Evaluating the Impact of the Jamuna Bridge

Downloads

Each dataset is available as a labeled Stata .dta and its source file.

⇩ Download all data (ZIP)stata_codebook.do

Run stata_codebook.do in Stata once to attach long-form per-variable notes to the .dta files.

Load directly in code

Every file loads straight from GitHub (raw URLs). Swap the file name to load any dataset.

Stata

* Stata 14+ : `use` reads an https URL directly
global BASE "https://raw.githubusercontent.com/cmg777/starter-academic-v501/master/content/post/python_bridge_impact/data/"
use "${BASE}bridge_nightlights.dta", clear
describe
notes

Python

!pip install -q pyreadstat
import pandas as pd
BASE = "https://raw.githubusercontent.com/cmg777/starter-academic-v501/master/content/post/python_bridge_impact/data/"
df = pd.read_stata(BASE + "bridge_nightlights.dta")

# load every dataset at once
files = ["bridge_nightlights", "bridge_employment", "bridge_yield", "bridge_dhs_household", "bridge_dhs_village"]
data = {f: pd.read_stata(BASE + f + ".dta") for f in files}

# pyreadstat (richest metadata) reads LOCAL files -> download first
import pyreadstat, urllib.request
urllib.request.urlretrieve(BASE + "bridge_nightlights.dta", "bridge_nightlights.dta")
df, meta = pyreadstat.read_dta("bridge_nightlights.dta")

Copy and paste this snippet in Google Colab app. https://colab.research.google.com/notebooks/empty.ipynb

R

# R : haven::read_dta auto-downloads an https URL
library(haven)
BASE <- "https://raw.githubusercontent.com/cmg777/starter-academic-v501/master/content/post/python_bridge_impact/data/"
df <- read_dta(paste0(BASE, "bridge_nightlights.dta"))

Overview & sources

Companion data for a hands-on Python tutorial that evaluates the economic impact of the Jamuna (Bangabandhu) Bridge, a 4.8 km crossing that opened in June 1998, cost about US\$985 million, connected roughly 26 million people in isolated northwest Bangladesh to Dhaka, and cut freight costs by about half. The design is a two-group, non-staggered difference-in-differences: the Jamuna hinterland is treated, and the Padma hinterland — a symmetric region cut off by the other great river, whose own bridge was not started until 2015 — is the comparison. The Dhaka–Chittagong core is excluded. These five files are tidy CSV exports of the authors' original Stata replication package, unchanged in value; only the file format and column selection differ. The tutorial reproduces all 122 published coefficients to three decimals.

Data sources

Cite this data

Please cite this dataset as follows.

APA

Mendez, C. (2026). Evaluating the Impact of Infrastructure: A Beginner's Guide to Difference-in-Differences with the Jamuna Bridge — companion data [Data set]. https://carlos-mendez.org/post/python_bridge_impact/

Data derived from the replication package of Blankespoor, B., Emran, M. S., Shilpi, F., & Xu, L. (2021), Journal of Economic Geography. Please cite the original authors for the data.

BibTeX

@misc{mendez2026pythonbridgeimpact,
  author       = {Mendez, C.},
  title        = {Evaluating the Impact of Infrastructure: A Beginner's Guide to Difference-in-Differences with the Jamuna Bridge — companion data},
  year         = {2026},
  howpublished = {\url{https://carlos-mendez.org/post/python_bridge_impact/}},
  note         = {Data set}
}

note = {Data derived from the replication package of Blankespoor, Emran, Shilpi and Xu (2021)}

Variable explorer search & filter all 51 variables

Type to filter by name or label, or use the chips to filter by type. Each row shows a mini distribution. Click a header to sort.

Cross-file variable index

Which file each variable appears in (● = present).

Construction & formulas

The bridge is treated as a quasi-natural experiment: which river got a bridge first was decided by political geography — President Ershad's base in Rangpur and Prime Minister Khaleda Zia's in Bogra, both in the Jamuna hinterland — rather than by economic prospects in the northwest. The comparison region stayed isolated for the whole window because the Padma bridge was not begun until 2015.

• Mean effect: Y_it = θ0 + μ_i + μ_t + θ1·(D_J × D_post) + Σ β_q X_qit + Σ π_m (Z_mi0 × t) + ε_it, with upazila fixed effects μ_i, period fixed effects μ_t, and the estimand the ATT.
• Short run vs long run: replace the single post switch with two, δ1·(D_J × D_SR) and δ2·(D_J × D_LR). This is where population density reverses sign, from −2.5% to +5.9%.
• Derived controls: mdist = min(jamuna_m, padma_m)/1000; lmdist = ln(mdist + 1); lpop91 = ln(pop91); lpop91_t = lpop91 × year; lmdist_t = lmdist × year. The trend interactions matter: unit fixed effects already absorb the time-invariant levels, so entering them alone would do nothing. Interacting them relaxes parallel trends from “all upazilas trend alike” to “upazilas that started alike trend alike”.
• Rainfall logs differ by file: ln(rainm + 1) for the nightlights panel (it contains zeros), plain ln(rainm) for the census and yield panels. This is not cosmetic — it changes which rows survive.
• Outcomes: lmn = ln(mn + 1) (luminosity is bottom-coded at 1.0); lyld = ln(yld); ldensity = ln(density); sagr = pop_agr/emp, sind = pop_ind/emp, sserv = pop_serv/emp.
• Two doubly-robust weights, both estimated on the smp1 sample: LWDR uses the propensity odds p/(1−p)·(1−π)/π; KOBDR uses Kline's Oaxaca-Blinder projection of the treated covariate mean onto the comparison design. Both put weight 1 on treated units, which is what makes them ATT weights.

The datasets

Switch datasets with the tabs. Each shows the full variable dictionary plus a sortable statistics table with mini distributions and data coverage.

Expand all datasets expand to search (Ctrl/⌘+F) or print across all datasets

Known limitations & caveats

• year is a period index, not a calendar year. Nightlights and yields are averaged into three-year blocks. The mapping is in the year row of each dataset below. Because the controls interact baseline traits with this index, substituting calendar years changes every coefficient.
• smp1 is the sample filter, and it is deliberately missing. It is missing for the Dhaka–Chittagong core region (treatd == 1), which is neither treated nor a credible comparison. Filtering on smp1.notna() is how the core is excluded from every regression. In the two DHS files there is no core, so smp1 is identical to treat.
• Distances are to the relevant crossing. mdist = min(jamuna_m, padma_m) gives each upazila its distance to the Jamuna bridge if it is in the treated hinterland, or to the planned Padma crossing if it is in the comparison hinterland. It is not distance to a single point.
• The yield panel is very thin. Sixteen former districts, of which eleven enter the estimation and nine survive the negative-weight drop. Cluster-robust inference with nine clusters is fragile; treat the yield magnitudes as indicative.
• The census panel has one pre-bridge year. Only 1991 precedes the bridge, so no pre-trend test is possible for population density or the employment shares — only a level-balance test. This includes the density variable that settles the paper's theoretical question.
• Village-questionnaire columns named like presence indicators are often distances. post_office and cinema are labelled as distances in the original Stata data despite being filed under “Presence of” in the published table. grameen_bank, co_operative_soc and NGO are genuine presence dummies.
• Missingness is heavy in the village file. dist_satellite_clinic (398 non-missing), NGO (413), dist_Thana (559) and dist_district (618) are far from complete, which is why the public-goods regressions have wildly different sample sizes.
• These are the authors' data, unchanged. Values are byte-faithful to the original .dta files; only the format and the column selection differ. Any error in the source data is preserved here.