The goal of this research is to jointly study physician prescription decisions and patient drug request behaviors. We have adopted a binary logit model and a multinomial logit (MNL) model to study patient drug request data with excessive zero requests and a MNL model to capture physician prescription decisions. These models are further joined by a flexible nonparametric multivariate distribution for their random effects. We also adopt an analytically consistent expression for interaction effects in our non-linear and joint modeling framework. We apply our model to a unique physician panel data set from the Erectile Dysfunction category. Our key empirical findings include the following: (1) the triggering of drug requests by DTCA is complex with category level DTCA reducing patients' probabilities of making drug requests and drug specific DTCA driving drug requests for the advertised drug; (2) patient characteristics may play a role in both the impact of DTCA on drug requests and the impact of patient requests on physician prescription decisions; (3) patient drug requests have a positive impact on physician prescription decisions and patients can be consistent with physicians in choosing a drug based on their diagnosis levels and some unobserved factors; (4) there are significant correlations among physician-level random effects that drive both patient drug requests and physician prescription decisions, which validates the joint modeling approach.