I. Introduction
Alcohol drinking has long been an integrated part of China’s food culture (Yen, Yuan and Liu, Reference Yen, Yuan and Liu2009), playing an essential role in many important social occasions, such as traditional festivals, family reunions, and business networking. Yet despite its social and cultural values, excessive alcohol drinking may impose a heavy health burden on alcohol consumers, raising the likelihood of developing chronic illnesses such as hypertension, hepatitis, and brain damage (Zahr, Kaufman and Harper, Reference Zahr, Kaufman and Harper2011; Zhao et al., Reference Zhao, Liu, Li, Feng, Chang and Lyu2020). Currently, the number of alcohol-related deaths in China ranks the highest among all countries worldwide (Griswold et al., Reference Griswold, Fullman, Hawley, Arian, Zimsen and Tymeson2018). Even so, per capita alcohol consumption among Chinese residents has been on the rise in recent decades. From 2005 to 2016, China’s per capita alcohol consumption rose from 4.1 liters to 7.2 liters, representing the largest increase globally during that period (World Health Organization, 2018). In 2020 alone, 4,048 million nine-liter cases of beer, 828 million nine-liter cases of Chinese spirits (Baijiu), and 414 million nine-liter cases of wine were consumed in China.Footnote 1 The rapidly rising trends of both alcohol consumption and chronic disease prevalence have led many experts to call for curbing alcohol consumption in China (Jiang, Room and Hao, Reference Jiang, Room and Hao2015).
Regarding the relationship between health and alcohol consumption, an important question remains largely unanswered: Will alcohol consumers alter their drinking behavior—especially in a healthy way—upon receiving notifications about their (worsened) health status, say, being diagnosed with a chronic condition? How chronic disease diagnoses may affect patients’ behavior is of both academic interest and policy relevance. If, for example, chronic disease diagnoses fail to serve as a wake-up call for alcohol drinkers to stop (heavy) drinking, effective early interventions and related policies may be implemented before their chronic conditions become more severe. However, identifying the causal effects of disease diagnoses on alcohol consumption is challenging due to two potential problems. First, there might be reverse causality operated from alcohol consumption to disease diagnoses, disguising the real effect of disease diagnoses. Second, there might exist unobserved confounding factors, such as genetic markups,Footnote 2 that affect both disease diagnoses and alcohol consumption, creating a spurious correlation between them.
The present study circumvents these problems by exploiting the unique medical criteria for hypertension diagnoses to identify the effect of diagnoses of this particular condition. Two unique features of hypertension render it an interesting case to study. First, despite the high prevalence of hypertension worldwide and the fact that it is relatively easy to detect during regular health checkups,Footnote 3 less than half (42%) of all hypertensive patients were diagnosed and treated (Zhou et al., Reference Zhou, Carrillo-Larco, Danaei, Riley, Paciorek and Stevens2021). Worse still, if not treated in time, hypertension may lead to other non-communicable chronic illnesses, such as stroke, coronary artery disease, and heart failure (Mendis, Puska and Norrving, Reference Mendis, Puska and Norrving2011). Thus, the effect of hypertension diagnosis, if convincingly identified, will inform not only hypertension management but also the prevention and treatment of other chronic illnesses. Secondly, unlike most other chronic diseases, hypertension has clear, medically set thresholds for diagnoses based on continuous biomarker (i.e., blood pressure) readings. More specifically, a person will be diagnosed with hypertension if his/her systolic blood pressure (SBP) reading attains 140 mmHg or his/her diastolic blood pressure (DBP) reading reaches 90 mmHg (Unger et al., Reference Unger, Borghi, Charchar, Khan, Poulter, Prabhakaran, Ramirez, Schlaich, Stergiou, Tomaszewski, Wainford, Williams and Schutte2020). These criteria provide a unique opportunity to apply a regression discontinuity (RD) design to identify the effects of hypertension diagnoses on one’s alcohol consumption behavior. To the extent that individuals with blood pressure readings just below and just above a given threshold are otherwise comparable, an RD design can provide credible estimates of hypertension-diagnosis effects.
Taking into account the fact that there are two thresholds for hypertension diagnoses, we follow Wong, Steiner, and Cook (Reference Wong, Steiner and Cook2013) and Dai et al. (Reference Dai, Jiang, Liu and Sun2022) and apply a two-dimensional RD design to a panel dataset involving 10,787 adults from the China Health and Nutrition Survey (CHNS), a large-scale household survey covering more than ten Chinese provinces. Our two-dimensional RD analyses reveal that first-ever hypertension diagnoses based on DBP readings exert some desirable effects, reducing hypertensive adults’ drinking frequency by 1.2 times/week, beer consumption by 518.6 ml/week, Chinese spirits (Baijiu) by 194.8 ml/week and the incidence of excessive drinking by 17.9 percentage points 3-4 years later. Interestingly, we also found modest evidence that hypertension diagnoses increase Chinese adults’ wine intake, suggesting a substitution pattern upon hypertension diagnoses based on DBP readings. In contrast, based on the SBP rule, no significant effects of hypertension diagnoses on alcohol consumption were found.
