Road Accident Fatalities across Indian States: An Applied Statistical Study of MoRTH Open Data

Reading India's Roads is an applied data-analytics project that studies five years of official, state-wise road-accident data published by the Ministry of Road Transport and Highways (MoRTH) and answers a question every Indian has personal experience of: where, and in what circumstances, do road crashes turn fatal?

India recorded about 4.81 lakh road accidents and 1.73 lakh deaths in 2023, and overspeeding alone accounts for roughly 68 per cent of accidents and fatalities. These numbers are published every year — but as hundreds of PDF tables and scattered CSVs that almost nobody joins together. This project turns that raw publication into a clean, queryable state-year dataset and interrogates it with the statistical tools of the B.Sc Data Science curriculum.

The work is organised in three layers. A data layer collects the MoRTH state-wise tables (2019–2023) via the data.gov.in open-data portal and its OpenCity CSV mirror, cleans them, and loads them into a MySQL database designed in 3NF. An analysis layer in Python (pandas, statsmodels, Matplotlib) performs exploratory analysis, chi-square tests of independence between crash causes and outcomes, and multiple regression on accident severity (persons killed per 100 accidents) against explanatory shares such as overspeeding, rural roads and two-wheeler involvement. A communication layer turns results into charts, a findings dashboard page and a policy-flavoured report.

Because every number traces back to a government publication, the project is deeply viva-defensible — and the honest treatment of aggregate-data limitations (the ecological fallacy) shows statistical maturity beyond a typical final-year EDA.

India loses about 1.7 lakh lives to road crashes every year, and Parliament, courts and state road-safety bodies keep asking the same question: which states and which conditions produce fatal outcomes? The answer exists in MoRTH's annual "Road Accidents in India" tables, but in a form unusable for analysis — dozens of per-year datasets with inconsistent columns and formats on data.gov.in, and hundreds of PDF tables for the rest. There is no clean, documented, multi-year state panel that a student team can query, test and model. As a result, common claims about road deaths circulate without statistical backing, and district- or state-level severity differences (deaths per 100 accidents vary widely across states) go unexamined. This project builds that missing dataset and applies the B.Sc statistics toolkit — cleaning, chi-square tests, correlation, multiple regression — to produce defensible, reproducible findings about where severity concentrates and which factors it travels with, while explicitly stating the limits of aggregate-level inference.

Abstract

Road accidents kill roughly one Indian every three minutes on average, yet the government's own state-wise statistics remain locked in report tables. This project builds a tidy state-year panel of road-accident outcomes for all States and Union Territories from the MoRTH "Road Accidents in India" annual reports (2019–2023), stores it in a MySQL database, and applies exploratory analysis, hypothesis testing and multiple regression to understand which factors travel with fatality severity. The study closes with state-grouping insights and a clear-eyed discussion of what aggregate data can and cannot prove.

Introduction

MoRTH publishes an annual statistical report with state/UT-wise counts of accidents, persons killed and injured, causes (overspeeding, drunken driving, wrong-side driving), road categories (national highways, state highways, other roads), and road-user categories (two-wheelers, pedestrians, pedal cyclists). The 2023 edition records 4,80,583 accidents and 1,72,890 deaths — a 4.2 per cent rise in accidents year-on-year — with two-wheeler riders forming the largest share of victims. The data are free and authoritative, but the publication format makes cross-state, cross-year comparison genuinely difficult for a student or journalist without a data pipeline.

Existing situation and the gap

Student analytics projects in this space typically stop at a single-year bar chart using scraped, unverified numbers. Official open-data releases on data.gov.in exist but are fragmented into dozens of single-year, single-topic datasets with differing column names, units and missing-UT conventions. The gap this project fills is methodological as much as analytical: a documented cleaning pipeline, a small normalised database, and a statistical layer that goes beyond counting into testing and modelling.

Proposed work

• Assemble state-year tables for 2019–2023: accidents, deaths, injuries, cause-wise and road-category shares, road-user fatalities, and report-computed rates.
• Design a 3NF MySQL schema (state, year, indicator, value plus reference tables), ingest with a Python loader, and answer slicing questions with SQL joins and aggregations.
• Explore: national and state trends, the pandemic-year break in 2020–21 and the recovery, worst-severity states relative to fleet and road network, and Karnataka's trajectory alongside its peers.
• Test: chi-square independence between cause mix and state groups; correlation and multiple regression of severity (killed per 100 accidents) on overspeeding share, rural-road share and two-wheeler fatality share, with diagnostics.
• Communicate: a chart pack, a one-page findings summary per region, and the ecological-fallacy discussion that frames every claim correctly.

