Senior Project final product

PRISM Alcohol Policy Impact Atlas

A transparent evidence dashboard for exploring alcohol-policy mechanisms, next-year risk forecasts, and the limits of causal interpretation.

Find Full Paper Here

Submitted for review at the SSRN journal.

Download Paper

Check out my Google Scholar page for other papers I've published.Google Scholar

Blog

Mechanism Macro-F1

0.962

Forecast RMSE

0.863

Forecast R²

0.454

State Forecast Workspace

Select a state, then compare the observed pattern with a policy scenario.

Loading chart

State

Arizona (AZ)

Latest impaired rate

1.91 per 100k

Beer tax

$0.16 / gallon

Text coverage

supplemental_only

Text quality

52%

Mechanism Profile

The state-year text signal projected into PRISM channels.

Price21%

Access42%

Enforcement61%

Predictive, not automatic proof.

Forecast changes show how the trained model responds to policy features. The event-study coefficient is a separate directional causal audit around beer-tax events.

Scenario Simulator

Change policy levers and compare baseline with scenario forecast.

Beer tax shift

$0.10 / gallon

-$0.20+$0.50

Sunday-sales/access shift

No access shift

StricterMore available

Underage-purchase enforcement

Stricter checks

SofterStricter

Baseline vs Scenario

Forecasted alcohol-impaired fatality rate per 100k for the next state-year.

Baseline

1.86

Scenario

2.41

Delta

+0.545

Loading chart

Price

+0.568

Access

0.000

Enforcement

-0.023

Scenario interval: 0.00 to 5.13 per 100k. This interval is carried through from the forecast artifact and shifted with the simulated point estimate.

3D Price / Access / Enforcement Space

Fixed baseline state-year mechanism positions for 2023.

Each dot is a selected state at the fixed 2023 baseline year. The axes stay fixed: Price on x, Enforcement on y, and Access on z.

Baseline Controls

Using 2023 data. Add up to 6 states and compare their mechanism coordinates side by side.

Baseline year: 2023

Add state

Selected states

State	Price	Access	Enforcement
AZ Arizona	21%	42%	61%
CA California	0%	43%	41%

Project Story

The SRP arc from messy policy data to a final public tool.

Week 1

Laws as Recipes

Framed alcohol statutes as combinations of price, access, and enforcement mechanisms rather than simple yes/no policy checkboxes.

Week 2

Panel Assembly

Merged APIS, FARS, FHWA, FRED, and YRBS into a state-year research panel with common state and year keys.

Week 3

Rates and Exposure

Moved beyond crash counts by using population-normalized and exposure-aware outcomes.

Week 4

Teen Data Humility

Identified sparse YRBS coverage, making teen drinking outcomes secondary rather than the primary evidence base.

Week 5

Text Becomes Signal

Turned policy language into mechanism features that forecasting models can use.

Week 6

Mechanism Split

Separated law text into price, access, and enforcement channels so policy changes can be interpreted.

Week 7

Forecasting Value

Showed that mechanism-aware text adds measurable signal compared with policy-only baselines.

Week 8

Enforcement Pattern

Found enforcement language was often the strongest interpretive signal, while avoiding causal overclaiming.

Week 9

Model Ladder

Compared forecasting models and selected Random Forest for next-year impaired-fatality prediction.

Week 10

Equations and Limits

Explained RMSE, MAE, R-squared, and why predictive accuracy is not the same as causal proof.

Week 11

Final Product Build

Turned the research package into a public, teacher-facing dashboard with simulations and transparent uncertainty.

Week 12

Reflection

reflection

Poster

High-resolution conference poster artifact for presentation and download.

Download Poster

Research Pipeline

How the final product connects the poster, deck, and repo.

Panel

51 jurisdictions across 2003-2023.

Text

Policy chunks scored as mechanisms.

Forecast

Random Forest selected by held-out RMSE.

Limitations

Built into the product instead of hidden in fine print.

Teen outcome coverage is sparse

YRBS teen current-use and binge-drinking data are useful secondary outcomes, but coverage is incomplete enough that fatality risk remains the main outcome.

Policy text can be supplemental

Many state-years rely on supplemental text sources. The dashboard surfaces coverage provenance and text quality so users know when the signal is thinner.

Beer-tax causal events are limited

The causal audit focuses on a small set of beer-tax increase events, so it is directional evidence rather than broad proof for every policy lever.

Forecasts are not policy promises

Scenario outputs show model movement under changed features. They do not prove what would happen after a real law passes.

Largest Scenario Responses

States with the biggest modeled movement in the curated PRISM scenario file.

Loading chart