Effect of Cigarette Price and Tax Increases on Smoking in Europe: A Difference-in-Differences Study with Double Machine Learning

Using Eurobarometer data from 27 EU countries (2012–2020) and a difference-in-differences approach with double machine learning, this study finds that cigarette price and tax increases significantly reduce smoking rates, particularly among daily smokers and individuals aged 15–24, though the magnitude of these effects depends on how the treatment is defined.

The Gist

Imagine smoking as a giant, stubborn garden in Europe. For decades, governments have tried to shrink this garden by raising the price of the "seeds" (cigarettes) through taxes. But is this strategy still working, especially now that new, shiny "fake seeds" (like e-cigarettes) have entered the market?

This paper is like a team of garden detectives (economists Andreas Stoller and Martin Huber) who went back to the garden between 2012 and 2020 to see if the price hikes actually stopped people from planting seeds. They didn't just use a ruler; they used a high-tech, super-smart magnifying glass called Double Machine Learning to get a clearer picture than ever before.

Here is the story of their findings, broken down simply:

1. The Old Way vs. The New Way

In the past, researchers tried to measure this by drawing a straight line on a graph. They assumed that if you raise the price a little, smoking drops a little, and if you raise it a lot, smoking drops a lot, all in a perfect, straight line.

The Problem: Life isn't a straight line. A small price hike might not bother a rich smoker, but a huge one might scare a teenager. Also, the relationship between price and behavior is messy.

The New Tool: The authors used Double Machine Learning (DML). Think of this as a super-smart AI detective that doesn't assume a straight line. Instead, it looks at millions of tiny details (age, income, location, other laws) to figure out exactly how different people react to price changes. It's like using a drone to map a forest instead of just guessing the tree count from the ground.

2. The Big Discovery: Taxes Work, But Prices are Tricky

The researchers compared two groups of countries:

• Group A: Countries that suddenly hiked cigarette taxes and prices significantly (the "Treatment" group).
• Group B: Countries where prices and taxes stayed mostly the same (the "Control" group).

The Result for Taxes:
When countries raised the tax share (the percentage of the price that goes to the government), smoking rates dropped.

• The Impact: About 3.4% fewer people were smoking after the tax hike. That's a big deal!
• The Magic Number: For every 1% the price went up because of the tax, smoking dropped significantly. This suggests that taxes are a very effective "weed killer" for the smoking garden.

The Result for Prices:
When they looked strictly at price increases (without isolating the tax part), the results were muddy. They couldn't prove with 100% certainty that the price hike caused the drop.

• Why? It's a "chicken and egg" problem. Sometimes, cigarette companies raise prices because people are smoking more (demand is high). So, the data gets confused. It's like seeing a fire and a fire truck and not knowing if the truck caused the fire or came to put it out.

3. Who is Most Sensitive? The Teenagers!

The study found that the "weed killer" (tax hikes) didn't work equally on everyone.

• The Winners (of the policy): Young people (ages 15–24). They are the most sensitive to price. When the cost goes up, they are the first to stop buying. This makes sense; they usually have less money in their pockets.
• The Losers (of the policy): Older adults and heavy smokers were less likely to quit just because of the price. They are more "addicted" to the habit and less sensitive to the cost.

4. The "E-Cigarette" Twist

The paper acknowledges that e-cigarettes are now everywhere. Imagine if you tried to stop people from buying apples by raising the price, but they just switched to buying pears.

• The authors worry that some young people might be swapping cigarettes for e-cigarettes.
• However, even with this "substitution," the tax hikes still successfully reduced the number of people smoking traditional cigarettes. The garden of traditional smoking is shrinking, even if some people are moving to a different patch of the garden.

5. Why the "Straight Line" Method Missed the Mark

The authors compared their fancy AI method with the old "straight line" method.

• The Old Method: Underestimated how effective taxes are. It assumed the effect was the same for everyone, so it missed the huge impact on young people.
• The New Method: Showed that taxes are actually much more powerful than we thought, especially when you look at big price jumps rather than tiny ones.

The Bottom Line

Taxes are a powerful tool. When European governments raise the tax on cigarettes, fewer people smoke, especially among the youth.

However, the study warns us:

1. Don't just look at the price tag: You have to look at the tax part of the price to see the real effect.
2. One size doesn't fit all: Policies work best on young people.
3. The old math is wrong: Using simple, straight-line math might make us think taxes aren't working as well as they actually are.

In short, if you want to shrink the smoking garden in Europe, raising taxes is a very effective weed killer, but you have to use a smart, flexible map (like Machine Learning) to see exactly where it's working best.

Technical Summary

1. Problem Statement and Motivation

Despite global declines in smoking rates, tobacco remains a leading preventable cause of death, particularly in Europe where smoking rates are the second highest globally. While tobacco taxation is a primary policy tool, its effectiveness is debated, especially given:

• High Tax Environments: Europe already has high tax-to-price ratios (often >60%), raising questions about diminishing returns as prices rise.
• Emerging Substitutes: The rise of e-cigarettes and other nicotine products may lead to substitution effects, potentially undermining traditional cigarette tax policies.
• Methodological Limitations: Existing literature predominantly relies on Two-Way Fixed Effects (TWFE) models with continuous treatment variables (prices/taxes). These parametric approaches impose strong functional form assumptions (linearity, additivity) and assume homogeneous treatment effects, which may be unrealistic. Furthermore, TWFE estimators can suffer from "negative weights" bias in staggered treatment settings.
• Data Gaps: There is a scarcity of quasi-experimental evidence from Europe, particularly using recent data (post-2012) that accounts for the emergence of new nicotine products.

