I think it is the difference of which tests are computed. car::Anova uses Wald tests, whereas drop1 refits the model dropping single terms. John Fox once wrote me, that Wald tests and tests from refitted models using likelihood ratio tests (i.e., the strategy from drop1) agree for linear but not necessarily non-linear models. Unfortunately this mail was offlist and did not contain any reference. But I know that his book has a chapter on Wald tests, which could contain the desired info.
The help to car::Anova says:
Type-II tests are calculated according to the
principle of marginality, testing each term after all others, except
ignoring the term's higher-order relatives; so-called type-III tests
violate marginality, testing each term in the model after all of the
others. This definition of Type-II tests corresponds to the tests
produced by SAS for analysis-of-variance models, where all of the
predictors are factors, but not more generally (i.e., when there are
quantitative predictors). Be very careful in formulating the model for
type-III tests, or the hypotheses tested will not make sense.
Unfortunately I cannot answer you second or third question as I also would like to know that.
Update reagrding comment:
There are no Wald, LR and F tests for generalized mixed models. Anova just allows for "chisq" and "F" tests for mixed models (i.e. "mer" objects as returned by lmer). The usage section says:
## S3 method for class 'mer'
Anova(mod, type=c("II","III", 2, 3),
test.statistic=c("chisq", "F"), vcov.=vcov(mod), singular.ok, ...)
But as the F-tests for mer objects are calculated by pbkrtest, which for my knowledge only works for linear mixed models, Anova for GLMMs should always return chisq (hence you see no difference).
Update regarding the question:
My previous answer just tried to respond to your main question, the difference between Anova() and drop1(). But now I understand that you want to test if certainf fixed effects are significant or not. The R-sig-mixed modeling FAQ says the following regarding this:
Tests of single parameters
From worst to best:
- Wald Z-tests
- For balanced, nested LMMs where df can be computed: Wald t-tests
- Likelihood ratio test, either by setting up the model so that the parameter can be isolated/dropped (via anova or drop1), or via
computing likelihood profiles
- MCMC or parametric bootstrap confidence intervals
Tests of effects (i.e. testing that several parameters are
simultaneously zero)
From worst to best:
- Wald chi-square tests (e.g. car::Anova)
- Likelihood ratio test (via anova or drop1)
- For balanced, nested LMMs where df can be computed: conditional F-tests
- For LMMs: conditional F-tests with df correction (e.g. Kenward-Roger in pbkrtest package)
- MCMC or parametric, or nonparametric, bootstrap comparisons (nonparametric bootstrapping must be implemented carefully to account
for grouping factors)
(emphasis added)
This indicates that your approach of using car::Anova() for GLMMs is generally not recommended, but an approach using MCMC or bootstrap should be used. I don't know if pvals.fnc from the languageR package woks with GLMMs, but it is worth a try.
