R help - Aug 2010 - what does it mean when my main effect 'disappears' when using lme4?

Hello,

Setup: I have data with ~10K observations. Observations come from 16
different laboratories (labs). I am interested in how a continuous factor,
X, affects my dependent variable, Y, but there are big differences in the
variance and mean across labs.

I run this model, which controls for mean but not variance differences
between the labs:
lm(Y ~ X + as.factor(labs)).
The effect of X is highly significant (p < .00001)

I then run this model using lme4:
lmer(Y~ X + (1|labs)) #controls for mean diffs bw labs
lmer(Y~X + (X|labs)) #and possible slope heterogeneity bw labs.

For both of these latter models, the effect of X is non-significant (|t| <
1.5).

What might this be telling me about my data? I guess the second (X|labs) may
tell me that there are big differences in the slope across labs, and that
the slope isn't significant against the backdrop of 16 slopes that differ
quite a bit between each other. Is that right? (Still, the enormous drop in
p-value is surprising!). I'm not clear on why the first (1|labs), however,
is so discrepant from just controlling for the mean effects of labs.

Any help in interpreting these data would be appreciated. When I first saw
the data, I jumped for joy, but now I'm muddled and uncertain if I'm
overlooking something. Is there still room for optimism (with respect to X
affecting Y)?

JJ

	[[alternative HTML version deleted]]

One difference is that the random effect in lmer is assumed --
implicitly constrained, as I understand it -- to
be a bell curve.  The fixed effect model does not have that constraint.

How are the values of "labs" effects distributed in your lm model?

On Tue, Aug 17, 2010 at 8:50 PM, Johan Jackson
<johan.h.jackson at gmail.com> wrote:
> Hello,
>
> Setup: I have data with ~10K observations. Observations come from 16
> different laboratories (labs). I am interested in how a continuous factor,
> X, affects my dependent variable, Y, but there are big differences in the
> variance and mean across labs.
>
> I run this model, which controls for mean but not variance differences
> between the labs:
> lm(Y ~ X + as.factor(labs)).
> The effect of X is highly significant (p < .00001)
>
> I then run this model using lme4:
> lmer(Y~ X + (1|labs)) #controls for mean diffs bw labs
> lmer(Y~X + (X|labs)) #and possible slope heterogeneity bw labs.
>
> For both of these latter models, the effect of X is non-significant (|t|
<
> 1.5).
>
> What might this be telling me about my data? I guess the second (X|labs)
may
> tell me that there are big differences in the slope across labs, and that
> the slope isn't significant against the backdrop of 16 slopes that
differ
> quite a bit between each other. Is that right? (Still, the enormous drop in
> p-value is surprising!). I'm not clear on why the first (1|labs),
however,
> is so discrepant from just controlling for the mean effects of labs.
>
> Any help in interpreting these data would be appreciated. When I first saw
> the data, I jumped for joy, but now I'm muddled and uncertain if
I'm
> overlooking something. Is there still room for optimism (with respect to X
> affecting Y)?
>
> JJ
>
> ? ? ? ?[[alternative HTML version deleted]]
>
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>