Loading FARS Auxiliary Accident and Vehicle .csv files, checking for missing values, joining datasets, and dropping duplicate records.

A_PED_F ‘1’ - involving pedestrian fatality; ‘2’ - not involving pedestrian fatality

Feature correlation heatmap. Highly correlated features (>.80) may result in quality issues during modeling.

Shuffling dataframe. The model could learn potential data entry patterns.

The outcome of a situation can be described statistically by the odds of success. A linear model cannot be used to set a threshold between classes, however. For that we use logistic regression.

Rewriting the logistic function in terms of the odds ratio and taking the natural log of both sides, the logit function converts values to the scale of a probability in the range of [0,1] (Bati, n.d.). Probabilities >0.50 are rounded up to 1, while probabilities <0.50 are rounded down to 0.

With the logit function, logistic regression is used in a statistical model to calculate predictor coefficient estimates.

With a supervised machine learning classification model, the logistic function is used to make classifications, or predictions on future data.

Theoretically, we can represent the discovery that speeding was a factor in a crash as the outcome, or response, making ‘speeding’, or ‘not speeding’ our binary logistic variable. The weights of predictors, or coefficient estimates, on a response will help us evaluate relationships within FARS.

A_SPCRA ‘1’ - not involving speeding, ‘0’ - involving speeding

Balancing the dataset, ensuring algorithm will not be skewed by abundance of ‘1’, incidents not involving speeding.

Test/Train Split

Using the statsmodels Logit function to calculate coefficients of predictors

Pseudo R-squared measures the variance of the error term (variance of errors), a residual variable representing the margin of error resulting from differences in a statistical model’s theoretical values and actual observed values.

The pseudo R-squared value, 0.14 of our logit model tells us that there are many other independent variables to take into consideration for understanding something so complex as a driver speeding while operating a motor vehicle.

A p-value of < the significance level of 0.05 for predictor, ‘A_PED_F’ supports evidence to reject the null hypothesis: here is no change in incidents involving speeding with a change in incidents of pedestrian fatality.

Making predictions on the test dataset

Known vs. Predicted

The ROC curve visualizes the true positive and false positive rate of change. The ideal curve of our classification model would pass through coordinate (0,1). This would be the point where true positive rate would be 1 and false positive rate would be 0. The area under the curve (AUC), in the range [0,1], measures the probability that a model will more likely predict one class over another.

Please see the presentations for recommendations and the report for findings here

1. Adjaye-Gbewonyo, D., & Briones, E. M. (2024, July). Walking for Leisure and Transportation Among Adults: United States, 2022.https://www.cdc.gov/nchs/products/databriefs/db504.htm

2. Bati, F. (n.d.) CMSC 437 Lecture 2c: Logistic Regression. University of Maryland Global Campus.

3. National Highway Traffic Safety Administration (NHTSA). (2021, February). Fatality Analysis Reporting System (FARS) auxiliary datasets analytical user’s manual, 1982-2019 (Report No. DOT HS 813 071). Washington, DC: Author.https://crashstats.nhtsa.dot.gov/Api/Public/ViewPublication/813071