warning: missing cells for some factors

[statmerkur]

When I work with numeric covariates instead of contrasts in mixed() I get the following warning message:

In createDesignMat(rho) :
  missing cells for some factors (combinations of factors) 
 care must be taken with type III  hypothesis.

I think this is a false positive because setting the same contrasts via set_sum_contrasts() doesn’t produce any warnings:

library(afex)

data(sk2011.1)
d <- aggregate(response ~ id + inference, data = sk2011.1, FUN = mean)
set_sum_contrasts()

contrast_mat <- contr.sum(4)
d$c1 <- contrast_mat[, 1][d$inference]
d$c2 <- contrast_mat[, 2][d$inference]
d$c3 <- contrast_mat[, 3][d$inference]
summary(mixed(response ~ c1 + c2 + c3 + (1|id), d))  # warning displayed
summary(mixed(response ~ inference + (1|id), d))  # runs without warning

Am I right in thinking that the warning can be ignored or is there something wrong with my contrast specification?

[henrik]

The question here really is what you want to achieve. The main goal of mixed is to provide tests of effects, such as main-effects or interactions (also called model terms). So splitting the variable into its part is somehow orthogonal to the intended goal. The reason this is not directly clear from your call is the use of summary. summary gives you output based on the parameters and not based on the terms, as the other functions that deal with mixed objects such as nice or anova (or even print). Invoking one of those also makes the difference between the two calls apparent:

library(afex)
set_sum_contrasts()
data(sk2011.1)
d <- aggregate(response ~ id + inference, data = sk2011.1, FUN = mean)
contrast_mat <- contr.sum(4)
d$c1 <- contrast_mat[, 1][d$inference]
d$c2 <- contrast_mat[, 2][d$inference]
d$c3 <- contrast_mat[, 3][d$inference]
m1 <- mixed(response ~ c1 + c2 + c3 + (1|id), d)
nice(m1)
# Mixed Model Anova Table (Type 3 tests, KR-method)
# 
# Model: response ~ c1 + c2 + c3 + (1 | id)
# Data: d
#   Effect     df        F p.value
# 1     c1 1, 117  7.79 **    .006
# 2     c2 1, 117     0.67     .41
# 3     c3 1, 117 10.52 **    .002
# ---
# Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘+’ 0.1 ‘ ’ 1

m2 <- mixed(response ~ inference + (1|id), d)
nice(m2)
# Mixed Model Anova Table (Type 3 tests, KR-method)
# 
# Model: response ~ inference + (1 | id)
# Data: d
#      Effect     df       F p.value
# 1 inference 3, 117 5.15 **    .002
# ---
# Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘+’ 0.1 ‘ ’ 1

If, for some reason, you are actually interested in p-values for the individual parameters (which seems quite questionable for factors with more than two levels as in the example), you can get this also via the lmerTest summary method:

lmerTest::summary(m2$full_model)
# [...]
# Fixed effects:
#             Estimate Std. Error      df t value Pr(>|t|)    
# (Intercept)   79.141      2.495  39.000  31.725  < 2e-16 ***
# inference1     8.372      3.000 117.000   2.791  0.00614 ** 
# inference2    -2.459      3.000 117.000  -0.820  0.41395    
# inference3    -9.728      3.000 117.000  -3.243  0.00154 ** 
# ---
# Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Note that for this to work you need to invoke set_sum_contrasts() earlier. Alternatively, set set_data_arg = FALSE. One final comment: You might also want to take a look at the per_parameter argument.

[statmerkur]

I am indeed interested in the individual parameters as I have specific hypotheses for different a priori contrasts. Maybe using contr.sum() was not the best example to illustrate this. But thanks for reminding me on the fact that afex is designed to provide tests of (main-)effects – it now became clear to me that using your package for this purpose seems to be inappropriate. However, when I use set_data_arg = FALSE I still get the same results.

[henrik]

There are several different ways to test prespecified contrasts. This can be done in afex in the way described here. As you correctly notice, the p-values of the different methoda re identical and the warnings appear to be inconsequential.

You could also fit the model with normal (i.e., sum-to-zero) contrasts and then setup the contrasts later via emmeans. There are surely further possibilities (e.g., fit the model with lmerTest::lmer and use summary). All of those should give you the same results (as long as you use the same method for calculating the df).

[Author]

Posts