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

Abstract

Endogenous consumption of advertising is common. Consumers choose to change channels to avoid TV ads, click away from paid online video ads, or discard direct mail without reading advertised details. As technological advances give firms improved abilities to target individual consumers through various media, it is becoming increasingly important for models to reflect the endogenous nature of ad consumption and to consider the implications that ad choice has for firms’ targeting strategies. With this motivation, we develop an empirical model of consumer demand for advertising in which demand for ads is jointly determined with demand for the advertised products. Building on Becker and Murphy (The Quarterly Journal of Economics, 108(4), 941–964 1993)’s ideas, the model treats advertising as a good over which consumers have utility and obtains demands as the outcome of a joint utility maximization problem. Leveraging new data that links household-level TV ad-viewing with product purchases, we provide empirical evidence that is consistent with the model: ad-skipping is found to be lower when a household has purchased more of the advertised brand, and purchases are higher when more ads have been watched recently, suggesting that advertising and product consumption are jointly determined. Fitting a structural model of joint demand to the data, we evaluate consumer welfare and advertiser profitability in advertising targeting counterfactuals motivated by an “addressable” future of TV. We find that targeting on the predicted ad-skip probability is an attractive strategy, as it indirectly selects consumers that value the product. Reflecting the positive view of advertising in the model, we also find that net consumer welfare may increase in several targeting scenarios. This occurs because under improved targeting, firms shift advertising to those who are likely to value it. At the same time, consumers that do not value the ads end up skipping them, mitigating possible welfare losses. Both forces are relevant to assessing advertising effects in a world with improved targeting and ad-skipping technology.

Appendices

Appendix A: Ad skipping and observable demographics

Deng and Mela (2017) report that household demographics are not a strong predictor of the observed variation in ad skipping behavior in their TiVo data.

In this appendix, we analyze how ad exposure and ad skipping are correlated with observed household demographics in our data. As shown in Table 14, we find that larger households tend to be exposed to more ads, but all else equal, household size does not correlate with skip rates. Homeowners, people over 50 and those with higher levels of income and education tend to see fewer exposures and have higher skip rates. The fact that wealthier, more educated people are more likely to skip an ad is consistent with the interpretation of the cost of an advertisement as the opportunity cost of one’s time. Like Deng and Mela, we find that observed household demographics explain little of the variation in ad skipping.

Table 14 Regression of Household Ad Exposure and Ad Skip Rate on Observed Characteristics

Appendix B: Price series construction

The purchase data records the price paid and package volume of transactions at the barcode level. We only observe the prices of purchased products, but in order to estimate the model, we need to reconstruct the price series of the alternatives that were not purchased. Additionally, our model is at the brand level, so we need to transform barcode level prices into brand level prices. Our approach is to reconstruct a barcode-chain-week level price series using all observed transactions and weight by purchase volume to create a brand-chain-week price per unit. As we do not model chain choice, the final step in the price series construction is to create a household-brand-week level price series by creating a weighted average of the chain-brand-week price series using the frequency of a household’s chain visits as the weights. The steps below describe the process used to construct the price series.

• 1. Although we only observe purchase and TV advertising data for 6,552 households, we observe purchase data for 22,670 households. The entire purchase database is used in the construction of the price series.

• 2. We observe at least one purchase of 58 different brands in the transaction data. In order to make the model more tractable, we restrict the analysis to the set of brands that have the largest purchase market shares. We focus on the brands that collectively cover 90% of the market. The “other” or smaller brands category has the largest purchase market share (61.72%). These brands generally do not advertise, and because we cannot be sure whether the ads we do observe in the data correspond to the same brands that were purchases in this category, we do not include the “other” brands in the analysis. This leaves us with 12 brands. The remaining brand with the largest market share is brand 195 with 8.49% of all purchases observed in the database.

• 3. We keep transactions for barcodes that make up at least 5% of brand sales. Brand 839 does not have any barcodes that have at least 5% of sales, so this brand is dropped from the analysis.

• 4. We keep transactions for chains that have at least 1,000 purchases of the barcodes identified in step 3.

• 5. A barcode-chain-week level price series is constructed by taking the median/mode observed price of the transactions in that week. If there are not any observed purchases in a week, we fill in the median/mode observed price from the previous week. If no transactions are observed to date, we fill in the median/mode observed price in the week of the first observed transaction.

• 6. Because barcodes correspond to different package volumes, we create a barcode-chain-week level price per unit by dividing weekly prices by package volume.

• 7. A brand-chain-week level price series is constructed by weighting the barcode level price per unit by the total volume of that barcode bought at that chain over the sample.

• 8. Finally, we create a brand-household-week price series by averaging the brand-chain-week level price series across chains and weighting by household i’s count of purchases at chain s over the sample period.

• 9. Using this chain weighting procedure, some households do not have a price series for all 11 brands because sometimes a household has never been to a chain where a given brand is sold. For example, household 135,075 only ever purchases at chain 1. Brand 195 is not sold at chain 1.

