We address a classic problem in observational seismology, namely the precise estimation of the focal depth of small to moderate earthquakes. The data available are three‐component records including Pn onsets and their coda. We believe that the Pn coda contains P‐type phases that originate in the source area and lag behind due to longer travel paths. Phases such as pP and sP are generally not included in seismic bulletins, and even cepstrum and array f‐k techniques have not been successful in extracting such arrivals. We use recent developments in polarization analysis (Roberts and Christoffersson, 1990; Fedorenko et al., 2008) to extract secondary P phases in local and regional records from five earthquakes. Phase characteristics are typically slowness and lag time relative to the preceding Pn phase. However, information is incomplete for confidently deciding whether an arriving wavelet should be classed as pP, sP, or something else. This validation problem is resolved by picking several coda arrivals that can uniquely be fitted to travel‐time curves for suites of Pn‐coda waves. Focal‐depth estimates, even for Ground Truth 05 events, are seldom more accurate than to 5 km, and therefore we compared our results with those from moment tensor analysis. Differences in depth estimates were less than 5 km and for H<12 km the first secondary Pn arrival is generally sPn. The Dt(Pn−pPn) lag time for shallow events is 2.0 s or less and hence difficult to detect in the coda. Most techniques for focal‐depth estimation are not user friendly. In contrast, the polarization schemes tested are easy to use and depth estimates are more accurate than those derived from ground truth studies based exclusively on P‐wave arrivals.