This paper demonstrates how the observed correlation between probability of default and loss given default depends on the fact that defaults in which collateral provides 100% recovery are not observed. Creditors see only the defaults of mortgagors who suffer from a fall in collateral value to less than the remaining loan principal. Consequently, the default data available to creditors amounts to a mere truncated sample from the underlying population of defaults. Correlation estimates based on such truncated samples are biased and differ substantially from estimates derived from representative non-truncated samples. Moreover, the observed correlation between default probability and loss given default is sensitive to the truncation point, which may explain the differences in correlation estimates found in the literature. This may also explain why correlation estimates seem to be specific to cycle phase.