Causal Impact of Plain Cigarette Packaging
Causal effect of plain packaging (2017) on cigarette consumption, healthcare utilisation, and smoking cessation — Master's project at TSE, in R.
- R
- Econometrics
- Causal inference
- Time series
France mandated plain packaging in 2017. Nearly a decade later, there is finally enough hindsight to measure its real effect — not from self-reported surveys, but from sales data and health records. Verdict: a decline in consumption that takes hold and deepens (≈ −3.7% at the baseline estimate, up to ≈ −7.6% by 2019), accompanied by a sharp rise in nicotine-replacement therapy use and smoking-cessation consultations.
Master's 1 project (Econometrics & Statistics) at Toulouse School of Economics, supervised by Angélique Acquatella. Team project (Nathan Stouque, Lara Le Bihan, Mathéo Trebosc), carried out in R.
The question
Did plain packaging reduce cigarette consumption in France, and shift health and cessation behaviours? The central challenge: isolating the packaging effect from concurrent measures (price increases, public campaigns) — in a context where French evidence was largely limited to self-reported data. We tackle the question from three angles.
1 — Consumption: Difference-in-Differences
Two-Way Fixed Effects (DiD) estimator on a panel of 13 European countries, 2010–2019: France treated from 2017, with 12 control countries that did not adopt plain packaging. Controls: unemployment rate, tobacco price index, average pack price. Country and year fixed effects, errors clustered by country.
- Average effect: ≈ −3.7% in per-capita sales (β = −0.0376, p = 0.045).
- Dynamics: zero in 2017, then −5.1% in 2018 (p = 0.002) and −7.6% in 2019 (p < 0.001) — an amplifying effect, consistent with declining addiction capital (Becker & Murphy, 1988).
- Parallel trends validated (Wald test, p = 0.33).
Robustness: event study (up to −10.7% at k = 2), leave-one-out (consistently negative, −2% to −5%), and placebos on false reform dates (2014–2016) all non-significant.
2 — Healthcare utilisation: DiD on French national health insurance (Assurance Maladie / Ameli) data
Administrative data from Ameli (2015–2023), population aged 15–54, three disease comparisons chosen for their responsiveness to tobacco:
- Acute coronary syndrome vs diabetes (fast response): decline of ≈ 10.7%, but not significant (p = 0.11) — still too early.
- Head and neck cancers vs appendicitis (falsification test): null effect (p = 0.38), expected given cancer latency → the model does not generate spurious signals.
- Smoking cessation vs alcohol consultations (behavioural mechanism): marked increase in tobacco consultations (+0.47 pp, p = 0.01). Pre-reform trends are not parallel here, so this is descriptive evidence of a behavioural shift, not strictly causal.
3 — Nicotine-replacement therapy: Interrupted Time Series (ITS)
ITS model on pharmacy sales of nicotine-replacement products (monthly, 2014–2024), controlling for anticipation effects (Mois Sans Tabac (national no-smoking month), late 2016), COVID, and prices. Strong autocorrelation addressed via HAC errors (Newey–West) and a GLS AR(1) model.
- The 2017 reform is associated with a rise of ≈ 136% to 171% in units sold, depending on the specification.
- But the effect loses intensity over time (negative post-reform trend).
Limitations
Residual autocorrelation not fully captured by the AR(1); a placebo revealing a structural break in 2016 (subsequently reintegrated into the model); and the general difficulty of isolating plain packaging from other anti-tobacco policies. ITS figures should therefore be read as orders of magnitude, not precise percentages.
Takeaways
Three complementary identification strategies (multi-country DiD, disease-based DiD, ITS), each with its robustness checks and placebos — and an honest reading of what the data does, and does not, allow us to conclude.