Estimator consistency measures whether an estimate converges toward a parameter’s true value as the amount of data increases. Because an estimator is computed from random data, it is itself a random variable with a distribution. Consistency asks how that distribution behaves as more observations become available. section 2.3.4 section 2.3.4
Use consistency to assess estimation methods alongside related statistical and computational properties:
- Distinguish consistency from unbiasedness: unbiasedness concerns the estimator’s mean, whereas consistency concerns convergence with more data. section 2.3.4
- Check whether the estimator scales with the dataset and can be parallelized, not merely whether it converges theoretically. section 2.3.4
- Treat computational behavior as a practical constraint when estimating causal parameters in applications. section 2.3.4
Convergence in an application
A consistent estimator may approach the true parameter with more data, yet still be unsuitable for an iPhone app if convergence takes too long or consumes excessive battery. section 2.3.4
