Content Sampling, Household Informedness, and the Consumption of Digital Information Goods

ABSTRACT

Technology and media are delivering content that is transforming society. Providers must compete for consumer attention to sell their digital information goods effectively. This is challenging, since there is a high level of uncertainty associated with the consumption of such goods. Service providers often use free programming to share product information. We examine the effectiveness of content sampling strategy used for on-demand series dramas, a unique class of entertainment goods. The data were extracted from a large set of household video-on-demand (VoD) viewing records and combined with external data sources. We extended a propensity score matching (PSM) approach to handle censored data, which permitted us to explore the main causal relationships. Relevant theories in the marketing and information systems disciplines informed our research on consumer involvement and informedness for decision making under uncertainty, the consumption of information goods, and seller strategies for digital content. The results show that content sampling stimulates higher demand for series dramas, but in a more nuanced way than was expected. Samples of the series reveal quality information to consumers and allow them to assess preference fit directly. As a result, they become more informed about their purchase decisions. Also, households seem to be willing to pay more to be better informed, and informed households tend to purchase more. This suggests that content providers should invest in strategies that help consumers to understand the preference fit of information goods.

Key words and phrases:

• content sampling
• data censoring
• digital entertainment
• household informedness
• information asymmetry
• information goods
• online content
• preference fit
• propensity score matching
• video-on-demand

Acknowledgments

This research was originally supported by the Singapore National Research Foundation under its International Research Centre @ Singapore Funding Initiative and administered by the Infocomm Development Authority. An earlier version was: A.P. Hoang and R.J. Kauffman, “Experience Me! The Impact of Content Sampling Strategies on the Marketing of Digital Entertainment Goods,” in R. Sprague and T. Bui (eds.), Proceedings of the 49th Hawaii International Conference on Systems Science (Kauai, HI: IEEE Computer Society Press, Washington, DC, 2016), where it was nominated for the Best Research Paper Award in the Organizational Systems Track. In addition, some aspects of the empirical methodology related to censored data and propensity score matching were presented at the 2017 Statistical Challenges in Electronic Commerce Workshop in Ho Chi Minh City, Vietnam, in July 2017. We are grateful to the late Steve Fienberg for his support and feedback in seminars at Carnegie Mellon University. Vladimir Zwass and the anonymous referees helped to shape our intellectual contributions in this research. We also acknowledge Steve Miller, Kapil Tuli, Qian Tang, and other academic participants in seminars at the Living Analytics Research Centre (LARC) at Singapore Management University, for their thoughtful comments. Eric Clemons, Atanu Lahiri, Jennifer Zhang, Yabing Jiang, Avi Seidmann and the participants of the “Integrating Business Operations, Information Technologies, and Consumer Behavior Mini-Track,” gave us many suggestions that have made this research stronger. Rong Zheng, Tuan Phan, Ting Li, and the participants of the 13th Statistical Challenges in E-Commerce Research (SCECR 2017) Workshop offered helpful suggestions, and Terence Saldanha and Zhoulun Li also gave us feedback. We are especially grateful for support and guidance from the anonymous reviewers. This research was conducted with the participation of a corporate sponsor under a binding nondisclosure agreement; thus, some details of the data and qualitative findings were disguised. The households and data used in this research were anonymized. Also, the identities of individual households and account holders cannot be traced back through our analysis. All errors and omissions are the responsibility of the authors.

Notes

1. A series drama consists of 10, 20, 30, or more episodes. Most American TV series, packaged since the 1960s with 20 to 26 episodes a season, are in this format. The economic importance of paid TV series revenue streams has increased, while providers have been fighting for profitability in the face of Internet delivery and digital convergence. Producing an original TV series requires a huge investment: about US$2 million to shoot a half-hour pilot and about US$5.5 million for an hour-long drama [62].

2. In this study, we consider a unitary model of the household in which the viewing time constraint, demand, and preferences of all household members are pooled [73]. The current technology in our setting did not permit tracking individual viewers.

3. Content sampling signals both horizontal and vertical differentiation on objective features of a series to consumers. If the content only signals vertical differentiation, then consumers just need to know such samples are available, and they do not actually need to watch any free-sample episodes.

4. There are some drawbacks to free content. A perception that free content is available may dissuade consumers from buying programs [44]. Also, unlimited access to free content makes other programs less attractive and decreases consumers’ willingness to pay [5]. Further, some consumers may sample with no intention to purchase anything, though this is unlikely for a majority of them in the VoD setting for several reasons. Series dramas are unique, so a viewer’s experience is not complete without seeing it all. So, after viewing the free sample of a series’ first episode, viewers may feel connected and want to view the rest of the content [78]. Those that sample a portion of the series are more likely to purchase the remainder of it. In addition, since households will have many channels in their TV subscriptions, they are unlikely to watch a free sample episode of a series if they have no prior topical interest.

