Latent Class Analysis for Ordinal Indicators

Loading the Data

To load the data, simply attach the tidySEM package. For convenience, we assign the indicator data to an object called df:

# Load required packages
library(tidySEM)
library(ggplot2)
# Load data
df <- maene_identity[1:5]

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 5-10 unique 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 see that there are no missing values, hence we proceed with our analysis. Note that the ethic identification variables are very right-skewed and some response categories have near-zero frequencies; this can cause problems in model specification and convergence. One potential solution would be to merge small adjacent categories. We will not do this here, however.

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 here consider 1-6 class models, but note that this may be overfit, as some of the indicators have only 5 response categories.

set.seed(123)
res <- mx_lca(data = df, classes = 1:6)

Class Enumeration

In class enumeration, we want to compare a sequence of LCA models fitted to the data. First, note that all models converged without issues. If this had not been the case, 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")]

Note that both the global and local ratio of cases to parameters is low; for models of 3 or more classes, there are just a few observations per parameter in the smallest class (see np_local). This is a good reason to disqualify those classes. We will eliminate classes 4-6 on those criteria, but in real data applications, the 3-class solution might also be disqualified.

In terms of classification diagnostics, note that the entropy and the minimum classification probabilities are high for all models. This indicates that all classes are distinct.

Class Enumeration

res_final <- res[[3]]

Interpreting the Final Class Solution

The 3-class model yielded classes of reasonable size; the largest class comprised 33%, and the smallest comprised 16% of cases. The entropy was high, S = .93, indicating good class separability. Furthermore, the posterior classification probability ranged from [.94, .99], which means that all classes had low classification error. 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)

Note that the first class (33%) has relatively high identification with the ethnic indicators and relatively low identification with Belgian and Flemish identity. The second class (16%) has moderate identification with Belgian and Flemish identity, and relatively low identification with ethnic identity. Finally, the third class (50%) has high identification with all identities.

Based on the probability plot, we can label class 1 as ethic identifiers, class 2 as low identifiers, and class 3 as high identifiers.

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 = maene_identity$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). The results indicate that there are no significant differences in depression across classes, ΔLL(5) = 4.32, p = .50.

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 <- maene_identity[, 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). As evident from the results table, the coefficients labeled class1.A[1,2] are the regression coefficients.

There are two ways to test the difference in regression coefficients across classes: using lr_test(aux_model, compare = "A"), to compare the ‘A matrix’ (regression coefficients) across classes, or wald_test(aux_model, "class1.A[1,2]=class2.A[1,2]&class1.A[1,2]=class3.A[1,2]"). The results indicate that there are no significant differences in the regression coefficients across classes, χ²(2) = 1.16, p = .56.