Skip to content
Prev 16615 / 20628 Next

What is the appropriate zero-correlation parameter model for factors in lmer?

I have to clarify that I was talking about fm5 in the way Rune Haubo
specified it:
fm5.2 = y ~ 1 + f + (1 | g) + (0 + f || g) which is, when using a
treatment coded factor with 3 levels, equivalent to
fm5.2 = y ~ 1 + f + (1 | g ) +
            (0 + dummy(f, "1st level") | g) +
            (0 + dummy(f, "2nd level") | g) +
            (0 + dummy(f, "3rd level") | g)
In the way Reinhold Kliegl specified fm5, i.e. with  (0 + f | g)
instead of (0 + f || g), it seems to me that fm5 is just an
over/re-parameterized version of fm6 with one additional parameter and
they both yield the same fit.

However, I think both fm5.2 and fm5 are difficult to understand
because they use (1 | g) and, at the same time, the 0 + f notation
within one formula. So my question remains the same: as Reinhold
Kliegl put it "if the factor levels are A, B, C, and the two contrasts
are c1 and c2, I thought I can specify either (1 + c1 + c2) or (0 + A
+ B + C)". But this doesn't seem to be the case for random effects -
or is it?
Maybe answering this question can also explain the difference between
fm3 and fm7.

Best regards,
Maarten

On Tue, May 22, 2018 at 1:17 PM, Reinhold Kliegl
<reinhold.kliegl at gmail.com> wrote: