Latent Class Analysis for Ordinal Indicators

This is an example of exploratory LCA with ordinal indicators using tidySEM, as explained in Van Lissa, C. J., Garnier-Villarreal, M., & Anadria, D. (2023). Recommended Practices in Latent Class Analysis using the Open-Source R-Package tidySEM. Structural Equation Modeling.

The present example uses synthetic data based on a study by Maene, D’hondt, Van Lissa, Thijs, & Stevens (2022). In a convenience sample of Flemish (Belgian) high-school students with a migration background, this study set out to identify distinct classes based on ordinal indicators of National, Regional, and Heritage identity. Sample size was not justified.

In the original study, the approach to class enumeration was semi-theory driven: The researchers expected to find profiles that were distinct on all three types of identity (national, regional, and heritage) - but the exact number of classes was not pre-specified (hypothesis 1).

Hypothesis 2 stated that adolescents who are nationally integrated would have lower depressive feelings than students from students with other combinations of identifications (hypothesis 2). Hypothesis 3 was that, for assimilated and separated adolescents, there would not be a significant effect of perceived teacher discrimination on depressive symptoms.

Use the command ?tidySEM::maene_identity to view the data documentation.

Loading the Data

To load the data, simply attach the tidySEM package. For convenience, we assign the indicator data to an object called df. As the analyses take a long time to run, this tutorial reduces the data to three identity indicators and collapses their ordered response categories into three levels. This is a teaching simplification, do not do this for substantive analyses: collapsing categories discards information and changes the estimated latent classes.

# Load required packages
library(tidySEM)
library(ggplot2)
# Load data Keep one indicator for each identity domain and
# collapse adjacent categories.  This deliberately smaller
# example makes ordinal mixture estimation faster.
ordinal3 <- function(x) {
    z <- as.integer(x)
    observed <- sort(unique(z[!is.na(z)]))
    if (length(observed) < 3L)
        stop("Indicator has fewer than three observed categories")
    # Collapse ordered response categories into three
    # approximately equal ranges.
    groups <- cut(seq_along(observed), breaks = 3L, labels = FALSE)
    ordered(groups[match(z, observed)], levels = 1:3)
}
df <- maene_identity[c("Ethnic_1", "Belgian", "Flemish")]
df[] <- lapply(df, ordinal3)
# Auxiliary variables must correspond to the same rows as
# the indicators.
aux_data <- maene_identity

Examining the Data

We use tidySEM::descriptives() to describe the data numerically. Because all items are categorical, we remove columns for continuous data to de-clutter the table:

desc <- tidySEM::descriptives(df)
desc <- desc[, c("name", "type", "n", "missing", "unique", "mode",
    "mode_value", "v")]
desc

Additionally, we can plot the data. The ggplot2 function geom_bar() is useful for ordinal data:

df_plot <- df
names(df_plot) <- paste0("Value.", names(df_plot))
df_plot <- reshape(df_plot, varying = names(df_plot), direction = "long")
ggplot(df_plot, aes(x = Value)) + geom_bar() + facet_wrap(~time,
    scales = "free") + theme_bw()

As we can see, the descriptives() table provides invaluable information about the measurement level of the indicators, which is used to specify the model correctly. If these data had not been coded as ordinal factors, these descriptive statistics would have revealed that each variable has only a few distinct values. The proportion of missing values is reported in the "missing" column. If any variables had missing values, we would report an MCAR test with mice::mcar() and explain that missing data are accounted for using FIML. In our example, we inspect missingness before proceeding with our analysis.

Conducting Latent Class Analysis

Before we fit a series of LCA models, we set a random seed using set.seed(). This is important because there is some inherent randomness in the estimation procedure, and using a seed ensures that we (and others) can exactly reproduce the results.

Next, we fit the LCA models. As all variables are ordinal, we can use the convenience function tidySEM::mx_lca(), which is a wrapper for the generic function mx_mixture() optimized for LCA with ordinal data. Any mixture model can be specified through mx_mixture(). At the time of writing, there are two other wrapper functions for special cases: mx_profiles(), for latent profile analysis, and mx_growth_mixture(), for latent growth analysis and growth mixture models. All of these functions have arguments data and number of classes. All variables in data are included in the analysis, so relevant variables must be selected first.

