The basic SEM supported by lessSEM assumes that the data is independently and identically distributed. That is, each subject in the data set comes from the same population. This assumption may be unrealistic, however. Researchers may suspect that subgroups within the data set are more similar to one another than to other subgroups, for example. That is, they differ in their parameter vectors.
If possible groupings within the data set are known beforehand, multi-group models are a convenient way to allow for group-specific parameters. Setting up such models with lavaan is explained here. Unfortunately, lessSEM does not support the same syntax at the moment. “Throwing” a multi-group SEM into lessSEM will just result in errors. Instead, lessSEM follows a slightly different approach: You can pass multiple lavaan models at once which are then combined into a multi-group model.
In the following, we will look at a two-group model to better understand how multi-group models are implemented in lessSEM.
To set up a multi-group model in lessSEM, we first have to fit separate models for each of the groups in lavaan:
library(lavaan)
# For simplicity, we will use a subset of the Holzinger Swineford data set
# which is also used at https://lavaan.ugent.be/tutorial/groups.html
# to demonstrate multi-group SEM
# To use mutli-group SEM in lessSEM, we have to set up a separate model
# for each of the groups:
# - Pasteur: Children attending the Pasteur school
# - Grant_White: Children attending the Grant-White school
data(HolzingerSwineford1939)
## Pasteur ##
Pasteur <- subset(HolzingerSwineford1939, school == "Pasteur")
model_Pasteur <- paste0('
visual =~ l1_Pasteur*x1 + l2_Pasteur*x2 + l3_Pasteur*x3
x1 ~~ v1*x1
x2 ~~ v2*x2
x3 ~~ v3*x3
visual ~~ lv1*visual
x1 ~ m1*1
x2 ~ m2*1
x3 ~ m3*1')
fit_Pasteur <- sem(model = model_Pasteur,
data = Pasteur,
std.lv = TRUE)
## Grant-White
Grant_White <- subset(HolzingerSwineford1939, school == "Grant-White")
model_Grant_White <- paste0('
visual =~ l1_Grant_White*x1 + l2_Grant_White*x2 + l3_Grant_White*x3
x1 ~~ v1*x1
x2 ~~ v2*x2
x3 ~~ v3*x3
visual ~~ lv1*visual
x1 ~ m1*1
x2 ~ m2*1
x3 ~ m3*1')
fit_Grant_White <- sem(model = model_Grant_White,
data = Grant_White,
std.lv = TRUE)Now that we have our group-specific models, we can pass them to lessSEM:
library(lessSEM)
# We will just estimate the parameters using the BFGS optimizer without any
# regularization.
# Note that we pass the two models as a vector. This is very important: lessSEM
# will then set up the multi-group model
fit <- bfgs(lavaanModel = c(fit_Pasteur, fit_Grant_White))Let’s have a look at the parameters:
coef(fit)
#>
#> Tuning ||--|| Estimates
#> ------- ------- ||--|| ---------- ---------- ---------- ---------- ----------
#> lambda alpha ||--|| l1_Pasteur l2_Pasteur l3_Pasteur v1 v2
#> ======= ======= ||--|| ========== ========== ========== ========== ==========
#> 0.0000 0.0000 ||--|| 0.7240 0.5610 0.8824 0.8449 1.0711
#>
#>
#> ---------- ---------- ---------- ---------- -------------- --------------
#> v3 m1 m2 m3 l1_Grant_White l2_Grant_White
#> ========== ========== ========== ========== ============== ==============
#> 0.6108 4.9212 6.0770 2.2281 0.7088 0.5536
#>
#>
#> --------------
#> l3_Grant_White
#> ==============
#> 0.7360That’s curious! There are group-specific parameters, but only for the parameters where we provided group-specific names! This is very important:
Important: If you set up a multi-group model with lessSEM, lessSEM will assume that all parameters with the same names should also have the same values. This includes parameters which you may have estimated, but for whom the names were provided by lavaan (e.g., variances).