  The idea behind weighting the chain price series by frequency of chain visits is intended to reflect a household’s best belief about a brand’s current price. In the instances where a household never visits a chain where a brand is sold, we assume that household’s beliefs about the price of that brand will be equal to the maximum price for that brand across all chains that week.

Appendix C: Within chain price variation over time

Figures 17, 18 and 19 in this appendix document the extensive price variation over time at the brand-level within chains.

Fig. 17

Time Series Plots of Chain-Level Prices - (A)

Fig. 18

Time Series Plots of Chain-Level Prices - Contd

Fig. 19

Time Series Plots of Chain-Level Prices - Contd

Appendix D: First stage IV regressions

This appendix presents the first stage IV regressions for both the specifications using chain level price series and those using household price series as instruments (Tables 15 and 16). In all specifications the coefficient on own-price is negative and highly significant.

Table 15 Regressions of Quantity on Price Instruments − Chains 8 and 9 Price Series

Table 16 Regressions of Quantity on Price Instruments − Household Price Series

Appendix E: Evidence that advertising shifts quantities purchased

This appendix reports on several specifications and data cuts to document that advertising has a positive effect on product demand.

First cut: Cross-sectional analysis relating advertising and product demand

We start by checking whether households who view more advertisements also purchase more on average. At a minimum, support for a model with complementarities requires seeing a positive covariation between quantities and ads in the data. We explore the joint distribution of total quantity purchase and total category ad consumption at the household level for different levels of ad consumption. To do this, we split the sample of households into three buckets − the lowest quartile, the middle two quartiles, and the upper quartile of the distribution of total ad consumption over the panel duration.²⁴ These buckets correspond to households who viewed between 0 and 65 ads, between 65 and 448 ads, and 448+ ads, respectively. Then, we estimate the density of purchase quantities for each of these groups. Table 17 summarizes the estimated kernel distributions. The quartiles of the purchase quantity distribution are larger for households in higher ad quartiles. Two-sample Kolmogorov-Smirnov tests reject the null hypotheses that these samples come from the same distribution.²⁵

Table 17 Conditional Distribution of Purchase Quantity by Ad Consumption Quartile

Panel analysis

Based on this cross-sectional analysis, we cannot conclude that advertising per se induces the shift out in the purchase density. We can use the panel variation to test if within-household variation over time in purchase quantity for a brand is related to cumulative past advertising consumption by that household of that brand.

We define cumulative past advertising consumption as the sum of the percentage watched of the advertisements to which the consumer was previously exposed. We construct this variable for the preceding 1, 2, 3, and 4 weeks, and regress household i’s day t purchase quantity of brand j on household i’s cumulative past advertising consumption of ads for brand j. Each observation in the regression is a household-brand-day. This regression is estimated unconditional on purchase, meaning that we include days with no-purchase in the analysis setting quantity equal to 0. We also control for the price per unit of brand j. Because we only observe prices when a purchase is made, we reconstruct the price series for the 11 most frequently purchased brands in the data and restrict all our analyses to these brands. Appendix B describes in detail how we constructed the price series for these brands.

Advertising is endogenous in this regression. Unfortunately, we do not have an instrument that moves ad consumption independently, which can be excluded from the propensity to buy more units. Given the lack of co-ordination of national TV advertising with local determinants of purchases that we described in Section 3, we believe the main source of endogeneity concerns is from unobserved heterogeneity and seasonality. The heterogeneity concern reflects the confound that consumers who like a brand also watch more of its ads, and the seasonality concern the confound that consumers may buy more and watch more ads during holiday seasons.

Given we have panel data over a long time horizon, we can include very rich controls for both, that address to a large extent these concerns. We include household-brand fixed effects to control for unobserved heterogeneity. Thus, our coefficients are estimated off within household-brand variation over time rather than across household variation and across brand variation. We also include week fixed effects in order to address the concern that there may be unobserved, time-varying shocks driving both purchases and ad consumption that remain even after including household-brand fixed effects. In particular, we estimate the following specification,

$$ q_{ijt}=\beta_{0ij}+\beta_{0t}+\beta_{1}A_{ijt}+\beta_{2}p_{ijt}+\epsilon_{ijt} $$

(38)

where q _{i j t} is daily purchase quantity in equivalent units and A _{i j t} is a cumulative ad-duration variable (defined more precisely in Table 18). Table 18 presents the results. Consistent with our cross-sectional findings, the effect of cumulative past advertising consumption is positive and statistically significant across all time windows we consider. To interpret the magnitudes of the ad-effect, we also report in the last row the effect on daily quantity demanded of a 1 SD increase in the cumulative ad consumption variables over the past 1, 2, 3, and 4 weeks. Across specifications, we find that a 1 SD increase in ad consumption over the past 1-4 weeks increases the mean daily quantity demanded by 3.4 −4.5%. For instance, looking at the last 3 rows of Table 18, the mean daily quantity demanded is 7.93 equivalent units. A 1 SD increase in A _{i j t,7} − the ad consumption over the last one week − increases the mean daily quantity demanded by 3.38%.

