Posterior inference of attitude-behaviour relationships using latent class choice models

Akshay Vij; Stephane Hess

arxiv: 2509.08373 · v2 · submitted 2025-09-10 · 💰 econ.EM

Posterior inference of attitude-behaviour relationships using latent class choice models

Akshay Vij , Stephane Hess This is my paper

Pith reviewed 2026-05-18 18:13 UTC · model grok-4.3

classification 💰 econ.EM

keywords latent class choice modelsattitude-behaviour relationshipsposterior inferencehybrid choice modelspreference heterogeneitychoice modelling

0 comments

The pith

Posterior inference in latent class models recovers attitude-behaviour relationships without hybrid complexity.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that posterior inference on latent class choice models offers a practical way to examine how attitudes relate to behaviours. It does so by profiling attitudinal indicators from class probabilities rather than integrating them directly into the choice model structure. A sympathetic reader would care if true because it promises behaviourally rich insights with less complexity and fewer estimation problems than traditional hybrid models, as shown in case studies on remote work preferences and vaccine acceptance where it outperforms or matches alternatives.

Core claim

The authors argue that by recovering class-specific attitudinal profiles through posterior inference in latent class choice models, one can explore attitude-behaviour associations empirically without the need to embed attitudinal constructs in the structural model, thereby avoiding complexity and convergence issues while still gaining rich insights into preference heterogeneity.

What carries the argument

Posterior inference applied to the class probabilities of latent class choice models to derive attitudinal profiles and their links to choice behaviour.

If this is right

When explaining preference heterogeneity the posterior approach retains interpretability and robustness without sacrificing behavioural depth.
Factor-based models may discard key attitudinal information.
Full-information hybrid models provide little additional explanatory power but require substantially more estimation effort.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The method could be applied to other choice contexts such as travel mode selection or product adoption to segment populations by inferred attitudes.
It raises the question of whether choice data alone suffices to proxy for latent attitudes in many domains.
Extensions might involve testing the stability of these posterior profiles across different model specifications or datasets.

Load-bearing premise

Latent classes estimated purely from observed choices align well enough with actual attitudinal differences that the posterior probabilities can reliably reveal the attitude-behaviour connections.

What would settle it

A clear falsifier would be finding that the attitudinal profiles obtained from posterior class probabilities show no meaningful relationship with independently measured attitudes or fail to improve understanding of choice variation beyond what the classes alone provide.

read the original abstract

The link between attitudes and behaviour has been a key topic in choice modelling for two decades, with the widespread application of ever more complex hybrid choice models. This paper proposes a pragmatic and computationally tractable alternative framework for empirically examining the relationship between attitudes and behaviours using latent class choice models (LCCMs). Rather than embedding attitudinal constructs within the structural model, as in hybrid choice frameworks, we recover class-specific attitudinal profiles through posterior inference. This approach enables analysts to explore attitude-behaviour associations without the complexity and convergence issues often associated with integrated estimation. Two case studies are used to demonstrate the framework: one on employee preferences for working from home, and another on public acceptance of COVID-19 vaccines. Across both studies, we compare posterior profiling of indicator means, fractional multinomial logit (FMNL) models, factor-based representations, and hybrid specifications. We find that posterior inference methods provide behaviourally rich insights with minimal additional complexity, while factor-based models risk discarding key attitudinal information, and full-information hybrid models offer little gain in explanatory power and incur substantially greater estimation burden. Our findings suggest that when the goal is to explain preference heterogeneity, posterior inference offers a practical alternative to hybrid models, one that retains interpretability and robustness without sacrificing behavioural depth.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Posterior profiling from standard LCCMs gives a workable shortcut for attitude-behavior links that avoids hybrid integration costs, though it rests on classes reflecting attitudinal differences.

