Selecting likely causal risk factors from high-throughput experiments using multivariable Mendelian randomization