Ethics

This study uses secondary data from Global Adult Tobacco Surveys (GATS) in 2010 and 2015, and Vietnam Household Living Standard Surveys (VHLSS) from 2010 to 2016. These data were provided by General Statistics Office of Vietnam and the World Health Organization. The data are fully anonymized, and there are no ethical issues and conflicts of interest related to the respondents in the surveys.

Estimation strategy

The first objective of this study is to measure the effect of tobacco price on daily cigarette smoking participation among male adults (aged from 15). We used the following regression model: (1) where Y_i,j,t is the status of daily smoking of male individual i in province j in year t. It is equal to 1 for daily smokers and 0 otherwise. Log(Price_j,t) is logarithm of the average price of Vinataba cigarette pack of province j in year t. This price is adjusted to the 2001 price using overall provincial-level CPI. X_i,j,t is a set of individual-level control variables, and P_j is a vector of province-level control variables. T_t is a vector of the time dummy variables.

The individual-level control variables include age, education, Kinh majority, urban dummy and wealth index. There are no data on income or expenditure of individuals in this model, and we construct a wealth index to measure the living standard of individuals. Higher value of the asset index means a higher access to the assets. The asset index is standardized to have a mean of zero and a standard deviation of one. We follow the principal components approach of Filmer and Pritchett [19] and Kolenikov et al. [20] to compute a wealth index. According to this approach, an index is constructed as the first principal component of a vector of assets of households such as durables and housing conditions. The principal component approach defines a wealth index in terms of the first principal component of the variables used. The wealth index, denoted by A_j, for household j is computed as follows: where x_p denotes the asset p, and denotes a mean of households in the sample. s is a standard deviation of asset x_p, and the p-dimensional vector of weight a is chosen to maximize the sample variance of A, subject to . The weight a is also called the vector of scores of asset variables, which can be estimated using principal component analysis. Based on the availability of data. The assets used to compute the wealth index consists of including flush latrine, telephone, television, radio, fridge, car, washing, air-conditioning, motorboat, electricity generator, grinder, computer, and internet connection. We also control for the year, and province dummies. The province-level control variables including overall CPI, per capita income, and log of population of density. S3 Table presents the definition of the control variables.

Model (1) is estimated using data from GATSs 2010 and 2015. We used data representing the smoking history of individuals and to compute the smoking status of individuals in a given year, using the GATS surveys. We constructed cohort data on smoking status for all individuals in the datasets. For example, for n male observation in the 2015 GATS, we know their smoking status not only in 2015 but also in previous years by combining information on their age to start and stop smoking. Suppose a man started smoking in 2012 but stopped smoking since 2014, then his smoking status is ‘not smoking’ before 2012, then ‘smoking’ in 2012 and 2014, and ‘not smoking’ in 2014 and 2015. By expanding observations over time, we have panel data on the smoking status, demographic variables such as age and gender, and the provincial-level cigarette prices. Using the cohort data increases the number of observations over time, which allowed us to have more variation in the cigarette price over time and across provinces. Since data on cigarette prices are available from 2001 to 2016 and the latest GATS is from 2015, we limited data on smoking among individuals to 2001 to 2015.

Since individuals are expanded over years, we need to cluster the standard errors at the individual level over time. We adopt the multiway clustering technique of Cameron and colleagues [21], which allows us to address the correlation of errors within individuals across yeas and between individuals within enumeration areas simultaneously. In the Stata, we use ivreg2 command which allows for regressions with two-way level cluster.

The second objective of this study is to examine the impact of the cigarette price on the tobacco expenditure of households. Using the household-level data from VHLSSs, we estimated the following model: (2) where Log(C_k,j,t) is logarithm of per capita expenditure on tobacco products during the past 12 months of household k in province j in year t. Log(Price_j,t) is logarithm of the average price of Vinataba cigarette pack of province j in year t. X_k,j,t is a set of household-level control variables, and P_j is a vector of province-level control variables. T_t is a vector of the time dummy variables. The price elasticity of tobacco expenditure is measured by θ₁. Eq (2) is widely used to model tobacco demand and the price elasticity of tobacco demand (e.g., see [22]).

