * code2, data = dat, family = poisson)
coef(M1)
#>             (Intercept)             code1Code 2             code2Code 2 
#>            1.098612e+00           -2.167798e-16            5.768888e-16 
#> code1Code 2:code2Code 2 
#>                      NA
M2 <- glm.cmp(formula=count ~ code1 * code2, data = dat)
#> Error in if (update_info_matrix < 0) {: missing value where TRUE/FALSE needed

^{Created on 2020-10-30 by the reprex package (v0.3.0)}

Need to update the main fitting function to make it behave more like the glm function.

Zero inflated version of mpcmp

As in the title.