Year: 2019
Type: dissertation
Abstract: This dissertation consists of three essays organized around Bayesian statistical and nonparametric methods to infer latent variables either unavailable or in-principle unobservable, yet nonignorable for robust data-driven marketing decision-making. This work examines applied contexts across ecommerce and online services using large-scale customer-level data from experimental and observational settings. Its development is motivated by the need to address the intersection of (1) marketing response models to tractably handle the high-dimensionality and volume of customer-level data sources faced by firms today, while (2) enabling robust inference on decision variables in the presence of nonignorable missingness, e.g., those arising from selection, truncation, and censoring. As these missingness typically take the form of partially observable response outcomes, they are unlikely to be fully overcome by the growth in data size alone, or even be feasibly collected (e.g., A/B tests without the existence of an outside option). Compounding this problem, classical methods for addressing nonignorable missingness (e.g., Heckman, covariate-based matching, parametric data fusion) are often hampered in their effectiveness and tractability by the vast scale of these modern datasets. To this end, this work focuses on the development of econometrically principled and computationally scalable Bayesian techniques for robust model-based inference on contemporary marketing data sources and problems, with specific focus on using Gaussian processes as a nonparametric prior over latent variables to flexibly and parismoniously infer the effect of missing data. The first essay explores how freemium users consider tradeoffs when faced with menu-based pricing using an orthogonalized pricing experiment. Here a data augmentation framework is fashioned for the purpose of recovering the joint distribution over plan choices by all users from the conditional distribution of plan choices by only upgraders, the latter inherently availed to us by the experiment. Scaffolding on the first essay’s parametric framework for missing data imputation, the second essay proposes a flexible Bayesian nonparametric data fusion framework over multiple data sources to infer the long-run subscription renewal outcomes of new users. The third essay addresses a prevalent form of unscored gaps in consumer credit scores, referred to as sparsely unscoreable, whereby a consumer’s credit score portfolio contains periods that contemporaneously exhibit a mix of observed and missing scores. To address this problem, a Bayesian nonparametric latent factor model is developed to impute credible intervals for gaps in individual score histories within a portfolio of dynamically and contemporaneously interrelated scores.
Author: Longxiu Tian
Language: English
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2988005828