Study | Country | Testing for misspecification | Sensitivity analyses | Results | Sources of support clearly acknowledged |
---|---|---|---|---|---|
Douglas, Hariharan (1994)10 | United States | Graphical assessments (predicted hazard function vs. nonparametric hazard functions). | Split-population duration models (probability of ever starting: probit; duration: lognormal); Log-logistic duration models. | Small effect sizes and not statistically significant. | No |
Douglas (1998)14 | United States | None reported. | Reports results with and without index of state regulations. | Small effect sizes and not statistically significant. When past prices are excluded, future prices are found to have a statistically significant effect (results not reported). | No |
DeCicca, Kenkel, Mathios (2000)15 | United States | None reported. | Models with/without state FE are compared; results are presented for the white, Hispanic and african-american sub-samples, but not for the full sample. | Prices not statistically significant for white and african-american samples but significant for Hispanic sub-sample. | Yes |
Forster, Jones (2001; 2003)16 23 | Britain | Graphical assessments (predicted survivor functions vs. KM survivor functions; Cox-Snell residuals); comparison of predicted proportions of starters vs. actual proportion of starters; LR tests to discriminate between pooled and sex-specific models. | Log-logistic duration models on subsamples of smokers. | Small effect sizes and not statistically significant. | No |
Hammar, Martinsson (2001)17 | Sweden | Ramsey's regression error specification test (RESET); AIC to discriminate between six specifications (lognormal, loglogistic, generalized gamma with and without heterogeneity). | Reports results for lognormal duration models with/without indicators of tobacco con policies and with/without price change last year and price chance next year. | Large wrongly signed effect sizes and not statistically significant. | Yes |
Tauras, O'Malley, Johnston (2001)18 | United States | None reported. | Ten models with and without state fixed effects with different covariates are estimated. Models in which clustering is based on zip codes are also estimated (results not reported). | Large effect sizes and statistically significant. Results robust to alternative specifications. | Yes |
DeCicca, Kenkel, Mathios (2002)19 | United States | None reported. | Models with/without state FE and with/without interactions between tax and survey year are compared; models with prices instead of taxes are examined but results not reported. | Taxes statistically significant without FE but not statistically significant with FE. | Yes |
None reported. | Models with prices instead of taxes are examined (results not reported). | Onset between 8th and 10th grade: Large effect sizes and statistically significant; Onset between 8th and 12h grade: Large effect sizes and not statistically significant. | |||
None reported. | Models with and without mean imputation are examined; Models with prices instead of taxes are examined (results not reported). | Large effect sizes and not statistically significant. | |||
None reported. | Models with and without mean imputation are examined; Models with prices instead of taxes are examined (results not reported). | Small, wrongly signed effect sizes and not statistically significant. | |||
Glied (2002)20 | United States | None reported. | Report results of specifications that include current taxes and two lags of current taxes. Models estimated on sub-samples of respondents: 1. whose family income in 1979 was below the sample median; 2. by sex. | Model 1: Large effect sizes and statistically significant. Model 2, 3: large positive effect sizes but not statistically significant. | No |
López Nicolás (2002)21 | Spain | Graphical assessments (predicted survivor functions vs. KM survivor functions; Cox-Snell residuals); LR tests to discriminate between pooled and sex-specific models. | Non-split log-logistic duration models. | Statistically significant but small effect size for both men and women. | Yes |
Cawley, Markowitz, Tauras (2004)24 | United States | None reported. | Models with alternative measures of onset are compared. | Large and statistically significant effect size for males only. Results robust to alternative specifications. | Yes |
Grignon, Pierrard (2004)25 | France | LR tests to discriminate between log-logistic and Weibull distribution. | None reported. | Price when 15 not statistically significant; Price when 18 statistically significant. | No |
Kidd, Hopkins (2004)26 | Australia | Graphical assessments (predicted survivor functions vs. KM survivor functions); LR tests to discriminate between split and non-split models and between pooled and sex-specific models. | Models are re-estimated using alternative age groups and alternative dataset; Log-logistic duration models on subsamples of smokers. | Large effect size for both men and women (elasticity = 0.162 and 0.122) and statistically significant. Results not robust to alternative specifications. | Yes |
