Two-level Poisson model with cross classified random effects

Dear All,

I have two-level data with individual as level-1, birth cohort and
community as level-2. All the level-2 covariates are generated from
the level-1 covariates by cross-classifying by cohort and community.

From what I read, an ordinary three-level model with individual nesed

within birth cohort nested within community, or individual nested
within community nested within birth cohort do not work well, neither
do model with individual nested within community by cohort. The right
way to go is to estimate a two-level model with two separate random
effects: within cohort and within community. The question I want to
ask is: how to do this using lmer?

I tried the following for a simple unconditioal model:

m1 <- lmer(count ~ offset(log(total)) + (1|comm) + (1|cohort), data, poisson)

where "count" is the dependent variable, "total" is the exposure
variable, "comm" is the community ID, and "cohort" is the birth cohort
ID. Will this be suffice? I got really smalle randome intercept
(5.0000e-10 for community and 4.4226e-05 for cohort), which got me a
bit nervous.

Thanks!

Best,
Shige

Shige Song a ??crit :

Dear All,

I have two-level data with individual as level-1, birth cohort and
community as level-2. All the level-2 covariates are generated from
the level-1 covariates by cross-classifying by cohort and community.

From what I read, an ordinary three-level model with individual nesed

within birth cohort nested within community, or individual nested
within community nested within birth cohort do not work well, neither
do model with individual nested within community by cohort. The right
way to go is to estimate a two-level model with two separate random
effects: within cohort and within community. The question I want to
ask is: how to do this using lmer?

I tried the following for a simple unconditioal model:

m1 <- lmer(count ~ offset(log(total)) + (1|comm) + (1|cohort), data, poisson)

where "count" is the dependent variable, "total" is the exposure
variable, "comm" is the community ID, and "cohort" is the birth cohort
ID. Will this be suffice? I got really smalle randome intercept
(5.0000e-10 for community and 4.4226e-05 for cohort), which got me a
bit nervous.

Thanks!

Best,
Shige

The default method to fit generalized mixed-effect models with lme4 is
"PQL" which is fast but not very accurate for the random effects (they
might be underestimated). Try other methods ("Laplace" or "AGQ") to see
if you get different results.

Best,

Renaud

--
Dr Renaud Lancelot, v??t??rinaire
Projet FSP r??gional ??pid??miologie v??t??rinaire
C/0 Ambassade de France - SCAC
BP 834 Antananarivo 101 - Madagascar

e-mail: renaud.lancelot at cirad.fr
tel.:   +261 32 40 165 53 (cell)
         +261 20 22 665 36 ext. 225 (work)