Description Usage Arguments Details Value Note Author(s) References See Also Examples
Fits and calculates p-values for all effects in a mixed model fitted with lmer
. The default behavior calculates type 3 like p-values using the Kenward-Roger approximation for degrees-of-freedom implemented in KRmodcomp
(for LMMs only), but also allows for parametric bootstrap (method = "PB"
), or likelihood ratio tests (the latter two for LMMs and GLMMs). print
, summary
, and anova
methods for the returned object of class "mixed"
are available (the last two return the same data.frame).
1 2 3 4 5 | mixed(formula, data, type = afex_options("type"),
method = afex_options("method_mixed"), per.parameter = NULL,
args.test = list(), test.intercept = FALSE,
check.contrasts = afex_options("check.contrasts"), set.data.arg = TRUE,
progress = TRUE, cl = NULL, ...)
|
formula |
a formula describing the full mixed-model to be fitted. As this formula is passed to |
data |
data.frame containing the data. Should have all the variables present in |
type |
type of test on which effects are based. Default is to use type 3 tests, taken from |
method |
character vector indicating which methods for obtaining p-values should be used: |
per.parameter |
|
args.test |
|
test.intercept |
logical. Whether or not the intercept should also be fitted and tested for significance. Default is |
check.contrasts |
|
set.data.arg |
|
progress |
if |
cl |
A vector identifying a cluster; used for distributing the estimation of the different models using several cores. See examples. If |
... |
further arguments (such as |
For an introduction to mixed-modeling for experimental designs see Barr, Levy, Scheepers, & Tily (2013; I highly recommend reading this paper if you use this function), arguments for using the Kenward-Roger approximation for obtaining p-values are given by Judd, Westfall, and Kenny (2012). Further introductions to mixed-modeling for experimental designs are given by Baayen and colleagues (Baayen, 2008; Baayen, Davidson & Bates, 2008; Baayen & Milin, 2010). Specific recommendations on which random effects structure to specify for confirmatory tests can be found in Barr and colleagues (2013) and Barr (2013).
p-values are per default calculated via methods from pbkrtest. When method = "KR"
(the default), the Kenward-Roger approximation for degrees-of-freedom is calculated using KRmodcomp
, which is only applicable to linear-mixed models. The test statistic in the output is a F-value (F
).
method = "PB"
calculates p-values using parametric bootstrap using PBmodcomp
. This can be used for linear and also generalized linear mixed models (GLMM) by specifying a family
argument to mixed
. Note that you should specify further arguments to PBmodcomp
via args.test
, especially nsim
(the number of simulations to form the reference distribution) or cl
(for using multiple cores). For other arguments see PBmodcomp
. Note that REML
(argument to [g]lmer
) will be set to FALSE
if method is PB
.
method = "LRT"
calculates p-values via likelihood ratio tests implemented in the anova
method for "merMod"
objects. This is recommended by Barr et al. (2013; which did not test the other methods implemented here). Using likelihood ratio tests is only recommended for models with many levels for the random effects (> 50), but can be pretty helpful in case the other methods fail (due to memory and/or time limitations). The lme4 faq also recommends the other methods over likelihood ratio tests.
Type 3 tests are obtained by comparing a model in which only the tested effect is excluded with the full model (containing all effects). This corresponds to the (type 3) Wald tests given by car::Anova
for "lmerMod"
models. The submodels in which the tested effect is excluded are obtained by manually creating a model matrix which si then fitted in "lme4"
. This is done to avoid R's "feature" to not allow this behavior.
Type 2 tests are truly sequential. They are obtained by comparing a model in which the tested effect and all higher oder effect (e.g., all three-way interactions for testing a two-way interaction) are excluded with a model in which only effects up to the order of the tested effect are present and all higher order effects absent. In other words, there are multiple full models, one for each order of effects. Consequently, the results for lower order effects are identical of whether or not higher order effects are part of the model or not. This latter feature is not consistent with classical ANOVA type 2 tests but a consequence of the sequential tests (and I didn't find a better way of implementing the Type 2 tests). This does not correspond to the (type 2) Wald test reported by car::Anova
. If you want type 2 Wald tests instead of truly sequential typde 2 tests, use car::Anova
with test = "F"
. Note that the order in which the effects are entered into the formula does not matter (in contrast to type 1 tests).