5. Netflix’s method of releasing a series—in its entirety—has helped the company to understand customer viewing behavior for the different series it offers across various market segments. This is relevant to our context, since it shows that a one-episode free sample may not be sufficient for the viewers [45].

6. In our research context, the households decided on the number of standard content clusters in their TV subscriptions at the beginning of long-term service contracts.

7. Households often finish watching an episode across multiple viewing sessions, as each episode takes more than 30 minutes. So, if a household had three free-sample sessions for a series, we only admitted the earliest session to our data set based on its timestamp, and removed other duplicates.This was normally not permissible.

8. Meaningful stratification is sometimes difficult with in big data analytics research. Even though the researcher may have access to a lot of data, often it is surprisingly hard to develop research designs to support causal analysis, such as researcher-designed field experiments, and quasi-experimental designs that have “just right” conditions that can be leveraged to produce undeniably correct managerial insights.

9. In the different count data models that we used, we did not include any household demographic characteristics as control variables. Instead, we used them in our propensity score matching approach, so this would have been double-counting to add them as control variables also. These variables include the demographic segmentation of the household, such as the region of the residence, age band, and gender of the residents. Other specifics regarding the ethnicity of the anonymized households are not included or reported, due to our nondisclosure agreement with the research sponsor. In fact though, these variables did not add much explanatory capability for the dependent variable of interest.

10. Hurdle models also relax the assumption that the zeros and nonzeros in the data set come from the same data-generating process. They use a Bernoulli probability that governs the binary outcome for the count variable with a 0 or a positive count. Once the hurdle or threshold is crossed, and a positive number occurs, the conditional distribution is represented by a truncated-at-zero count data model. Since we had prior knowledge of the cause of the excess zeros, we chose to proceed with zero-inflated models.

11. In censored data, the total number of observations is known but full information is not available for some [17]. Left-censoring arises when the events of interest occurred before the study period; right-censoring refers to events that might or might not have occurred after the period of observation ended. Data without censoring are ideal for empirical testing.

12. We justify the use of the NB model by showing that the data are overdispersed. The Poisson model is nested in the NB model. It relaxes the assumption that the conditional variance is equal to the conditional mean. We use a likelihood ratio test to assess the null hypotheses to see if this restriction is true: λ = –2 · (LL_NB – LL_Poisson). We rejected the null hypothesis that it is appropriate in favor of the NB model, based on χ² = 394.29. This exceeds 2.71 (p < 0.001), so overall the evidence suggested the data are overdispersed.

13. We show that the ZINB model fits the data better than the null intercept-only model does. The associated χ² value for the difference between the model-level log likelihoods, λ = –2 (LL_ZINB – LL_Null) is 408.64. So the ZINB model is preferred over the null intercept-only model.

Additional information

Notes on contributors

Ai-Phuong Hoang

Ai-phuong Hoang ([email protected]; corresponding author) is a Ph.D. candidate in the Interdisciplinary Doctoral Programme in Information Systems and Marketing at Singapore Management University. Her research interests involve consumer behavior, social network and marketing strategy for digital information goods, and methodology innovations for causal inference with consumers’ digital trace data. Her research work appears in Electronic Commerce Research and Applications and in several leading IS conferences and workshops. From 2015 to 2016, she was a visiting Ph.D. student at the Heinz College of Information and Public Policy, Carnegie Mellon University.

Robert J. Kauffman

Robert J. Kauffman ([email protected]) is a Professor of Information Systems at the School of Information Systems, Singapore Management University. He also serves as Associate Dean (Faculty). He was a Distinguished Visiting Fellow at the Center for Digital Strategies at the Tuck School of Business at Dartmouth. He has served on the faculty of the business schools of New York University, University of Rochester, University of Minnesota, and Arizona State University. His graduate degrees are from Cornell and Carnegie Mellon. His interdisciplinary research spans strategy, information, information technology, economics, marketing and consumer behaviour, the fintech revolution, environmental sustainability, and computational social science and data analytics. He is a frequent keynote speaker on these subjects. His research has appeared in Management Science, Information Systems Research, Journal of Management Information Systems, MIS Quarterly, Organization Science, and the Review of Economics and Statistics, among others.