Bayesian two-part multilevel model for longitudinal media use data

Abstract

Multilevel models are effective marketing analytic tools that can test for consumer differences in longitudinal data. A two-part multilevel model is a special case of a multilevel model developed for semi-continuous data, such as data that include a combination of zeros and continuous values. For repeated measures of media use data, a two-part multilevel model informs market research about consumer-specific likeliness to use media, level of use across time, and variation in use over time. These models are typically estimated using maximum likelihood. There are, however, tremendous advantages to using a Bayesian framework, including the ease at which the analyst can take into account information learned from previous investigations. This paper develops a Bayesian approach to estimating a two-part multilevel model and illustrates its use by applying the model to daily diary measures of television use in a large US sample.

Mapping 2022 in Journal of Marketing Analytics: what lies ahead?

“…The time and longitudinal evolution theme includes analytics and data analysis methods over time (Blozis 2022), as well as the insights of bibliometrics and their relevance in marketing research and analytics (Caputo and Kargina 2022;Ghorbani et al 2022). JMA studies also reflected on an essential topic in marketing analytics-i.e., prediction-in relation to performance and profitability, as well as machine learning methods (Valluri et al 2022;Vriens et al 2022).…”

How to really quantify the economic value of customer information in corporate databases

“…Normally, they are used to study different business processes, such as customer acquisition, retention, and profitability, among others (Kumar and Reinartz, 2016). For analyses that aim to study differences between customers, such as segmentation research, data collection at a single point in time is appropriate (Blozis, 2022). This model is based on the multiplicative effect of each construct as a function of a potential transactional CLV that supports the economic value of a client's information.…”

Fast estimation of mixed‐effects location‐scale regression models

“…Moreover, due to the computational cost, using resampling methods like the bootstrap for inference on the model parameters is not practical, leaving researchers without a robust way to estimate standard errors or determine significance of model parameters. Researchers have also used Bayesian methods to fit the MELS model 17–20 . These techniques are natural given the hierarchical structure of the MELS model, but they are not meant to reduce computational burden.…”

“…Researchers have also used Bayesian methods to fit the MELS model. [17][18][19][20] These techniques are natural given the hierarchical structure of the MELS model, but they are not meant to reduce computational burden. We will restrict ourselves to the frequentist interpretation of the model.…”