Television ad-skipping, consumption complementarities and the consumer demand for advertising

Past frameworks for handling the micro-foundations of advertising include the informative model that posits that advertising affects demand by communicating information about products to consumers (Nelson 1970, 1974; Butters 1977; Grossman-and-Shapiro 1984), and the so-called persuasive model, in which advertising is incorporated into the utility from product consumption and viewed as a means of creating brand loyalty (please see Bagwell 2007 for a comprehensive review of the literature). The informative view is not a good description of ad consumption in our study. The product category we study is a fast moving consumer packaged good that has been on the market for years with no new brand entry during the time period of our data. Like Ackerberg (2001), we find that advertising continues to affect the purchase behavior of experienced consumers in the data even after significant product trial, suggesting its primary role is not to convey information about existence, attributes or match values. In the persuasive stream, advertising is usually treated as a taste shifter in utility, and there is usually no specific theoretical justification for its inclusion in the utility function.

Evidence in the literature suggests a link between those who respond to advertising and risk in such markets. Using randomized trials on direct-mail advertising, Ausubel (1999) documents that customer pools resulting from credit card offers with inferior terms (e.g., a higher introductory interest rate, a shorter duration for the introductory offer) have worse observable credit-risk characteristics and are more likely to default than those drawn in by solicitations offering superior terms.

For more reading, see Wilbur et al. (2013) who discusses the effect ad-skipping can have on audience composition and how networks might think about incorporating audience externalities into the pricing of ad spots.

For an example of such data, see http://​www.​finnpanel.​fi/​en/​tv.​php (not necessarily the same as the country in our sample.) Examples of the types of devices used here include the UNITAM meter by Nielsen Media Research (Unitam 2009) and the RapidMeter by Kantar (2012). For a video demonstrating how Nielsen’s system works, see https://​www.​youtube.​com/​watch?​v=​jYrVijea0UM.

If a consumer finds an advertisement immediately after switching channels, that advertisement is also coded as partially watched. The inclusion of such ‘accidental’ consumption is unlikely to be correlated with the consumer’s purchase patterns, especially taking into account the time controls we include in the analysis.

TV ads also cannot be targeted to an individual − all individuals watching the same show see the same ad.

TVision Insights uses a Microsoft Kinect device to identify who is in the room and whose face and eyes are directed towards the TV screen. A handful of other companies have created mobile apps that passively monitor surrounding noise and can recognize shows and ads through audio tags called “watermarks” that are contained in TV content. This technology can be used to determine if an individual is in the same room as the TV, but it does not resolve the question of whether the individual is paying attention.

“Eye-tracking” data (e.g., Teixeira et al., 2010) like that collected by TVision holds promise in improving measurement of advertising attention, but we have not seen these data collected at scale in field-settings and matched to purchases. To give a sense of scale, TVision devices are currently installed in 2,000 homes in the US as part of an opt-in study, while Nielsen devices are installed in 42,000 homes in the US.

The distribution of household ad-skip rates we observe is similar to the distribution reported in Wilbur (2016, Figure 8).

Typically, reported skip-rates of TV ads are lower than skip-rates of online ads. This difference may arise because the effort required to skip an ad online (ignoring a banner ad or clicking to skip a YouTube ad) is generally less than the effort required to skip a TV commercial (changing the channel and monitoring when to return to the program). Some advertising executives we spoke to stated that it could be because TV technology typically requires active avoidance: the passive default option for an online consumer is to ignore the ad, while the action that involves some effort on his part is to click on it. In television advertising, this is reversed: the passive default option for the consumer is to view the ad, while the action that involves some effort on his part is to change the channel.

Variables considered include the brand of the ad, show genre, network in which the ad airs, product category, location of the commercial break within the show and the slot within the break, day of week and hour of show.

What could be the psychological underpinnings of such complementarity with product purchases? While we cannot provide data-based support for more micro-explanations, we conjecture one reason consumers may watch more ads of products purchased recently may derive from “licensing,” wherein users watch ads of others consuming the product as a way to justify to themselves their own consumption (Shafir et al., 1993). Behavioral researchers such as Prelec and Lowenstein (1998) have pointed out that such behaviors are likely when consumption of the product evokes a sense of guilt (e.g., expensive luxury items, junk food, indulgent goods). Another explanation is that watching ads of products purchased recently may provide utility to users from re-living the felt-utility from enjoyable past consumption of the product. Past literature (e.g., Lowenstein and Elster 1992) has pointed out that reliving and contemplating past experiences is a source of significant utility for human beings. Mere repeated exposure may also predict a causal relationship between purchase histories and subsequent advertising consumption. The reason is that both purchasing and consuming a product usually imply exposure to that same product. Repeated exposures, including the ones occurring during consumption, may lead to a higher liking of the product as well as of content related to the product, including its advertisements. For example, the results from Bornstein and D’Agostino (1994) suggest that repeated exposure may increase a consumer’s processing fluency towards product-related material, which she may misattribute to the merits of the ad copy or of the product itself. Finally, a body of literature has documented the fact that decision-makers display selective attention (Cherry 1953; Deutsch and Deutsch 1963; Wolford and Morrison 1980; Tacikowski and Nowicka 2010, among others), often directing it towards aspects that are relevant to the self. While we are unaware of work specifically linking purchase histories to subsequent attention selection, it is possible that this mechanism leads consumers to focus more on advertisements about products they have bought in the past. In particular, an advertisement featuring a previously bought product may merit the attention of the viewer, leading her to be less likely to think about immediate alternatives, such as skipping the advertisement by switching the channel.

