How Where You Move From Shapes What Happens to You — A Multi-Shock Analysis from HILDA
Australian households experience three substantively different types of shock — climate disasters (cyclones, floods, bushfires), health shocks (serious illness or injury), and property crime — and the household-finance literature has treated each in isolation. What has not been jointly estimated across shock types is the financial cost of moving in response, distinct from the cost of the shock itself. The decision to relocate is endogenous to both the shock and the household's pre-shock circumstances, so the empirical question splits into two: which households select into moving, and what is the financial trajectory of movers conditional on the same shock?
The conceptual frame is the selection–response decomposition. Under pure selection, the mover–non-mover hardship gap reflects pre-shock differences in income, assets, household composition, and other endowments. Under pure response, the gap reflects what moving does to households who would otherwise be observationally similar. The decomposition matters for policy: if the gap is selection, post-shock support should be targeted by pre-shock circumstances; if response, support should be targeted at the move itself (relocation grants, transition support, settlement services).
The empirical setting is HILDA Release 22, waves 1–22 (2001–2022), 454,861 person-wave rows, from which a 7,606-row analytical sample is constructed across three shock types: climate disaster (n = 810), health shock (n = 4,446), and property crime fallback (n = 2,350). The newly extracted data file (property_movement_variables.parquet, 454,861 × 112) captures 61 HILDA variables across eight groups — disaster shocks, mobility indicators, dwelling type, geographic location, reasons-for-moving (checkbox-encoded mhrea items), welfare-and-insurance, shocks, and prosperity — with careful corrections for the checkbox encoding that the HILDA documentation under-flags. A second parquet (analytical_sample.parquet, 7,606 × 50) houses the three shock samples with a shock_type indicator and a 5-category mobility encoding (local move / state move / interstate move / household-composition move / no move) applied per a priority rule.
The analysis pipeline runs three layers. The classical layer applies logistic regression for P(move), OLS for the hardship / net-worth / income trajectories at t−1, t, t+1..t+4, and Oaxaca-Blinder decomposition of the mover–non-mover hardship gap. The ML layer applies a DML-approximation causal forest for the ATE and HTE of moving on financial hardship per shock type, and a predictive comparison of logistic regression versus random forest versus gradient boosting for P(move). Figures include three dwelling-tenure Sankey diagrams, three pre→post state Sankey diagrams, two geomap movement-flow charts, and a hardship-trajectory line chart.
The key findings, extracted from the full-pipeline run completed 2026-05-10: the causal-forest ATE of moving on hardship_increased is +12.3 pp for climate disasters (z = 8.9), three times the health-shock cost of +3.8 pp (z = 6.8) and roughly double the property-crime cost of +5.6 pp (z = 6.8). The Oaxaca-Blinder explained share for climate is only 5.7 percent — the gap is dominated by response heterogeneity, not selection — versus 39 percent for health and 32 percent for property crime. Interstate movers fare better on hardship (mobility category 3, −5.0 pp, p = 0.06), consistent with opportunity migration rather than distress migration. Predictive AUC for P(move) is LR 0.703, RF 0.729, GB 0.735 — gradient boosting modestly outperforms. The paper anchors a new Property Movement research line. Submission target: HILDA Survey Research Conference 2026 (1–2 October 2026, Melbourne), with post-conference journal target Journal of Banking & Finance.