If check.contrasts = TRUE
, contrasts will be set to "contr.sum"
for all factors in the formula if default contrasts are not equal to "contr.sum"
or attrib(factor, "contrasts") != "contr.sum"
. Furthermore, the current contrasts (obtained via getOption("contrasts")
) will be set at the cluster nodes if cl
is not NULL
.
An object of class "mixed"
(i.e., a list) with the following elements:
anova_table
a data.frame containing the statistics returned from KRmodcomp
. The stat
column in this data.frame gives the value of the test statistic, an F-value for method = "KR"
and a chi-square value for the other two methods.
full.model
the "lmerMod"
object returned from fitting the full mixed model.
restricted.models
a list of "lmerMod"
objects from fitting the restricted models (i.e., each model lacks the corresponding effect)
tests
a list of objects returned by the function for obtaining the p-values.
type
The type
argument used when calling this function.
method
The method
argument used when calling this function.
Two similar methods exist for objects of class "mixed"
: print
and anova
. They print a nice version of the anova_table
element of the returned object (which is also invisibly returned). This methods omit some columns and nicely round the other columns. The following columns are always printed:
Effect
name of effect
p.value
estimated p-value for the effect
For LMMs with method="KR"
the following further columns are returned (note: the Kenward-Roger correction does two separate things: (1) it computes an effective number for the denominator df; (2) it scales the statistic by a calculated amount, see also http://stackoverflow.com/a/25612960/289572):
F
computed F statistic
ndf
numerator degrees of freedom (number of parameters used for the effect)
ddf
denominator degrees of freedom (effective residual degrees of freedom for testing the effect), computed from the Kenward-Roger correction using pbkrtest::KRmodcomp
F.scaling
scaling of F-statistic computing from Kenward-Roger approximation.
For models with method="LRT"
the following further columns are returned:
df.large
degrees of freedom (i.e., estimated paramaters) for full model (i.e., model containing the corresponding effect)
df.small
degrees of freedom (i.e., estimated paramaters) for restricted model (i.e., model without the corresponding effect)
chisq
2 times the difference in likelihood (obtained with logLik
) between full and restricted model
df
difference in degrees of freedom between full and restricted model (p-value is based on these df).
For models with method="PB"
the following further column is returned:
stat
2 times the difference in likelihood (obtained with logLik
) between full and restricted model (i.e., a chi-square value).
The summary
method for objects of class mixed
simply calls summary.merMod
on the full model.
When method = "KR"
, obtaining p-values is known to crash due too insufficient memory or other computational limitations (especially with complex random effects structures). In these cases, the other methods should be used. The RAM demand is a problem especially on 32 bit Windows which only supports up to 2 or 3GB RAM (see R Windows FAQ). Then it is probably a good idea to use methods "LRT" or "PB".
"mixed"
will throw a message if numerical variables are not centered on 0, as main effects (of other variables then the numeric one) can be hard to interpret if numerical variables appear in interactions. See Dalal & Zickar (2012).
Please report bugs or unexpected behavior to maintainer via email!
Henrik Singmann with contributions from Ben Bolker and Joshua Wiley.
Baayen, R. H. (2008). Analyzing linguistic data: a practical introduction to statistics using R. Cambridge, UK; New York: Cambridge University Press.
Baayen, R. H., Davidson, D. J., & Bates, D. M. (2008). Mixed-effects modeling with crossed random effects for subjects and items. Journal of Memory and Language, 59(4), 390-412. doi:10.1016/j.jml.2007.12.005
Baayen, R. H., & Milin, P. (2010). Analyzing Reaction Times. International Journal of Psychological Research, 3(2), 12-28.