Explanatory variables include household composition, characteristics of household head, log of per capita income, and geographic variables. We tend to use exogenous explanatory variables, since the explanatory variables should not be affected by the treatment variable [23, 24]. The summary statistics of the outcome and control variables are presented in S4 Table.

An important issue in estimating the effect of the tobacco price is endogeneity bias. The cigarette price is jointly determined by demand and supply of cigarettes. A higher cigarette price reduces demand for cigarettes, but a lower cigarette demand will subsequently lead to a lower cigarette price. In addition, the cigarette price can be also correlated with unobserved variables, especially province-level variables. To address this issue, we controlled for province dummies in the regression (province fixed effects). The cigarette price is at the provincial level, and controlling for province fixed-effects is expected to address the endogeneity bias. In addition, we used instrumental variable regression, in which the instrument for the price of cigarette is the lagged cigarette price. The instrumental variable regression is a traditional econometric method that can provide a consistent estimate of the effect of an endogenous variable (the price of cigarette in this study). An instrument is required to be correlated with the endogenous variable but not the error term in the outcome equation (for detailed presentation of this method, see any econometric textbook, e.g., [25]).

A widely-used instrument for the cigarette price is lags of cigarette price (e.g. [22, 26]). In this study, we use cigarette price lagged by one, two and three years as the instrumental variables for the current cigarette price. The main advantage of using lags as the instrumental variables is that they are strongly correlated with the current price, therefore avoiding the weak instrument bias problem. However, it is possible that the instruments can be correlated with the error term in the outcome equation. Using different lags of cigarette price, we can conduct an over-identification test of the validity of the instruments [27, 28]. A better instrumental variable for cigarette price is cigarette taxes. For the case of Vietnam, the special consumption tax on cigarettes increased only two times (in 2006 and 2008) during the 2006–2016 period. Moreover, there is a national cigarette tax rate in Vietnam. There are no variations in the tax rate across provinces. Thus we cannot measure the effect of the cigarette tax rate on cigarette price and use it as the instrumental variable for the cigarette price. Another possible instrumental variable for cigarette price is the lead price. In this study, we tried this instrumental variable, but the estimation result shows a positive and insignificant effect of the cigarette price. It is possible that the lead price can be affected by the cigarette demand in the current period. If people consume less (or more) cigarette in the current period, the price of cigarettes can be lower or higher in the next period. The exclusion assumption can be violated when the lead prices are used as the instrument. Thus we do not use the lead price as the instrumental variable in this study.

The effect of tobacco price on smoking participation

Table 1 presents the effect of cigarette price on daily smoking participation based on the regression Eq (1). The dependent variable is a dummy variable indicating daily smoking among men. It should be noted that binary models such as logit and probit, which treat the smoking status as a binary event, are often used when modelling tobacco participation. In this study, we use both linear probability regression and probit regression. The 2SLS estimators are consistently and widely used for the binary model with endogenous variables [29, 30]. S5 Table reports marginal effects from the probit regression. We used the results from the linear probability regression in Table 1 for interpretation.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Regression of the smoking participation.

https://doi.org/10.1371/journal.pone.0260415.t001

As mentioned in the estimation method section, we used information on age of smoking initiation and cessation to construct cohort data on smoking status from 2001 to 2015. The total number of individuals in the two GATS surveys used in the regression was 8,339. After being expanded over time, the total number of observations was 93,662. Since we expanded the data, there are no data on several time-variant explanatory variables, such as education and wealth index, for years other than 2010 and 2015. We used education and wealth index computed from the 2010 GATS as proxy measures of education and wealth index for years prior to 2010. For 2011 to 2015, we used education and wealth index computed from the 2015 GATS as proxy measures. In S6 Table, we ran regressions without controlling for education and wealth index. The results were very similar to those from regressions which controlled for these variables.

