I am running a cox proportional hazard model to analyze the association between air pollution and risk of mortality. In the subgroup analysis, I obtained some confusing results. For example, when stratified by age (<60 vs >=60), the hazard ratio (95% confidence interval) between air pollution and mortality were 1.12(1.06-1.18) for <60 and 1.14(1.08-1.20) for >=60, and interaction P value for air pollution * age is 0.008. However, when stratified by alcohol drinking status (never vs ever), the HR(95% CI) were 1.12(1.07-1.16) for never drinkers and 1.19(1.09-1.29) for ever drinkers, and the interaction P value for air pollution * drinking is 0.606. I feel confused that why the groups with similar effect estimates have significant interaction, while large differences in HR between groups have non significant interaction.