Table 18 Regression of Daily Purchase Quantity on Cumulative Ad Consumption

Even though we have included a rich set of controls, concerns may remain about individual-specific correlated time-varying shocks to both product and ad consumption. One consideration is a form of “activity bias” − that the consumer is busy during some time periods, so is likely both to purchase less and skip ads, which manifests as spurious co-movement in joint consumption. For example, when a consumer goes out of town, we might observe zero purchases and zero ad consumption, which could create spurious correlation between purchase quantity and ad consumption. To assess this, we re-estimate the same model, restricting the data to days in which a household purchased at least one brand and is thus observed to be “active” in the data. We use our preferred specification in which past ad consumption is defined over the preceding two weeks (Column 2, Table 18). Again, we continue to estimate a positive relationship between purchase quantity and cumulative ad consumption (see Table 19).

Table 19 Activity Bias Robustness Check of Regression of Daily Purchase Quantity on Cumulative Ad Consumption

Appendix F: Are complementarities at the brand or category level?

In this appendix we explore whether the complementarities between ad consumption and product consumption occur at the brand or the category level. In column 2 of Table 20 we regress ad consumption for brand j on the quantity purchased of brand j and the quantity purchased of all other brands -j. Column 3 includes total product consumption across all brands as the independent variable. Cross-brand effects are estimated to be negative, though not statistically significant. Table 21 runs the reverse regressions of quantity purchased of brand j on cumulative own and cross advertising of all other brands. After controlling for own effects, the cross-effects are not significant. The last column uses cumulative advertising for all brands as the independent variable; this effect is marginally significant. These results suggest that complementarities between product and ad consumption operate at the brand level as opposed to the category level in these data.

Table 20 Regressions of Ad Consumption on Cumulative Quantity

Table 21 Regression of Daily Purchase Quantity on Cumulative Ad Consumption

Appendix G: Most ad-consumption occurs after 5:00 PM

The maintained assumption is that product purchase occurs earlier in the day compared to ad-exposures. The advertising dataset included a timestamp indicating the exact second of each exposure.²⁶ Looking at Fig. 20 below, we see that 70% of those ad exposures occurred after 5 PM. Unfortunately, we do not observe a time stamp in the purchase data to test the time of purchase of products, but it is not unreasonable to assume that many in-store purchases occur during the day.

Fig. 20

Ad Exposures by Time of Day

Appendix H: Ad consumption model sensitivity estimates

In our model, we consider an ad to be skipped if it is not watched to completion. In this section we explore the sensitivity of our results to different definitions of skipping. Table 22 reports the results of a logit model in which we regress the binary decision of whether to watch an ad on cumulative purchase quantity in the previous two weeks. We consider alternative definitions of ad consumption where an ad is considered skipped if a) less than 100% of the ad is watched (the ad is not watched to completion), b) less than 95% of the ad is watched, and 3) less than 75% of the ad is watched. The regression is estimated at the household-brand-day level and household-brand random effects are included to control for heterogeneity across households. The magnitudes of the coefficients on product quantity are similar, showing that our results are not sensitive to our specific definition of ad skipping.

Table 22 Logit Regression of Ad Watched Dummy on Cumulative Purchase Quantity

Appendix I: Simulation procedure

Here, we discuss in more detail how we implement the simulations in Section 5.1 to measure “long-run” elasticities in the model. For each household we simulate purchase and advertising consumption outcomes over the last 3 months in the data at the observed levels of prices and advertising exposures. We compare the model-predicted results to the results of a second simulation in which we allocate additional exposures to each household. The steps below outline the simulation procedure.

• 1. Restrict the sample to the set of households who viewed at least one ad for brand j.

• 2. Allocate the additional ad exposures for brand j to the first household, spreading the additional exposures evenly across days in which the household viewed an ad for brand j.

• 3. Conditional on the initial observed stock of product consumption, take error draws for the ad-skip model and predict ad consumption decisions for each of the s ^′ exposures for all brands on day t.

• 4. If an ad is skipped, draw the percentage of the ad watched independently from the observed distribution of ad durations in the data. If an ad is not skipped, set the percentage equal to 1.

• 5. Update the advertising stock $\overrightarrow {A}$ using the simulated ad percentages in t.

• 6. Conditional on the observed prices in the data and the advertising stock $\overrightarrow {A}$, take error draws for the product purchase model and predict product consumption for all brands on day t.

• 7. Update the consumption stock $\overrightarrow {Q}$ using the simulated purchase quantities in t. Set t = t + 1. Return to step 3.

• 8. Repeat the forward simulation procedure in steps 3 - 7 for R = 100 paths of error shocks and average the statistics of interest (total purchased quantity, ad consumption, and consumer surplus) over all simulations. Repeat this procedure for all households.