A study of P- and Lg-based source size estimators has been conducted for 101 presumed underground nuclear explosions at the Shagan River test site near Semipalatinsk, USSR. A systematic variation in the relative patterns across the test site has been observed when comparing the P-wave estimators (world-wide mb and Ψ) and the Lg-based observations (RMS Lg determined from NORSAR and Gräfenberg array recordings). These variations correlate well with geological features determined from satellite imagery, and suggest in particular that the northeast and southwest portions of the test site, divided by the Chinrau fault, are characterized by distinct geophysical properties. A difference in mb(P) –mLg bias of as much as 0.15 magnitude units exists between these two subareas. Available data indicate that the mLg observations are very consistent with published yields, and suggest that the magnitude-yield relationship for P-wave based estimators can be significantly improved by introducing regional bias corrections. We have determined such corrections for three subareas of the test site, and used the corrected mb and Ψ observations in combination with the Lg data in a weighted averaging scheme to arrive at yield estimates for all events in the database. For the explosion of 1988 September 14, we obtain yield estimates from the three estimators ranging from 106 to 118 kt, which is consistent with the independently derived estimates of 115–122 kt quoted by Sykes & Ekström (1989). A further assessment of the validity of our results would require access to independently measured yields of additional explosions.