How the Monty Hall problem is similar to the false discovery rate in high-throughput data analysis