[R-meta] When to remove the intercept and other questions

Viechtbauer, Wolfgang (SP) wolfg@ng@viechtb@uer @ending from m@@@trichtuniver@ity@nl
Sun May 20 19:28:15 CEST 2018

Dear Rafael,

Please always cc the mailing list.

Since your question is about the fixed effects part of the model, whether you are fitting a model with rma() or rma.mv() doesn't change how one deals with categorical moderators (or a combination of categorical and continuous ones). So, while those links lead to discussions of models fitted with rma(), the underlying principles are the same.


-----Original Message-----
From: Rafael Rios [mailto:biorafaelrm at gmail.com] 
Sent: Saturday, 19 May, 2018 19:40
To: Viechtbauer, Wolfgang (SP)
Subject: Re: When to remove the intercept and other questions

Dear Wolfgang,

Thank you a lot for the quick answer. If I understood correctly, the test with intercept term is used to evaluate if the mean effect size for each subgroup of the categorical predictor is the same. However, when I remove the intercept, I am testing if each mean outcome is 0. Thus, I need to use the intercept in my meta-analytic models, because I want to compare  outcome differences between subgroups of the moderator.

Your links with examples about how to calculate differences between subgroups will be helpful too, but my doubt is related to a different model. I am using rma.mv function with R argument to control for phylogenetic non-independence among effect sizes. The moderators are a categorical variable (with four subgroups) and a continuous one. Consequetly, I will obtain four straghts. I need intercept and slope values for each straight line to plot in a graph. How can I obtain such information?

Best wishes,


Em sáb, 19 de mai de 2018 12:02, Viechtbauer, Wolfgang (SP) <wolfgang.viechtbauer at maastrichtuniversity.nl> escreveu:
Dear Rafael,

Let's take a simple example:

### load BCG vaccine data
dat <- get(data(dat.bcg))

### calculate log risk ratios and corresponding sampling variances
dat <- escalc(measure="RR", ai=tpos, bi=tneg, ci=cpos, di=cneg, data=dat.bcg)

### mixed-effects meta-regression model with categorical moderator
res <- rma(yi, vi, mods = ~ alloc, data=dat)

You will find:

Test of Moderators (coefficients 2:3):
QM(df = 2) = 1.7675, p-val = 0.4132

Model Results:

                 estimate      se     zval    pval    ci.lb   ci.ub 
intrcpt           -0.5180  0.4412  -1.1740  0.2404  -1.3827  0.3468    
allocrandom       -0.4478  0.5158  -0.8682  0.3853  -1.4588  0.5632    
allocsystematic    0.0890  0.5600   0.1590  0.8737  -1.0086  1.1867    

The QM-test tests coefficients 2 and 3, which are the *differences* between the reference level (in this case 'alternate') and the other two levels ('random' and 'systematic'). So, this gives you a test whether the mean outcome is the same for the three levels.

### model with intercept removed
res <- rma(yi, vi, mods = ~ alloc - 1, data=dat)

You will find:

Test of Moderators (coefficients 1:3):
QM(df = 3) = 15.9842, p-val = 0.0011

Model Results:

                 estimate      se     zval    pval    ci.lb    ci.ub 
allocalternate    -0.5180  0.4412  -1.1740  0.2404  -1.3827   0.3468      
allocrandom       -0.9658  0.2672  -3.6138  0.0003  -1.4896  -0.4420  *** 
allocsystematic   -0.4289  0.3449  -1.2434  0.2137  -1.1050   0.2472      

In this case, the coefficients are the mean outcomes for each level. The QM-test tests coefficients 1, 2, and 3, so it tests whether the mean outcome is 0 for all three levels.

You might also want to work through these examples:



-----Original Message-----
From: Rafael Rios [mailto:biorafaelrm at gmail.com] 
Sent: Saturday, 19 May, 2018 3:57
To: r-sig-meta-analysis at r-project.org; Viechtbauer, Wolfgang (SP)
Subject: When to remove the intercept and other questions

Dear Dr. Wolfgang and All,

I have some doubts that you could help me to clarify. I read e-mails in R-meta list of a conversation between Samuel Knapp and James Pustejovsky with the title "Testing of moderators in rma()". James clarified the following:

"When the model includes an intercept term, the omnibus test does *not* include the intercept. So the null hypothesis is b1 = 0 and b2 =0.

If you fit the model without the intercept, then the equivalent to the omnibus test from the model with an intercept would be that the average effect sizes are all equal, as in b1 = b2 = b3."

I want to compare the differences between effect sizes of four subgroups from a categorical moderator. According James, I should remove the intercept from the meta-analysis. I have had different results including or removing the intercept. That is why I am insecure about to use this approach. What do you think? Is it reasonable to remove the intercept in my case?

I also have some questions about how to obtain the intercept and slope from multi-level meta-analysis with two moderators, a categorical moderator and a continuous). How can I estimate the intercept and slope of each subgroup to include the straights in a graph.

Every help is welcome!

Best wishes,

Rafael Rios Moura.
scientia amabilis

PhD in Ecology and Conservation
Postdoctoral Researcher
Universidade Estadual de Campinas (UNICAMP)
Campinas, São Paulo, Brazil
ORCID: http://orcid.org/0000-0002-7911-4734
Research Gate: https://www.researchgate.net/profile/Rafael_Rios_Moura2
Currículo Lattes: http://buscatextual.cnpq.br/buscatextual/visualizacv.do?id=K4244908A8

More information about the R-sig-meta-analysis mailing list