Searches

We searched, from January 1, 2010, two electronic bibliographic databases (MEDLINE and EconLit), two search engines (Google and Google Scholar), and two working paper repositories (RePEc [Research Papers in Economics] and the National Bureau of Economic Research Working Paper series). The database searches were last updated on April 18, 2022. Additionally, we examined references of included studies and used Google Scholar to examine studies that referenced included studies. The search strategy is provided in the S2 Appendix. At least two reviewers independently screened the articles for inclusion, extracted detailed characteristics, and assessed the risk of bias of each included study. Conflicts were resolved by involving additional reviewers.

Inclusion and exclusion criteria

We included all quantitative studies that examined the association between prices, taxes and tobacco use and reported socioeconomic differences in price responsiveness to tobacco use in LMICs. We included all measures of tobacco use (initiation or onset, participation, consumption, cessation, substitution, escalation, or persistence) and excluded aggregate outcome measures, such as national cigarette consumption or sales. We defined socioeconomic status broadly and included measures such as income, wealth, education, and occupation and excluded studies that solely examined differences between urban and rural areas. We included all measures of associations, including but not limited to: own-price elasticities, odds ratios, hazard ratios and marginal effects.

We used the World Bank Analytical Classifications to exclude high-income countries. As countries can move between categories over time, we used the classification for the year(s) in which data were collected. We excluded cross-country studies if it were not possible to disentangle results between LMICs and high-income countries. Given IARC’s compressive review was published in 2011, we only included studies published from 2010. Table A1 in S1 Appendix presents detailed results for each LMIC study included in the IARC review.

Data extraction, risk of bias assessment, and data synthesis

Studies that examine associations between prices, taxes and tobacco use often use methodological approaches that are overlooked in risk of bias assessment tools. Examples include demand system approaches such as the Almost Ideal demand system, instrumental variable and difference-in-differences approaches and, duration/survival analyses. We used appraisal elements proposed for social experiments and quasi-experiments, and borrowed critical appraisal criteria from the Maryland Scientific Methods Scale and the Cochrane Effective Practice and Organisation of Care (EPOC) Risk of bias tools [25–28]. Due to heterogeneity between studies, a meta-analysis was not possible. Important differences between studies included outcomes (e.g., initiation, participation, consumption), population (e.g., youth, adults, smokers), measures of prices and taxes (e.g., contemporaneous/spacial differences in prices, temporal changes prices or taxes), and analytical approaches (e.g., demand systems using budget shares and unit values, instrumental variable estimation, survival analysis, two-part models).

The following study characteristics were extracted and summarized: citation, country, competing interests and funding disclosure, data type, method, outcomes, price/tax measure, and covariates, sensitivity or robustness checks, and detailed results for each outcome. We assessed the following elements, if applicable:

• clear reporting of the dependent variable(s), explanatory variables, adjustment for inflation; methods, and results (including measures of uncertainty); the source of funding and competing interests;
• clear reporting of the source and extent of price or tax variation (i.e., was there enough variation to identify the effect of price or tax on tobacco use);
• misspecification tests and/or sensitivity analyses that were conducted;
• the extent of missing data, outliers, and attrition and how they were handled;
• was measurement error, quality, or endogeneity taken into account;
• when duration analyses were used, was there any informative censoring, was the functional form of time/duration dependency clearly reported and appropriate, were prices matched to retrospective data appropriately, was the assumption that everyone eventually fails examined;
• was there a formal assessment of socioeconomic differences in price responsiveness;
• precision: how narrow/wide were uncertainty intervals;
• conclusions: were the study conclusions supported by results.

We did not contact authors of primary studies if information was missing or unclear. We did not use scores or scales for assessing the overall risk of bias, as empirical evidence does not support them [29]. For example, blindly adding distinct elements assumes that each element is worth the same, while assigning weights to each element disallows any variation in the bias of each element. Although unclear reporting may be a signal for poor execution, more weight was assigned to our assessment of the methodological approaches. We followed ROBINS-I (a tool for assessing risk of bias in non-randomised studies of interventions) and assessed studies as ‘low risk of bias’ (the study is comparable to a well-performed randomized trial), ‘moderate risk of bias’ (the study is sound for a non-randomized study but cannot be considered comparable to a well-performed randomized trial), ’serious risk of bias’ (the study has some important problems), and ‘critical risk of bias’ (the study is too problematic to provide any useful evidence on the effects of intervention) [30].

