Should We Put Ice in Wine? A Difference-in-Differences Approach from Switzerland Alexandre Mondoux KOF Swiss Economic Institute, ETH Zurich 11th Annual AAWE Conference - Padua 2017 Thursday, 29 June 2017 1 / 30
Table of Contents 1 Introduction 2 Data 3 Identification strategy: Difference-in-Differences 4 Results 5 Robustness checks 6 Conclusion 2 / 30
Introduction Motivations Current theme Weather shocks have regained their relevance in recent weeks with a spring frost in most Swiss vineyards. Natural experiment: We exploit a natural experiment (weather hail shock affecting the grape harvest) from the Swiss wine region Three Lakes in 2013 on the consumption in the Swiss retail market. Presentation from the study: Mondoux, A. (2017): Should We Put Ice in Wine? A Difference-in-Differences Approach from Switzerland, KOF Working Paper Series (soon to be published). 3 / 30
Introduction Strategy Idea: Using a Difference-in-Differences (DID) approach, we are able to estimate the Average Treatment Effect (ATE) of this exogenous supply shock. Literature: Seminal and recent works of this econometric method as well as in the field of wine economics: Ashenfelter (1978), Card and Krueger (1994), Ashenfelter et al. (2007), Egger et al. (2006) and Angrist and Pischke (2008). Key findings: We find negative and statistical significant effects of the hail weather shock on Three Lakes wines in terms of consumption (-22.8%) and price (+2.8%). 4 / 30
Introduction The six Swiss wine regions Figure 1: The six Swiss wine regions. (SWP, 2015) 5 / 30
Introduction General considerations Panel data constructed from the database Nielsen that Swiss Wine Market Observatory (www.osmv.ch) made available retail market, data scanned at the till about 1/3 of whole Swiss wine consumption Main variables: income, quantity and price (interpreted as equilibrium between supply and demand) Other channels of distribution: direct sale, Horeca, wholesalers (see Mercuriale of OSMV) Lack of wine vintage information (year) of the bottles we assume that the Production t (example: from the harvest 2013) will enter the retail market in t + 1 (more precisely, 04/2014) 6 / 30
Data Data Structure of the panel 4 weekly data (13 observations per year) 5 years of time span (2012-2016) updated every trimester Identification variables: color, region, subregion, AOC (protected designation of origin), foreign, etc. 80 Swiss AOC types of wines (examples: Merlot red Ticino, Fendant white Valais) For identification, we use only AOC wines Type of variables Dependent variables: income, quantity, price and promotions facts Economic variables: exchange rates, Swiss CPI, wine import prices Climatic variables: temperature, sunshine and rainfall 7 / 30
Data Descriptive statistics Table 1: Descriptive statistics by treatment and control group Treatment group Control group Variable Mean Before After Mean Before After Dependent variable Income 44.771 48.301 41.099 356.521 371.584 339.014 Quantity 2.986 3.346 2.611 30.262 31.882 28.378 Price 19.725 19.521 19.937 15.388 15.298 15.494 Income Promo 12.417 12.238 12.563 148.682 154.572 142.675 Quantity Promo 0.939 0.995 0.892 14.047 14.818 13.260 Price Promo 15.790 15.175 16.276 13.672 13.533 13.812 Economic variable Exchange Rate EUR/CHF 1.180 1.219 1.131 1.180 1.219 1.131 Exchange Rate GBP/CHF 1.478 1.469 1.490 1.478 1.469 1.490 Exchange Rate USD/CHF 0.936 0.928 0.945 0.936 0.928 0.945 Consumer Price Index 101.549 101.867 101.147 101.549 