Barr, D. J. (2013). Random effects structure for testing interactions in linear mixed-effects models. Frontiers in Quantitative Psychology and Measurement, 328. doi:10.3389/fpsyg.2013.00328
Barr, D. J., Levy, R., Scheepers, C., & Tily, H. J. (2013). Random effects structure for confirmatory hypothesis testing: Keep it maximal. Journal of Memory and Language, 68(3), 255-278. doi:10.1016/j.jml.2012.11.001
Dalal, D. K., & Zickar, M. J. (2012). Some Common Myths About Centering Predictor Variables in Moderated Multiple Regression and Polynomial Regression. Organizational Research Methods, 15(3), 339-362. doi:10.1177/1094428111430540
Judd, C. M., Westfall, J., & Kenny, D. A. (2012). Treating stimuli as a random factor in social psychology: A new and comprehensive solution to a pervasive but largely ignored problem. Journal of Personality and Social Psychology, 103(1), 54-69. doi:10.1037/a0028347
Maxwell, S. E., & Delaney, H. D. (2004). Designing experiments and analyzing data: a model-comparisons perspective. Mahwah, N.J.: Lawrence Erlbaum Associates.
aov_ez
and aov_car
for convenience functions to analyze experimental deisgns with classical ANOVA or ANCOVA wrapping Anova
.
see the following for the data sets from Maxwell and Delaney (2004) used and more examples: md_15.1
, md_16.1
, and md_16.4
.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 | ### replicate results from Table 15.4 (Maxwell & Delaney, 2004, p. 789)
data(md_15.1)
# random intercept plus slope
(t15.4a <- mixed(iq ~ timecat + (1+time|id),data=md_15.1))
# to also replicate exact parameters use treatment.contrasts and the last level as base level:
contrasts(md_15.1$timecat) <- contr.treatment(4, base = 4)
(t15.4b <- mixed(iq ~ timecat + (1+time|id),data=md_15.1, check.contrasts=FALSE))
summary(t15.4a) # gives "wrong" parameters extimates
summary(t15.4b) # identical parameters estimates
# for more examples from chapter 15 see ?md_15.1
### replicate results from Table 16.3 (Maxwell & Delaney, 2004, p. 837)
data(md_16.1)
# original results need treatment contrasts:
(mixed1_orig <- mixed(severity ~ sex + (1|id), md_16.1, check.contrasts=FALSE))
summary(mixed1_orig$full.model)
# p-value stays the same with afex default contrasts (contr.sum),
# but estimates and t-values for the fixed effects parameters change.
(mixed1 <- mixed(severity ~ sex + (1|id), md_16.1))
summary(mixed1$full.model)
# data for next examples (Maxwell & Delaney, Table 16.4)
data(md_16.4)
str(md_16.4)
### replicate results from Table 16.6 (Maxwell & Delaney, 2004, p. 845)
# Note that (1|room:cond) is needed because room is nested within cond.
# p-value (almost) holds.
(mixed2 <- mixed(induct ~ cond + (1|room:cond), md_16.4))