These analyses make two contributions to the literature. First, to the best of our knowledge, our study is among the very few that identify the causal effects of hypertension diagnosis on alcohol-drinking behaviors, likely the first in China. The existing literature primarily concerns the impact of health information on individuals’ nutrient intake, smoking behavior, and physical activities (Zhao, Konishi and Glewwe, Reference Zhao, Konishi and Glewwe2013; Slade and Kim, Reference Slade and Kim2014; Rabel et al., Reference Rabel, Mess, Karl, Pedron, Schwettmann, Peters, Heier and Laxy2019; Zeng and Shimokawa, Reference Zeng and Shimokawa2020; Dai et al., Reference Dai, Jiang, Liu and Sun2022), paying little attention to its impact on alcohol consumption, a key component of China’s food culture that has potential health concerns. Our study fills this gap. Second, beyond identifying the effects of hypertension diagnoses, we demonstrate how these effects may vary across consumers’ drinking behaviors (whether to drink and drinking frequency) and types of alcoholic beverages (beer, Baijiu, and wine), which have rarely been examined in previous studies on alcohol consumption, especially in the context of China.
The remainder of the paper is structured as follows. The next section briefly reviews the relevant literature. Section 3 describes our data. Section 4 develops a two-dimensional RD framework for identifying the effects of hypertension diagnoses. Section 5 reports and discusses our main findings. The final section draws conclusions and offers several policy implications.
II. Relevant literature
Behavioral theories have long emphasized the importance of “cues” in motivating behavioral changes (Hochbaum, Reference Hochbaum1958; Rosenstock, Reference Rosenstock1974). A specific cue labeled “teachable moment” describes an unexpected health event that could induce individuals to adopt risk-reducing behaviors (McBride, Reference McBride2003). The diagnosis of a chronic condition serves as one such teachable moment (Keenan, Reference Keenan2009; Xiang, Reference Xiang2016; Oster, Reference Oster2018; Hu, Chen and Zhang, Reference Hu, Chen and Zhang2021; Qin, Reference Qin2022). Among known chronic conditions, hypertension has received considerable attention because of its high prevalence and its role as a risk factor for other chronic diseases, including other cardiovascular and cerebrovascular diseases that impose significant threats to the patient’s health status and quality of life (Cheungpasitporn et al., Reference Cheungpasitporn, Thongprayoon, Srivali, Vijayvargiya, Andersen, Kittanamongkolchai, Sathick, Caples and Erickson2016). As such, hypertension diagnoses provide a valuable opportunity to understand how an individual’s health behavior (alcohol drinking in our context) may alter in response to a teachable moment.
Previous studies have examined the impact of hypertension diagnoses on individuals’ smoking behaviors, dietary patterns, and physical activities (Zhao, Konishi and Glewwe, Reference Zhao, Konishi and Glewwe2013; Slade and Kim, Reference Slade and Kim2014; Rabel et al., Reference Rabel, Mess, Karl, Pedron, Schwettmann, Peters, Heier and Laxy2019; Zeng and Shimokawa, Reference Zeng and Shimokawa2020; Dai et al., Reference Dai, Jiang, Liu and Sun2022). Although conducted in different countries, these studies consistently found that individuals reduced their cigarette consumption after a hypertension diagnosis (Zeng and Shimokawa, Reference Zeng and Shimokawa2020; Dai et al., Reference Dai, Jiang, Liu and Sun2022). In contrast, the effects of hypertension diagnosis on individuals’ dietary patterns and physical activities found in the literature are less conclusive. While its impacts on protein and total energy intakes are not statistically significant, hypertension diagnosis has been found to significantly reduce patients’ daily fat and sodium intakes (Zhao, Konishi and Glewwe, Reference Zhao, Konishi and Glewwe2013; Slade and Kim, Reference Slade and Kim2014; Dai et al., Reference Dai, Jiang, Liu and Sun2022). Regarding physical activities, Rabel et al. (Reference Rabel, Mess, Karl, Pedron, Schwettmann, Peters, Heier and Laxy2019) found that individuals who were already physically active tended to remain active after being diagnosed with hypertension; however, the diagnosis did not necessarily make physically inactive individuals more active.
While controlling alcohol consumption is a key component of the medical advice for hypertensive patients, studies investigating whether hypertensive patients actually adhere to this recommendation are still lacking. To the best of our knowledge, only two studies have explored the hypertension-alcohol drinking nexus, and the findings of these studies are also mixed. Liang and Chikritzhs (Reference Liang and Chikritzhs2011), applying multivariate logistic regressions to analyze a cross-sectional dataset from Australia, found that respondents diagnosed with hypertension were more likely to reduce or stop alcohol drinking. In contrast, Kerr et al. (Reference Kerr, Ye, Greenfield, Williams, Lui, Li and Lown2017), fitting generalized estimating equation (GEE) models using retrospective data from the United States, found no statistically significant association between hypertension diagnosis and the likelihood of heavy drinking.
Research examining the hypertension diagnosis-alcohol consumption relationship in China, where the world’s largest hypertensive population resides (World Health Organization, 2003),Footnote 4 is largely nonexistent. Even though this relationship has been examined in Australia and the U.S., findings from these countries may not inform the case of China, given the different drinking preferences and food cultures across countries. For example, beer is the most commonly consumed alcoholic beverage in Australia and the U.S.—as of 2016, beer consumption accounted for 40% and 47% of total alcohol intake in these two countries, respectively. In contrast, Baijiu is the most demanded alcoholic beverage in China, accounting for 67% of China’s total (pure) alcohol intake in 2016 (World Health Organization, 2018). These differences suggest that the impact of health information on alcohol consumption in China might also differ from those found in Australia and the U.S.
Note also that the hypertension diagnosis-alcohol consumption relationships found in the Australian (Liang and Chikritzhs, Reference Liang and Chikritzhs2011) and American studies (Kerr et al., Reference Kerr, Ye, Greenfield, Williams, Lui, Li and Lown2017) discussed above may not be causal. For example, the findings of both studies might be subject to self-selection bias due to their reliance on self-reported data collected through drop-and-collect or computer-assisted telephone interviews. It is possible that individuals who consume more alcohol are more concerned about their hypertension status and are more likely to participate in the survey, which might have biased the estimates of the hypertension-diagnosis effect in these studies.
In light of these knowledge gaps, our study attempts to estimate the causal effect of hypertension diagnosis on Chinese adults’ alcohol consumption by applying an RD design to data drawn from the CHNS, a large-scale longitudinal household survey covering more than ten Chinese provinces.
III. Data
a. Survey and sampling
The China Health and Nutrition Survey (CHNS) is an ongoing survey jointly designed, implemented, and managed by the Carolina Population Center at the University of North Carolina and the National Institute for Nutrition and Health at the Chinese Center for Disease Control and Prevention. The original survey was launched in 1989, covering nine Chinese provinces that vary substantially in geographical conditions, economic development, public resources, and health indicators: Liaoning, Heilongjiang, Jiangsu, Shandong, Henan, Hubei, Hunan, Guangxi, and Guizhou. The project was later expanded to include three municipalities (Beijing, Chongqing, and Shanghai) in 2011 and three more provinces (Shaanxi, Yunnan, and Zhejiang) in 2015.Footnote 5 Nine follow-up surveys were conducted between 1991 and 2015.Footnote 6 Approximately 7,200 households and over 30,000 individuals have participated in the project.Footnote 7
In the first wave (conducted in 1989), the CHNS adopted a multistage, random sampling procedure to select target households. Cities and rural counties were first stratified by income levels (low, middle, and high) in each of the nine original project provinces. Four counties from each province were then randomly selected by a weighted sampling scheme; the provincial capital and a lower-income city were selected whenever feasible.Footnote 8 Villages and townships within the selected counties and urban/suburban communities within the selected cities were chosen randomly. Information on sampled individuals’ socio-demographic characteristics, economic activities, food consumption, and nutrition intake, as well as health behaviors, was collected through face-to-face interviews. Physical examinations, including blood pressure measurements, have also been performed by professionally trained investigators during the survey since the 1993 wave (Zeng and Shimokawa, Reference Zeng and Shimokawa2020; Dai et al., Reference Dai, Jiang, Liu and Sun2022).
To obtain the “cleanest” estimates of the effects of hypertension diagnoses, we focus on identifying the effects of first-ever hypertension diagnoses, as subsequent diagnoses are likely to have weaker effects or pick up the effects of earlier diagnoses. The panel structure of the CHNS data allows us to tell whether a hypertension diagnosis recorded in a given wave t was the first-ever hypertension diagnosis for a sampled individual. In each wave, the respondent was asked, “Have you ever been diagnosed with [X],” where X refers to a particular chronic disease. With this information, we identified the first-ever hypertension diagnosis if a person reported being diagnosed with hypertension in wave t but answered “No” to the same question above before t. Note also that since blood pressure readings were measured during the survey, hypertension diagnoses recorded in a given wave will not affect a respondent’s alcohol consumption behavior reported in the same wave, as the information on alcohol consumption recorded during a given wave refers to the situation sometime before that wave (“last year”). Given such a data structure, we focus on examining the (relatively long-term) effect of first-ever hypertension diagnoses in wave t on individuals’ alcohol consumption observed in the next wave t+1 (about three years later).
Our empirical analysis focuses on data collected from eight CHNS waves (1993, 1997, 2000, 2004, 2006, 2009, 2011 and 2015). We excluded the first two waves (1989 and 1991) from the analysis because the 1989 wave did not collect information on respondents’ blood pressure readings or their drinking behavior (Zeng and Shimokawa, Reference Zeng and Shimokawa2020; Hu, Chen and Zhang, Reference Hu, Chen and Zhang2022), and the 1991 survey only collected data on respondents’ drinking frequency but not the amounts of alcoholic beverages (beer, wine, and Baijiu) consumed. Given these limitations, we chose 1993 as the starting year of analysis. Information collected in 1993-2011 was used to define individuals’ hypertension status (at t) and that in 1997-2015 to construct measures of their alcohol drinking behaviors (detailed below) about three years later (at t +1).
Several sample restrictions were further applied to form the final analytical sample: (1) we limited our attention to respondents who reported consuming alcohol at t; (2) we excluded individuals under age 18 at the time of the survey, given the extremely low incidence of hypertension among them (Zhao, Konishi and Glewwe, Reference Zhao, Konishi and Glewwe2013) and the fact that they are not supposed to consume alcoholic beverages;Footnote 9 (3) given our focus placed on the effects of first-ever hypertension diagnoses at t, we excluded individuals diagnosed with hypertension at t-1 (the previous wave). Applying these restrictions yielded an analytical sample with 10,787 observations. The actual sample sizes used in the regressions are somewhat smaller, as they are determined by the optimal bandwidths (usually within the ±15-mmHg interval around the diagnostic threshold) chosen by the algorithm proposed by Calonico, Cattaneo, and Titiunik (Reference Calonico, Cattaneo and Titiunik2014) and Imbens and Kalyanaraman (Reference Imbens and Kalyanaraman2012).
