Skip to content

[R-meta] Guidance for interpreting fixed effects in multilevel models

3 messages · Simon Harmel, Jack Solomon

Original
Simon Harmel
# 
Dear Colleagues,

I want to use two multilevel meta-regression models (below), but was
wondering what the correct interpretation of my fixed effects in each one
is? (given that the models' random effects are different).

In both models, *time* is a factor ranging from 0 to 4;  *ktype* is a
factor (0=direct,1=indirect) that can vary between studies and between
groups in each study (but not between outcomes and time points nested
within those groups in each study);  and *treat* is a study-level
continuous variable (# of treatments in each study).

Model 1: rma.mv(yi ~ 0 + time + ktype + treat, random = list(~ ktype |
study, ~time | interaction(study, group, outcome), ~1 | esID),struct =
c("HAR","HAR"))

Model 2: rma.mv(yi ~ 0 + time + ktype + treat, random =  list(~ ktype |
interaction(study,group,outcome), ~time | study, ~1 | esID), struct =
c("UN","UN"))

I highly appreciate your expertise and assistance,
Simon
Simon Harmel
# 
A typo correction:

Model 1: rma.mv(yi ~ 0 + time + ktype + treat, random = list(~ ktype |
study, ~time | interaction(study, group, outcome), ~1 | esID),struct =
c("HCS","HAR"))
On Sun, Aug 8, 2021 at 10:12 PM Simon Harmel <sim.harmel at gmail.com> wrote:

            

      
      
      
    

        

  
  


  


      

    

    
2 days later

    

      
      

        


  

        

      Jack Solomon
    

    

      
      #
    

  


  

      
Dear Simon,

Given that you have not provided any output, I can attempt to answer your
question generally. Based on your data description, you likely have an
output similar to what I'm showing further below.

*time0*: Shows the average true effect at time 0 given ktype of 0 and 0
number of treatments (Right off the bat, you may want to center "treats").
Simply change "at time 0" with another value for time, and you'll get the
interpretation of other time values.

*ktype1*: Shows the difference in average true effects between ktype 1 and
ktype 0 at time 0 given 0 number of treatments (again, treats calls for
centering perhaps by its [rounded] mean, or median)

*treats*: Shows how the magnitude of the average true effect is associated
with the number of treatments employed in studies holding time and treats
at their reference values (0 and 0 respectively). Specifically, controlling
these variables, how size of average true effect changes for 1 unit of
increase in treats.

HTH,
Jack

          estimate   se  zval  pval
time0
time1
time2
time3
time4
ktype1
treat

        
On Sun, Aug 8, 2021 at 10:23 PM Simon Harmel <sim.harmel at gmail.com> wrote: