[R-meta] multilevel glmm meta-analysis question

Fri, Mar 5, 2021 12:04 AM

Correct. And just to be precise: N should be the number of sampled cats on which a particular prevalence is based. A study might have sampled 200 cats, 100 of which were examined for having parasite A, the other for parasite B. Then N would be 100 for the two prevalences. On the other hand, if all 200 cats were examined for having parasite A and also for having parasite B, then the two prevalences are both based on N=200. Maybe the former scenario is unrealistic, but it's not the total number of sampled cats in each study that matters, but the number of cats based on which a particular prevalence was computed.

tau^2 is just a name for a particular variance component in certain types of models. Your model will have multiple variance components. Would be nice to report them.

Similarly, the Q-test is a particular type of test for heterogeneity. You can do something analogous by leaving out the variance components and running a likelihood ratio test comparing the model with versus without the variance components.

Best,
Wolfgang

From: Viechtbauer, Wolfgang (SP) <wolfgang.viechtbauer at maastrichtuniversity.nl>
Sent: Wednesday, March 3, 2021 2:59 AM
To: Kayleigh Chalkowski <kzc0061 at auburn.edu>; r-sig-meta-analysis at r-project.org
<r-sig-meta-analysis at r-project.org>
Subject: RE: multilevel glmm meta-analysis question

Dear Kayleigh,

Do not use 'weights' in this manner in glmer(). In generalized linear mixed-
effects models, we do not specify sampling variances / weights; the model takes
care of the appropriate weighting for you. The weights argument also serves a very
different function in glmer(). For family=binomial, if 'prevalence' is the
proportion of infected cats, then weights should be the actual number of cats
studied (so, if 20 out of 100 cats are infected and prevalence = 0.2, then weights
= 100 for this study).

The way you are adding random effects to the model leads to crossed random
effects. Whether this is appropriate or not (as opposed to using nested random
effects) is debatable.

Strictly speaking, this model cannot properly account for the dependence in
multiple prevalences (for different parasites) for the same group of cats. If you
have prevalence = 0.2 for parasite A and prevalence = 0.2 for parasite B for those
100 cats above, then you do not know if there are 20 cats infected with both
parasites, 40 cats are infected (20 with A and 20 with B), or anything in between.
Without the raw data, you do not know this. I think one could still go ahead with
such an analysis, using a random effect at the study level (as you use) to account
for the some of the dependence, but this doesn't fully capture it. The results
should therefore be treated with caution.

Best,
Wolfgang

-----Original Message-----
From: R-sig-meta-analysis [mailto:r-sig-meta-analysis-bounces at r-project.org] On
Behalf Of Kayleigh Chalkowski
Sent: Tuesday, 02 March, 2021 20:30
To: r-sig-meta-analysis at r-project.org
Subject: [R-meta] multilevel glmm meta-analysis question

Dear all,

I am undertaking a multilevel generalized linear mixed model meta-analysis and,
following the advice here<http://www.metafor-project.org/doku.php/todo> regarding
multilevel glmms in metafor, I am using the glmer function of the lme4 package. I
am wondering 1) if that I am specifying my weights correctly and 2) how to get
values that are important to report for this meta-analysis, like heterogeneity.

In this meta-analysis, I hypothesize that both socioeconomic and ecological
variables are important in predicting parasite prevalence in free-roaming cats

and

dogs-- so I'm mainly interested in the effects of moderators (an average
prevalence isn't very informative). Since each study gives a proportion of
dogs/cats infected out of a total, I have chosen a binomial model.

Here is my code for one of my univariable models:
res_san_3<-glmer(prevalence ~ san_10 + (1 | Species) + (1 | country) + (1 |

study)

+ (1 | uniq), weights = 1/vi, data=feral, family=binomial, na.action=na.fail)

prevalence is the total infected out of the total sampled dogs/cats of each

study,

and each listed random effect there are the different nested levels. Nested

levels

include species of parasite, followed by country, then study, then each sample
within each study (because many studies sampled multiple parasites). I used
inverse variance for the weights here.

I would greatly appreciate any thoughts, or any helpful information that could be
referred to me. I've searched the web extensively for help understanding
multilevel glmms and was unable to find answers to my questions.

Thank you so much,
Kayleigh

Below here is the output from that model in case it is helpful:

Generalized linear mixed model fit by maximum likelihood (Laplace Approximation)
['glmerMod']
Family: binomial? ( logit )
Formula: prevalence ~ san_10 + (1 | Species) + (1 | country) + (1 | study) +
(1 | uniq)
?? Data: feral
Weights: 1/vi

???? AIC????? BIC?? logLik deviance df.resid
? 5131.5?? 5162.8? -2559.7?? 5119.5???? 1374

Scaled residuals:
??? Min????? 1Q? Median????? 3Q???? Max
-1.3271 -0.4905 -0.2745? 0.1395? 1.8286

Random effects:
Groups? Name??????? Variance Std.Dev.
uniq??? (Intercept) 0.650879 0.80677
study?? (Intercept) 0.386573 0.62175
Species (Intercept) 0.245294 0.49527
country (Intercept) 0.007934 0.08907
Number of obs: 1380, groups:? uniq, 1380; study, 449; Species, 204; country, 70

Fixed effects:
??????????? Estimate Std. Error z value Pr(>|z|)
(Intercept) -0.49574??? 0.29256? -1.694 0.090176 .
san_10????? -0.11641??? 0.03191? -3.648 0.000264 ***
---
Signif. codes:? 0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1

Kayleigh Chalkowski, M.Sc.
PhD Student
Fulbright Madagascar 2020-2021
School of Forestry and Wildlife Sciences
Auburn University
(607) 319-6342

[R-meta] multilevel glmm meta-analysis question

Thread (4 messages)