Table 1, panel B, depicts the profile of sample respondents in wave t. In the analytical sample, 84.2% are male—the high proportion of males resulted from our sample restriction to include only those who consumed alcohol at t, as noted above. The average sampled individual was 44.5 years old, completed 7.7 years of formal education, and came from a four-membered family with a household income per capita of roughly 8,428 Yuan (
$ \approx $1,352 U.S. dollars at constant 2015 prices).Footnote 10 About 85.5% of the sampled individuals were married, and approximately 24.4% held an official urban residential permit (Hukou). Slightly more than 80% of the sampled individuals were working at the time of the survey.
Table 1. Summary statistics
Source: Author’s calculation using data from the China Health and Nutrition Survey (1993-2015).
Note: Total alcohol intake is computed as “total alcohol intake = beer consumption × 0.04 g/ml+ wine consumption × 0.10 g/ml+ Baijiu consumption × 0.52 g/ml,” but due to missing values of some of the variables involved, the relationship between the mean values of these variables may not exactly follow this formula.
*** p<0.01, **p<0.05, *p<0.1.
b. Blood pressure readings and hypertension diagnoses
Blood pressure readings, the treatment assignment variable (or “forcing” variable or “running” variable used in the literature) in our RD design, were obtained through detailed physical examinations performed by the CHNS team. In each survey wave, medically trained investigators recorded each respondent’s systolic and diastolic blood pressure readings on three consecutive days and took the average of the three readings. An individual is diagnosed with hypertension if his/her average SBP reaches 140 mmHg or his/her average DBP attains 90 mmHg (Zeng and Shimokawa, Reference Zeng and Shimokawa2020; Dai et al., Reference Dai, Jiang, Liu and Sun2022).Footnote 11 After completing the physical examinations, the respondents were informed of their blood pressure readings and the result of their hypertension diagnosis. However, those diagnosed with hypertension were not explicitly informed of the severity of their condition. Nor were they prescribed specific treatments during the survey (Zhao et al., Reference Zhao, Liu, Li, Feng, Chang and Lyu2020). As such, the effects identified in this study are those of hypertension diagnoses rather than specific medical advice or prescribed treatments usually associated with them. It is possible that hypertensive respondents sought medical advice or treatments after the diagnosis, but the effects of these follow-up procedures should be interpreted as part of the effect of hypertension diagnosis.
c. Outcome variables
The CHNS recorded detailed information on sampled respondents’ general drinking behaviors (drinking incidence and drinking frequency) and their consumption of the three most commonly consumed alcoholic beverages in China (beer, Baijiu, and wine), greatly facilitating our construction of indicators that capture different aspects of Chinese adults’ alcohol consumption behavior.
1. Drinking behavior
The first set of outcome variables of interest concerns one’s general drinking behavior. Specifically, two variables were constructed to measure whether a respondent (still) consumed alcoholic beverages around the time of the survey at t+1 given his/her hypertension diagnosis result at t and, if yes, his/her drinking frequency. In each survey wave, the CHNS team asked each respondent two questions regarding his/her general drinking behavior. The first question was: “Did you drink beer, wine,Footnote 12 or Baijiu last year?” The answer to this question was used to create a binary indicator of drinking incidence (=1 if “Yes” and =0 if “No”). If the answer was “Yes,” the respondent was further asked: “How often did you drink?” The responses to choose from include: “Almost every day,” “3-4 times a week,” “1-2 times a week,” “1-2 times a month,” “less than once a month,” and “I don’t know.” Since the respondents provided their answers by selecting one of the aforementioned intervals, for ease of analysis and interpretation, we created a continuous variable measuring “weekly drinking frequency” by assigning the midpoints of the chosen intervals as its values (—a value of “missing” was assigned to the response of “I don’t know”).
2. Alcohol intake
The second set of outcome variables concerns Chinese adults’ weekly consumption of beer, Baijiu, and wine. Since in the raw data, beer consumption was measured in bottles (one bottle = 600 ml), while wine and Baijiu consumption in liang (one liang = 50 grams), we converted the units of weekly consumption of these three alcoholic beverages uniformly into grams (g) for ease of comparison. To further calculate one’s total alcohol intake, we followed Ren, Castro Campos and Loy (Reference Ren, Castro Campos and Loy2020) and multiplied the weekly consumption of these beverages by the percentage of their alcohol content (i.e., 0.04 g/ml for beer, 0.52 g/ml for Baijiu, and 0.10 g/ml for wine), to obtain the weekly alcohol intake associated with the consumption of each of these three beverages. Summing across the three beverages yields the total alcohol intake, based upon which an “excessive drinking” variable was further constructed. As suggested by the dietary guidelines for Chinese residents (Chinese Nutrition Society, 2016), a male (female) with an average daily total alcohol intake of more than 25 g (15 g) is considered an excessive alcohol drinker.
