This paper describes a multivariate analysis of a 3-year prospective study of the covariates of age at menopause among women who were randomly sampled from a general population. This approach avoids three problems that may account for the inconsistent results of prior studies: recall bias in retrospectively collected data, failure to control confounding in data analyses, and self-selection in the target populations. At baseline, the 2014 subjects in this study were 46.5-56.5 years old and still menstruating. Median age at last menstruation estimated from a model of the marginal distribution of age at menopause, was 50.7 years. When covariates were considered singly, age at menopause varied with smoking, education and income but not with marital status, parity, location, height, weight or use of either oral contraceptives or menopausal estrogens. Multivariate analyses showed that education and income were confounded with smoking status. The results indicate that many previously identified sources of variation in age at menopause reflect confounding with smoking and self-selection in target populations, rather than real effects.