Skip to content
Prev 16450 / 20628 Next

High correlation among random effects for longitudinal model

Dear John,

Thank you so much--will use poly in the future (even in cases in which I
might *not* use orthogonal polynomials), as it appears that the summary
function returns helpful output (i.e., correlations fo the fixed effects),
in addition to its other benefits.

When I use the poly function, the random effects correlations are lower:

Random effects:
 Formula: ~+poly(wave, 2) | student_ID
 Structure: General positive-definite, Log-Cholesky parametrization
               StdDev     Corr
(Intercept)      2.178734 (Intr) p(,2)1
poly(wave, 2)1 279.521839 0.834
poly(wave, 2)2 281.979199 0.751  0.987
Residual        16.292629

As are the fixed effects:

Fixed effects: stwm ~ +poly(wave, 2)
                   Value Std.Error   DF   t-value p-value
(Intercept)      5.30155  0.231096 7070 22.940957       0
poly(wave, 2)1 196.33516 23.481935 7070  8.361115       0
poly(wave, 2)2 113.69005 23.591988 7070  4.819011       0
 Correlation:
               (Intr) p(,2)1
poly(wave, 2)1 0.344
poly(wave, 2)2 0.311  0.516

However, when I calculate individual (i.e., group)-specific predicted
values (i.e., BLUPs, using the predict() method, with level = 1), they are
(very) highly correlated:

# A tibble: 3 x 4
  rowname   intercept linear quadratic
  <chr>         <dbl>  <dbl>     <dbl>
1 intercept    NA      0.960     0.960
2 linear        0.960 NA         1.00
3 quadratic     0.960  1.00     NA

When I calculate the same individual-specific predicted values using the
non-orthogonal (raw) polynomials, these correlations are very nearly as
high. At this point, I'm curious how / why these predictions are so highly
correlated.

Thank you again for your help, John, and for yours, Ben, Stuart, Jorg, and
Thierry. Sorry if this is a beginner or otherwise ignorant question as I
learn to work with these longitudinal / polynomial models.
josh
On Tue, Apr 3, 2018 at 1:32 PM, Fox, John <jfox at mcmaster.ca> wrote: