3
$\begingroup$

Let's say each participant does multiple trials in a study, where x represents trial difficulty and y represents the score on that trial. Since the participant is represented by data points spread across the x and y scales, I can see the sense in using a linear mixed-effects regression, modelling random slopes and intercepts for participant.

But what if each participant is represented by just one x value, and several y values? Let's say the x is participant age and y is participant score on a series of trials. It seems nonsensical now to model participant as a random effect - if their data points are just represented by one x value I don't see how they could be modelled as a random slope or intercept.

  1. Am I right in thinking an LMER with participant as random effect is not appropriate in such a case? (putting aside the question of whether to model trial as a random effect)

  2. Is the 0 variance for participant in the random effects matrix below diagnostic of this specific problem when modelling a random intercept?

  3. In the random-intercept example below, a regular lm() has the same output as the lmer(). Will that always be the case in such a situation?

  4. Is the correlation of -1 in the random effects structure below diagnostic of this problem with modelling a random slope?

# Generate random data
df = data.frame(participant = rep(letters[seq( from = 1, to = 8)], each=3),
     x = rep(c(1,2,3,4), each=6), 
     offset = runif(24, -0.5, 0.5)) 
 %>% mutate(y=x+offset)

# Model random intercept
Linear mixed model fit by REML ['lmerMod']
Formula: y ~ x + (1 | participant)
   Data: df

REML criterion at convergence: 17.1

Scaled residuals: 
     Min       1Q   Median       3Q      Max 
-1.45511 -0.85896 -0.03722  0.72266  1.63586 

Random effects:
 Groups      Name        Variance Std.Dev.
 participant (Intercept) 0.00000  0.0000  
 Residual                0.09449  0.3074  
Number of obs: 24, groups:  participant, 8

Fixed effects:
            Estimate Std. Error t value
(Intercept) -0.11465    0.15370  -0.746
x            1.02373    0.05612  18.241

Correlation of Fixed Effects:
  (Intr)
x -0.913

# Model random slope
Linear mixed model fit by REML ['lmerMod']
Formula: y ~ x + (1 + x | participant)
   Data: df

REML criterion at convergence: 17.1

Scaled residuals: 
     Min       1Q   Median       3Q      Max 
-1.45511 -0.85896 -0.03722  0.72266  1.63586 

Random effects:
 Groups      Name        Variance  Std.Dev.  Corr 
 participant (Intercept) 6.469e-16 2.543e-08      
             x           4.309e-17 6.564e-09 -1.00
 Residual                9.449e-02 3.074e-01      
Number of obs: 24, groups:  participant, 8

Fixed effects:
            Estimate Std. Error t value
(Intercept) -0.11465    0.15370  -0.746
x            1.02373    0.05612  18.241

Correlation of Fixed Effects:
  (Intr)
x -0.913
$\endgroup$
1
$\begingroup$

Just because each subject only receives one level of a particular treatment doesn't mean a random intercept by subject is inappropriate. It would be completely useless to do so if there is was no systematic variation by subject or if there was no replication, i.e. for any level of the treatment there was at most one subject receiving it.

In the case of your output, there is no subject to subject variation, so it isn't surprising that it is estimated to be zero. Here is a modified snippet that serves as a counterexample:

set.seed(1)
n=100
r=3
df = data.frame(participant = rep(seq( from = 1, to = n), each=r),
                p_int = rep(runif(n, -.5,.5), each=r),
                x = rep(1:(n/2), each=r*2), 
                offset = runif(300, -0.5, 0.5)) %>% 
  mutate(y = x+offset+p_int)

lmer(y ~ x + (1 | participant), data=df)
Linear mixed model fit by REML ['lmerMod']
Formula: y ~ x + (1 | participant)
   Data: df
REML criterion at convergence: 248.3894
Random effects:
 Groups      Name        Std.Dev.
 participant (Intercept) 0.2714  
 Residual                0.2884  
Number of obs: 300, groups:  participant, 100
Fixed Effects:
(Intercept)            x  
    0.02259      0.99923  

On the other hand, a random slope is useless because there is no way to measure a slope when subsetting on a single participant. The negative one estimated covariance that you observe may be due to the random effects being effectively redundant, but I don't know if this is a reliable diagnostic.

$\endgroup$

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.