Marc Dotson, Marketing & Global Supply Chain
Academic Objectives
I set out to apply the MEG Grant to a project entitled “Validating Market Segmentation Solutions.” Early on in the process, it became clear that, while this project was and still is of managerial importance, I could do more academically relevant project work with more students if I were to switch gears to a number of other projects. Instead of the single project, the grant has been used over the past two years to help fund mentoring activities with students on the following projects:
• “Counting the Cockroaches in the Walls: Assessing the Severity and Diffusion of Service Failures Through Social Chatter”
• “Accommodating Multiple Data Pathologies in Conjoint Studies via Clever Randomization and Ensembling Strategies”
• “An Empirical Generalization of the Effects of Category Captainship”
• “Identifying the Drivers of Individual NGO Donations using Tradeoff Analysis”
Student work on these projects has satisfied the overall objective stated in my proposal, namely that “the intent is to work with students in the development of … modeling components.” Instead of market segmentation, on-the-fly classification, and validation, many students across all four projects have worked on and had exposure to advanced predictive modeling techniques, modern Bayesian model estimation and probabilistic programming, large-scale data cleaning and manipulation, and current tools for conjoint analysis and market simulation. All of this was relevant for students in their pursuit of employment and graduate studies.
Mentoring Environment
This has been a great learning experience for working with so many students.
Given the number of research assistants, I’ve had to task the senior students with helping to lead the project and train fellow students. We’ve been able to meet off and on for training and updates, but I’ve largely come to rely on project leads for the student-led components of each of the projects.
Students and Deliverables
• Adriel Casellas – Wordvec-based classification of complaints on Twitter.
• Christopher Wallace – Data cleaning and complaint classification.
• Kyle Adams – Data cleaning and complaint classification.
• Megan Hopkins – Survey and conjoint experiment.
• Mitchell Kimball – Survey and conjoint experiment.
• Derek Miller – Ensembling-based approach to estimation; co-authoring paper.
• Samuel Sorensen – Data cleaning and manipulation and regression.
• Morgan Bale – Data cleaning and manipulation and regression.
• Cameron Bale – Data cleaning and manipulation and regression.
Results/Findings
The work on each of these four projects is ongoing, with a collection of some of the same students and a new batch of students, but here is a brief summary of the current status and objectives of each of the projects.
Counting the Cockroaches in the Walls: Assessing the Severity and Diffusion of Service Failures Through Social Chatter – The objective is to make inference using text data from social media, which is typically used for description only. In particular, we want to estimate the size, extent, and severity of a service failure based only on the non-representative sample of complaints posted on social media. We have data, have built a classification tool for the complaints, and are in the process of fitting an initial model derived from ecology and its population size estimation literature using modern probabilistic programming.
Accommodating Multiple Data Pathologies in Conjoint Studies via Clever Randomization and Ensembling Strategies – The objective is to improve predictive performance for market simulators by running an ensemble of standard models where the ensemble is built in such a way so that it covers potential issues with the data.
These data issues result from different respondent behaviors and are understood in the literature, with specialized, hard-to-fit models developed to address them. These hard- to-fit, specialized models are difficult to use in practice for a number of reasons, including because it’s rarely obvious when they are needed and there is reason to believe that multiple of these issues or pathologies may be present at any one time, without any specialized model to deal with such multiple pathologies. We have all the data we need, working code of the ensembling approach, and we are finalizing a working paper with initial results based on simulated datasets.
An Empirical Generalization of the Effects of Category Captainship – Using large- scale data, the objective is to generalize the findings of category captainship across retailers and not only for a few specific retailers, which was the empirical application in the original paper. We have data and have conducted extensive data cleaning and initial modeling.
Identifying the Drivers of Individual NGO Donations using Tradeoff Analysis – The objective is to understand the tradeoffs donors make when deciding if and how much to donate to various non-governmental organizations. This is work in the public policy space. We have a draft survey and conjoint experiment and will be going to field within the month.
Budget Spend
The budget was spent to pay research assistants, purchase data and survey panel respondents, and fund student attendance at four conferences to get training and exposure to the latest tools in pr babilistic programming, statistical programming, andpredictive modeling techniques.