Tool · Research Design

Close-election RD power calculator.

Most close-election regression discontinuity studies are severely underpowered — often detecting fewer than one in five effects that exist. This tool computes exact power for U.S. Senate and House races using real election margins, showing how many races fall within your bandwidth and what effect sizes you can realistically detect.

Key finding from Stommes, Aronow & Sävje (2023): only 22% of published RD studies achieve 80% power for a small standardized effect (Cohen's d = 0.2). Among election-data studies, the figure falls to 13%.

Step 1 Chamber

Both chambers use actual contested general election results to compute how many races fall within your chosen bandwidth.

U.S. Senate 793 contested general elections, 1976–2024. ~32 races per 2-year cycle. U.S. House 9,348 contested general elections, 1976–2024. ~374 races per 2-year cycle.

Step 2 Bandwidth

The margin-of-victory window around 50% that defines your quasi-experiment. Narrower bandwidths improve the RD validity assumption but shrink your sample — and the sample is already very small.

5.0 pp

Races within bandwidth: —

Step 3 Effect size

The minimum effect you want to detect, in Cohen's d (standardized by the SD of the control group's outcome). d = 0.2 is "small" by Cohen's conventions. For vote share (SD ≈ 18 pp), d = 0.2 corresponds to ~3.6 pp.

0.20

In vote-share units (SD ≈ 18 pp): —

Step 4 Estimation correction

Local linear regression at a boundary inflates variance relative to a simple group comparison. A factor of 2× is typical for triangular-kernel local linear; 4× is conservative for uniform kernels or sparse support.

Variance inflation (C)

1× (none) 2× (typical) 4× (conservative)

Step 5 Significance level

α (two-tailed)

0.05 0.01

Power curve

Power to detect each effect size at your selected bandwidth and correction. Orange line marks your current d. Dashed line = 80% target.

Sensitivity table

Power (%) at each bandwidth and effect size, for your chosen chamber and correction factor. Highlighted row = current bandwidth.

Data. U.S. Senate general election results 1976–2024: MIT Election Data Science Lab via Harvard Dataverse (doi:10.7910/DVN/PEJ5QU). Two-party margin computed as |Democratic vote share − 0.50| × 200; non-partisan, multi-candidate, and special elections excluded. U.S. House data: MIT Election Data Science Lab (doi:10.7910/DVN/IG0UN2); requires one-time guestbook registration.

Power formula. Power = Φ(d√N / (2C) − z_α/2), where N = races within bandwidth, C = local-linear variance inflation factor, and d = Cohen's d. This approximation treats each side of the cutoff as an equal group; C accounts for the efficiency loss of boundary estimation in local linear regression relative to a simple mean comparison. Stommes, D., Aronow, P.M., & Sävje, F. (2023). On the reliability of published findings using the regression discontinuity design in political science. Research & Politics, 10(2). doi:10.1177/20531680231166457

Working paper. Holbein, J.B. (2026). How underpowered are close-election regression discontinuity designs? Evidence from U.S. Senate, House, gubernatorial, state legislative, and local races. Manuscript in preparation.