Abstract

Aging bridges are a long-standing public safety and infrastructure concern in the United States. National summaries show how many bridges are in poor condition, but they hide large differences between places and mix together factors such as age, traffic, and structure type. This project uses the National Bridge Inventory (NBI), a federal dataset of more than 600,000 bridges, to ask how bridge condition varies across states and counties and whether places with unusually many poor-rated bridges remain unusual after accounting for age and traffic. The analysis will first describe the national picture, then rank places using methods that account for small sample sizes, and finally model the probability that a bridge is rated poor. Two or three case studies, chosen from the best and worst performing places, will examine local patterns in more detail. I expect bridge age to be the strongest predictor of poor condition, but I also expect meaningful geographic differences that age and traffic alone do not explain. Results will be descriptive and associational, not causal, and will help show where further inspection and investment questions are most worth asking.

Introduction

National summaries of bridge condition do not show which U.S. places have unusually many poor bridges, or whether older and busier bridges explain those differences. The United States depends on more than 600,000 bridges, and a large share of them were built decades ago. Bridge condition is a public safety issue, an economic issue, and a budget issue, because bridges in fair condition can be preserved at lower cost than bridges allowed to fall into poor condition (American Society of Civil Engineers [ASCE], 2025). The National Bridge Inventory records a standardized condition rating for every public road bridge, which makes national comparison possible.

In an earlier lab, I used this dataset to show two things: most bridges are rated Fair or Good, and bridges rated Poor tend to have been built earlier than bridges rated Good. That result is a useful starting point, but it is only a national average and it raises more questions than it answers. A nationwide dataset combines very different settings, such as dense cities, rural counties, coastal areas, and cold-climate states, and many factors are tangled together. Older bridges may also carry more traffic, have been rebuilt more often, or sit in places with different funding. A simple relationship between age and condition therefore cannot tell us where the real problems are. Earlier studies of NBI data identify age and traffic as predictors of bridge component ratings (Lu et al., 2019; Mia & Kameshwar, 2023), and climate factors such as freeze-thaw cycles and snowfall also matter (Lamei et al., 2025).

The project is organized using the seven domains of the INFORMS Analytics Framework, which underlies the Certified Analytics Professional certification (INFORMS, n.d.). The business problem (Domain I) is the gap described above, which leaves decision makers without a way to identify places doing unusually well or badly. The analytics problem (Domain II) is to turn this into measurable questions: estimate each place’s share of Poor bridges, and compare that share with what bridge age, traffic, and structure type would predict. Success means stable, defensible rankings and a few case studies, not a causal explanation. I will start nationally and then drill down into places at the extremes.

Research question: How does bridge condition vary across U.S. states and counties, and which places have a higher or lower share of Poor-rated bridges than expected after accounting for bridge age, traffic, and structure type?

Data

The project uses the National Bridge Inventory as published in the Bureau of Transportation Statistics’ National Transportation Atlas Database (Federal Highway Administration & Bureau of Transportation Statistics, 2025; https://geodata.bts.gov/datasets/national-bridge-inventory/about), dated June 20, 2025. Coding definitions come from the Federal Highway Administration’s coding guide (Federal Highway Administration [FHWA], 1995; https://www.fhwa.dot.gov/bridge/mtguide.pdf). The file I am using has 624,193 bridge records and about 130 columns.

Key variables are overall condition (BRIDGE_CONDITION: Good, Fair, Poor), the lowest component rating (LOWEST_RATING), year built (YEAR_BUILT_027), year reconstructed, average daily traffic (ADT_029), structure kind and type, and state, county, and place codes. Records are individual bridges, which can be aggregated to state or county.

Known limitations are:

  • The data are a single snapshot, so I can compare places but cannot measure change over time.
  • Condition ratings come from human inspectors and may not be applied identically everywhere.
  • The overall condition category is based on the lowest component rating, so one weak part can classify a whole bridge as Poor.
  • Culverts are rated on a separate scale and need separate handling.
  • Some fields have missing values or placeholder codes, such as a year built of zero, that must be cleaned before analysis.