2. Methodology

The authors employ a Difference-in-Differences (DiD) framework enhanced by Double Machine Learning (DML), specifically the DiDDML estimator proposed by Zimmert (2020).

Data Sources

• Survey Data: Eurobarometer surveys (2012, 2014, 2018, 2020) covering 27 EU countries ($N \approx 107,258$ individuals).
• Policy Data: WHO data on cigarette prices (PPP-adjusted) and tax shares (tax as a % of retail price).
• Control Variables: Socio-demographics (age, gender, education, etc.) and MPOWER policy indicators (smoke-free laws, warnings, etc.).

Treatment Definition

Instead of treating price and tax as continuous variables, the authors define a binary treatment:

• Treatment Group: Individuals in countries experiencing "large" increases (Price increase >15% or Tax-share increase >2%) between pre- and post-treatment periods.
• Control Group: Individuals in countries with stable prices/taxes (changes within $\pm 5\%$ for price; 0% for tax).
• Periods: Two distinct DiD windows: 2012–2014 and 2018–2020.

Estimation Strategy

• DiDDML Estimator: Uses machine learning (Random Forests) to estimate nuisance parameters (propensity scores and outcome regressions) in a data-driven, non-parametric way.
  • Doubly Robust: Combines outcome regression and inverse probability weighting.
  • Neyman Orthogonal: Makes the estimator robust to approximation errors in the ML steps.
  • Cross-fitting: Prevents overfitting by estimating nuisance parameters on one fold of data and the treatment effect on another.
• Comparison Models: The study compares DiDDML results against:
  1. Parametric DiD with binary treatments.
  2. Parametric DiD with continuous treatments (standard TWFE approach).

Identification Assumptions

The study relies on the standard DiD assumptions:

1. No Anticipation: Treatment assignment does not affect pre-treatment outcomes.
2. Conditional Common Trends: Trends in smoking rates are parallel between treated and control groups conditional on covariates.
3. Common Support: Overlap in covariate distributions between groups.
   Note: The authors acknowledge that parallel trends cannot be tested via standard placebo tests due to the binary treatment definition and high variation in prices, so they use a specific placebo test inspired by Callison and Kaestner (2014).

3. Key Contributions

1. Methodological Innovation: First application of DiDDML to tobacco taxation in Europe. This relaxes functional form assumptions, allowing for non-linear and non-additive relationships between confounders and smoking outcomes.
2. Treatment Definition Sensitivity: A rigorous comparison of binary vs. continuous treatment definitions. The paper argues that binary definitions (comparing large shocks to stability) provide a cleaner counterfactual than continuous linear models.
3. Contextual Relevance: Provides evidence from a high-tax European environment during the rise of e-cigarettes (2012–2020), updating the understanding of tax efficacy in modern contexts.

4. Key Results

Main Findings (DiDDML Estimates)

• Tax Increases: Significant reduction in smoking rates.
  • Current Smokers: Reduced by 3.44 percentage points (pp) (15% reduction; $p=0.04$).
  • Daily Smokers: Reduced by 3.15 pp (15% reduction; $p=0.09$).
  • Pass-through: Tax increases led to a 77% pass-through to retail prices.
• Price Increases: No statistically significant effect at the 5% level. The authors attribute this to endogeneity (prices may rise in response to high demand), which biases estimates upward toward zero.

Elasticity Estimates

• The estimated price elasticity at the extensive margin is large (point estimates around -5.6 to -5.7), though confidence intervals are wide.
• Conservative Lower Bounds: Even conservative estimates suggest an elasticity of -0.3 (DiDDML) to -1.6 (Parametric Binary), which is higher (more elastic) than the traditional literature range of -0.1 to -0.3.

Treatment Definition Sensitivity

• Binary vs. Continuous: Estimates based on binary treatments (large shocks) yield significantly larger effect sizes than those based on continuous treatments (linear models).
• Implication: Standard TWFE models using continuous variables and full samples likely underestimate the effectiveness of tax increases because they assume linearity and dilute the impact of large shocks with small/no changes.

Heterogeneity

• Age: The effect is driven primarily by individuals aged 15–24, who show statistically significant reductions. This aligns with the hypothesis that youth have lower income elasticity and higher substitution potential (e.g., to e-cigarettes).
• Gender/Education: Reductions are observed across all groups, but are generally stronger among women and those with lower-to-medium education levels.

5. Significance and Policy Implications

• Policy Effectiveness: Contrary to theories suggesting diminishing returns in high-tax environments, the study confirms that tax increases remain effective in reducing smoking in Europe, even with high baseline taxes and the presence of e-cigarettes.
• Methodological Shift: The findings suggest that previous literature relying on continuous TWFE models may have underestimated the true impact of tobacco taxes. Policymakers should consider that large, discrete tax hikes may be more effective than marginal adjustments, and evaluation methods should account for non-linearities.
• Youth Focus: The disproportionate impact on youth (15–24) highlights that tax policy is a potent tool for preventing smoking initiation and encouraging cessation among young people, though substitution to e-cigarettes remains a concern to monitor.
• Robustness: Results are robust to different covariate sets, standard error clustering, and placebo tests, reinforcing the validity of the causal inference.

In conclusion, the paper provides robust evidence that cigarette tax increases are a highly effective tool for reducing smoking in Europe, particularly among youth, and demonstrates that modern causal machine learning methods can reveal stronger effects than traditional parametric approaches.