The sign of an association measure between two varibles may be strongly affected and even be reversed after marginalization over a backgruoud variable, which is the well-known Yule-Simpson paradox.Odds ratios are stro...The sign of an association measure between two varibles may be strongly affected and even be reversed after marginalization over a backgruoud variable, which is the well-known Yule-Simpson paradox.Odds ratios are strongly collapsible over a background variable if they remain unchanged no matter how the background variable is partially pooled.In this paper, we firstly give some definitions and notations about odds ratios between a dichotomous explanatory variable and a continuous response variable.Then, we present conditions for simple collapsibility of odds ratios.Further, necessary and sufficient conditions are given for strong collapsibility of odds ratios for continuous outcome variable.展开更多
Simpson' s paradox reminds people that the statistical inference in a low-dimensional space probably distorts the reality in a high one seriously.To study the paradox with respect to Yule's measure, this paper...Simpson' s paradox reminds people that the statistical inference in a low-dimensional space probably distorts the reality in a high one seriously.To study the paradox with respect to Yule's measure, this paper discusses simple collapsibility, strong collapsibility and consecutive collapsibility, and presents necessary and sufficient conditions of them.In fact, these conditions are of great importance for observational and experimental designs, eliminating confounding bias, categorizing discrete variables and so on.展开更多
基金Funded by Fundamental Research Funds for the Central Universities (Grant No.BUPT2012RC0708)
文摘The sign of an association measure between two varibles may be strongly affected and even be reversed after marginalization over a backgruoud variable, which is the well-known Yule-Simpson paradox.Odds ratios are strongly collapsible over a background variable if they remain unchanged no matter how the background variable is partially pooled.In this paper, we firstly give some definitions and notations about odds ratios between a dichotomous explanatory variable and a continuous response variable.Then, we present conditions for simple collapsibility of odds ratios.Further, necessary and sufficient conditions are given for strong collapsibility of odds ratios for continuous outcome variable.
基金This work was supported by the National Natural Science Foundation of China (Grant Nos. 19831010 and10001008) the National Excellent Youth Science Foundation of China.
文摘Simpson' s paradox reminds people that the statistical inference in a low-dimensional space probably distorts the reality in a high one seriously.To study the paradox with respect to Yule's measure, this paper discusses simple collapsibility, strong collapsibility and consecutive collapsibility, and presents necessary and sufficient conditions of them.In fact, these conditions are of great importance for observational and experimental designs, eliminating confounding bias, categorizing discrete variables and so on.
