[R-sig-ME] p-values with lmer

[Christoph Scherber]

Dear Doris,

I assume you performed a likelihood ratio test on the two models?

The model containing X1 has the lower AIC, so it is actually pointless to
perform a likelihood ratio test on that model and a model containing both
X1 and X2.

Rather, to assess the significanced of X1, you should compare the model to
a null model (containing the intercept only).

The algorithm is like this:

- Write down your minimal adequate model
- Write down the corresponding null model
- Perform likelihood ratio tests, starting with the null model and adding
terms one at a time until you arrive at the minimal adequate model.

Example:

- minimal adequate model contains X1+X2+X3
- Null model is ~1
- Now do LRT´s with:

~1
~1+X1
~1+X1+X2
~1+X1+X2+X3

And be sure to use;
(1) QAIC in case of overdispersion
(2) AICc in case of small sample sizes

All the best
Christoph









> Dear lmer users,
>
> I compute several models with lmer (poisson distribution, correction for
> overdispersion).
>
> I run several models with the same random effect and different structures
> for
> fixed effect (a full model and different models with only some of the
> factors of
> the full model).
>
> I select the model with the minimal AIC. In fact, two models are within 2
> AIC
> units:
> - a model (with lowest AIC) with one factor, X1.
> - another model (at 0.8 of the first) with two factors, X1 and X2.
>
> Looking at the p-values, X1 is significant in both models while X2 is not
> significant, despite the low AIC associated to the model. How to interpret
> the
> apparent contradiction between the non-significant p-value and the low
> AIC?
>
>
>>From an old message on the list, I can see that p-values  are not good.
> But the bayesian approach mcmcsample does not run either with lmer.
>
> How to get the p-values associated with the fixed effects?
> Thank you for your help
> Doris Gomez
>
>
>
> 	[[alternative HTML version deleted]]
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>