read the letter

The main things here are that fitting a plain latent class choice model on choice data and then profiling attitudinal indicators via posterior probabilities can recover useful associations without embedding the attitudes in the structural model, and that in the two applications this route delivers comparable or better behavioral detail than hybrids while cutting estimation burden substantially. Factor approaches lose more information along the way. The paper walks through this on employee work-from-home preferences and COVID vaccine acceptance, running the same data through posterior means, fractional multinomial logit on the posteriors, factor representations, and full hybrid specs, then reports the trade-offs in explanatory power and computational load. That side-by-side setup is the concrete contribution, and it lands as a practical recommendation for analysts who already run LCCMs and want to bring in attitudinal indicators without convergence headaches. The comparisons are straightforward and the conclusions follow directly from the reported patterns. The soft spot is the assumption that the latent classes estimated from choices alone line up with attitudinal constructs rather than other unobserved heterogeneity such as demographics or context. If that alignment is weak in a given dataset, the posterior profiles will not cleanly isolate attitude-behavior relationships, which undercuts the claim that the method reliably substitutes for hybrid models. The stress-test note flags exactly this, and the paper would benefit from more explicit checks on class-attitude correspondence. This is aimed at applied discrete choice researchers who use latent class models and need a lighter way to incorporate attitudes. Readers working on transportation, health, or similar empirical questions will find the implementation details and the two worked examples directly usable. I would send it for peer review because the empirical comparisons are specific enough to get useful referee feedback on the assumption and on how general the findings are.

Referee Report

3 major / 2 minor

Summary. The paper proposes posterior inference on latent class choice models (LCCMs) estimated from choice data alone as a pragmatic alternative to hybrid choice models for recovering attitude-behaviour relationships. Rather than structurally integrating attitudinal indicators, the approach uses post-estimation class probabilities to profile attitudinal means or run fractional multinomial logit regressions. Two case studies (working-from-home preferences and COVID-19 vaccine acceptance) compare this to factor-based and full-information hybrid specifications, concluding that posterior methods deliver rich insights at low complexity while hybrids add little explanatory power at high estimation cost.

Significance. If the empirical comparisons hold, the work offers choice modellers a computationally lighter route to attitudinal insights that preserves interpretability and avoids common hybrid convergence problems. The dual-case-study design and explicit benchmarking against factor and hybrid alternatives are strengths; the emphasis on minimal added complexity could influence applied practice in transportation and health economics.

major comments (3)

[§3] §3 (Methodology): the central claim that posterior class probabilities reliably recover attitude-behaviour associations rests on the untested assumption that LCCM classes estimated from choice data primarily reflect attitudinal heterogeneity rather than other unobserved factors (demographics, context, or measurement error). No diagnostic, balance test, or robustness check is reported to support this alignment, which is load-bearing for the superiority claim over hybrid models.
[§5] §5 (Empirical results): the statement that hybrid models offer 'little gain in explanatory power' is not supported by quantitative evidence such as log-likelihood differences, AIC/BIC values, or out-of-sample predictive metrics relative to the baseline LCCM; without these, the conclusion cannot be evaluated.
[Table 3] Table 3 or equivalent comparison table: reported fit statistics for the four approaches across periods or subsamples are needed to substantiate the cross-method claims; if only point estimates without standard errors or likelihood-ratio tests are shown, the 'minimal additional complexity' advantage remains qualitative.

minor comments (2)

[Abstract] Abstract: sample sizes, number of latent classes retained, and exact fit criteria used in each case study should be stated to allow readers to gauge the scale of the comparisons.
[§3.2] Notation in §3.2: distinguish clearly between the class-conditional indicator means and the FMNL coefficient vectors to prevent reader confusion when interpreting the posterior profiling results.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments on our manuscript. We address each major comment below in a point-by-point manner and indicate where revisions will be made to strengthen the paper.

read point-by-point responses

Referee: §3 (Methodology): the central claim that posterior class probabilities reliably recover attitude-behaviour associations rests on the untested assumption that LCCM classes estimated from choice data primarily reflect attitudinal heterogeneity rather than other unobserved factors (demographics, context, or measurement error). No diagnostic, balance test, or robustness check is reported to support this alignment, which is load-bearing for the superiority claim over hybrid models.

Authors: We thank the referee for this observation. The proposed framework does not posit that LCCM classes estimated from choice data are driven primarily or exclusively by attitudinal heterogeneity. Rather, the classes represent unobserved preference heterogeneity recovered from the choice data alone; the subsequent posterior inference step is then used to profile attitudinal indicators across classes and thereby document empirical associations. This is distinct from the structural integration attempted in hybrid models. We agree that a clearer statement of this scope, together with explicit discussion of possible confounding sources, would improve the manuscript. We will revise §3 to include this clarification and add a robustness check that examines the relationship between class membership probabilities and observed demographic covariates. revision: yes
Referee: §5 (Empirical results): the statement that hybrid models offer 'little gain in explanatory power' is not supported by quantitative evidence such as log-likelihood differences, AIC/BIC values, or out-of-sample predictive metrics relative to the baseline LCCM; without these, the conclusion cannot be evaluated.