We considered cumulative quantity measures ranging from quantity purchased in the preceding day, \(\tilde {Q}_{ijt,1}\), to quantity purchased over the preceding four weeks, \(\tilde {Q}_{ijt,28}\). We found a positive and significant (α = 0.05) relationship between ad consumption and cumulative quantity for [\(\tilde {Q}_{ijt,11},\tilde {Q}_{ijt,15}].\) Outside of this range the relationship between cumulative quantity and ad consumption was consistently positive, but not statistically significant.

This viewpoint has parallels in the applied econometric literature. For example, Angrist and Krueger (1991) estimate the effect of schooling on earnings using quarter of birth as an instrument for years of education. The typical expectation is that those of higher ability will find schooling easier and will obtain more schooling to signal their ability. Thus, a priori we may expect that OLS estimates of earnings on years of schooling are upward biased because of omitted unobserved ability that is positively correlated with earnings and schooling. Alternatively, it may be possible that there is no signaling, or that some individuals with higher earning potential drop out of school earlier to pursue their own endeavors. On instrumenting for years of schooling, Angrist and Kreuger find the IV coefficient to be positive and slightly larger than the OLS estimate in several specifications, indicating if anything that OLS is slightly biased downward.

In Appendix H, we show that our results are not sensitive to local changes to how we define an ad as “skipped”. Our results remain unaffected if we define an ad as “skipped” if the fraction viewed is < 1, 0.95, 0.9, 0.8 or 0.75. Similarly, the descriptive results from our main specification in Column 1 of Table 4 do not change qualitatively if we instead employ the binary skip or watch indicator as the dependent variable.

Equation (4) is quasilinear because by dividing through by \(e^{\gamma _{0}+\mu \epsilon _{0t}^{G}}\), a monotone transformation of \(U^{G}\left (.\right )\), we can write \(U\left (\left .x_{0t}...x_{Jt}\right |\boldsymbol {A}_{t-1}\right )=x_{0t}+U\left (\left .x_{1t}...x_{Jt}\right |\boldsymbol {A}_{t-1}\right )\).

τ is not identified separately from the intercept and is absorbed into α ₀.

These results are meant to illustrate the importance of considering demand-side complementarities and the value of endogenizing the decision to consume advertising in assessing these targeting scenarios. A caveat is we do not accommodate competitive price and advertising response in reaction to the improved price and advertising targeting by the focal advertiser. Thus, the simulations do not speak to equilibrium outcomes in a market with improved addressability and targeting. Doing this would require specifying a supply-side model of price and advertising competition, which is beyond the scope of the current analysis.

For example, suppose we observe b _{j t} exposures by advertiser j on day t = 1,.., 106 in Fall 2011 in the data. In our counterfactuals, we hold b _{j t} fixed for each t and vary how the b _{j t} exposures are allocated across different sets of consumers. Thus, the ad-side control variable for the firm in all our counterfactuals is a set of indicators \(\left \{ \tilde {b}_{ijt};i=1,..,N\right \} \) such that \(\tilde {b}_{ijt}=1\) if consumer i gets allocated ads on day t, and 0 otherwise, and such that \({\sum }_{i=1}^{N}\tilde {b}_{ijt}=b_{jt}\,\forall t\).

We also analyzed the case in which firms engage exclusively in targeted pricing. We find that this policy dominates targeting based on ad-viewing behavior for all firms. This does not imply that firms should focus their efforts on targeted pricing, however, because such policies require a great deal of knowledge about each household’s demand curve. In contrast, our focal targeting policy only requires data on households’ ad-viewing behaviors.

For simplicity, we do not account for the fact that some sophisticated consumers could change their ad viewing behavior in order to influence future prices in the price discrimination scenario.

For the remainder of the regressions reported here, we restrict our analyses to only include the households who made at least one purchase and were exposed to at least one ad.

We reject the null hypothesis that the observed purchase quantities for households in the bottom quartile and middle two quartiles of the ad consumption distribution are drawn from the same distribution (p ≈ 0); we also reject the null for the comparison between the middle two quartiles and the upper quartile of the ad consumption distribution (p ≈ 0).

This field was included in the first version of the dataset released by WCAI, but not in the final version of the data released by WCAI. We use it only to assess informally our timing assumptions.