Goal: use GEE or GLMM to analyze repeated measures data in R
GEE problem: can’t find a way to do GEE with negative binomial family in R GLMM problem: not sure if I’m specifying random effect correctly Study question: Does the interaction of director and recipient group affect rates of a behavior? Data: Animals (n = 38) in one of 3 groups (life stages): B or C. Some individuals (~5) transitioned between groups between observation periods (2010, 2011, 2012), e.g., transitioning from B -> C. I gathered data on individuals in groups B and C, recording how often they directed a behavior to individuals in groups A, B, or C. I have multiple measures for each director (both within and between years). For example, for an individual who was alive for the entire study, I have one count of the behavior directed toward groups A, B, and C for each of the three years of study (total = 9 counts). Some individuals were observed all three years, others were only observed one or two years. Approach 1: Initially I used GEE in SPSS, the software I initially learned on (but no longer have access to) Outcome variable: counts of directed behaviors (“Diract”) from directors (“Dir” in group B or C) to members of groups A, B, C (“Rec"). Values range from 0-4, with overdispersion. Offset: the amount of time the focal individual was observed with individuals from the recipient group (natural log transformation applied, “LnScan”) Explanatory variable: the interaction of director and recipient group (“Dir*Rec”) Fixed effect: “Year” Family: Negative binomial with log link function Exchangeable working correlation matrix I hoped to rerun the analyses in R, which I am now learning to use. However, I cannot find a straightforward way to run a GEE in R with a negative binomial family. I am able to code what I want using a Poisson distribution using package geeglm: library("geeglm") m1 <- geeglm(Diract ~ Dir*Rec + Year + offset(LnScan), family = poisson("log"), data=Direct, id=ID, corstr="exchangeable") The lack of a negative binomial option for GEE in R has been addressed in the past few years here: https://www.researchgate.net/post/Does_anyone_know_how_to_undertake_Generalized_Estimating_Equation_GEE_modelling_using_the_negative_binomial_distribution_in_R and here: http://r.789695.n4.nabble.com/Negative-Binomial-Regression-td861977.html A similar question here is unanswered: https://stats.stackexchange.com/questions/83957/fit-negbin-glm-model-with-autoregressive-correlation-structure I wonder if there are any newer developments, since the posts are a few years old. I’m not very advanced in R, so if the solution involves a lot of creative coding, I won’t likely be able to figure it out. I have already tried using: library("sos") findFn("{generalized estimating equation}") and researching every package listed. Most seem to leverage gee (JGEE) or geepack (wgeesel), or lack a negative binomial family (PGEE, spind). Approach 2: GLMM I have become more familiar with GLMMs in R, so perhaps that is a better approach. I tried running GLMMs with package glmmTMB, but I am not sure I specified the random effect correctly (am I properly accounting for the repeated measures within AND between years?): m2 <- glmmTMB(Diract ~ DirPar*RecPar + offset(LnScan) + Year + (1|ID), data=Direct, family=list(family="nbinom1",link="log")) I further tried to specify a compound symmetry covariance structure with glmmTMB, but this failed: m2a <- glmmTMB(Diract ~ DirPar*RecPar + offset(LnScan) + Year + cs(1|ID), data=Direct,family=list(family="nbinom1",link="log")) Warning message: In fitTMB(TMBStruc) : Model convergence problem; non-positive-definite Hessian matrix. See vignette('troubleshooting') I also tried Ben Bolker’s suggestion (posted on Nabble; see link above) to use glmmPQL, but I got an error message: m3 <- glmmPQL(Diract ~ Dir*Rec + offset(LnScan) + Year, random = ~ 1 | ID, family = negative.binomial(1), data = Direct, correlation=corCompSymm(form=~1|ID)) Error in glmmPQL(Diract ~ DirPar * RecPar + offset(LnScan) + Year, random = ~1 | : could not find function "corCompSymm" If anyone has tips for a) a GEE with a negative binomial family and/or b) making sure I am specifying the random effect in a GLMM correctly to account for multiple measures within and across years, that would be greatly appreciated. Thank you from a self-taught but passionate R user! Bethany K. Hansen, PhD Chimp Haven sanctuary [[alternative HTML version deleted]] ______________________________________________ [hidden email] mailing list -- To UNSUBSCRIBE and more, see https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code. |
Free forum by Nabble | Edit this page |