Feasibility

• Technical: all tooling is open source and taught in the programme — Python/pandas for the pipeline, MySQL (CS31T) for storage, statsmodels for inference. The dataset is a few thousand rows; a normal laptop is more than enough.
• Data: public domain government publications; no permissions, no cost, no privacy concerns (no individual-level records at all).
• Operational: the 8 h/week Sem-6 slot comfortably fits the 14-week plan for two members; every technique maps to a course already completed.

In scope

• State/UT-level aggregate tables from MoRTH annual reports, 2019–2023 (about 36 states/UTs × 5 years × 10–15 indicators).
• Cleaning decisions: renaming schemes, missing-UT conventions, unit harmonisation, and flagging years where a state's reporting changes.
• MySQL storage (3NF), Python analysis in Jupyter notebooks, and Matplotlib/Seaborn-style static charts.
• Statistics: descriptive trends, chi-square independence tests, correlation analysis, one multiple linear regression model with full diagnostics.

Out of scope

• District-level or individual-crash (FIR-level) analysis — the source data is state-aggregate only.
• Causal claims: the report frames associations, never causes, and says why.
• Forecasting of future accident counts (a natural follow-on project).
• Interactive web dashboards with a server backend; the deliverable is static charts and a summary page, sized to the Sem-6 slot.

The project follows an adapted CRISP-DM analytics lifecycle — business understanding, data understanding, data preparation, analysis, evaluation, communication — which fits an applied statistics project better than a construction-style SDLC because the risk sits in data quality and inference, not in code volume.

Quality assurance: every computed statistic is cross-checked against figures printed in the MoRTH report itself (e.g., national totals must match the report's all-India row before any further analysis); notebooks are re-run top-to-bottom from raw CSVs to prove reproducibility; and the two members cross-review each other's notebooks before the analysis is accepted into the report.

Because this is an analysis project rather than a software product, the architecture is the study design itself, drawn as a one-way pipeline in which no stage feeds back silently: every correction returns through the documented cleaning log, never through ad-hoc edits downstream.

flowchart TD
  A["data.gov.in datasets +<br/>OpenCity CSV mirror"] --> B["Raw store:<br/>one file per table, per year"]
  C["MoRTH annual report PDF<br/>(authoritative check)"] --> B
  B --> D["Cleaning pipeline:<br/>rename, harmonise, handle missing UTs"]
  D --> E["Tidy state-year panel:<br/>CSV + data dictionary"]
  E --> F[("MySQL database:<br/>3NF - state, year, indicator, value")]
  F --> G["EDA notebooks:<br/>trends, rates, rankings"]
  F --> H["Inferential layer:<br/>chi-square, correlation, regression"]
  G --> I["Chart pack and<br/>findings summary"]
  H --> I
  I --> J["Report chapters with<br/>ecological-fallacy framing"]

Three design choices deserve explanation in the viva. First, the authoritative check sits upstream: the MoRTH report PDF is the referee, and any number in the OpenCity mirror or a data.gov.in dataset that disagrees with the report's all-India row is quarantined and investigated — open-data portals re-publish with occasional transcription errors. Second, the tidy panel (one row per state-year-indicator) is deliberately boring: it makes the MySQL schema trivially normalised and lets every analysis, however exotic, begin from the same six-column query. Third, the inferential layer is separated from the exploratory layer so that hypothesis-driven work (chi-square, regression) cannot be quietly tuned after peeking at the exploratory charts — a small but real guard against p-hacking in a student project.

The database itself is intentionally small: a STATE table (code, name, region), an INDICATOR table (code, label, unit, source table), and a FACT table (state, year, indicator, value, flag), giving clean 3NF with no update anomalies — a direct application of the DBMS (CS31T) course.

• Add the newest MoRTH edition each year so the panel extends rather than restarts.
• District-level deep-dive for one home state, using state police or state road-safety-cell publications.
• Time-series forecasting of national or state fatality counts as a companion project.
• A small interactive map page for the chart pack using the Web Programming (DSCS61) course skills.