Table 1 reports results from OLS and 2SLS regression. In 2SLS, the instrumental variables for the log of current cigarette price are logs of lagged cigarette prices. When there is more than one instrument for the endogenous variable, a widely-used method to compute the instrumental estimates is two-stage least squares (2SLS). In the first stage, we regress the endogenous variable (the cigarette price in this study) on all of the control variables and the instruments and estimate the predicted values of the dependent variable. In the second stage, we regress the outcome variable (the cigarette smoking in this study) on the control variable and the predicted values of the endogenous variable from the first stage (see [25] for more detailed discussion). For comparison, we present 2SLS estimates using one instrument (columns 2 and 6), two instruments (columns 3 and 7) and three instruments (columns 4 and 8). The instruments are positively and strongly correlated with the current price. We performed Cragg-Donald and Kleibergen-Paap weak identification tests on the instruments and received the test statistics of over eight thousand (see Table 1). We use command ivreg2 in Stata to estimate the instrumental variable regression and conduct weak instrument and over-identification tests (see [28]). This result indicates extremely strong instruments (as a rule of thumb, if an F- statistic is under 10, the instruments might be weak [31]), which play an important role in reducing the bias in estimating the effect of the endogenous variable. Since the lagged variables are used as the instrument, the number of observations in the OLS regression is smaller than that in the 2SLS regression.

Using more than one instrument, we can conduct an over-identification test of the validity of the instruments [27, 28]. The joint null hypothesis under this test is that the instruments are valid instruments, i.e., they are uncorrelated with the error term. The last row of Table 1 shows the overidentification test statistics and the corresponding P-value in parentheses. It shows that the null hypothesis is not rejected at the 5% significance level. Estimates using different sets of instruments are very similar. For interpretation, we will use the estimate from using one instrument (columns 2 and 6 of Table 1), since this estimate is more efficient with a smaller standard error.

Both OLS and 2SLS regressions showed a negative effect of cigarette price on smoking. The point estimate from 2SLS was larger than that from OLS, but the difference was not statistically significant at the conventional level. The fact that the OLS point estimates are smaller than the 2SLS estimates suggests a possible downward bias in the OLS estimates. A possible reason is that the OLS estimates tend to reflect the interaction between the price and demand of cigarettes. As we know, an exogenous increase in the cigarette price leads to a decrease in the cigarette demand. However, the decrease in the cigarette demand can subsequently reduce the cigarette price. Thus the association between the cigarette price and the cigarette demand might be closer to zero than the causal effect of the cigarette price on the cigarette demand.

Column (2) of Table 1 shows that a one-percent increase in the cigarette price reduces the probability of cigarette smoking of individuals by 0.08 percentage points (the 95% CI from -0.06 to -0.10). In 2015, the daily smoking rate of men was 30.7%. A reduction of 0.08 percentage points is equal to 0.26% of the smoking rate. It means that the elasticity of the daily smoking prevalence of men with respect to the cigarette price was -0.26 (the 95% CI from -0.16% to -0.33%).

Most empirical studies estimate the price elasticity of demand for tobacco products. There is less empirical evidence on price elasticity of the smoking prevalence, especially in low-income countries. In Australia, Siahpush and colleagues [32] estimated the prevalence price elasticities for low-, medium-, and high-income groups at −0.32, −0.04, and −0.02, respectively. According to Sharbaugh and colleagues, an additional $0.25 per-pack tax was associated with a 0.6 percentage point reduction in the smoking prevalence in the United States [33]. Using the average price and the smoking prevalence, we can compute the prevalence of smoking elasticity to the price at 0.95. Compared with these two studies, the price elasticity of the smoking prevalence for Vietnam is slightly lower. Possibly, the cigarette price is rather low in Vietnam, and tobacco products remain affordable by a large number of people.

In addition to the cigarette smoking, we also estimated the effect of the cigarette price on daily tobacco smoking. i.e., the dependent variable is measured by smoking of any tobacco products. Columns from (5) to (8) of Table 1 show that the effect of the cigarette price on tobacco smoking was also negative and statistically significant. According to 2SLS regression in column (6), a one-percent increase in the cigarette price resulted in a 0.043 percentage point decrease in the probability of tobacco smoking, equivalent to 0.11% of the tobacco smoking rate. The point estimates of the effect of cigarette price on tobacco smoking was smaller than those on cigarette smoking, since tobacco products include not only cigarettes, but also other products such as traditional and shisha water-pipes, smokeless tobacco, and these non-cigarette products might be less sensitive to cigarette price.

