I am running the following logistic mixed model in lme4:

AllCond_logme <- glmer(Response ~ Condition + (Condition | Pt_ID), 
                       family = binomial(), data = lme.df,
                       control = glmerControl(optimizer = "bobyqa"))

where Response is a binary outcome (1 vs 0).

I computed repeated contrasts and applied them to the Condition predictor, which is a factor with 4 levels (0, 25%, 50%, 75%). The contrasts are:

  1. cH01: 25% - 0
  2. cH02: 50% - 25%
  3. cH03: 75% - 50%

To create the contrasts, I first computed the following hypothesis matrix:

0 25% 50% 75%
Intercept 0.25 0.25 0.25 0.25
cH01 -1 1 0 0
cH02 0 -1 1 0
cH03 0 0 -1 1

where the intercept is the mean probability of Response=1 across all conditions.

I then use ginv() to convert to a contrast matrix and apply this to 'Condition'.

*Note: I now realise that this was equivalent to using the contr.sdif() function in the MASS package.

The mixed model results are as follows:

Fixed effects:
              Estimate Std. Error z value Pr(>|z|)    
(Intercept)    -0.1197     0.1021  -1.173    0.241    
ConditioncH01   1.6971     0.1587  10.692   <2e-16 ***
ConditioncH02   1.9622     0.1108  17.713   <2e-16 ***
ConditioncH03   1.5400     0.1370  11.240   <2e-16 ***

The results seem plausible except for my result for the intercept.

I believe that the intercept should represent the mean probability of Response=1 across all conditions, and the mixed model tests whether this is significantly different to 0. The probability of Response=1 should be significantly different to 0 (across all conditions, about 40% of responses are Response=1). Is my interpretation of the intercept correct, and why is it not significant?

I've also attached a plot of Response= 1 ('Tone Present' response) across the conditions as a visual aid to understand the analysis and also to demonstrate how the intercept should differ from 0.

enter image description here


1 Answer 1


To make an attempt at answering my own question...

I believe that the intercept here is indeed the mean (log likelihood) of Response=1 across all conditions. The p-value for the intercept tests whether this significantly differs from 0. While the mean probability of Response=1 across conditions is ~48%, the standard deviation for this value is ~35%. I think this large standard deviation may be the cause of the non-significant intercept in this case.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.