101.867 101.147 Import Price Italy Red 8.152 8.204 8.087 8.152 8.204 8.087 Import Price Italy White 4.881 5.016 4.711 4.881 5.016 4.711 Import Price France Red 14.569 16.601 12.008 14.569 16.601 12.008 Import Price France White 9.848 10.108 9.521 9.848 10.108 9.521 Import Price Spain Red 6.784 6.747 6.830 6.784 6.747 6.830 Import Price Spain White 4.903 5.113 4.637 4.903 5.113 4.637 Climatic variable Temperature Mean ( C) 10.764 9.673 12.139 11.129 10.068 12.467 Temperature Minimum 1.599-0.021 3.641 1.755 0.624 3.181 Temperature Maximum 21.519 20.821 22.399 22.838 22.095 23.774 Wind Mean (km/h) 11.934 11.876 12.008 9.678 9.721 9.623 Wind Maximum 78.704 79.059 78.257 63.988 64.773 62.998 Sunshine (hours) 153.296 145.976 162.525 165.738 160.593 172.226 Rainfall (mm) 77.872 85.082 68.781 86.167 90.711 80.437 Pressure Minimum (hpa) 1000.868 1000.630 1001.168 1000.526 999.994 1001.196 Pressure Maximum (hpa) 1029.842 1029.304 1030.519 1029.838 1029.233 1030.602 Observations 780 435 345 3276 1827 1449 No. of labels 15 15 15 63 63 63 8 / 30
Data Swiss wine is a heterogeneous good Figure 2: Proportion of wine color by region. 9 / 30
Data Examples of types of wines (a) Neuchâtel (b) Vaud (c) Ticino 10 / 30
Data Swiss wines experience strong seasonality in consumption Figure 3: Seasonality consumption of three treated wines. 11 / 30
Data Swiss wine consumption by group (retail market) (a) Treated group (OSMV, 2016) (b) Control group (OSMV, 2016) 12 / 30
Data Three Lakes wine consumption (FOAG, 2016) Figure 4: Volume for all channels of distribution. (FOAG, 2016) 13 / 30
Identification strategy: Difference-in-Differences Identification strategy Appropriate econometric model: Difference-in-Differences (DID) Treated group (Three Lakes) Hail weather shock date (20/06/2013) Cut-off point (04/2014) Limitations? Start date (01/2012) End date (03/2015) Wine data from the retail market only (about 1/3 of the whole wine consumption) Strong assumption: lack of vintage information 14 / 30
Identification strategy: Difference-in-Differences Theoretical approach What do we expect from a negative exogenous supply shock? Consumption of treated wines Prices of treated wines 15 / 30
Identification strategy: Difference-in-Differences Econometric model We use the following DID regression equations: Outcome i,t = β 0 + β 1 Treat i + β 2 Post t + β 3 (Treat i Post t ) i,t + ɛ i,t (1) ln(q i,t ) = β 0 + β 3 (Treat i Post t ) i,t + S i,tθ + z tγ + c i + δ t + ɛ i,t (2) ln(p i,t ) = β 0 + β 3 (Treat i Post t ) i,t + S i,tθ + z tγ + c i + δ t + ɛ i,t (3) With: { 0 if control group Treat i = 1 if treated group (Three Lakes) { 0 if before-treatment period (01/2012-03/2014) Post t = 1 if post-treatment period (04/2014-03/2015) 16 / 30
Identification strategy: Difference-in-Differences Parallel trend assumption (visual evidence) Figure 6: Time trend for quantity. 17 / 30
Results Results (1/3): ln(quantity) Table 2: Regression results for ln(quantity) (1) (2) (3) (4) RE RE FE FE Treat i Post t -0.2369** -0.2336** -0.2300** -0.2277*** (0.0991) (0.0988) (0.0991) (0.0691) Covariates no no no yes Individual FE no no yes yes Time FE no yes yes yes Constant 1.6010*** 0.9382*** 1.1580*** -88.7061** (0.2724) (0.2820) (0.0650) (41.5889) Observations 2819 2819 2819 2319 No. of labels 73 73 73 70 R-squared 0.0636 0.0076 0.0074 0.3245 Note: *** p<0.01, ** p<0.05, * p<0.1; clustered robust standard errors (individual) in parentheses; RE=random effect (individual); FE=fixed effect (individual). 18 / 30