To assess confidence in our outcome (i.e., socioeconomic differences in the impact of prices and taxes on tobacco use in LMICs), we considered precision of the effect estimate, consistency of findings across studies, and study design limitations (in particular if socioeconomic differences in price responsiveness were formally assessed using a statistical test). A review protocol was prepared in advance as a part of a funding proposal but was not publicly registered. No amendment were made to the review protocol. At least two reviewers independently extracted detailed characteristics, and assessed the risk of bias of each included study. We used structured summaries to present results (due to heterogeneity between studies a meta-analysis was not possible). A completed PRISMA 2020 Checklist is available in S1 Checklist [31].

Study characteristics

Table 1 presents the characteristics and risk of bias of included studies. Studies are presented by geographic regions, and countries (in alphabetic order), in reverse chronological order (by date of publication). Most studies used cross-sectional or repeated cross-sectional data, while four studies used retrospective data constructed from cross-sectional data. Only one study used longitudinal data. Fifteen studies estimated two-part models, while three studies estimated one of the two parts (i.e., participation or consumption).

Five studies used demand system approaches; four used Deaton’s two-equation system of budget shares and unit values that corrects for quality and measurement error, while two studies used the Almost Ideal Demand System (AIDS) of Deaton and Muellbauer. A further seven studies used Deaton’s two-equation system of budget shares and unit values, but only for the consumption component of the two-part model. Four studies used duration analyses to examine the association between cigarette prices and smoking onset and/or cessation.

Most studies used unit values (expenditures/quantity consumed) or self-reported cigarette prices to construct a measure of price that varied in space. Four studies used time series of cigarette prices used in the construction of consumer or wholesale price indices. Only one study examined a change in tax.

Most studies clearly disclosed competing interests and funding; 11 studies clearly disclosed funding but not competing interests, while one study did not report competing interests or funding. Thirteen studies reported funding by the Bloomberg Philanthropies, a New York City based philanthropic organization and nine reported funding by the International Development Research Center, a Canadian Crown corporation that funds research in LMICs. None of the studies reported funding from tobacco manufacturers or any of their affiliates.

Study findings

Key results and a summary of findings are presented in Table 2. Of 32 studies, only seven studies formally examined socioeconomic differences in price responsiveness and three found that lower-SES individuals or households were more responsive to cigarette prices than those of higher-SES. One study, rated at moderate risk of bias, found that households in Bosnia and Herzegovina with lower total expenditures were more responsive to cigarette prices, with differences large enough to be economically meaningful [63]. Another study, rated at serious risk of bias, found that low- and middle-SES households were more responsive to price than high-SES households in Montenegro [62]. A third study, rated at critical risk of bias, found some statistically significant socioeconomic differences in price responsiveness for cigarette/bidi smoking initiation in India; differences, however, were very small and unlikely to be meaningful [57].

Download:

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

Table 2. Summary of key findings.

https://doi.org/10.1371/journal.pgph.0002342.t002

Own-price elasticities estimates for a further eight studies suggest a potential gradient in SES. Studies rated at serious or critical risk of bias from Albania, Bangladesh, India, North Macedonia, Peru, the Philippines, and Serbia found that participation and/or consumption own-price elasticities tended to be higher for lower-SES households [40, 47, 49, 52, 55, 58, 59, 61]. However, none of the studies formally examined if socioeconomic differences were statistically significantly different.

An additional four studies, one assessed at serious risk and three at critical risk of bias, provided unclear evidence that own-price effects differed between SES groups in Bangladesh [51], Pakistan [54], the Philippines [48], and Tanzania [32].

