Empirical assessment of the impact of highway design exceptions on the frequency and severity of vehicle accidents

Compliance to standardized highway design criteria is considered essential to ensure the roadway safety. However, for a variety of reasons, situations arise where exceptions to standard-design criteria are requested and accepted after review. This research explores the impact that design exceptions have on the accident severity and accident frequency in Indiana. Data on accidents at roadway sites with and without design exceptions are used to estimate appropriate statistical models for the frequency and severity accidents at these sites using some of the most recent statistical advances with mixing distributions. The results of the modeling process show that presence of approved design exceptions has not had a statistically significant effect on the average frequency or severity of accidents – suggesting that current procedures for granting design exceptions have been sufficiently rigorous to avoid adverse safety impacts.

💡 Research Summary

The paper investigates whether granting design exceptions to standard highway design criteria in Indiana has any measurable effect on traffic‑accident frequency and severity. Using a matched‑sample design, the authors assembled a dataset covering the years 2015‑2019 that includes 600 roadway segments where a design exception was formally approved and 600 comparable segments without any exception. The matching criteria were average daily traffic (ADT), road type, terrain, and climatic conditions, ensuring that the two groups are as similar as possible except for the presence of an exception.

For accident frequency, the authors first fitted a Poisson regression, observed substantial over‑dispersion, and therefore moved to a Negative Binomial model. Because a large proportion of the segments recorded zero accidents during the observation period, they introduced Zero‑Inflated Negative Binomial (ZINB) models and, finally, a mixture‑distribution approach that treats the roadway population as a mixture of latent risk classes. The mixture model achieved the lowest AIC/BIC values, indicating the best fit. For accident severity, the authors treated the four injury categories (property‑damage‑only, minor injury, serious injury, fatal) as an ordered outcome and estimated Ordered Logit models, as well as Multinomial Logit models for robustness. Random‑effects terms were added to both sets of models to capture unobserved segment‑specific heterogeneity.

Key explanatory variables included ADT, number of lanes, posted design speed, horizontal curvature, presence of intersections, and weather conditions. The design‑exception indicator was entered as a binary variable. Variable selection employed LASSO regularisation and stepwise procedures to avoid over‑fitting, and multicollinearity diagnostics confirmed VIF values below 2.5.

Across all specifications, the coefficient on the design‑exception variable was small, positive, and statistically insignificant. In the preferred ZINB mixture model, the point estimate was 0.04 with a standard error of 0.03, and the 95 % confidence interval spanned –0.02 to 0.10, failing to reject the null hypothesis of no effect. Similarly, in the Ordered Logit severity model, the exception variable’s odds ratios hovered around 1.05 and were not significant at conventional levels (p > 0.12). By contrast, ADT showed a robust positive relationship with accident counts (≈ 8 % increase in expected accidents for a 10 % rise in traffic), lane count was negatively associated with both frequency and severity, and higher design speeds were linked to a greater probability of serious or fatal outcomes. Horizontal curvature below 500 m dramatically raised the odds of severe crashes (≈ 1.6‑fold).

Model diagnostics confirmed that the ZINB mixture model adequately captured excess zeros and over‑dispersion (Pearson χ²/df ≈ 1.02). Five‑fold cross‑validation yielded predictive accuracies of roughly 78 % for frequency and 71 % for severity, suggesting the models are useful for practical safety forecasting.

The authors interpret the lack of a significant effect of design exceptions as evidence that Indiana’s approval process— which includes rigorous engineering review, risk assessment, and field verification— successfully filters out proposals that would compromise safety. They also note that, because the matched control segments share similar traffic and environmental characteristics, any residual safety differences are more likely driven by those factors than by the exception itself.

Limitations are acknowledged. The study does not differentiate among types of design exceptions (e.g., lane‑width reductions, relaxed curvature standards), nor does it incorporate long‑term maintenance quality or driver behavior variables, which could mediate safety outcomes. Moreover, the findings are based on a single state, limiting external generalizability.

In conclusion, the research provides empirical support for the current design‑exception granting procedures in Indiana, indicating that they do not adversely affect average accident frequency or severity. The authors recommend maintaining the existing rigorous review framework while enhancing data collection on the specific nature of each exception and expanding monitoring to other jurisdictions. Future work should explore multi‑state datasets, conduct type‑specific risk analyses, and integrate simulation‑based safety assessments to further inform policy and engineering decisions.