All multi-group models can be regularized similar to the standard
SEM: Instead of using the bfgs-function, we use (for
instance), the lasso-function:
fit <- lasso(lavaanModel = c(fit_Pasteur, fit_Grant_White),
regularized = c("l1_Pasteur"),
nLambdas = 20)The coefficients can be extracted as usual:
coef(fit, criterion = "AIC")
#>
#> Tuning ||--|| Estimates
#> ------- ------- ||--|| ---------- ---------- ---------- ---------- ----------
#> lambda alpha ||--|| l1_Pasteur l2_Pasteur l3_Pasteur v1 v2
#> ======= ======= ||--|| ========== ========== ========== ========== ==========
#> 0.0000 1.0000 ||--|| 0.7241 0.5611 0.8822 0.8447 1.0710
#>
#>
#> ---------- ---------- ---------- ---------- -------------- --------------
#> v3 m1 m2 m3 l1_Grant_White l2_Grant_White
#> ========== ========== ========== ========== ============== ==============
#> 0.6110 4.9211 6.0768 2.2279 0.7089 0.5537
#>
#>
#> --------------
#> l3_Grant_White
#> ==============
#> 0.7358Where regularized multi-group models shine is when automatically
testing for group-differences. This was proposed by Huang (2018) and
provides a convenient way to decide which of the parameters should be
group-specific. To this end, differences between parameters must be
regularized. Say, we are interested in the loading l1 and
wonder if we do indeed need separate loadings for students attending the
Pasteur school (l1_Pasteur) and the Grant-White school
(l1_Grant_White). Using the Pasteur school as baseline
group (see Huang, 2018, for more details), we can define
l1_Grant_White = l1_Pasteur + l1_delta, where
l1_delta is the difference between the two schools. If
l1_delta is zero, then both schools have the same loading
(i.e., we have measurement invariance). Within lessSEM,
we can regularize such differences using transformations (see
vignette(topic = "Parameter-transformations", package = "lessSEM")
for more details). Therefore, the first step is to define the
transformation:
transformation <- "
parameters: l1_Pasteur, l1_Grant_White, l1_delta
l1_Grant_White = l1_Pasteur + l1_delta;
"Next, we pass this transformation to our model:
fit <- lasso(lavaanModel = c(fit_Pasteur, fit_Grant_White),
regularized = c("l1_delta"), # we want to regularize the difference!
nLambdas = 20,
modifyModel = modifyModel(transformations = transformation))Now, let’s look at the parameter estimates:
coef(fit, criterion = "AIC")@estimates[,c("l1_Pasteur", "l1_delta")]
#> l1_Pasteur l1_delta
#> 0.716718 0.000000As the l1_delta parameter has been set to zero, we can
assume measurement invariance. Note that you won’t find
l1_Grant_White in the parameters of the model. This is
because l1_Grant_White is a deterministic function of the
actual parameters l1_Pasteur and l1_delta. If
you want to find the value for l1_Grant_White, have a look
at:
fit@transformations
#> lambda alpha l1_Grant_White
#> 1 0.0056517504 1 0.7167576
#> 2 0.0053542898 1 0.7167180
#> 3 0.0050568293 1 0.7166009
#> 4 0.0047593687 1 0.7161368
#> 5 0.0044619082 1 0.7155761
#> 6 0.0041644476 1 0.7151688
#> 7 0.0038669871 1 0.7147807
#> 8 0.0035695265 1 0.7141878
#> 9 0.0032720660 1 0.7137695
#> 10 0.0029746054 1 0.7132645
#> 11 0.0026771449 1 0.7128818
#> 12 0.0023796844 1 0.7125411
#> 13 0.0020822238 1 0.7120493
#> 14 0.0017847633 1 0.7115889
#> 15 0.0014873027 1 0.7110626
#> 16 0.0011898422 1 0.7107554
#> 17 0.0008923816 1 0.7102373
#> 18 0.0005949211 1 0.7097089
#> 19 0.0002974605 1 0.7094042
#> 20 0.0000000000 1 0.7088852Note that lslx (Huang, 2020) supports different
penalties for the delta parameter (l1_delta) and the
baseline parameter (l1_Pasteur). This is currently not
supported by lessSEM.
Automatic cross-validation for multi-group models with, for instance,
cvLasso is not yet implemented. This is because it
can be difficult to decide how to split up the data set in each
submodel. If you want to use cross-validation, you will unfortunately
have to set up the procedure manually.
Models with definition variables are basically the same as multi-group models, with the sole exception that the group-specific parameters are not estimated but fixed to specific values.
If your main interest is in setting up a multi-group SEM with lessSEM and you don’t care about the details, the lessTemplates package (https://github.com/jhorzek/lessTemplates) provides means to easily set up such models (see SEMWithDefinitionVariables function in lessTemplates).
In the following, we will look in detail at how definition variables can be used in lessSEM
Unfortunately, lavaan does not allow us to set up models for \(N=1\), however. This is required for many definition variable applications, such as latent growth curve models with subject-specific measurement occasions. In the following, we will use a workaround.