Table 1, panel A, presents descriptive statistics of the outcome variables discussed above for all respondents in the analytical sample (column 1) and separately for non-hypertensive (column 2) and hypertensive respondents (column 3). In the full sample (column 1), 69.0% of the respondents reported consuming alcoholic beverages at t+1, with an average frequency of 3.5 times/week. Among the three types of alcoholic beverages considered, beer is the most consumed by volume, with an average intake of 699.2 ml/week, followed by Baijiu, with an average of 325.8 ml/week; wine is the least consumed, with an average intake of 19.9 ml/week. It should be noted that the alcohol content in Baijiu is much higher than in beer. As such, translated from the above volume measures, sample respondents’ alcohol intake from drinking Baijiu (169.4 g/week) is much higher than that from drinking beer (28.0 g/week); that from drinking wine (2.0 g/week) is the smallest.
A quick comparison between columns (2) and (3) suggests that, in general, hypertensive respondents drink significantly more than non-hypertensive ones in terms of drinking incidence, drinking frequency, Baijiu consumption, total alcohol intake, and incidence of excessive drinking. However, it bears mentioning that these comparisons were made without addressing potential confounding issues. As shown in Table 1, panel B, there are significant differences in socioeconomic characteristics, including age, gender, ethnicity, household size, years of education, household income per capita, and working status, between hypertensive and non-hypertensive respondents. To identify the causal effect of hypertension diagnoses on alcohol consumption, methods that can adequately address potential confounding issues are needed. The following section develops such a method.
IV. Empirical methods
The fact that a person’s hypertensive status is “switched on” when his/her DBP (diastolic blood pressure) or SBP (systolic blood pressure) reading, a continuous treatment assignment variable, smoothly passes a clearly specified threshold naturally suggests an RD framework for identifying the effects of hypertension diagnoses. To the extent that individuals with blood pressure readings just above and just below the threshold are comparable in all aspects except their diagnosis results, an RD design effectively addresses unobserved confounding. But unlike the standard RD design that involves only one threshold, there are two thresholds for hypertension diagnoses: 90 mmHg for DBP readings and 140 mmHg for SBP readings. As such, a two-dimensional RD design is needed.
Four approaches to implementing a two-dimensional RD design have been proposed in the literature. The first approach simply picks one set of blood pressure readings as the treatment assignment variable, ignoring the potential contamination of diagnoses based on the other set of blood pressure readings. For example, Zhao, Konishi and Glewwe (Reference Zhao, Konishi and Glewwe2013) focus on SBP readings in their study:
\begin{equation}{\delta ^{S|D\_all}} = \mathop {\lim }\limits_{{S_ \searrow }140} E\left[ {{y_{i,t + 1}}{\text{|}}{S_{i,t}} = 140} \right] - \mathop {\lim }\limits_{S \nearrow 140} E\left[\, {{y_{i,t + 1}}{\text{|}}{S_{i,t}} = 140} \right],\end{equation} where 
${y_{i,t + 1}}$ is an outcome variable of interest (e.g., total alcohol intake) observed at time t+1 for individual i; 
${S_{i,t}}$ is his/her SBP readings measured at t; 
$\mathop {\lim }\limits_{{S_ \nearrow }140} E\left[\, {{y_{i,t + 1}}{\text{|}}{S_{i,t}} = 140} \right]$ and 
$\mathop {\lim }\limits_{S \searrow 140} E\left[\, {{y_{i,t + 1}}{\text{|}}{S_{i,t}} = 140} \right]$ are, respectively, the left-hand and right-hand limits of the outcome variable 
$y$, as the value of the treatment assignment variable approaches 140 mmHg, the SBP-based threshold for hypertension diagnosis.
To circumvent potential contamination of “treatment” based on DBP readings, a refinement, adopted by Zeng and Shimokawa (Reference Zeng and Shimokawa2020), is to exclude observations with SBP readings over 90 mmHg while estimating equation (1):
\begin{equation*}{\delta ^{S|D \lt 90}} = \mathop {\lim }\limits_{{S_ \searrow }140} E\left[ {{y_{i,t + 1}}{\text{|}}{S_{i,t}} = 140,{D_{i,t}} \lt 90} \right]\end{equation*}
\begin{equation} - \mathop {\lim }\limits_{S \nearrow 140} E\left[ {{y_{i,t + 1}}{\text{|}}{S_{i,t}} = 140,{D_{i,t}} \lt 90} \right],\end{equation} where 
${D_{i,t}}$ is individual i’s DBP readings at time t. But, as with the estimator 
${\delta ^{S|D\_all}}$ defined in (1), 
${\delta ^{S|D \lt 90}}$ lacks the power to identify the effect of hypertension diagnosis based on DBP readings.
The third approach, recommended by Wong, Steiner, and Cook (Reference Wong, Steiner and Cook2013) and recently adopted by Dai et al. (Reference Dai, Jiang, Liu and Sun2022), identifies the effects of hypertension diagnosis based on both sets of blood pressure readings. In essence, this approach performs two separate RD estimations, each focusing on one of the two diagnostic thresholds while controlling for potential contamination of “treatment” defined by the other threshold. Specifically, one first implements a one-dimensional RD design, say, 
${\delta ^{S|D \lt 90}}$ defined in (2), using SBP readings as the treatment assignment variable (and 140 mmHg as the threshold), excluding all observations with DBP readings above 90 mmHg (i.e., those in quadrants I and II of Figure 1). Then, one switches the roles of DBP and SBP readings and performs another RD analysis along the DBP dimension:
\begin{equation*}{\delta ^{D|S \lt 140}} = \mathop {\lim }\limits_{{D_ \searrow }90} E\left[ {{y_{i,t + 1}}{\text{|}}{D_{i,t}} = 90,{S_{i,t}} \lt 140} \right]\end{equation*}
\begin{equation} - \mathop {\lim }\limits_{D \nearrow 90} E\left[ {{y_{i,t + 1}}{\text{|}}{D_{i,t}} = 90,{S_{i,t}} \lt 140} \right].\end{equation}
Figure 1. Illustration of identification strategies.