Methodology

The analysis has four steps in R, mapped to the framework domains (Domain III, data, is described above).

  1. Data preparation (III). I will read and standardize the data with readr and janitor, recode implausible years, convert codes to place names, calculate bridge age, log-transform traffic, and decide how to treat culverts, reporting records dropped at each step.
  2. National description. Using dplyr and ggplot2 (Wickham et al., 2019), I will summarize condition by age, traffic, and structure type as a baseline.
  3. Place-level comparison. I will calculate the percent of bridges rated Poor by state and county, requiring a minimum bridge count, and compare shrinkage estimates with raw percentages to handle noisy small counties. Maps will use sf and tigris.
  4. Modeling and case studies (IV, V). I will fit a logistic regression for the probability that a bridge is Poor, building on earlier logistic models of NBI ratings (Lu et al., 2019). Predictors are age, log traffic, and structure type, with state (possibly county) as a random effect in lme4 (Bates et al., 2015), because bridges are nested within places. Unlike earlier studies of deck, superstructure, and substructure ratings, I model the overall Poor class. Place effects show which places differ from what bridge characteristics predict, and I will compare two or three case studies from the extremes.

For deployment and iteration (VI, VII), the deliverable is a reproducible R Markdown report with parameters (minimum bridge count, condition measure, places studied) and reusable functions, so new data or case studies run without rewriting the analysis.

The main challenge is confounding. Unmeasured factors such as climate, maintenance funding, and de-icing salt may drive place differences (Lamei et al., 2025), so I will report associations, not causes, and check sensitivity to the minimum bridge count and culvert treatment.

Anticipated Results

I expect the probability of a Poor rating to rise with age, and that traffic and structure type will add some, but smaller, explanatory power. I expect substantial variation between states and counties, with some places having more Poor bridges than their age and traffic would predict, though the size of this unexplained difference is unknown. Shrinkage will likely soften the extremes among small counties. The case studies should show whether high-Poor places are mainly places with older bridges or places with other distinguishing features.

References

American Society of Civil Engineers. (2025). 2025 report card for America’s infrastructure: Bridges. https://infrastructurereportcard.org/cat-item/bridges-infrastructure/

Bates, D., Mächler, M., Bolker, B., & Walker, S. (2015). Fitting linear mixed-effects models using lme4. Journal of Statistical Software, 67(1), 1-48. https://doi.org/10.18637/jss.v067.i01

Federal Highway Administration. (1995). Recording and coding guide for the structure inventory and appraisal of the nation’s bridges (Report No. FHWA-PD-96-001). https://www.fhwa.dot.gov/bridge/mtguide.pdf

Federal Highway Administration & Bureau of Transportation Statistics. (2025). National Bridge Inventory [Data set]. National Transportation Atlas Database, U.S. Department of Transportation. https://geodata.bts.gov/datasets/national-bridge-inventory/about

INFORMS. (n.d.). INFORMS Analytics Framework. https://www.informs.org/Professional-Development/Professional-Development-Classes/INFORMS-Analytics-Framework

Lamei, T., Abdel-Rahim, A., & Ibrahim, A. (2025). Development of performance-based deterioration models for bridge components. Innovative Infrastructure Solutions, 10(6), Article 230. https://doi.org/10.1007/s41062-025-02025-7

Lu, P., Wang, H., & Tolliver, D. (2019). Prediction of bridge component ratings using ordinal logistic regression model. Mathematical Problems in Engineering, 2019(1), Article 9797584. https://doi.org/10.1155/2019/9797584

Mia, M. M., & Kameshwar, S. (2023). Machine learning approach for predicting bridge components’ condition ratings. Frontiers in Built Environment, 9, Article 1254269. https://doi.org/10.3389/fbuil.2023.1254269

Wickham, H., Averick, M., Bryan, J., Chang, W., McGowan, L. D., François, R., … Yutani, H. (2019). Welcome to the tidyverse. Journal of Open Source Software, 4(43), 1686. https://doi.org/10.21105/joss.01686