Model selection based on a predictive performance measure is compared to model selection based on maximum likelihood. Both procedures exhibit unstable relative performance when predictive performance is measured over time periods that may contain overly small proportions of total failures. Both procedures may be stabilized by bounding this proportion away from zero. In that case, both procedures exhibit similar predictive performance relative to the other. This provides evidence that model selection based on the commonly used maximum likelihood approach may in fact have good predictive performance.< &gt;