Skip to contents

This guide uses the package’s example data to estimate a four-cluster model.

data <- mmgsem::data_example

Fit the measurement model

The measurement model is estimated using lavaan:

S1_model <- '
    F1 =~ V1 + V2 + V3 + V4 + V5
    F2 =~ V6 + V7 + V8 + V9 + V10
    F3 =~ V11 + V12 + V13 + V14 + V15
    F4 =~ V16 + V17 + V18 + V19 + V20
'

MM_fit <- lavaan::cfa(
  model         = S1_model,
  data          = data,
  group         = "group",
  group.equal   = "loadings",
  group.partial = c("F1 =~ V2",
                    "F2 =~ V7",
                    "F3 =~ V12",
                    "F4 =~ V17")
)

Fit the structural model

The structural model is then clustered by mmgsem():

S2_model <- '
    F4 ~ F1 + F3
    F3 ~ F1 + F2
'

mmgsem_fit <- mmgsem::mmgsem(
  dat     = data,
  S1      = S1_model,
  S2      = S2_model,
  group   = "group",
  nclus   = 4,
  seed    = 1,
  s1_type = "lavaan",
  s1_fit  = MM_fit
)

summary(mmgsem_fit)
print(mmgsem_fit$posteriors)

Select the number of clusters

ModelSelection() fits each cluster count in the range specified by nclus. Its computation can take considerably longer than fitting a single model.

modelSelection_fit <- mmgsem::ModelSelection(
  dat     = data,
  S1      = S1_model,
  S2      = S2_model,
  group   = "group",
  nclus   = c(1, 6),
  seed    = 1,
  s1_type = "lavaan",
  s1_fit  = MM_fit
)

summary(modelSelection_fit, model_selection = TRUE)
selected_fit <- mmgsem::extract(object = modelSelection_fit, nclus = 4)

Compute standard errors

Computing standard errors uses numerical differentiation and can take a long time, especially for larger models. Run this step when you need standard errors; precomputed standard errors for this example data are also available.

# standard_errors <- mmgsem::standard_errors
standard_errors <- mmgsem::compute_se(object = selected_fit)
summary(selected_fit, se = standard_errors)

Longitudinal example

For longitudinal data, lmmgsem() tracks clusters (latent states) across time. Supply a long-format data frame with group and time columns and the same observed indicators at each group-time combination. Every group must have observations at every time point, with at least two time points. Replace longitudinal_data below with your data; adapt the measurement and structural model syntax if your indicators or model differ.

The measurement and structural models from the cross-sectional example can be reused:

mi_results <- mmgsem::compute_mi_long_stepwise_score(
  data = longitudinal_data,
  model = S1_model,
  group_var = "group",
  time_var = "time",
  method = "test_statistic",
  specific_noninv = FALSE
)
s1_fit <- mi_results$final_fit

This estimates a measurement model in which each group-time combination is treated as a separate group. Alternatively, omit s1_fit below to have lmmgsem() estimate the measurement-invariance model automatically.

lmmgsem_fit <- mmgsem::lmmgsem(
  dat = longitudinal_data,
  S1 = S1_model,
  S2 = S2_model,
  group = "group",
  time = "time",
  nclus = 4,
  seed = 1,
  s1_fit = s1_fit
)

summary(lmmgsem_fit)

To compare longitudinal models with different numbers of states, include time in ModelSelection(). The range must not exceed the number of groups.

longitudinal_selection <- mmgsem::ModelSelection(
  dat = longitudinal_data,
  S1 = S1_model,
  S2 = S2_model,
  group = "group",
  time = "time",
  nclus = c(1, 4),
  seed = 1
)

summary(longitudinal_selection, model_selection = TRUE)
selected_longitudinal_fit <- mmgsem::extract(
  longitudinal_selection,
  nclus = 4
)

Standard errors for lmmgsem() are not currently implemented.