All other included studies provide no evidence that SES modified the association between cigarette prices and total cigarette consumption or cigarette smoking participation/consumption (seven studies assessed at serious risk of bias using data from Ecuador [36], El Salvador [44], Mexico [35, 42], Peru [40], China [46], and India [50], and seven studies assessed as critical risk of bias conducted in Brazil [41], Colombia [43], China [45], Thailand [49], Bangladesh [53], Iran [56], and Kosovo [60].

Similarly, included studies provide no evidence of any statistically or economically significant socioeconomic differences in price responsiveness for cigarette smoking initiation in four studies assessed at moderate (Argentina [38] and Chile [39]), serious (Kenya [33]) and critical risk of bias (Colombia [43]) or cessation in three studies assessed at serious (Mexico [34] and Kenya [33]) and critical risk of bias (Colombia [43]). Results from Argentina, if anything, suggest that individuals with lower education may have been less responsive to price than those with higher education attainment [38].

With the exception of one study, assessed at critical risk of bias [52], studies that examined tobacco products other than cigarettes such as bidis, roll-your-own, and smokeless tobacco found that SES was not an effect modifier [49, 50, 53, 54, 57].

A list of excluded studies and reasons for exclusion is provided in S4 Appendix. Of note, we excluded several recent World Bank working papers which reported own-price elasticities by income deciles or quintiles. We excluded these studies because too little methodological information was provided to allow us to assess the risk of bias, or because no original own-price elasticity estimates were provided. Additionally, measures of uncertainty and statistical significance were generally not provided, while unit values were often naively treated as market prices.

Risk of bias of included studies

Our assessment of risk of bias is summarized in Table 1 (last column). We did not rate any studies at low risk of bias. We rated three studies at moderate, 13 studies at serious, and 16 studies at critical risk of bias.

Nine studies did not provide a clear description of the dependent (outcome) while seven did not do so for the independent (explanatory) variables. Twenty-four studies did not clearly report how missing data and/or outliers were handled, and 16 studies failed to describe if, and if so how, prices or proxy measures of prices were adjusted for inflation. Studies that relied on spatial variation in unit values or self-reported prices often failed to define or report the number of clusters. Thirteen studies did not report having conducted any sensitivity analyses and only five reported having conducted any testing for misspecification. Of the four studies that used duration analyses, two did not clearly report the functional form of duration/time dependency and how prices were matched to retrospective individual-level data and two likely had informative censoring among younger survey respondents. Lastly, the most important limitation was the lack of formal assessment of socioeconomic differences in price responsiveness. Only seven of 32 studies formally assessed statistically whether own-price effects were modified by SES; three used interactions [38, 39, 46] while four used seemingly unrelated regressions to conduct Wald tests of equality of coefficients across equations [33, 57, 62, 63].

Main findings

As the World Bank in 1999 and IARC in 2011, we found the evidence was not sufficient to conclude that socioeconomic status modified the relationship between prices or taxes and tobacco use in LMICs. It is important to note that we did not find evidence of no effect. Rather, we conclude that the evidence in LMICs was too limited and methodologically weak to make any conclusive statements.

Our review highlights a number of data and methodological limitations in existing studies. First, as mentioned earlier, most studies used unit values (expenditures/quantity consumed) to construct a measure of price that varied in space (with no or little time dimension). The use of unit values relies on the assumption that there are no price variations within each cluster (e.g., village) which requires both geographical proximity of the households and that households be interviewed at approximately the same time [64]. The variation in unit values within each cluster reflects differences in tobacco product quality and measurement errors. Unit values, unlike market prices, are, at least to some extent, affected by the quality of the product (e.g., higher-SES households may purchase higher priced cigarettes so that unit values may be related to income). Consequently, aggregating unit values over large regions can be problematic. Additionally, because unit values are derived from reported spending and quantity purchased, measurement error in quantity and expenditure will introduce measurement error in the unit value [65]. A recent study found that the own-price elasticity of quantity demand was overstated by a factor of four, on average, if the response of quality to price was ignored [66]. In Papua New Guinea, it was found that just one third of the response to price was from changes in the quantity of tobacco products and two thirds was from quality [67]. Given that the response of quality to price may vary between rich and poor tobacco users, results from studies that ignore quality may under or overestimate differences in price responsiveness between rich and poor. Two studies included in our review presented quality elasticities that varied substantially by SES categories [37, 44]. Moreover, there is evidence the Deaton approach, which takes quality into account, understates its response [66–68]. An additional difficulty is that to take quality into account when examining socioeconomic differences in price responsiveness, one needs to partition the sample into SES groups which reduces the number of households with positive tobacco expenditures in each cluster, as each cluster should contain at least two consuming households for all goods including in the demand system. It was also often unclear whether all households, at cluster-level, faced the same mean unit values, or a mean unit value specific to each household SES category. In the latter case, households in the same cluster may end-up with different unit values.

