35.6 Advertising Response Measurement
- Structural, Experimental and Quasi Experimental Approaches
35.6.1 (Terui, Ban, and Allenby 2011)
Previous studies assume that advertising has a direct and lagged effect on consumer utility
This study found evidence that there is no direct effect of advertising on consumer utility for mature brands.
Data: scanner panel (laundry detergent and instant coffee)
Advertising affect consideration sets, not the marginal utility of offering (i.e., previous studies did not account for consideration set formation, and just take the advertising effect on customer utility, later underestimate the effect of advertising on sales)
- Hence, we should use brand consideration as the dependent variable when studying the advertising effect.
Periodic advertising is still beneficial because it raises the advertising stock to be above the threshold level for brand inclusion in the consideration set.
Contribution:
Account for heterogeneous consumer response to advertising and consideration set formation
Include a hard constraint on brand inclusion in the consideration set which helps distinguish considerations from choice in the model likelihood.
Base model: (Gilbride and Allenby 2004)
Future research: can use this paper for structural models of consideration for price.
Question: is it applicable to high-involvement products?
Model Development:
Let \(N\) be the number of choice alternatives
Consumer \(h\) has advertising stock \(AS_{jht}\) for each alternative (\(j = 1, \dots, N\))
Alternative \(j\) can be in the consideration set, \(C_{ht}^{AS_{jht} \ge r_h}\) , of consumer \(h\) at time \(t\) when \(AS_{jht} \ge r_h\) (where \(r_h\) is the threshold value of consumer \(h\) across choice alternative and time invariant. also known as effective advertising stock)
Elements in \(C_{ht}^{AS_{jht} \ge r_h}\) can change over time with changes in \(AS\)
Consumer \(h\) utility for the alternatives in the consideration set is
\[ u_{jht} = x'_{jht} \beta_h + \epsilon_{jht} \]
where
- \(\epsilon_{jht} \sim N(0, \sigma^2_j = 1)\) for \(j \in C_{ht}^{AS_{jht} \ge r_h}\)
The choice probability of an alternative in the consideration set is
\[ P(j)_{ht} = P\{ u_{jht} = \max \{ u_{kht} : k \in C_{ht}^{AS_{jht} \ge r_h} \}\} \]
To make the model solvable, if a person did not watch any ad, but still purchase a brand, then his or her \(r_h \approx 0\)
Advertising stock is modeled based on (Bass and Clarke 1972), (Clarke 1976):
\[ AS_{jht} = \sum_{g=0} ^ \infty \alpha_{jht- g} \rho_h^g \]
where
\(\alpha_{jht-g}\) is when consumer \(h\) is exposed to adverting for brand \(j\) at time \(t-g\)
\(\rho_h\) is advertising diminishing effect (\(0 \le \rho_h <1\))
Advertising effect occurs instantly and diminished exponentially (to the \(g\) order), which was evidenced in experiential research design (Lodish et al. 1995) (Little 1979)
Two other stock variables:
Brand Loyalty (Peter M. Guadagni and Little 2008) (Tülin Erdem 1996):
\[ BL_{jht} = \sum_{g=1}^\infty y_{jht-g} \tau_h^g \]
where
\(y_{jht-g}\) is the purchase variable for brand \(j\)
\(0 \le \tau <1\)
Threshold \(\lambda_h\)
Display Stock
\[ DS_{jht} = \sum_{g=9}^\infty d_{jht- g} \phi_h^g \]
where
\(0 \le \phi_h <1\)
Threshold \(\kappa_h\)
35.6.2 (Narayanan and Kalyanam 2015)
Causal effect of position in search engine advertising listing on click-through rates and sales
Because of selection bias, causal inference is difficult (experiments can’t model bidding behavior).
Without addressing for these selection biases, position effects on click-through rates and sales are huge, but with RD design, the estimates are smaller.
Position effects are
stronger for small advertiser, or consumer with little experience with the keyword for the advertiser.
weaker brand or production info is included in the keyword, on weekends compared to weekdays.
Position could affect click-through rate and purchase behavior via
signalling: advertising expenses signal product quality
consumer expectation
sequential search: learned experience by costumers that better results are higher in the search engine.
attention: consumers only pay attention certain parts of the screen.
Endogeneity problems:
Brands target keywords with high conversions. (inflate the causal effect of viewing ad on conversions)
Position is determined by online auction. (randomization of bid would not lead to randomization of position)
- Cannot use parametric selection equations because positioning is determined by complex processes
Solution: RD
Running variable: adrank = f(advertisers’ bids, quality score) -> (sharp cutoff)
nonobservability of competitors’ adrank prevents selection into treatment by the focal firms. Hence, unless you have both focal and competitors bids and Adrank, you can’t do RD here (but the authors they have both a focal advertiser and its main competitors - before M&A).
Moderators (Decision by the advertisers):
Match: Exact vs. Broad (for keywords)
Advertisers: e.g., higher vs. lower quality firms
Experience and advertising are substitute (Narayanan and Manchanda 2009): recent consumers are not going to change their probability of buying when exposed to ads, as compared to those who have not recently experienced the product.
Category vs. brand terms: prior literature shows category terms precede brand terms (people use broad search terms = novices = rely more on ad position).
Weekday vs. weekend: search cost lower on the weekends. Thus position effects are stronger on weekdays.
Selection Issues:
Selection on observables:
- Differences in keywords, match types, advertisers
Selection on unobserveable:
Bidding behaviors by advertisers (both ways: positive - higher CRT invest more and negative - higher CTR invest less).
Competition:
Possible Solutions:
Experiments: but cannot control/randomize competitors.
Model selection parametrically: hard to believe
Latent Instrument: but rely on a single latent instrument, outcomes are normal (hard to believe)
RD:
Assumptions:
Brands can’t manipulate its position: unobservability of competitors Adrank even ex-post
Forcing variable is continuous: Adrank
Procedure:
Selection of observation (those close to the cutoff)
Selection of the bandwidth (how wide the window, bias and variance trade-off)
Use local linear regression within the bandwidth
Test different bandwidth using “leave-one-out cross validation”
Data: 28.5 mil daily obs -> 13.1 mils (with 2 firms involved) -> 414,310 obs with adjacent observations.
Results:
- Both OLS and Fixed Effect inflate the effect of position.
OLS estimates are positively biased (selection on observables and unobservables),
Fixed effects correct for selection of observables (a little lower than OLS) (selection on unobservables causes negative bias)
With varying selection bias by position, it’s unlikely that parametric approaches or instrumental variables can accommodate for this.
The effect of position on CTR and later on sales is not straight forward. And only moving from 6 to 5 has a significant difference to sales. (which is right above the page fold, or it might be due to consumer perceive top 5 as higher quality).
35.6.3 (Lewis and Rao 2015)
Individual sales data are volatile which leads to high experiments cost to require precise estimate.
Data on 25 field experiments (cost $2.8 mil in digital marketing)
Evidence that observational methods (i.e., control for observables) are untrustworthy to measure returns to advertising.
Economic universe
Weak evidence of advertising effectiveness.
35.6.4 (Gordon, Zettelmeyer, et al. 2019)
Compare experiments results with observational models, where observational methods do not show the same effect as the randomized experiments.
Demand (click-through rate) universe