Power Outages and Population Density in the United States
2026-06-05
Using 1,534 major U.S. power outage events from 2000–2016, I looked at whether population density predicts how long an outage lasts — do high-density states recover faster because they have more infrastructure and resources nearby, or does the sheer number of people and equipment involved work against them?
The question
Central question: Are states with higher population density more vulnerable to prolonged power outages, or do they recover faster?
This matters for resource allocation during emergencies — if low-density states systematically take longer to recover, utility companies and emergency planners have a concrete reason to weight support toward them rather than assuming a one-size-fits-all response time.
States were split into "High Density" / "Low Density" groups using the median urban population density (2,380 persons per square mile) as the cutoff.
Exploratory analysis
Outage duration is heavily right-skewed — most outages resolve in under 5,000 minutes, but some extreme cases run past 100,000 minutes.
Comparing the two density groups directly, high-density states show slightly shorter outage durations on average, but both groups have wide spread:
One aggregate result stood out: outages caused by equipment failure last roughly 10x longer in low-density states (4,672.6 min) than high-density states (427.7 min) — suggesting rural areas may lack the repair resources that dense areas have on hand. Fuel-supply-emergency outages were the longest category for both groups.
Is the missing data a problem?
CUSTOMERS.AFFECTED has missing values, and whether that's an issue worth
worrying about depends on why it's missing. I performed two permutation
tests:
- Missingness does depend on
CAUSE.CATEGORY(p < 0.001), where outages from causes like intentional attacks are less likely to have a reported customer count. - Missingness does not depend on a state's year-over-year GDP change (p ≈ 0.600), which has no logical link, and the test found none.
Hypothesis test: does density actually predict duration?
H0: Outage duration is the same for high and low density states. Any difference is random. H1: High-density states have shorter outage durations.
A permutation test (10,000 permutations, difference in means) gave p ≈ 0.065. At α = 0.05, we fail to reject the null. The density difference in the raw data doesn't clear the bar for statistical significance, even though it looked suggestive in the EDA above.
Predicting outage duration
Problem: predict OUTAGE.DURATION (regression) using only information
available at the moment an outage starts, using location, cause category, and time of
year. Post-outage fields like CUSTOMERS.AFFECTED were not included.
Baseline model: Decision Tree Regressor on two features
(POPDEN_URBAN, one-hot CAUSE.CATEGORY) — test RMSE 7,418.97 min,
versus a mean-only baseline of 7,838.99 min.
Final model: Random Forest Regressor, tuned via 5-fold
GridSearchCV, adding five features (CLIMATE.REGION, POPPCT_URBAN,
MONTH, ANOMALY.LEVEL, NERC.REGION) — test RMSE 6,888.27 min, a
further ~530-minute improvement over the baseline.
Is the model fair across density groups?
H0: the model's RMSE is the same for high- and low-density states. H1: RMSE is higher for low-density states (the model is less accurate where it matters more).
Result: p = 0.35 — fail to reject. The model performs comparably across both groups; it isn't systematically worse for the states that might need accurate predictions most.
Full analysis
The complete notebook, cleaning steps, and modeling code can be found on GitHub: power-outages-population-density.