Let’s first simulate some data:
#### Population parameters ####
intercept_mu <- 0
intercept_sigma <- 1
slope_mu <- .3
slope_sigma <- 1
#### data set ####
N <- 50
intercepts <- rnorm(n = N,
mean = intercept_mu,
sd = intercept_sigma)
slopes <- rnorm(n = N,
mean = slope_mu,
sd = slope_sigma)
times <- matrix(seq(0,5,1),
nrow = N,
ncol = 6,
byrow = TRUE) +
cbind(0,matrix(round(runif(n = N*5, min = -.2,max = .2),2),
nrow = N,
ncol = 5,
byrow = TRUE)) # we add some jitter to make the times person-specific
lgcData <- matrix(NA, nrow = N, ncol = ncol(times), dimnames = list(NULL, paste0("x", 0:5)))
for(i in 1:N){
lgcData[i,] <- intercepts[i] + times[i,]* slopes[i] + rnorm(ncol(lgcData),0,.3)
}
lgcData <- as.data.frame(lgcData)
head(lgcData)
#> x0 x1 x2 x3 x4 x5
#> 1 0.26801876 -0.5353028 -1.0262295 -2.533564 -2.890246 -4.655061
#> 2 0.26260181 -0.4802635 -2.0790754 -2.686983 -3.677022 -4.540501
#> 3 -0.71054447 -0.5477048 -0.5747618 -1.231821 -1.229388 -1.418401
#> 4 0.14564354 1.0690979 1.8919916 2.842025 2.953471 4.654879
#> 5 0.09501816 1.6350899 2.6331440 4.729328 5.423381 6.799296
#> 6 2.51013193 3.0917904 4.1193617 5.749542 6.222599 7.716919
head(times)
#> [,1] [,2] [,3] [,4] [,5] [,6]
#> [1,] 0 0.94 2.13 3.01 3.91 5.12
#> [2,] 0 0.84 2.13 2.91 4.10 5.19
#> [3,] 0 0.83 2.14 3.12 3.95 4.93
#> [4,] 0 0.88 2.03 3.16 4.01 5.03
#> [5,] 0 1.07 2.01 3.00 3.81 4.82
#> [6,] 0 1.17 2.11 2.88 4.06 5.06Note that the times are random and subject-specific. We need a separate model for each subject. Because lavaan won’t let us set up such models, we will instead set up models using the entire data set and replace the data post-hoc.
models <- c()
for(i in 1:N){
model_i <- paste0(
"
int =~ 1*x0 + 1*x1 + 1*x2 + 1*x3 + 1*x4 + 1*x5
slope =~ ",times[i,1],"*x0 +
",times[i,2],"*x1 +
",times[i,3],"*x2 +
",times[i,4],"*x3 +
",times[i,5],"*x4 +
",times[i,6],"*x5
int ~ intMean*1
slope ~ slopeMean*1
int ~~ intVar*int + 0*slope
slope ~~ slopeVar*slope
x0 ~~ v*x0
x1 ~~ v*x1
x2 ~~ v*x2
x3 ~~ v*x3
x4 ~~ v*x4
x5 ~~ v*x5
x0 ~ 0*1
x1 ~ 0*1
x2 ~ 0*1
x3 ~ 0*1
x4 ~ 0*1
x5 ~ 0*1
"
)
fit_i <- sem(model = model_i,
data = lgcData,
do.fit = FALSE)
internalData <- lavInspect(fit_i, "data")
# replace the data set
fit_i@Data@X[[1]] <- as.matrix(lgcData[i,colnames(internalData),drop = FALSE])
models <- c(models,
fit_i)
}Exemplarily, it makes sense to look at one of the models:
cat(model_i)
#>
#> int =~ 1*x0 + 1*x1 + 1*x2 + 1*x3 + 1*x4 + 1*x5
#> slope =~ 0*x0 +
#> 1.13*x1 +
#> 1.81*x2 +
#> 3.06*x3 +
#> 4.12*x4 +
#> 4.98*x5
#>
#> int ~ intMean*1
#> slope ~ slopeMean*1
#>
#> int ~~ intVar*int + 0*slope
#> slope ~~ slopeVar*slope
#>
#> x0 ~~ v*x0
#> x1 ~~ v*x1
#> x2 ~~ v*x2
#> x3 ~~ v*x3
#> x4 ~~ v*x4
#> x5 ~~ v*x5
#>
#> x0 ~ 0*1
#> x1 ~ 0*1
#> x2 ~ 0*1
#> x3 ~ 0*1
#> x4 ~ 0*1
#> x5 ~ 0*1Note that the loadings of the slope are fixed to the time points at
which person i provided data.
Now we can pass the models to lessSEM:
fit <- bfgs(lavaanModel = models)The parameters are given by:
coef(fit)
#>
#> Tuning ||--|| Estimates
#> ------- ------- ||--|| ---------- ---------- ---------- ---------- ----------
#> lambda alpha ||--|| intMean slopeMean intVar slopeVar v
#> ======= ======= ||--|| ========== ========== ========== ========== ==========
#> 0.0000 0.0000 ||--|| -0.0152 0.2921 0.7703 0.9411 0.0851