What happens to a poor region when you finally connect it to a rich one?
In June 1998 a 4.8-kilometre bridge opened over the Jamuna river in Bangladesh. It cost about US$985 million, connected roughly 26 million people in the isolated northwest to Dhaka, and cut freight costs by about half. A truck from Bogra to Dhaka went from twenty hours to six.
Economists disagree — sharply — about what should happen next. This app lets you work through the evidence yourself: turn the dials on the assumption that carries the whole design, walk the effect period by period, and see where in space the gains actually landed.
Three theories, and two of them predict the same thing about factories
This is the trap the study had to escape. A study that measured only manufacturing would see the share fall, write the word "deindustrialisation", and declare backwash. It would be wrong, and it would have no way of knowing.
| Theory | Manufacturing share | Population density |
|---|---|---|
| Big push Integration raises efficiency |
up, or flat | up |
| Backwash Myrdal 1957, Krugman 1991 |
down | down |
| Comparative advantage The region specialises |
down | up or flat |
Backwash and comparative advantage agree on factories and disagree on people. That one extra outcome — from a census that was already sitting there — turns an ambiguous finding into a decisive one.
The discriminating test, live
Press the button to reveal what actually happened to each of the two outcomes, and watch which theories survive.
Why the Padma hinterland is a fair comparison
Bangladesh is cut into three by two enormous rivers. The Jamuna separated the poor northwest from Dhaka; the Padma cut off the south. Only one got a bridge. The Padma crossing was not begun until 2015 — two years after these data end — so the comparison region stayed isolated for the entire window.
Which river came first was decided by political geography, not economics: President Ershad's base was in Rangpur and Prime Minister Khaleda Zia's in Bogra, both in the Jamuna hinterland.
The 2×2, and the assumption it rests on
Difference-in-differences is four numbers and two subtractions. The first difference removes anything permanent about a place; the second removes anything that hit the whole country. What survives both is the bridge — provided the two regions would have moved together without it.
Drag the counterfactual and watch the estimate break
The teal dashed line is what difference-in-differences assumes: the treated region would have grown at the comparison region's rate. Use the slider to suppose that assumption is wrong by some amount, and watch the estimated effect move away from the truth.
Notice that the bias is one-for-one: every unit of violation in the counterfactual trend passes straight into the estimate. That is why so much of the analysis is spent bounding how large the violation could plausibly be, rather than testing whether it is exactly zero.
Three estimators, three columns
The published table reports the same specification three ways: unweighted OLS, and two doubly robust estimators that reweight the comparison group to look like the treated one. Toggle between them.
Both reweighted estimates are larger than the unweighted one. Adjustment does not always shrink an effect — a useful thing to see once.
Every period gets its own coefficient
A single post-treatment dummy compresses fifteen years into one number. Give every period its own coefficient instead, measured against the last pre-bridge window. The coefficients before the bridge are a test of the assumption; the ones after are the answer.
Walk the effect period by period
What would a confounder have to look like?
For nighttime lights the pre-bridge coefficient sits on zero, and then the effect climbs monotonically across all five post-bridge periods. A confounder producing that shape would have to be absent before June 1998, appear at exactly the right moment, and grow steadily for fifteen years without ever reversing.
Compare the rice-yield panel. Its pre-bridge coefficients are also insignificant — but with standard errors so wide that the test has very little power. With nine clusters there is not much that panel can rule out, and it would be wrong to present its clean pre-period as strong evidence.
The average effect hides almost everything
Split the treated region into three bands by distance from the bridge foot, and two of the outcomes reverse sign across bands. The upazilas nearest the bridge got the largest proportional cut in travel time — and gained the least.
Effects by distance band
Why do the distant places gain more?
A 40 percent cut on a ten-dollar taxi ride saves four dollars. A 17 percent cut on a five-hundred-dollar flight saves eighty-five. The percentage is smaller, the base is enormous, and the saving is much larger.
Upazilas near the bridge foot were already reasonably connected — the ferry was an inconvenience, not a wall. Upazilas 250 kilometres out were close to autarky, where fertiliser rarely arrived and rice rarely left. Trade responds to the level of the barrier, not the percentage change in it.
Before the bridge existed, the agriculture share rose with distance while manufacturing and services fell. The places furthest out had the most agricultural output to ship — and therefore the most to gain when shipping got cheaper.
How hard can you push before the answer changes?
HonestDiD: bounding the violation instead of testing for it
Rather than asking whether the pre-trend is exactly zero, allow the post-treatment violation of parallel trends to be up to M times the largest violation observed before treatment, and report the widest confidence set consistent with that.
The breakdown value sits just under M = 1. That is a moderate robustness margin, not a spectacular one. A result surviving to M = 3 would be much stronger; one breaking at M = 0.3 would be fragile. Saying which is more useful than dressing it up.
The political-economy placebo
The most serious rival explanation is that a prime minister with roots in the Jamuna hinterland simply sent more schools, clinics and electricity there. If so, the "bridge effect" would be a public-spending effect in disguise.
Twenty-one estimates, and not one is significant at 5 percent. The closest is the long-run distance to a high school — and it has the wrong sign for the story, since it says schools got farther away in treated villages.
One undefined macro, two different papers
The most instructive thing in the original replication package is a bug that produces no error message. An undefined Stata global made a trimming cutoff missing; because any number is less than missing in Stata, the trim then fired for every comparison unit, and the regression silently ran on treated units only.
The tell is not the coefficient — it is the footer. 124 upazilas where there should be 239. Read the sample size first.
All 122 published coefficients reproduce
Every headline coefficient in the paper's four tables, re-estimated in Python and compared with the authors' own Stata output. Maximum absolute deviation 0.0005 — inside the tolerance implied by three-decimal printing.