Goodman (2002) Latent Class Analysis: The Empirical Study of Latent Types, Latent Variables, and Latent Structures

Summary

Chapter 1 of Hagenaars and McCutcheon (2002), Applied Latent Class Analysis (Cambridge University Press). Goodman provides a pedagogical treatment of latent class models as tools for examining whether observed relationships between manifest variables are spurious: latent class analysis asks why a relationship exists (or whether it is real), not merely whether it is significant, which is all that association/correlation/loglinear models can do. The chapter develops a typology of three causal roles for a latent variable, demonstrates the "explain away vs. explain" distinction with two worked examples, and shows how restricted 3-class models can be as parsimonious as a null independence model while fitting 97% better. [Note: Available excerpt covers only pp. 3–18 of a 53-page chapter; pp. 19–55 not available in this PDF.]

Key Claims

• Three causal roles of a latent variable $X$ with respect to two manifest variables $S$ and $M$:

  1. Antecedent (Fig. 1a): $X$ causes both $S$ and $M$; the $S$–$M$ relationship is spurious.
  2. Intervening (Fig. 1b): $S$ causes $X$, $X$ causes $M$; $S$–$M$ is mediated.
  3. Coincident/reciprocal (Fig. 1c): $S$ and $X$ reciprocally affect each other; $X$ causes $M$. In all three cases $S$ and $M$ are conditionally independent given $X$, so the latent class model cannot distinguish among them — the causal interpretation is the researcher's choice, not a statistical outcome.

• "Explain away" vs. "explain": Finding that a latent class model $H_1$ fits the data does not automatically mean the $S$–$M$ relationship is spurious. The same fit is consistent with $X$ explaining (mediating or enabling) the relationship rather than explaining it away (rendering it spurious). Goodman explicitly warns against the automatic spuriousness interpretation.

• Model: For $m$ manifest variables and latent variable $X$ with $T$ classes: $$\pi_{ijklt}^{ABCDX} = \pi_t^X \pi_{it}^{AX} \pi_{jt}^{BX} \pi_{kt}^{CX} \pi_{lt}^{DX}$$ Within each latent class, manifest variables are mutually independent (local independence). The observed marginal cell probability is $\pi_{ijkl} = \sum_t \pi_{ijklt}^{ABCDX}$.

• Example 1 — Socioeconomic Status (SES) × Mental Health ($n=1{,}660$, $6\times4$ table; Srole et al. 1962):

  • Independence $H_0$: $X^2=45.99$, $df=15$ (poor fit; correlation $r=0.16$ but significant)
  • 2-class $H_1$: $X^2=2.74$, $df=8$ → 94% reduction in $X^2$; model fits very well
  • The significant SES–mental health relationship is fully explained by a latent "endowment" variable.

• Example 2 — Stouffer-Toby role conflict ($n=216$, $2^4$; Stouffer and Toby 1951):

  • $M_0$ (independence): $X^2=104.11$, $df=11$
  • $M_1$ (2-class, unrestricted): $X^2=2.72$, $df=6$ → 97% reduction
  • $M_2$ (3-class, unrestricted): $X^2=0.42$, $df=2$
  • Restricted 3-class models: $M_3$ ($df=9$, $X^2=2.28$); $M_4$ ($df=10$, $X^2=2.42$); $M_5$ ($df=11$, $X^2=2.85$)
  • $M_5$ has the same degrees of freedom as $M_0$ (mutual independence) yet fits 97% better — a parsimonious structure with genuine latent interpretability.
  • Item probabilities under $M_1$: universalistically inclined class (28%): $P(+|A)=0.993$, $P(+|B)=0.940$, $P(+|C)=0.927$, $P(+|D)=0.769$; particularistically inclined class (72%): $P(+|A)=0.714$, $P(+|B)=0.330$, $P(+|C)=0.354$, $P(+|D)=0.132$.
  • Under $M_3$: strictly universalistic (17%, all $P(+)=1$), mixed (78%, $P(+|A)=0.796$, $P(+|B)=0.420$, $P(+|C)=0.437$, $P(+|D)=0.175$), strictly particularistic (5%, all $P(+)=0$).

• Parsimony principle: Restricted 3-class models ($M_3$–$M_5$) are more parsimonious than $M_1$ while fitting essentially as well. The best model is not necessarily the one with most degrees of freedom recovered, but the one with the most substantively interpretable class structure.

• Historical note: Some 19th-century models (Peirce 1884 on measuring prediction success; Goodman and Kruskal 1959) can be reinterpreted as special cases of latent class or latent structure models.

Concepts Introduced or Extended

• Latent Class Model
• EM Algorithm

Entities Mentioned

• Leo A. Goodman
• Hagenaars and McCutcheon (2002) Applied Latent Class Analysis

Quotes

"None (or almost none) of the usual measures of the nonindependence between variables A and B can help the researcher to determine whether the observed relationship (the nonindependence) between variables A and B can be explained away by some other variable, say, variable X." (p. 4)

"The latent class model M5 in Table 5b turns out to be as parsimonious as the simple null model M0 of mutual independence among the four observed variables (A, B, C, and D), with 11 degrees of freedom corresponding to model M0 and also to model M5; and the goodness-of-fit chi-square value obtained under M5 is also 97% less than the corresponding chi-square value obtained under M0." (p. 17)

My Take

A lucid pedagogical chapter that complements the more technical Goodman (1974). The key conceptual advance over 1974 is the explicit typology of causal roles (antecedent/intervening/coincident) and the warning against reflexive "spuriousness" interpretation — a point still routinely overlooked in applied work. The parsimony result ($M_5$ with $df=11$ fitting as well as $M_1$ with $df=6$) is striking: equality constraints that seem to weaken the model actually recover degrees of freedom without meaningfully degrading fit, because the data have a simple underlying structure. The limited excerpt means the full chapter's contribution cannot be assessed; later sections presumably address multiple latent variables, identifiability, and further examples. The Peirce (1884) historical note hints at a broader treatment of latent structure precursors not visible in the excerpt.