When modeling the relationship between marketing actions (e.g., price) and an outcome (e.g., sales), possibly the most important choice that researchers make is the model specification (e.g., which functional form to use, which covariates to include). To this end, researchers often compare in-sample fit across models and select the best. When following this common practice, researchers think they are selecting only the model, but this is not the only decision that they are making. In fact, researchers unknowingly make two other important modeling decisions: data granularity and parameter granularity. Data granularity denotes decisions about whether to use the most granular data (e.g., daily) or to aggregate it to a coarser level (e.g., weekly). In a similar vein, parameter granularity denotes decisions about whether to assume that parameters in a model vary across the most granular units (e.g., products) or at a more aggregate level (e.g., brands). In this dissertation, which consists of two essays, we highlight the importance of these two decisions and propose an integrative solution that selects both data and parameter granularities. We refer to the solution as Bayesian dual-graph clustering. The method, as the name suggests, represents data and parameters as two separate graphs with nodes being the unit of analysis (e.g., time, space, SKU, person) and selects data and parameter granularities by clustering the two graphs accordingly. We conclude by discussing the generalizability of the proposed method and its potential to address a wide range of marketing problems.