Some conceptual problems in multivariable analyses of perinatal mortality.

Biostatisticians, epidemiologists and other researchers in maternal and child health have often used multivariable analysis to investigate associations with perinatal and infant mortality. A review of reports of such multivariable analyses published over the last 35 years revealed four problems that occur repeatedly. (1) Variables that are intermediate in the causal pathway between some study variables and perinatal death ('intervening variables') are controlled as though they were confounders. (2) When birthweight is analysed as an intervening variable, it is entered into the analysis in a few large categories, such as above and below 2500 g. This is not an adequate way of controlling for birthweight. (3) Social and demographic variables often interact statistically with birthweight in their effects on perinatal mortality, but these interactions have not been analysed in most multivariable studies. (4) Highly intercorrelated variables that represent similar theoretical constructs are entered simultaneously into one regression analysis. Solutions to these problems are suggested. Analytical approaches in which investigators use knowledge of biological and medical subject matter to make judgements about confounding and causal inference are encouraged.