Heterogeneous treatment efficacy across special subgroups in clinical trials has been frequently observed in various therapeutic areas. Recent multi-region studies and effect modifying biomarkers push this issue to a more prominent position than ever. The impact of (1) quantitative or qualitative treatment by subgroup interaction, (2) subgroup sample size, and (3) non-random sampling on the behavior of subgroup analyses in terms of false-positive rate, false negative rate and estimation bias, are examined using simulation studies. Rules of planning, analysis, and interpreting of subgroup analyses in clinical trials, especially in oncology trials, are proposed.