In implementing 
${\delta ^{D|S \lt 140}}$ defined in (3), observations with SBP readings above 140 mmHg (i.e., those in quadrants II and IV of Figure 1) are excluded. Estimating 
${\delta ^{D|S \lt 140}}$ and 
${\delta ^{S|D \lt 90}}$ one at a time, this approach identifies the impact of hypertension diagnosis based on each diagnostic rule separately; individuals who are hypertensive by both diagnostic rules (i.e., those in quadrant II of Figure 1) are also naturally excluded from the analysis.
The final approach, proposed by Imbens and Zajonc (Reference Imbens and Zajonc2009), accommodates both diagnostic rules simultaneously. This approach uses the “distance” to the diagnostic “boundary” jointly defined by the two diagnostic rules as the treatment assignment variable. Specifically, the distance from the diagnostic boundary for individual i at time t, 
$Dis{t_{i,t}}$, is defined as:
\begin{equation}Dis{t_{i,t}}{ } = \left\{ {\begin{array}{*{20}{c}}
{\begin{array}{*{20}{c}}
{\sqrt {{{\left( {{S_{i,t}} - 140} \right)}^2} + {{\left( {{D_{i,t}} - 90} \right)}^2}} ,if\,{S_{i,t}} \geq 140,{D_{i,t}} \geq 90;\,} \\
{{D_{i,t}} - 90,if\,{S_{i,t}} \lt 140,{D_{i,t}} \geq 90;\,}
\end{array}} \\
{ - \min \left( {140 - {S_{i,t}},90 - {D_{i,t}}} \right),if\,{S_{i,t}} \lt 140,{D_{i,t}} \lt 90;} \\
{{S_{i,t}} - 140,\,if{S_{i,t}} \geq 140,{D_{i,t}} \lt 90.}
\end{array}} \right.\end{equation}The corresponding RD estimator is:
\begin{equation}{\delta ^{S\& D}} = \mathop {\lim }\limits_{Dis{t_ \searrow }0} E\left[ {{y_{i,t + 1}}{\text{|}}Dis{t_{i,t}} = 0} \right] - \mathop {\lim }\limits_{Dist \nearrow 0} E\left[ {{y_{i,t + 1}}{\text{|}}Dis{t_{i,t}} = 0} \right].\end{equation} Of all four estimators discussed above, our preferred ones are 
${\delta ^{S|D \lt 90}}$, defined in (2), and 
${\delta ^{D|S \lt 140}}$, defined in (3). Compared with one-dimensional RD estimators that exploit only one diagnostic threshold for identification (e.g., 
${\delta ^{S|D\_all}}$ defined in (1)), the preferred estimators have the advantage of avoiding potential contamination of “treatment” defined by the other threshold. Compared with the “distance-to-joint-diagnostic-boundary” estimator defined in (4) and (5), our preferred estimators, 
${\delta ^{S|D \lt 90}}$ and 
${\delta ^{D|S \lt 140}}$, show how individuals may respond to different treatment assignment rules differently. For comparison purposes, we report estimation results based on all other estimators in Appendix Table A1.
All estimations reported in this study were performed using the statistical package STATA (Version 17). All estimates reported are non-parametric estimates. Based on the common practice in recent RD studies, the actual sample size used in each of the models reported below is determined by the optimal bandwidth chosen to minimize the mean squared error (MSE) at the diagnostic threshold (Imbens and Kalyanaraman, Reference Imbens and Kalyanaraman2012; Calonico, Cattaneo and Titiunik, Reference Calonico, Cattaneo and Titiunik2014; Calonico et al., Reference Calonico, Cattaneo, Farrell and Titiunik2017). Finally, recall from section 3.1 that each sampled respondent diagnosed with hypertension at time t has a “baseline” observation at t-1 with normal blood pressure readings (with both SBP and DBP readings below their respective thresholds) in the sample. To account for these individuals’ repeated appearances in the analytical sample, we follow Liang and Zeger’s (Reference Liang and Zeger1986) “one-level-up” strategy and cluster the standard errors at the community level to avoid serial correlation issues. Results remain very similar when the standard errors are clustered at the individual level or the blood pressure reading level.Footnote 13
V. Results
a. Main results
This section reports our main results of estimating the effects of first-ever hypertension diagnoses on Chinese adults’ alcohol consumption behavior. Table 2 presents non-parametric RD estimates based on our preferred estimators, i.e., 
${\delta ^{S|D \lt 90}}$ defined in (2), and 
${\delta ^{D|S \lt 140}}$ defined in (3). The general finding is that first-ever hypertension diagnoses by the DBP diagnostic rule significantly impact Chinese adults’ alcohol consumption in various aspects; in contrast, those by the SBP rule have no significant impact.
Table 2. Non-parametric RD estimates of the effects of first-ever hypertension diagnosis on Chinese adults’ alcohol consumption
Source: Author’s analysis using data from the China Health and Nutrition Survey (1993-2015).
