AI summary by TwinMind
Lecture 24: Advertising
MIT OpenCourseWare
1h 24m 43s
Course Introduction and Marketing Field Overview
- Lecturer resuming advertising topic with gap between sessions
- Marketing field overview: empirical-focused with many practitioners demanding ROI measurement
- Multiple consulting firms help measure advertising effectiveness
- Academic interest in basic economic questions: taste changes vs information provision vs signaling
- Different welfare effects depending on advertising mechanism
- Methodological approaches: experimental work, structural demand models, recent rise in causal inference techniques
Milyo and Waldfogel Rhode Island Liquor Study (1996)
- Background: Rhode Island banned alcohol price advertising 1956-1996
- Supreme Court overturned ban May 13, 1996 as free speech violation (44 Liquor Mart case)
- Built on Benham's 1970s eyeglass study showing 20% lower prices in advertising-allowed states
Data Collection Methods
- Manual price collection at 115 liquor stores for 33 alcoholic beverages
- Started June 1995 (gambling on Supreme Court decision)
- Control group: Massachusetts stores near/far from Rhode Island border
- Creative collection techniques: memorization, voice recorders, bringing children as cover
- Additional data: wholesale prices, print/window advertising usage, lottery ticket sales as quantity proxy
Key Findings
- Minimal price effects: 0.5-1% average price decline after advertising allowed
- Advertising adoption: <10% of stores chose to advertise post-ban
- Advertised vs non-advertised prices: Advertising stores set 20% lower prices on advertised items
- No loss-leader effect: Non-advertised prices at advertising stores unchanged or slightly lower
- Limited competitive response: Non-advertising stores only 0.2% price reduction
- Quantity effects: Advertising stores' lottery ticket share increased 16.4% to 18.4%
- Pre-existing differences: Advertising stores were already 7% cheaper before ban lifted
Implications
- Search costs for alcohol may not be major markup determinant
- Advertising effectiveness limited in changing consumer costs
- Small aggregate consumer welfare effects
Lewis and Riley Yahoo Advertising Experiment (2000s)
- Company background: Yahoo had 25% of internet page views in 2000
- Ad types: Large rectangle (LRE) ads on front page, various targeted ads on inside pages
- Ad sales methods: Daily front page ads, demographic targeting, content targeting, retargeting
Experimental Design
- Large offline retailer (department store-type) with 1.5M customer database
- Matched retailer database to Yahoo user accounts
- 80/20 treatment/control split (80% shown ads, 20% control group)
- Intent-to-treat design: treatment group had ads bid for them, but display depended on auction outcomes
Campaign Scale and Results
- Early fall campaign: 32M ads displayed to 814K people (average 39.6 ads per viewer)
- Click-through rate: 0.28%
- Ad engagement: 7.2% of viewers clicked at least one ad during campaign
- Sales measurement: Both online and offline purchases tracked via credit card matching
Key Challenges and Findings
- Statistical power issues: Sales standard deviation ($1,900) 10x higher than mean ($189)
- Campaign cost 1% of sales, requiring precise estimates to measure ROI
- Selection bias: Ad viewers were lower-value customers (pre-period sales $172 vs $181)
- Intent-to-treat estimates: $5.30 treatment effect, $3.80 standard error (40% ROI ± 100%)
- Difference-in-differences: $8.30 ± $5.90 effect accounting for pre-period differences
- Long-term effects: Potential benefits extending weeks after campaign end
Methodological Lessons
- Estimating advertising ROI extremely difficult even with massive scale
- Viewing ads not exogenous treatment - browsing behavior correlated with purchase likelihood
- Need tens/hundreds of millions of observations for precise estimates
- Individual fixed effects problematic when people are "different people" at different times
Aridor, Che, and Saeedi GDPR Study (2018)
- Regulation background: EU General Data Protection Regulation adopted April 2016, effective May 2018
- Required explicit opt-in consent for tracking European consumers
- Privacy vs business model tension: Targeted ads sell for orders of magnitude more than untargeted
Data and Methodology
- Travel industry intermediary with contracts with "almost all top EU and US travel sites"
- Company function: Machine learning algorithm predicts purchase probability, runs auctions for low-probability users
- Clickstream data aggregated to website-country-week level
- Treatment countries: UK, France, Germany, Italy, Spain
- Control countries: US, Canada, Russia
- Difference-in-differences with website, country, and time fixed effects
Key Findings
- Consumer behavior: Most consumers accept all cookies despite privacy concerns
- Tracking reduction: 12% drop in unique cookies, 11% drop in recorded searches
- Selection effects: Consumers opting out were often using cookie-blocking technology (fewer "first-time" visitors)
- Revenue impact: Initial 13.3% drop in clicks, 16% drop in revenue
- Bid price recovery: Steady increase in auction bids over 7 weeks, offsetting ~50% of revenue loss
- Data quality improvement: Better prediction accuracy on remaining users after algorithm adjustment
Economic Implications
- GDPR removes consumers who don't want tracking and those obfuscating data
- Remaining users more valuable and easier to predict
- Demonstrates potential offsetting effects of privacy regulation on business models
Shapiro Border Discontinuity Study: Antidepressant Advertising (1999-2003)
- Background: SSRI antidepressants under patent protection, TV advertising began 1999 after FDA rule changes
- Previous restriction: No prescription drug TV ads in US before 1997
Data and Identification Strategy
- Prescription data: 5% random sample of physicians prescribing antidepressants with exact office addresses
- Advertising data: TV expenditures by brand across 101 US markets (DMAs), September 1999-December 2003
- Border design: Analysis limited to counties along DMA borders to ensure similar populations
- Fixed effects: product-border-quarter and product-border-DMA to control for unobserved heterogeneity
Key Results
- Own advertising effects: 1 cent per person spending increases sales 2.4%
- Spillover effects: Rival advertising increases own sales by 1.6% (2/3 of own effect)
- Decreasing returns: Quadratic terms negative, optimal advertising levels above observed
- Strategic interactions: Marginal return to advertising decreases when rivals advertise more
Structural Model Results
- Three-level nested logit: Inside/outside good choice, category choice, brand choice within category
- Advertising persistence: Effects die out quickly - 90% gone within 3 months at brand level, 6 months at aggregate level
- Spillover magnitudes: Other firms gain 2.5-5.5x more prescriptions than advertiser from single campaign
- Optimization analysis: Firms advertising 4x less than optimal level in early period
- Cooperative potential: Industry cooperative would advertise 10x current levels, increase profits 15%
Economic Insights
- Substantial positive spillovers create free-rider problems
- Firms greatly under-advertise relative to collective interest
- New advertising medium leads to learning period before optimal strategies