To examine how cigarette smoking among men changes as a result of a price increase, we simulated the effect of the cigarette price on the prevalence of cigarette smoking and the number of male cigarette smokers. Using the estimated effect of the log of cigarette price on the cigarette smoking participation at -0.0799 (column 2 in Table 1), we predicted how smoking prevalence changes as a result of a change in the cigarette price. The daily cigarette smoking rate in 2015 was 30.7%. The national average price of a Vinataba pack was 20.6 thousand VND in 2015. If this price is increased by 10%, the cigarette smoking prevalence was predicted to decrease from 30.7% to 29.9% (Panel A of Fig 3). It is estimated that the number of cigarettes declines from 10.53 to 10.26 million (Panel B of Fig 3), a reduction of 270 thousand smokers. A larger increase in the cigarette price led to a larger decrease in smoking prevalence and the number of smokers. If the cigarette price is increased by 100%, it is predicted that the cigarette smoking rate will decrease by 5.5 percentage points and the number of cigarette smokers will decline by 1.89 million.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Simulated effects of the price increase on daily male cigarette smoking.

Panel A. The daily male cigarette smoking rate (%) corresponding the percentage change in the price. Panel B. The number of daily male smokers (million) corresponding the percentage change in the price. Note: This figure presents the estimated prevalence and its 95 confidence intervals of daily cigarette smoking among men (in percent) and the number of males with daily cigarette smoking (in million) corresponding to the increase in the cigarette price. Source: Estimation from GATS 2010 and 2015.

https://doi.org/10.1371/journal.pone.0260415.g003

We used the same model specification as column 2 in Table 1 to estimate the effect of log of the cigarette price on cigarette smoking for different population sub-groups. Fig 4 presents the point estimates and their 95% confidence interval of the log of the cigarette price on the cigarette smoking. All the point estimates were negative. However, the effect was not statistically significant for some groups. By wealth quintile, the effect of the cigarette price on smoking was largest for people in the second wealth quintile. The effect of the cigarette price was small and not statistically significant for people in the lowest wealth quintile. It could be explained by the availability of variety of cigarette brands in Vietnam market [34], and the low wealth quintiles tend to use the economy brands. Another possible reason for lower elasticity of the smoking prevalence with respect to the cigarette price among the disadvantaged groups is that these groups are more likely to use homemade tobacco products which can substitute for cigarettes. The effect tended to be higher for urban and Kinh people compared to rural and ethnic minority people, though the difference in the effect between these groups was not statistically significant.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Heterogeneous effect of the cigarette prices on the smoking probability.

Source: Estimation from GATS 2010 and 2015.

https://doi.org/10.1371/journal.pone.0260415.g004

The effect of cigarette price tended to be larger on people with upper-secondary education compared to people with other levels of education. By age groups, the effect of the cigarette price was significant for people below 31 years of age and those aged 51–60. The significant effect observed for young people implies that higher cigarette prices might delay smoking onset. This is consistent with findings from Guindon [16] and Nguyen and colleagues [17] on the smoking-onset delaying effect of cigarette prices.

Table 1 also reveals several interesting findings regarding the effect of different variables affecting smoking. We found there to be an inverted-U shaped relation between smoking and age. The probability of smoking increased with age, but after achieving a peak at around age 47 it started to decrease. People with higher education levels were less likely to smoke than those with lower education levels. There was a negative association between smoking and wealth index, indicating a lower smoking rate among wealthier people. However, urban and Kinh people had a higher probability of smoking compared to rural and ethnic minority people.

