Empirical analyses of observed pairs of variables often suggest that the underlying joint distribution of these variables is highly dependent and each one of them is marginally unimodal. Exploiting marginal shape information can improve the bivariate density estimation. We develop a nonparametric density estimation method where the joint density of the bivariate outcome variables is estimated using a flexible class of mixtures of scaled Beta distributions and the mixing weights are suitably constrained to ensure marginal unimodality. Large sample consistency of the proposed density estimator is established using the method of sieves and Argmax Continuous Mapping Theorem. The proposed method enlarges the scope of previous studies in two important aspects: (i) joint density provides a simultaneous estimation of higher-order moments beyond mean and variance functions, and (ii) marginal shape constraints provide more efficient estimates. The proposed procedures are illustrated using simulated and real case study datasets on gestational age and infant birth weight.
Read full abstract