Spectral amplitudes of direct P and P coda between 0.5 and 6 Hz have been measured for 14 Eastern Kazakh test explosions using 8 hydrophones of the Wake Island Array, and a new model for these data has been developed to estimate relative yields of those explosions. Each log spectral amplitude is considered to be the sum of four parameters and an error term. Three parameters are frequency-dependent and represent: (1) the average Eastern Kazakh spectrum at the Wake Island Array; (2) the spectral shape characteristics unique to each event; and (3) the station or hydrophone corrections. The fourth parameter represents relative yield. A total of 346 parameters, including the 14 relative yields, were needed to model all 901 spectral amplitudes that were measured, and first-order linear regression techniques were used to invert the data for these parameters. Standard deviations of the computed relative yields are very small, averaging only 0.015 mb units. Deviations between the relative yields and their corresponding relative NEIS body-wave magnitudes, however, are larger, averaging 0.14 mb units. This difference is interpreted to be an indicator of the level of inhomogeneity in the worldwide pattern of energy radiated from Eastern Kazakh tests. There is evidence that this pattern can be correlated with the precise epicentral location of each explosion within the test site and is due to focusing and defocusing near the source. It is proposed that the yield of a large explosion might be disguised by siting the explosion at a location that selectively defocuses energy towards continents where most seismic stations are located. The precision with which a relative yield is determined from the data of any single hydrophone averages 0.035 mb units. Thus, the model appears to provide a very stable measure of yield. Errors associated with frequencies greater than 2 Hz are only slightly greater than errors associated with frequencies less than 2 Hz, verifying the utility of using the higher frequencies observed in direct P and P coda to estimate yield.