Second, duration models are well-suited to study the effects of prices or taxes on smoking onset or cessation. However, incorporating the time or duration dependency is crucial as both theory and data suggest that the hazard rates of smoking onset and cessation are not constant with respect to time [69]. Additionally, a key assumption of duration models is noninformative censoring. Censoring occurs when incomplete information is available about the duration time of some individuals (e.g., an individual being interviewed at 16 years old, before she started smoking at 18) [70]. Noninformative censoring requires that the mechanism that caused censoring was not related to the probability of an event occurring (e.g., starting or quitting smoking). Younger individuals at interview are less likely to have initiated smoking and those who have already started, are less likely to have quit. Hence, younger respondents are more likely to be censored, and censoring is likely to be informative for younger respondents.

Third, the standard statistical approach for potential effect modifiers such as SES is a test for interaction. However, interactions cannot be easily incorporated in demand system approaches that correct for quality and measurement error. Additionally, the interpretation of interaction terms in nonlinear models such as logistic or probit regressions is more complex than in linear models [71, 72]. Most included studies partitioned the sample into two or more SES groups but failed to formally assess if differences in price elasticity estimates were statistically significant. To test whether the coefficients are equal across SES groups, a Wald test can be used [73]. However, standard tests of the equality of coefficients may not be valid in models with binary outcomes such as logistic and probit regressions [74, 75]. Moreover, some measures of associations such as odds ratios should not be compared when they were obtained from different samples from different populations [72].

Our review highlighted a number of data and methodological limitations in existing studies which can inform future research. First, more attention ought to be given to the source and extent of price or tax variation. Second, studies relying on budget shares and unit values should pay particular attention to product quality and measurement errors, avoid averaging unit values over large geographical regions, and clearly report the total number of clusters, the number of clusters with positive expenditures, and the average number (per cluster) of households with positive expenditures for each good examined. Third, for studies that use duration analyses, more attention needs to be given to informative censoring, the functional form of time/duration dependency, how prices are matched to retrospective data, and the assumption that everyone eventually will fail (e.g., start smoking). Lastly, differences in price responsiveness should be formally assessed.

Limitations

First, it is difficult to predict the direction of the effect of the overall risk of bias on average own-price effects, let alone on differences in own-price effects between SES categories. To the extent that ill-suited data and weak methods likely yielded estimates that were more imprecise, it is likely the case that weaker studies were less likely to find own-price effects estimates that differed between SES categories. Second, three included studies were coauthored by an author of this review and eight by recent collaborators of an author of this review. As a result, readers are urged not to rely solely and uncritically on the risk of bias assessment we presented. Third, although a number of the studies reviewed did not clearly provide important methodological information, which rendered risk of bias assessment difficult, we did not contact authors of included studies. Lastly, a review protocol was not made publicly available before starting to work on the review. Risk of bias and quality assessment tools such as ROBIS (Risk of Bias in Systematic reviews) and AMSTAR (A MeaSurement Tool to Assess systematic Reviews) recommend that a review protocol that predates the start of the review be made publicly available [76, 77]. Without such availability, it difficult to assess if eligibility and risk of bias criteria were actually stipulated in advance. We are hopeful that the comprehensiveness and transparency of our search and risk of bias assessment (we documented reasons behind all our risk of bias assessments in detail) and that we did not exclude any studies that met our inclusion criteria because of poor quality will alleviate concerns related to the lack of pre-registration.