Statistics of earthquakes in California show linear frequency-magnitude relationships in the range of M2.0 to M5.5 for various data sets. Assuming Gutenberg-Richter distributions, there is a systematic decrease in b value with increasing depth of earthquakes. We find consistent results for various data sets from northern and southern California that both include and exclude the larger aftershock sequences. We suggest that at shallow depth (∼0 to 6 km) conditions with more heterogeneous material properties and lower lithospheric stress prevail. Rupture initiations are more likely to stop before growing into large earthquakes, producing relatively more smaller earthquakes and consequently higher b values. These ideas help to explain the depth-dependent observations of foreshocks in the western United States. The higher occurrence rate of foreshocks preceding shallow earthquakes can be interpreted in terms of rupture initiations that are stopped before growing into the mainshock. At greater depth (9–15 km), any rupture initiation is more likely to continue growing into a larger event, so there are fewer foreshocks. If one assumes that frequency-magnitude statistics can be used to estimate probabilities of a small rupture initiation growing into a larger earthquake, then a small (M2) rupture initiation at 9 to 12 km depth is 18 times more likely to grow into a M5.5 or larger event, compared to the same small rupture initiation at 0 to 3 km.