Laxminarayan, Deolalikar (2004)27 | Vietnam | None reported. | None reported. | Changes in the price of cigarettes (but not waterpipe tobacco) are significantly and negatively associated with smoking onset. | No |
Arzhenovsky (2006)28 | Russia | Graphical assessments (Cox-Snell residuals); comparison of predicted proportions of starters vs. actual proportion of starters; Schoenfeld residuals tests; | Models are re-estimated using alternative age groups and alternative price measure. | Price of 'cheap' brands: Statistically significant effect. Price of 'expensive' brands: Statistically significant effect (wrongly signed). | Yes |
Cawley, Markowitz, Tauras (2006)29 | United States | None reported. | Models estimated using: 1. three separate indicators of obesity; 2. probit instead of linear probability; 3. IV in which the weight of a sibling is used as the instrument; 4. IV in which the in which the endogenous variable is an indicator for whether the respondent was clinically underweight (instead of overweight). | Large and statistically significant effect size for males only. Results robust to alternative specifications. | Yes |
Kim and Clark (2006)30 | United States | None reported. | Results presented for low, middle, and high SES. | Fairly large effect sizes for low and high SES but not statistically significant. | Yes |
Zhang, Cohen, Ferrence, Rehm (2006)31 | Canada | None reported. | Models estimated on sub-samples of respondents with primary policy data, of respondents not living within 40 km of border to province with lower prices (<$5 per carton), and of respondents who remained in same province. | Large effect sizes and statistically significant. | Yes |
Coppejans, Gilleskie, Sieg, Strumpf (2007)32 | United States | None reported. | Models are re-estimated using alternative measures of price volatility. | Price levels and price volatility statistically significantly associated with the hazard of starting. | Yes |
Grignon (2007)33 | France | None reported. | Models are re-estimated with correction for recall errors (heaped values) using dummy variables; Models separately estimated for individuals who report different time preference. | Large effect size and statistically significant. | No |
Madden (2007)22 | Ireland | Graphical assessments (Cox-Snell residuals). | Models are re-estimated using alternative specifications: 1. split-population loglogistic-probit models with interaction between tax and education; 2. Non-split loglogistic duration models (with and without frailty). | Small effect sizes and not statistically significant. Limited evidence that education modifies the effect of taxes on smoking onset. | No |
DeCicca, Kenkel, Mathios, Shin, Lim (2008)34 | United States | None reported. | Models with and without state FE are compared. | Prices statistically significant without FE but not statistically significant with FE | No |
DeCicca, Kenkel, Mathios (2008)37 | United States | None reported. | Models estimated on sub-samples of individuals who lived in the same state in 1992 and 2000 (stayers), of individuals who lived in different states 1992 and 2000 (movers). | Large effect size and statistically significant. When movers are excluded, effect size is large and statistically significant. | No |
Malhotra, Boudarbat (2009)36 | Canada | Graphical assessments (predicted hazard function vs. nonparametric hazard functions). | None reported. | Price when 15 statistically significantly associated with participation but not duration. | No |
Kenkel, Lillard, Liu (2009)35 | China | None reported. | Models estimated using alternative time trends (results not reported). | Small effect sizes and not statistically significant. Results sensitive to alternative specifications. | Yes |
Liu (2010)38 | United States | None reported. | Models estimated using alternative specifications: by age (15-24; 25-44) for baseline; state FE; antismoking sentiment index. | Age 15-24: large effect sizes; only statistically significant in baseline specification. Age 25-44: not statistically significant; small effect size or wrongly signed. | No |
Étilé, Jones (2010)40 | France | Two specification tests to assess the robustness of the IV procedure. | Several alternative models estimated. | Large and statistically significant effect size for women only. Results robust to alternative specifications. | Yes |
Nonnemaker, Farelly (2011)39 | United States | None reported. | Alternative models estimated with additional covariates: time variant measure of state-level prevalence; antismoking sentiment index; state FE; state-level. | Taxes: statistical significance and effect sizes vary across specifications; effect sizes largest for black youth Prices: generally statistically significant and moderately large; effect sizes generally largest for black youth. | Yes |
AIC, Akaike information criterion; FE, fixed effects; IV, instrumental variables; KM, Kaplan–Meier; LR, likelihood ratio; SES, socioeconomic status.