The effect of tobacco price on tobacco consumption expenditure

This section reports and discusses the estimation of the cigarette price on tobacco expenditure among households in Vietnam. Table 2 reports the regression of the log of per capita tobacco expenditure on log of cigarette prices. We used 2SLS results for interpretation. The instrumental variable for the cigarette price is the lagged price (in the previous year). The lagged price is very strongly and positively correlated with the current price. The test statistics from Cragg-Donald and Kleibergen-Paap weak identification tests were higher than ten thousand, indicating a very strong instrument (as a rule of thumb, if an F-statistic is under 10, the instruments may be weak [31]).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Regression of the log of per capita tobacco consumption expenditure.

https://doi.org/10.1371/journal.pone.0260415.t002

There are two dependent variables. The first is the log of total expenditure of all tobacco products, including homemade tobacco products (reported in columns 1 to 4 of Table 2), and the second is the log of purchased tobacco products (reported in columns 5 to 8 of Table 2). The effect of cigarette prices was very similar in regressions of the two dependent variables. Similar to Table 1, the OLS point estimates are smaller than the 2SLS estimates, suggesting a possible downward bias in the OLS estimates (though the differences between the OLS and 2SLS estimates are not statistically significant). We use the 2SLS estimates for interpretation. Our estimate of the price elasticity of tobacco expenditure is consistent with previous studies. It is quite close to the estimate from Eozenou and Fishburn [14], which estimated the price elasticity for cigarette demand at around -0.53, using Vietnam Living Standard Survey 1998. Our estimated elasticity is similar to those found in Latin American countries, where own-price elasticity for cigarettes was estimated at -0.4 [9]. A meta-analysis of elasticities from Gallet and List [35] estimates an average price elasticity of –0.48 based on 86 studies. Jha and Chaloupka [36] found the price elasticity of tobacco expenditure to be approximately -4% in high-income countries and approximately -8% in low-income countries. Chaloupka and colleagues [6] also concluded that tobacco demand is more elastic to the price in low-income countries compared to high-income countries.

We used a similar model specification as column 6 in Table 2 to estimate the elasticity of tobacco expenditure with respect to prices among different groups of households. Fig 5 shows that the price elasticity of tobacco expenditure was higher for households in higher income quintiles. The price elasticity was highest among households with household heads who have obtained upper-secondary education. This finding is consistent with our results on the highest effect of cigarette prices on the smoking participation among men with upper-secondary education (Fig 4). We also found cigarette price to have a larger effect on tobacco expenditure among urban and Kinh households, compared to rural and ethnic minority households.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Heterogeneous effect of cigarette prices on the tobacco expenditure.

Source: Estimation from VHLSSs 2006–2016.

https://doi.org/10.1371/journal.pone.0260415.g005

Finally, Table 2 shows several important findings on the association between tobacco expenditure and household characteristics. Households with more members had lower per capita tobacco expenditure compared to households with fewer members. We found that households with more children and older people spend less on tobacco. The amount of tobacco expenditure among urban households was approximately 11 percent higher than that of rural households. This is consistent with findings from Table 1 which indicate that urban people had a higher smoking rate compared to their rural counterparts. As expected, male-headed households were much more likely to consume tobacco than households headed by female. This is likely because smoking in Vietnam is mainly prevalent among men. In female-headed households, household heads are more likely to be widowed, divorced, or have migrating husbands.

We found a strong correlation between education of household heads and tobacco expenditure. The tobacco expenditure was much lower in households with highly-educated heads than in households with lowly-educated heads. Higher income was associated with higher tobacco expenditure. The income elasticity of tobacco expenditure was estimated at 0.35 (column 4), suggesting that tobacco is a normal good. As per capita income increased by 1%, the demand for tobacco increased by around 0.35%. This estimate is very similar to the estimates of income elasticity of tobacco demand at 0.3 to 0.4 in other studies such as Gallet and List (2003) [35], Martinez and colleagues (2015) [37], Álvarez and colleagues (2020) [38].

It should be noted that Table 1 indicates that people in lower wealth quintiles are more likely to smoke than those in higher wealth quintiles, while Table 2 shows a positive correlation between income and expenditure on tobaccos of households. Our finding from Table 1 is consistent with a well-known evidence that the poor tends to smoke more than the rich in low- and middle-income countries [39]. For the case of Vietnam, Kien and colleagues (2017) also find the proportion of adult male hardcore smokers is higher among the poor than among the rich [40]. However, in Table 2, We estimate the elasticity of tobacco consumption of households with respect to the current income. Put it differently, we limit the sample to households with positive tobacco consumption. Our findings on income elasticity is also consistent with previous studies which suggest that tobacco is a normal good with positive income elasticity.