# (differences are dut to the use of Kenward-Roger approximation here,
# whereas M&W's p-values are based on uncorrected df.)
# again, to obtain identical parameter and t-values, use treatment contrasts:
summary(mixed2) # not identical
# prepare new data.frame with contrasts:
md_16.4b <- within(md_16.4, cond <- C(cond, contr.treatment, base = 2))
str(md_16.4b)
# p-value stays identical:
(mixed2_orig <- mixed(induct ~ cond + (1|room:cond), md_16.4b, check.contrasts=FALSE))
summary(mixed2_orig$full.model) # replicates parameters
### replicate results from Table 16.7 (Maxwell & Delaney, 2004, p. 851)
# F-values (almost) hold, p-values (especially for skill) are off
(mixed3 <- mixed(induct ~ cond + skill + (1|room:cond), md_16.4))
# however, parameters are perfectly recovered when using the original contrasts:
mixed3_orig <- mixed(induct ~ cond + skill + (1|room:cond), md_16.4b, check.contrasts=FALSE)
summary(mixed3_orig)
### replicate results from Table 16.10 (Maxwell & Delaney, 2004, p. 862)
# for this we need to center cog:
md_16.4b$cog <- scale(md_16.4b$cog, scale=FALSE)
# F-values and p-values are relatively off:
(mixed4 <- mixed(induct ~ cond*cog + (cog|room:cond), md_16.4b))
# contrast has a relatively important influence on cog
(mixed4_orig <- mixed(induct ~ cond*cog + (cog|room:cond), md_16.4b, check.contrasts=FALSE))
# parameters are again almost perfectly recovered:
summary(mixed4_orig)
## Not run:
# use the obk.long data (not reasonable, no random slopes)
data(obk.long)
mixed(value ~ treatment * phase + (1|id), obk.long)
# Examples for using the per.parammeter argument:
data(obk.long, package = "afex")
obk.long$hour <- ordered(obk.long$hour)
# tests only the main effect parameters of hour individually per parameter.
mixed(value ~ treatment*phase*hour +(1|id), per.parameter = "^hour$", data = obk.long)
# tests all parameters including hour individually
mixed(value ~ treatment*phase*hour +(1|id), per.parameter = "hour", data = obk.long)
# tests all parameters individually
mixed(value ~ treatment*phase*hour +(1|id), per.parameter = ".", data = obk.long)
# example data from package languageR:
# Lexical decision latencies elicited from 21 subjects for 79 English concrete nouns,
# with variables linked to subject or word.
data(lexdec, package = "languageR")
# using the simplest model
m1 <- mixed(RT ~ Correct + Trial + PrevType * meanWeight +
Frequency + NativeLanguage * Length + (1|Subject) + (1|Word), data = lexdec)
m1
## Effect stat ndf ddf F.scaling p.value
## 1 Correct 8.15 1 1627.73 1.00 .004
## 2 Trial 7.57 1 1592.43 1.00 .006
## 3 PrevType 0.17 1 1605.39 1.00 .68
## 4 meanWeight 14.85 1 75.39 1.00 .0002
## 5 Frequency 56.53 1 76.08 1.00 <.0001
## 6 NativeLanguage 0.70 1 27.11 1.00 .41
## 7 Length 8.70 1 75.83 1.00 .004
## 8 PrevType:meanWeight 6.18 1 1601.18 1.00 .01
## 9 NativeLanguage:Length 14.24 1 1555.49 1.00 .0002
# Fitting a GLMM using parametric bootstrap:
require("mlmRev") # for the data, see ?Contraception
gm1 <- mixed(use ~ age + I(age^2) + urban + livch + (1 | district), method = "PB",
family = binomial, data = Contraception, args.test = list(nsim = 10))
#######################
### using multicore ###
#######################
require(parallel)
(nc <- detectCores()) # number of cores
cl <- makeCluster(rep("localhost", nc)) # make cluster
# to keep track of what the function is doindg redirect output to outfile:
# cl <- makeCluster(rep("localhost", nc), outfile = "cl.log.txt")
## There are two ways to use multicore:
# 1. Obtain fits with multicore:
mixed(value ~ treatment*phase*hour +(1|id), data = obk.long, method = "LRT", cl = cl)
# 2. Obtain PB samples via multicore:
mixed(use ~ age + I(age^2) + urban + livch + (1 | district), family = binomial,
method = "PB", data = Contraception, args.test = list(nsim = 10, cl = cl))
## Both ways can be combined:
mixed(use ~ age + I(age^2) + urban + livch + (1 | district), family = binomial,
method = "PB", data = Contraception, args.test = list(nsim = 10, cl = cl), cl = cl)
stopCluster(cl)
## End(Not run)
|
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.