When I fit any model in lmer(), summary identifies a correlation between fixed effect(s) and the (fixed) intercept. How should I interpret this, and why is it not $0$? In my way of thinking any correlation between the intercept and a fixed effect should be zero, since, in the design matrix $X$, the intercept term is just a column of $1$s, and therefore constant. There is a related post on this topic (lmer interpretation of correlation), but it doesn't help me much. Below is model summary output, so you can see what I mean. See the block labeled Correlation of Fixed Effects:, at the very bottom.

Linear mixed model fit by REML ['lmerMod']
Formula: RT_log ~ Condition + (Condition | Subject) + (Condition | Item)
   Data: exp2

REML criterion at convergence: -1978.6

Scaled residuals: 
    Min      1Q  Median      3Q     Max 
-3.7080 -0.5775 -0.0634  0.4801  7.8186 

Random effects:
 Groups   Name            Variance  Std.Dev. Corr       
 Subject  (Intercept)     0.0261472 0.16170             
          Conditionoddman 0.0028821 0.05369  -0.21      
          Conditionhetero 0.0037356 0.06112  -0.46  0.80
 Item     (Intercept)     0.0018914 0.04349             
          Conditionoddman 0.0002885 0.01699  -0.97      
          Conditionhetero 0.0010140 0.03184  -0.65  0.81
 Residual                 0.0320147 0.17893             
Number of obs: 3600, groups:  Subject, 20; Item, 12

Fixed effects:
                 Estimate Std. Error t value
(Intercept)      6.583310   0.038622  170.46
Conditionoddman -0.009109   0.014883   -0.61
Conditionhetero  0.021487   0.018018    1.19

Correlation of Fixed Effects:
            (Intr) Cndtnd
Conditnddmn -0.309       
Conditinhtr -0.472  0.722

This is a general issue of parameter estimates and not limited to mixed models (compare How to interpret parameter estimates correlated with the intercept parameter estimate?).

For your case with the $1$-vector and one covariate $x$ (both with length $n$), let's define the $n \times 2$ matrix $X = \begin{bmatrix}1 & x\end{bmatrix}$. The covariance for the fixed effects is $\sigma \cdot (X'X)^{-1}$ (see for example eq. 34 in https://web.stanford.edu/~mrosenfe/soc_meth_proj3/matrix_OLS_NYU_notes.pdf). In your case this would be $\sigma \cdot \begin{bmatrix} \sum x^2 & -\sum x \\ -\sum x & n\end{bmatrix}$. Here you see the clear correlation between the estimates for the intercept and slope parameter (and also that it becomes $0$ if you center $x$ around $0$).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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