Political Portfolio

Thesis

Campaign finance is usually studied as a causal question: does spending affect elections?

In practice, campaign committees face a different problem: given a fixed budget, where does the next dollar generate the highest expected increase in seats?

This reframes campaign spending as a capital allocation problem under uncertainty, not a simple outcome regression.

Approach

I built a three-layer framework to estimate marginal seat gain per dollar.

1. Causal Spending → Vote Margin
Using Levitt's (1994) repeat-challenger design, I isolate the causal effect of spending on vote margin within matched incumbent–challenger pairs (2012–2022). This reduces endogeneity from strategic spending decisions.

2. Vote Margin → Win Probability
Vote margin is mapped nonlinearly into win probability, reflecting the fact that dollars are more valuable in competitive races and marginal effects are convex near 50/50 contests. This converts spending into expected seat contribution rather than raw vote share impact.

3. Portfolio Allocation Across Races
House races are not independent. National swings introduce correlated risk through the generic ballot. I model this explicitly using a shared factor structure across races, constructing a covariance-aware allocation model that treats the House map as a portfolio rather than isolated bets.

Key Insight (Open Seats)

Open seats behave differently from incumbency races. They resemble higher-variance assets:

• Wider outcome uncertainty increases baseline win probability
• But reduces marginal return to additional spending

This creates a risk/return tradeoff analogous to volatility exposure in options pricing, which is naturally captured through a Bayesian uncertainty model.

Empirical Test

If spending is efficient, observed allocations should be positively correlated with estimated marginal seat gain. I test this among competitively matched races controlling for partisan lean and incumbency.

Results

• ρ = −0.597 (p < 0.001) across 53 competitive 2024 races → spending is systematically misallocated relative to marginal seat gain
• +5.3 expected seats from reallocating the same budget using model ranking
• Out-of-sample validation (trained on 2012–2020, tested on 2022): ρ = −0.380, +5.5 seats improvement
• Calibration benchmark — Brier score: 0.028 (model) vs 0.038 (Cook PVI baseline)

Contribution

• Reframes campaign finance as a constrained optimization problem over expected seat gain
• Builds a causal estimate of spending → vote margin using repeat-challenger identification
• Extends inference into a risk-aware allocation model across correlated races
• Produces a falsifiable, out-of-sample claim about allocation inefficiency

Methods

Repeat-challenger causal design, Bayesian shrinkage, bounded treatment effects (Oster-style sensitivity analysis), factor-based covariance modeling, nonlinear optimization, bootstrap inference.

Data: MIT Election Lab, FEC filings, OpenSecrets, Cook PVI, Census ACS (all public sources)