Results Results (2/3): ln(price) Table 3: Regression results for ln(price) (1) (2) Treat i Post t 0.0462* 0.0277** (0.0277) (0.0135) Covariates no yes Constant 2.7289*** 24.5889*** (0.0157) (8.7366) Observations 2819 2319 No. of labels 73 70 R-squared 0.0213 0.7026 Note: *** p<0.01, ** p<0.05, * p<0.1; clustered robust standard errors (individual) in parentheses. 19 / 30
Results Results (3/3): ln(income) Table 4: Regression results for ln(income) (1) (2) FE FE Treat i Post t -0.1838** -0.2000*** (0.0839) (0.0656) Covariates no yes Constant 3.8869*** -64.1173 (0.0577) (35.8422) Observations 2819 2319 No. of labels 73 70 R-squared 0.0055 0.2375 Note: *** p<0.01, ** p<0.05, * p<0.1; clustered robust standard errors (individual) in parentheses. 20 / 30
Robustness checks Robustness checks (1/4): ln(quantity) Table 5: Placebo pre-post treatment regressions (quantity) (1) (1) Pre-Treatment Post-Treatment Treat i Post t -0.1265 0.0748 (0.0883) (-0.1024) Constant -191.3315* -107.1857 (92.2082) (189.7410) Observations 1464 2340 No. of labels 70 71 R-squared 0.3026 0.3287 Note: *** p<0.01, ** p<0.05, * p<0.1; clustered robust standard errors (individual) in parentheses. 21 / 30
Robustness checks Robustness checks (2/4): ln(quantity) Table 6: Placebo control regions regression (quantity) (1) (2) (3) (4) (5) Placebo Region Valais Vaud Geneva G-CH Ticino Treat i Post t -0.0771 0.0379 0.1217 0.0642 0.1117 (0.0497) (0.0833) (0.0927) (0.0714) (0.0746) Constant -81.3427** -76.9864* -80.5901* -78.5024* -95.4196** (40.2871) (41.4338) (41.2078) (40.6861) (41.7297) Observations 2319 2319 2319 2319 2319 No. of labels 70 70 70 70 70 R-squared 0.2828 0.2970 0.2918 0.3010 0.2996 Note: *** p<0.01, ** p<0.05, * p<0.1; clustered robust standard errors (individual) in parentheses. 22 / 30
Robustness checks Robustness checks (3/4): ln(quantity) Table 7: Regressions for different configurations of the control group (quantity) (1) (2) (3) (4) (5) Taken Out Valais Vaud Geneva G-CH Ticino Treat i Post t -0.2578*** -0.2190*** -0.2197*** -0.2260*** -0.2270*** (0.0712) (0.0720) (0.07097) (0.0708) (0.0704) Constant -142.4845*** -92.0549* -79.4692* -72.7472-113.4583* (47.1435) (46.9406) (45.0053) (45.2509) (57.4230) Observations 1769 1869 1927 1937 2096 No. of labels 56 55 57 60 63 R-squared 0.3002 0.3250 0.3000 0.3221 0.3800 Note: *** p<0.01, ** p<0.05, * p<0.1; clustered robust standard errors (individual) in parentheses. 23 / 30
Robustness checks Robustness checks (4/4): ln(quantity) Table 8: Regressions by color type (quantity) (1) (2) (3) Treat i Post t -0.2726*** -0.2363*** -0.2765*** (0.0826) (0.0765) (0.0870) Color Treated White All White Color Control All Red & Rosé Red & Rosé Constant -60.3588-107.3692-65.2067 (42.6202) (51.4292) (52.1732) Observations 2145 1504 1330 No. of labels 64 46 40 R-squared 0.3175 0.2894 0.2759 Note: *** p<0.01, ** p<0.05, * p<0.1; clustered robust standard errors (individual) in parentheses. 24 / 30
Robustness checks DID regression with leads and lags Autor (2003) analyses the effect of increased employment protection on the firm s use of temporary help workers, including both leads and lags in a DID regression: ln(q i,t ) = β 0 + 1 t= q m β3(treat t i δ t ) i,t + β3(treat t i δ t ) i,t +S i,tθ+c i +ɛ i,t (4) t=0 We allow β 3 (coefficient of interest) to vary across time: β t 3 We set t = 0 the starting of the treatment (cut-off date) δ t could be monthly, trimester, semester or year dummies q leads 1 t= q βt 3 (Treat i δ t ) i,t : analysis of pre-trends characteristics m lags m t=0 βt 3 (Treat i δ t ) i,t : analysis of treatment effect changes (if any) over time after treatment 25 / 30
