Spatial econometric models are increasingly popular across political science but require specifying a dependence network (W) before estimation. The true structure of W is rarely known, and theories often provide little guidance, producing a form of model uncertainty labeled here as network uncertainty.
🧭 What This Study Tests
This study evaluates how uncertainty about the network W affects inference in spatial models and assesses whether Bayesian model averaging (BMA) provides a robust remedy that balances theoretical priors with empirical evidence.
🧪 How Model Sensitivity Was Evaluated
- Monte Carlo experiments were used to simulate a range of plausible networks and model specifications.
- Two real-world replication studies from different subfields of political science were reanalyzed to show applied impacts.
- Comparisons were made between standard spatial autoregressive approaches, other common alternatives, and BMA when faced with misspecified networks.
🔍 Key Findings
- Effect estimates are generally robust to misspecification of the functional form of W (for example, choice of weighting scheme).
- Uncertainty in neighborhood definition — which observations count as neighbors — can produce biased effect estimates in spatial autoregressive models.
- BMA directly addresses network uncertainty by averaging over a set of feasible W specifications, correctly identifying the true network when it is included among alternatives and yielding unbiased effect estimates.
- In contrast to alternative techniques, BMA both acknowledges uncertainty about W and leverages the data to reduce bias.
💡 Why This Matters for Applied Research
- Network misspecification can meaningfully distort conclusions in spatial analyses; attention to neighborhood definition is essential.
- BMA offers a practical, theory-informed strategy to mitigate network uncertainty and improve inference in spatial econometrics.
- The replication examples demonstrate that adopting BMA can change substantive conclusions in empirical political science work.
