The polymerase chain reaction (PCR) is a procedure by which the DNA in a single cell can be made to replicate many times in a test tube. By amplifying the DNA from individual sperm cells and typing the results, estimates of male recombination fractions can be made, which are valuable for creating genetic maps and locating regions of unusually intense crossover activity on the human genome. Because PCR typing results are subject to random error, stochastic models must be constructed to obtain accurate results. In practice, to obtain enough information to accurately estimate small recombination fractions, it is necessary to combine data from several PCR experiments. Stochastic models in common use assume that PCR error rates are constant across experiments. We show by analysis of a dataset that PCR error rates can vary considerably from experiment to experiment, and that models that fail to take this heterogeneity into account can produce biased estimators. We present two new estimators and show with simulation studies that they perform better than conventional methods under realistic conditions. These estimators may be appropriate whenever PCR data from several experiments are combined.