Skip to content

Random intercepts and/or slopes

4 messages · Maria Helena Mourino Silva Nunes, Gabriel Baud-Bovy, Viechtbauer Wolfgang (STAT)

Original
Maria Helena Mourino Silva Nunes
# 
Dear all,

I?m new to mixed models, and I?m analysing a patients? data set whose response variable is adherence/no adherence to medication (binary variable, 0/1). For each patient there are different types of independent variables: sociodemographic; clinical; blood pressure control; etc. Additionally, patients are grouped (nested) within doctors. Due to the sampling design, I only have one replicate for each patient.

I want to study the impact of the independent variables on the adherence/no adherence to medication.  I?m not interested in studying the patients by themselves.
My question is: should I consider the simplest model     response  ~  Indep. Variables +  (1|Doctor),
or  should I explicitly introduce the nested structure of patient within doctor,   response ~ Indep. Variables + (Patient|Doctor) ?


Thanks in advanced!

Best regards,
Helena.
Gabriel Baud-Bovy
#
On 10/10/2013 12:39 PM, Maria Helena Mourino Silva Nunes wrote:
Hi Helena,

Here a some syntax for your random effects assuming that all IV are
between-subject factors

  response  ~  Indep. Variables +   (1|Patient)

response ~ Indep. Variables + (1|Doctor) + (1|Patient)

The last one allow to define random effects for differences
across doctors as well as patients (with lmer, patients need to be uniquely
identified)

If you have repeated-measure (within-subject factors), then
things become more complex as the effect of some factor might
also be different across patients.  With lme, one advise in this
case is to include a maximal random structures (Barr et al., 2012, 2013)

(all within subject factor |Patient)

it one is interested only in fixed effects. However, other expert
recommend introducing random effects only for factors with a sufficient
large number of levels because variance estimates might not be
reliable. In any case, the situation with binomial model might be
different.

As you will see, there is not (yet) a generally agreed methodology
with mixed-effect models, the way you have it with old-fashioned ANOVA.
It takes a deep understanding of the methods and your data to
decide exactly what to do. There are now quite a bit of books and
papers on GLMM that might help to learn more on this topic.

Best,

Gabriel

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Viechtbauer Wolfgang (STAT)
    

    

      
      #
    

  


  

      
Since there is only a single replicate per patient, it is not possible to estimate the patient-level variance in these models:

response ~ Indep. Variables + (1|Patient)
response ~ Indep. Variables + (1|Doctor) + (1|Patient)

So, the only model that is really applicable here is:

response ~ Indep. Variables + (1|Doctor)

This model accounts for the nesting of patients within doctors.

Best,
Wolfgang

--   
Wolfgang Viechtbauer, Ph.D., Statistician   
Department of Psychiatry and Psychology   
School for Mental Health and Neuroscience   
Faculty of Health, Medicine, and Life Sciences   
Maastricht University, P.O. Box 616 (VIJV1)   
6200 MD Maastricht, The Netherlands   
+31 (43) 388-4170 | http://www.wvbauer.com

            

      
      
      
    

  
  


  


      

    

    

      
      

        


  

        

      Maria Helena Mourino Silva Nunes
    

    

      
      #
    

  


  

      
Dear list members
Dear Wolfgang, Gabriel and Hans,

thanks a lot for your suggestions! It has been very useful.

I will use the model   response ~ Indep. Variables + (1|Doctor)   because I only have one set of observations for each patient (response variable + independent variables).


Best regards,
Helena.