Falling Through the Cracks

Abstract

Approximately one in twelve working-age adults in the OECD lives with a diagnosed learning disability — dyslexia, dyscalculia, attention-deficit and related cognitive-processing differences — and the household-finance literature has been slow to converge on a credible cumulative estimate of the financial-hardship consequences. Within the Australian setting, a positive learning-disability gradient on hardship survives standard income, education, and household-composition controls.

The conceptual backbone is the **disability-as-friction model of household financial fragility**: the cognitive differences that learning disability denotes operate as a friction in the standard household life-cycle decision problem, raising the cost of optimal financial behaviour at every stage. The paper's contribution is a recovery-dynamics interpretation: the disability friction does not just elevate the probability of falling into hardship — it also elevates the persistence of hardship spells once they begin, because the same cognitive-processing differences that complicated pre-event financial planning complicate the recovery-period rebuilding work as well.

The empirical setting is the HILDA panel, restricted to working-age adults with the validated learning-disability indicator from the Health module. The existing analysis applies the classical identification stack: propensity-score matching balancing pre-event covariates, difference-in-differences around diagnosis-eligible age windows, panel fixed-effects, probit, Cox proportional-hazards modelling for hardship-spell duration, Oaxaca-Blinder decomposition, Oster endogeneity bounds, and NLP topic-modelling on open-text disability-experience responses. The classical evidence already documents a measurable LD-hardship gradient with substantial persistence channel.

We propose extending the existing analysis with three ML layers: a Causal Forest that would use the PSM scores as nuisance Pr(LD) and recover the heterogeneous treatment effect by gender and comorbid mental-health; a Random Survival Forest that would estimate hardship-spell duration without the proportional-hazards constraint; and an Imai–Keele debiased-ML mediation analysis decomposing the total LD effect into the employment-mediator and mental-health-mediator channels under high-dimensional partialling-out. Cross-country comparison panels (BHPS, PSID, SOEP) are flagged as exploratory follow-up rather than the primary HILDA-based contribution. The paper is the flagship of the disability-and-financial-recovery research line.

Data & Methods