Skip to content
Prev 12826 / 20628 Next
Original Raw

standard error and statistical significance in lmer versus lm

Peter Claussen 
Vincenzo,

My comments are strictly about the computations, as I understand them, and not about the appropriateness of the models.

First, I want to be clear that I am commenting on your concern about having two different answers. My thinking is that you have two different answers because you  are not asking the same question in each case.

What you seem to be asking is, when using either lm or lmer, is the significance genotype:week interaction different?

But what you are really asking is, if I include probeset in the model, is the significance genotype:week interaction different?

That?s because the model specify for lmer includes probeset, but the model you specify for lm does not. These are two different statistical models, and you would expect different inferences about the significance of genotype:week, regardless of which computational engine you use.

I have not doubt that you are not interested in the specific effect of each gene. Nonetheless, by including probeset in the formula you pass to lmer, you are getting estimates for each effect; they?re just not being reported in the summary for lmer objects. IIRC, you would use ranef to view the for each gene by week (that?s a best guess, I?d really have to see the data to be sure about the interpretation)

You can also model probeset as a random effect using lm. Including a random effect in a linear model does not change the nature of the effect, it simply changes how the effects are estimated - in this case, ordinary least squares vs maximum likelihood. The effective difference is how you would compute F values in the resulting AOV table. aov() in R will only run F-tests against residual error; if you have other random effects in the model you need to compute F values manually.

My suggestion is that you run, on a subset of your data, both
	aov(lm(values ~ week * genotype))
	aov(lm(values ~ week * genotype + week:probesets))
	
then compare the residual error to the residual error of 
	lmer(values ~ week * genotype + (week | probesets))

I suspect the second lm model to be very similar to lmer, and that you might rest assured that you are not getting two different answers. I think your concern arises from not comparing like things.

Peter

Thread (4 messages)

Peter Claussen standard error and statistical significance in lmer versus lm Jan 2 Vincenzo Lagani standard error and statistical significance in lmer versus lm Jan 2 Peter Claussen standard error and statistical significance in lmer versus lm Jan 2 Vincenzo Lagani standard error and statistical significance in lmer versus lm Jan 3