Aceh Tsunami — Interactive Causal-Inference Lab

A pedagogical companion to Evaluating the Economic Impact of Natural Disasters ↗ Back to the post

Did a disaster leave Aceh richer — and how would we ever know?

The 2004 tsunami flooded some Aceh districts and spared others by accident of coastal geography. Comparing the two — before and after — is a natural experiment. Four ideas run through this lab:

The four takeaways

  • Disaster → a higher long-run path. Flooded districts lost ~8% of output in 2005 but grew +6.3%/yr in 2006–08, ending permanently higher — confirmed by a synthetic control (+18%).
  • A single "after" hides the story. The pooled 2×2 estimate was an insignificant +0.0125; only splitting time into event-time windows revealed the dip-then-overshoot.
  • Clustered treatment demands honest inference. With all 10 treated districts in one corner of the map, Conley spatial standard errors roughly double the recovery SE — a spurious *** becomes an honest **.
  • Intensity concentrates the effect. Only the worst-hit places rebound the most — the average is driven by where damage, and reconstruction, was greatest.
Treated (flooded) vs control group-mean GDP growth, 2000–2012. Parallel before 2005, then the treated line dives (−0.03 in 2005) and overshoots (+0.12 in 2007); the shaded area is the difference-in-differences gap.

Key concepts

Difference-in-Differences (DiD)
Compare the change in the treated group to the change in the control group; the difference of those differences is the causal estimate. It nets out anything permanent about a district and any trend shared by the whole country.
Parallel trends
DiD's identifying assumption: absent the tsunami, flooded and non-flooded districts would have grown by the same amount on average. Tested here by the flat pre-2005 lines and the near-zero pre-tsunami coefficient (+0.0172, ns).
ATT — average treatment effect on the treated
The effect on the flooded districts, not on a district chosen at random. Both DiD and synthetic control target the ATT.
Conley spatial-HAC standard errors
Standard errors that let a district's errors correlate with nearby districts in the same year (spatial) and with itself over time (serial). They are larger — and more honest — than naive errors when the treated units cluster in space. Explore them in Tab 4.

The real estimates, with honest confidence intervals

Every coefficient from the post, with its Conley spatial-HAC 95% interval. A bar that crosses the dashed zero line is not statistically distinguishable from "no effect". Toggle the rows; hover a bar for its exact estimate, SE, and t-statistic.

Teal = significantly positive, orange = significantly negative, grey = not significant (95%). Estimates are dynamic-DiD coefficients on annual growth.

What to look for

  • The 2005 shock and the recovery are both significant for district GDP — but the recovery bar sits much further from zero relative to its width.
  • Per-capita shows no significant 2005 loss (output and population fell together) yet a strong recovery gain — not a denominator artifact.
  • The placebo (neighbours of flooded districts) bars straddle zero — exactly what a credible design should find.
  • Rural districts took the 2005 hit; city districts led the recovery (but with only 2 flooded cities, that bar is wide).

Simulate a tsunami: can DiD recover the truth?

Generate a synthetic district panel with a known 2005 shock and recovery boom, then let the dynamic DiD estimate them. Shrink the number of treated districts toward the real study's 10 and watch the estimate stay centered on the truth but scatter more — the small-sample fragility behind the post's wide standard errors.

The real study had only 10 — that is the whole precision problem.
The destruction effect (paper: ≈ −0.079).
The reconstruction boom (paper: ≈ +0.063).
Idiosyncratic district-year variation.
Recovery — your estimate
—
one simulated draw
Recovery — the truth
—
what you set
2005 shock — estimate
—
one simulated draw
One simulated draw: the dynamic-DiD estimate (orange) against the true injected effect (steel dashed). Move a slider to redraw.

The point estimate never moves — only your honesty about it

All 10 treated districts sit in one corner of Sumatra, so their shocks are not independent (Moran's I = +0.065, p = 0.003). As you dial up the spatial correlation, the recovery effect's standard error inflates from naive toward Conley spatial-HAC — and the significance verdict changes — while the estimate stays fixed at +0.0628.

0 = treat every district as independent (naive). 1 = the paper's full Conley spatial-HAC adjustment.
Recovery estimate
+0.0628
fixed — never changes
Effective SE
—
naive → Conley-HAC
t-statistic
—
estimate ÷ SE
Significance
—
The 95% confidence interval for the recovery effect widens as spatial correlation rises; the four bars are the standard-error recipes computed in the post.

What to look for

  • Drag the slider to 0: the naive SE (0.0146) gives a t of 4.3 and *** — a confidence the data do not support.
  • Drag to 1: the Conley-HAC SE (0.0244) gives a t of 2.57 and an honest **.
  • The orange dot — the point estimate — never moves. Spatial standard errors change inference, not the estimate.