Robustness checks Estimated shock effect over time Figure 7: Estimated shock effect over time (semester dummies). 26 / 30
Conclusion Conclusion Unique natural experiment: hail weather shock in the wine region Three Lakes in 2013 with panel data monthly quantity and price for different types of Swiss wines AOC in the retail market (Nielsen). Strong visual evidence of the parallel trend assumption, before the weather shock, between the treated group and the control group. Statistically significant ATE effects on Three Lakes wines in terms of consumption (-22.8%) and price (+2.8%) several robustness checks confirm the validity and stability of the results. This study could give a contribution to predict and forecast wine outcomes in support of appropriate economic policy decisions when a supply shock occurs in this specific market. Insurance (Suisse grêle), stock management (climate reserve), cantonal and federal subsidies, implication for the wine industry and producers. Internal validity (causal effect through DID framework) versus external validity (similar weather shocks in other wine regions or other agricultural commodities). 27 / 30
Conclusion Thank you for your attention! (a) Lavaux (Vaud) (b) Valais (c) Boccalino of Merlot (Ticino) (d) Château d Auvernier (Neuchâtel) 28 / 30
Conclusion References Angrist, J. D. and J.-S. Pischke (2008): Mostly Harmless Econometrics: An Empiricist s Companion, Princeton University Press, 52, 503 504. Ashenfelter, O., S. Ciccarella, and H. J. Shatz (2007): French Wine and the U.S. Boycott of 2003: Does Politics Really Affect Commerce? NBER Working Papers 13258, National Bureau of Economic Research, Inc. Ashenfelter, O. C. (1978): Estimating the Effect of Training Programs on Earnings, The Review of Economics and Statistics, 60, 47 57. Autor, D. H. (2003): Outsourcing at Will: The Contribution of Unjust Dismissal Doctrine to the Growth of Employment Outsourcing, Journal of Labor Economics, 21, 1 42. Card, D. and A. B. Krueger (1994): Minimum Wages and Employment: A Case Study of the Fast-Food Industry in New Jersey and Pennsylvania, American Economic Review, 84, 772 93. Egger, P., M. Larch, M. Pfaffermayr, and H. Winner (2006): The impact of endogenous tax treaties on foreign direct investment: theory and evidence, Canadian Journal of Economics, 39, 901 931. FOAG (2016): Année viticole 2015, Federal Office for Agriculture, Statistiques vitivinicoles. OSMV (2016): Observatoire suisse du marché des vins, Rapport N05, Année 2015, Changins Haute école en viticulture et oenologie, Swiss wine promotion. SWP (2015): Swiss Wine Promotion,. 29 / 30
Conclusion Definition of the Average Treatment Effect (ATE) E{Y T =t i D i = d} = β 0 + β 1 d + β 2 t + β 3 (dt) + ɛ i,t (5) { 0 if control group d = 1 if treated group (Three Lakes) { 0 if before-treatment period (01/2012-03/2014) t = 1 if post-treatment period (04/2014-03/2015) β 0 = E{Y 0 i D i = 0} β 1 = E{Y 0 i D i = 1} β 0 = E{Y 0 i D i = 1} E{Y 0 i D i = 0} β 2 = E{Y 1 i D i = 0} β 0 = E{Y 1 i D i = 0} E{Y 0 i D i = 0} β 3 = E{Yi 1 D i = 1} β 0 β 1 β 2 = E{Yi 1 D i = 1} E{Yi 0 D i = 0} E{Yi 0 D i = 1} + E{Yi 0 D i = 0} E{Yi 1 D i = 0} + E{Yi 0 D i = 0} = E{Yi 1 D i = 1} E{Yi 0 D i = 1} (E{Yi 1 D i = 0} E{Yi 0 D i = 0}) 30 / 30