Notes: Non-parametric RD estimates with the triangular kernel and the associated optimal bandwidths are reported. The analytical sample size is determined by the optimal bandwidth chosen to minimize the mean squared error at the threshold (Calonico, Cattaneo and Titiunik, Reference Calonico, Cattaneo and Titiunik2014).
Standard deviations in brackets; standard errors in parentheses, clustered at the community level.
*** p<0.01, **p<0.05, *p<0.1.
Column 2 of Table 2 reports the estimates based on the DBP rule. Although first-ever hypertension diagnoses by the DBP rule did not significantly alter Chinese adults’ drinking incidence (panel A), they significantly reduced their drinking frequency by 1.2 times/week about three years later (i.e., observed in the next wave) (panel B). The reduction in drinking frequency translates into a reduction in beer consumption by 518.6 ml/week (panel C) and a reduction in Baijiu consumption by 194.8 ml/week (panel D). Compared with the mean amounts of beer (690.7 ml/week) and Baijiu consumption (323.5 ml/week) among non-hypertensive individuals, these reductions are rather sizable. Although these reductions did not lead to a statistically significant drop in total alcohol intake (panel F)—presumably due to an insignificant increase (by 21.6 ml/week) in wine consumption (panel E)—they did lead to a significant drop (by 17.9 percentage points) in the incidence of excessive drinking (panel G).
The left panels of Figure 2 visualize these findings: a drop at the 90-mmHg threshold is clearly shown for drinking frequency (panel c), beer consumption (panel e), Baijiu consumption (panel g), and the proportion of excessive drinking (panel m). Meanwhile, no significant jump was revealed for drinking incidence (panel a), wine consumption (panel i), or total alcohol intake (panel k), at the threshold.Footnote 14
Figure 2. Two-way relationships between blood pressure readings and alcohol consumption.
Column (3) of Table 2 reports the estimates based on the SBP diagnostic rule. Unlike hypertension diagnoses based on the DBP rule, which exerts several significant impacts on Chinese adults’ alcohol consumption (column 2), those based on the SBP rule do not seem to impact Chinese adults’ alcohol consumption in a notable way. All non-parametric RD estimates reported in this column are statistically insignificant. The corresponding RD figures (Figure 2, right panels) also reveal no significant jumps at the 140-mmHg threshold for SBP readings.
Non-parametric RD estimates based on other estimators discussed in section 4 are reported in Appendix Table A1. The one-dimensional RD estimates reported in columns (1) and (2) are quite comparable to their counterparts reported in Table 2, albeit less statistically significant, presumably due to contamination of treatment defined by the other diagnostic rule. The estimates based on the “distance-to-joint-diagnostic-boundary” estimator 
${\delta ^{S\& D}}$, defined in equations (4) and (5), are mostly insignificant (column 3). This is not surprising because the estimates based on 
${\delta ^{S\& D}}$ are the “weighted averages” of the estimates based on 
${\delta ^{S|D \lt 90}}$ (the SBP-based estimator) and 
${\delta ^{D|S \lt 140}}$ (the DBP-based estimator). Since the estimates based on the former are mostly insignificant (Table 2, column 3), the inclusion of systolic hypertensive individuals in the estimation introduces a certain amount of noise in the “weighted averages,” making them mostly insignificant.
b. Validity of the identifying assumption
The validity of the above results (Table 2) hinges on the plausibility of the identification assumption needed for our RD design, i.e., all factors except the hypertension diagnosis status vary continuously when the treatment assignment variable, blood pressure readings, passes the thresholds smoothly. If this assumption is plausible, one would expect to see little impact of hypertension diagnoses on characteristics determined prior to the diagnoses. To verify this, we ran a set of non-parametric RD regressions using a set of pre-determined socio-demographic characteristics as dependent variables. Subject to data availability, the pre-determined characteristics examined include respondents’ age, gender, years of schooling, residential area (urban versus rural), household income, and household size. The results, visualized in Appendix Figure A1, reveal essentially no “jumps” on these characteristics at either the DBP (left panels) or the SBP threshold (right panels). It is impossible to perform a similar check on unobserved characteristics, but the continuity in observed pre-determined characteristics at the two diagnostic thresholds greatly strengthens the validity of our RD design and findings.
A valid RD design also requires that sample respondents did not manipulate the values of the treatment assignment variable so as to land on one side of the diagnostic threshold (Lee and Lemieux, Reference Lee and Lemieux2010), which implies that there be no “clustering” of blood pressure readings on either side of a given threshold. Appendix Figure A2, plotting the densities of the two blood pressure readings, reveals no evidence of such clustering around the thresholds.
