Introduction
Finite mixture models are essential tools in causal inference, particularly within the Principal Stratification framework. This framework allows researchers to estimate treatment effects in the presence of post-randomization complications, such as non-compliance or attrition. By conceptualizing observed data as a mixture of latent groups (e.g., "Compliers" and "Never Takers"), researchers aim to disentangle overlapping distributions to identify group-specific effects.
However, standard statistical theory for Maximum Likelihood Estimation (MLE) assumes these latent groups are well-separated. In many social science applications, these groups overlap significantly, leading to what is known as "weak separation." This paper explores the "Twilight Zone of Statistics," demonstrating that in such regimes, the MLE suffers from severe pathologies, including glacial convergence rates and systematic estimation failures.
The "Pile Up" Phenomenon and Slow Convergence
A primary finding of this research is the extreme sensitivity of convergence rates to group separation. While standard estimators typically converge at a rate of $O(n^{-1/2})$, weakly separated mixtures can exhibit rates as slow as $O(n^{-1/6})$ or even $O(n^{-1/8})$. This implies that achieving a marginal increase in precision requires an exponentially larger sample size.
Beyond slow convergence, the MLE often exhibits "pile up"—a phenomenon where the estimator identifies zero difference between component means even when a true difference exists. This occurs when the observed variance of the data is less than the expected within-component variance, causing the likelihood to favor a single-component solution.
Likelihood Pathologies and Sign Errors
The interaction between location (mean) and scale (variance) parameters in these models creates a complex, often "flat" likelihood surface. When variances are unknown, the MLE can easily become trapped in local maxima or produce "sign errors," where the estimated effect direction is the opposite of the truth. In many realistic weakly separated settings, the sign of the estimated effect is statistically no more reliable than a coin flip.
Empirical Re-analysis: JOBS II and Job Corps
The practical implications of these findings are significant. Re-analyzing the JOBS II study ($N=132$), which evaluated a job search intervention, the authors found a 63% probability of pile up. The original finding of a "zero effect" in some literature may not reflect reality but rather the estimator's inability to resolve the mixture.
Similarly, in the Job Corps evaluation ($N=3,371$), despite a larger sample size, a highly skewed mixing proportion (6%) led the MLE to produce implausibly large effect sizes. These results suggest that many published findings using Principal Stratification in the social sciences may be artifacts of these statistical pathologies rather than true causal effects.
Conclusion
When separation between latent groups is weak, parametric mixture models can be deceptive. The authors suggest that researchers should utilize moment-based diagnostics to assess the risk of pile up and sign errors. In cases of high risk, relying on non-parametric bounds may provide a more honest assessment of the data's informational content than the potentially spurious precision of the MLE.
Summary updated