We consider one- through three-class models. Limiting the number of classes and categories makes this tutorial faster to reproduce. A substantive analysis should investigate a broader range of plausible models and use the original measurement scale where feasible.

set.seed(123)
res <- mx_lca(data = df, classes = 1:3)

Class Enumeration

In class enumeration, we want to compare a sequence of LCA models fitted to the data. Convergence must be checked for each fitted model. If a model with ordinal indicators does not converge, it is possible to aid convergence using mxTryHardOrdinal(), which expands the search for optimal parameter values for models with ordinal indicators. It is part of the family of functions based on mxTryHard().

Next, we create a model fit table using table_fit() and retain relevant columns. We also determine whether any models can be disqualified.

fit <- table_fit(res)
fit[, c("Name", "LL", "n", "Parameters", "BIC", "Entropy", "prob_min",
    "n_min", "np_ratio", "np_local")]

Inspect the ratio of cases to parameters (np_ratio and np_local) alongside classification diagnostics such as entropy and minimum classification probability. Models with small classes or unstable estimates require particular caution.

Class Enumeration

Using BIC

Compare the BIC values in the model-fit table. Lower BIC values indicate a preferred balance of fit and complexity. The fitted model ranking may differ from that obtained with the full, uncollapsed dataset.

Theoretical considerations

For illustration, we examine the three-class solution. In a substantive analysis, class selection should consider theoretical interpretability, convergence, and class sizes alongside information criteria.

res_final <- res[[3]]

Interpreting the Final Class Solution

Assess class sizes, entropy, and classification probabilities for the fitted three-class solution. The results below are based on the simplified dataset and should not be compared numerically with the original full-data example. We can produce a table of results using table_results(res_final). However, the results are thresholds, indicating quantiles of a standardized normal distribution. These may be difficult to interpret. Therefore, we ask for the results in probability scale:

tab <- table_prob(res_final)
reshape(tab, direction = "wide", v.names = "Probability", timevar = "group",
    idvar = c("Variable", "Category"))

The results can also be interpreted by plotting the response probabilities:

plot_prob(res_final, bw = TRUE)

Use the response-probability plot to compare the identity patterns across classes. Class labels should be chosen based on the newly estimated probabilities rather than carried over from the original analysis.

Auxiliary Analyses

To address the remaining two hypotheses, we will perform auxiliary analyses. Hypothesis 2 stated that adolescents who are nationally integrated would have lower depressive feelings than students from students with other combinations of identifications (hypothesis 2).

To test this hypothesis, we can call the BCH function and supply the auxiliary variable depression to the data argument, omitting the model argument. Below, we estimate an auxiliary model to compare depressive symptoms across classes:

aux_dep <- BCH(res_final, data = aux_data$depression)

To obtain an omnibus likelihood ratio test of the significance of depression differences across classes, as well as pairwise comparisons between classes, use lr_test(aux_dep). Interpret the omnibus and pairwise tests using the results from the simplified example; numerical conclusions from the original analysis do not necessarily carry over.

Hypothesis 3 was that, for assimilated and separated adolescents, there would not be a significant effect of perceived teacher discrimination on depressive symptoms. To test this hypothesis, we will compare the regression coefficient of discrimination on depressive symptoms across classes.

df_aux <- aux_data[, c("vict_teacher", "depression")]
# Dummy-code vict_teacher
df_aux$vict_teacher <- (as.integer(df_aux$vict_teacher) - 1)
aux_model <- BCH(res_final, model = "depression ~ vict_teacher",
    data = df_aux)

To view the coefficients of this model, we can use either coef(aux_model) or table_results(aux_model, columns = NULL). The regression coefficients are represented by paths in the class-specific A matrices.

table_results(aux_model, columns = c("label", "est_sig", "se",
    "pval", "confint", "group", "matrix"))

There are two ways to test whether the regression coefficients differ across classes.

First, we can use a likelihood-ratio test to compare the regression coefficients in the class-specific A matrices:

lr_test(aux_model, compare = "A")

Alternatively, we can use a Wald test to evaluate the equality constraints directly:

wald_test(aux_model, "aux1.A[1,2]=aux2.A[1,2]&aux1.A[1,2]=aux3.A[1,2]")

Both tests evaluate the null hypothesis that the regression coefficient of perceived teacher discrimination on depressive symptoms is equal across the three latent classes. Interpret the results based on the newly fitted auxiliary model; the original test statistics do not apply to this reduced example.