Latent Classes in a 7-item Hardship Module and What They Predict for Subsequent Recovery
The 7-item hardship battery of a national household panel has been treated as an aggregate score (count of items endorsed) or a binary any-hardship indicator in essentially every published Australian household-finance study that uses the panel. This paper begins from the observation that the seven items measure substantively different states, asking friends for help is qualitatively different from being unable to pay for heat, and the aggregate score conflates these states in ways that may matter for predicting subsequent recovery.
The behavioural-economics hypothesis under test is that the FIRST hardship item endorsed at hardship onset determines the subsequent trajectory, distinct from the aggregate hardship score. Three substantive states are proposed: strategic help-seeking (asking friends/family or welfare organisations for help), passive bill failure (couldn't pay bills, rent), and severe deprivation (couldn't heat home, went without meals, pawned/sold items). The central claim is that people whose first hardship is 'asked for help' do better than people whose first hardship is 'couldn't pay bills', even though the standard aggregate index treats them as equivalent.
The empirical setting is the 7-item hardship module of a national household panel, restricted to person-waves observable on all seven items, with the LCA sample formed from those endorsing at least one. First-landing events are transitions from hardship_count = 0 to hardship_count ≥ 1.
The analysis pipeline applies Bernoulli-mixture latent class analysis (LCA) via EM with K = 2, 3, 4, 5 and BIC-based model selection; first-landing detection at the 0 → ≥1 transition; trajectory recovery analysis at t+1 to t+4 by first-landing class; and heterogeneity stratification by income, age, education, household, and era.
The feasibility exploration completed 2026-05-10 returns a clear REVISE verdict. The latent-class model selection, the first-landing class assignments and the recovery trajectories at t+1 to t+4 are all estimated, and the paper is being rewritten around what they showed rather than around the prior it set out to test.