US traffic accident analysis

The Work

Choose the questions before loading everything

The source contains records collected through traffic-event APIs covering 49 states. I loaded the subset of fields needed for the analysis, including incident times, location, weather, and the traffic-impact score, rather than bringing every available column into memory.

I checked the meaning of the outcome before using it. The dataset’s Severity field measures traffic disruption on a one-to-four scale. It is not an injury scale. I grouped levels three and four for comparisons of higher disruption and kept that definition attached to the findings.

Make the comparisons consistent

I checked missingness and required start-time, severity, and state fields for the analysis. I converted timestamps and derived hour, weekday, month, season, and time-of-day categories. A city-and-state label helped keep cities with the same name from being treated as one place.

Weather descriptions needed grouping before they could support a readable comparison. I mapped variants into categories such as Fair, Cloudy, Rain, and Snow/Ice, while retaining an Unknown category for missing values. Unknown was excluded from weather-association tests rather than presented as an actual condition.

I used the same time and weather categories across the charts and statistical comparisons. That kept a change in grouping from being mistaken for a change in the underlying pattern.

Separate volume from the disruption share

The notebook’s selected commute windows, 7–9 a.m. and 3–6 p.m., contained 33.5% of reported incidents across all days. That calculation includes records starting in hours 7, 8, 15, 16, and 17. Five states contained 50.9%. Those are concentrations within the records, useful for identifying where to investigate further, rather than rates per vehicle or mile traveled.

For weather, I compared both incident counts and the share with higher disruption. Rain’s higher-disruption share was 23.5%, compared with 16.9% for Fair weather. A category can have a large total without the largest disruption share, so the two charts answer different questions.

I used chi-square tests and Cramer’s V to examine categorical associations and their strength. The tests were statistically significant, but the associations were weak: Cramer’s V was 0.0972 for time of day and 0.0645 for weather. With millions of records, that distinction mattered when interpreting the patterns. I also compared adverse and fair-weather proportions, with a sensitivity check excluding Fog/Haze from the adverse group.

Turn patterns into response priorities

I proposed time-based staffing and safety messaging, weather-responsive warnings, and further investigation of geographic hotspots. The presentation connects each proposal to a finding and a metric that could be monitored if the idea were piloted.

A pilot would need local traffic-volume data, a baseline, and consistent reporting to evaluate whether a response improved conditions. The patterns identify places and times for further investigation; the next step is to test the proposed response in that local context.

THE DELIVERABLES

What I delivered

• Python analysis notebook with preparation steps, visual exploration, statistical comparisons, and limitations.
• Exported analysis data for Tableau and a set of temporal, weather, and geographic charts.
• Interactive dashboard and presentation with three response priorities.

THE OUTCOME

Where the work stands

Completed a notebook, Tableau dashboard, and presentation connecting temporal, weather, and geographic patterns to three response priorities.