Authors: We accept that the current presentation of this conclusion would benefit from direct quantitative support. While the manuscript already notes the estimation difficulties and convergence issues of the hybrid specifications, we will augment §5 with explicit model-fit statistics (log-likelihood, AIC, BIC) for the LCCM baseline, the factor-based approach, and the hybrid models. Where the data permit, we will also report likelihood-ratio tests and a brief out-of-sample predictive comparison. revision: yes
Referee: Table 3 or equivalent comparison table: reported fit statistics for the four approaches across periods or subsamples are needed to substantiate the cross-method claims; if only point estimates without standard errors or likelihood-ratio tests are shown, the 'minimal additional complexity' advantage remains qualitative.

Authors: We agree that a consolidated comparison table would make the cross-method evaluation more transparent. We will revise or supplement the existing table (or create a new table in the main text or appendix) to report log-likelihood, number of parameters, AIC, BIC, and, where feasible, standard errors or likelihood-ratio statistics for all four modeling approaches. This will allow readers to assess both explanatory power and complexity on a common quantitative footing. revision: yes

Circularity Check

0 steps flagged

No circularity detected in the posterior inference framework

full rationale

The paper proposes estimating standard latent class choice models from choice data alone, then applying post-hoc posterior class probabilities to profile attitudinal indicators via means or FMNL regressions. This sequence does not reduce any reported attitude-behaviour associations to quantities defined solely by the fitted parameters by construction. No equations are presented that create self-definitional loops, and the comparisons to hybrid and factor models rely on separate estimations rather than self-citations or imported uniqueness theorems. The framework remains self-contained against external benchmarks of model fit and behavioural interpretability in the two case studies.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The framework rests on standard discrete-choice modelling assumptions plus the domain-specific premise that latent classes derived from choice data can serve as proxies for attitudinal segmentation.

free parameters (1)

Number of latent classes
The number of classes is a modelling choice typically selected by information criteria; its value is not reported in the abstract.

axioms (1)

domain assumption Latent classes estimated from observed choices capture systematic preference heterogeneity that correlates with unobserved attitudinal constructs
This premise underpins the validity of using posterior class probabilities to profile attitudes.

pith-pipeline@v0.9.0 · 5750 in / 1332 out tokens · 63669 ms · 2026-05-18T18:13:43.737816+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean absolute_floor_iff_bare_distinguishability unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

posterior class membership probabilities ... profile class-specific mean responses to Likert-scale indicators ... fractional multinomial logit (FMNL) model
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We find that posterior inference methods provide behaviourally rich insights with minimal additional complexity

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

[1]

Abou-Zeid, M., & Ben-Akiva, M. (2024). Hybrid choice models. Handbook of choice modelling, 489-521. Bahamonde-Birke, F., & de Dios Ortúzar, J. (2014). Is sequential estimation a suitable second best for estimation of hybrid choice models?. Transportation Research Record, 2429(1), 51-58. Ben-Akiva, M., McFadden, D., Train, K., Walker, J., Bhat, C., Bierlai...

work page 2024
[2]

F., Bolduc, D., & de Dios Ortúzar, J

Raveau, S., Álvarez-Daziano, R., Yáñez, M. F., Bolduc, D., & de Dios Ortúzar, J. (2010). Sequential and simultaneous estimation of hybrid discrete choice models: Some new findings. Transportation Research Record, 2156(1), 131-139. Sohn, K. (2017). An expectation -maximization algorithm to estimate the integrated choice and latent variable model. Transport...

work page 2010

[1] [1]

Abou-Zeid, M., & Ben-Akiva, M. (2024). Hybrid choice models. Handbook of choice modelling, 489-521. Bahamonde-Birke, F., & de Dios Ortúzar, J. (2014). Is sequential estimation a suitable second best for estimation of hybrid choice models?. Transportation Research Record, 2429(1), 51-58. Ben-Akiva, M., McFadden, D., Train, K., Walker, J., Bhat, C., Bierlai...

work page 2024

[2] [2]

F., Bolduc, D., & de Dios Ortúzar, J

Raveau, S., Álvarez-Daziano, R., Yáñez, M. F., Bolduc, D., & de Dios Ortúzar, J. (2010). Sequential and simultaneous estimation of hybrid discrete choice models: Some new findings. Transportation Research Record, 2156(1), 131-139. Sohn, K. (2017). An expectation -maximization algorithm to estimate the integrated choice and latent variable model. Transport...

work page 2010