Real-time wireless link reliability estimation is a fundamental building block for self-organization of multihop sensor networks. Observed connectivity at low power is more chaotic and unpredictable than in wireless LANs, and available resources are severely constrained. We seek estimators that react quickly to large changes, yet are stable, have a small memory footprint, and are simple to compute. We create a simple model that generates link loss characteristics similar to empirical traces collected under different contexts. With this model, we simulate a variety of estimators, and use the simple exponentially weighted moving average (EWMA) estimator as a basis for comparison. We find that recently proposed flip-flop estimators are not superior, however, our cascaded EWMA on windowed averaging is very effective.