c. Alternative bandwidths
Recall that the RD estimates reported in Table 2 were obtained based on the optimal bandwidths chosen by the algorithm devised by Calonico, Cattaneo and Titiunik (Reference Calonico, Cattaneo and Titiunik2014). These estimates could be biased if the corresponding estimation sample (determined by the optimal bandwidth) still includes too many observations with blood pressure readings far away from the diagnostic thresholds—in some cases, the optimal bandwidth exceeds 15 mmHg. Arguably, individuals with a DBP reading of 75 mmHg and those with a reading of 105 mmHg may not have comparable health status (at least in terms of the severity of hypertension). Even though we have controlled for a flexible (non-parametric) function of blood pressure readings, the treatment assignment variable, on either side of the threshold in all regressions reported above, the fitted function may not be sufficiently flexible to eliminate the incomparability problem. To see how the inclusion of “far-away” observations may impact our estimation results, Appendix Table A2 reports non-parametric RD estimates using observations within half (column 1) and twice (column 2) of the optimal bandwidth based on the DBP diagnostic rule (i.e., using the estimator 
${\delta ^{D|S \lt 140}}$, defined in equation 3).Footnote 15 Reassuringly, the new estimates are all quantitatively comparable to their counterparts reported in Table 2, column 2, suggesting that the choice of bandwidth is not a major concern in our context.
d. Other confounding factors
Even though the above checks help strengthen the validity of our RD estimates of the effects of first-ever (DBP-based) hypertension diagnoses on Chinese adults’ alcohol consumption behavior, our estimates might still pick up the influence of other founding factors. One possibility is that our estimated effects of DBP-based hypertension diagnoses reflect the effects of other chronic conditions that are triggered by hypertension. As noted above, if not detected early and appropriately controlled, hypertension may lead to other chronic conditions, such as stroke, coronary artery disease, and heart failure (Mendis, Puska and Norrving, Reference Mendis, Puska and Norrving2011). Many of these other chronic illnesses may also affect one’s food and beverage consumption behavior (Oster, Reference Oster2018; Krämer, Kumar and Vollmer, Reference Krämer, Kumar and Vollmer2021; Hu, Chen and Zhang, Reference Hu, Chen and Zhang2022). Since our outcome variables were observed a few years after the sample respondents’ hypertension diagnoses, our RD estimates may capture the influence of other chronic conditions developed during this time window rather than the impact of hypertension diagnoses per se.
To check this possibility, we ran additional RD regressions treating the incidence of any of the other chronic conditions recorded in the CHNS data (i.e., diabetes, myocardial infarction, apoplexy, and asthma) as the outcome variable. The results, reported in panel A of Table 3, reveal virtually no associations between hypertension diagnoses and the incidence of any other chronic conditions. While it is impossible to check for other chronic diseases (e.g., cancer and hepatitis) that are unavailable in the CHNS data, the finding reported in Table 3, panel A, alleviates the concern about the confounding effects of other chronic conditions. A related concern is that the inclusion of sampled individuals with other chronic conditions in the estimation sample might “contaminate” the effects of hypertension diagnoses in that earlier diagnoses of other chronic conditions may have induced changes in hypertensive individuals’ alcohol consumption behavior. To check this possibility, we excluded individuals who had been diagnosed with other chronic diseases before their (first-ever) hypertension diagnoses and re-estimated our main RD models reported in Table 2. The results, not reported here but made available upon request, remained almost identical to those reported in Table 2.
Table 3. Influence of confounding factors at the hypotension diagnosis thresholds
Source: Author’s analysis using data from the China Health and Nutrition Survey (1993-2015).
Notes: One U.S. dollar ≈ 6.23 Yuan in 2015. Non-parametric RD estimates with the triangular kernel and the associated optimal bandwidths are reported. The analytical sample size is determined by the optimal bandwidth chosen by minimizing the mean squared error at the threshold (Calonico, Cattaneo and Titiunik, Reference Calonico, Cattaneo and Titiunik2014).
Standard errors in parentheses, clustered at the community level.
The effects of hypertension diagnoses reported in Table 2 may also pick up the influence of non-medical factors. Again, since the outcome variables discussed above were observed about three years after one’s hypertension diagnosis result was revealed, it is possible that the diagnosis had caused non-medical factors related to one’s alcohol consumption behavior, such as work status and income, to change. Table 3, panels B-D, reports the results of examining several of such factors. These results suggest that hypertension diagnoses had essentially no effects on whether one was working at the time of the survey (panel B), one’s annual household income (panel C), or household expenditure (panel D). Again, the insignificant correlations between hypertension diagnoses and these factors help reduce the potential concern about confounding factors.
e. Differential effects based on different diagnostic rules: the role of age
Recall that, unlike hypertension diagnoses based on the DBP rule (Table 2, column 2), those based on the SBP rule did not appear to significantly affect Chinese adults’ drinking habits (Table 2, column 3). Yet why do Chinese adults respond differently to hypertension diagnoses based on different diagnosis rules? Age plays an important role in two regards. First, the incidence of diastolic and systolic hypertension varies with age but exhibits distinct patterns. Many studies have shown that “high DBP but normal SBP” readings are more commonly observed among young and middle-aged individuals, whereas “high SBP but normal DBP” readings are more common among older adults (Fang et al., Reference Fang, Madhavan, Cohen and Alderman1995; Tate et al., Reference Tate, Manfreda, Krahn and Edward Cuddy1995; Nürnberger et al., Reference Nürnberger, Dammer, Opazo Saez, Philipp and Schäfers2003; Kanegae et al., Reference Kanegae, Oikawa, Okawara, Hoshide and Kario2017). Consistent with this pattern, panel A of Figure 3 shows that high DBP readings predominate among our sample individuals (roughly) under the age of 60. In contrast, among older adults, the proportion with high SBP readings surpasses that with high DBP readings and increases rapidly with age.Footnote 16 These patterns suggest that the differences in the effects of hypertension diagnoses based on different rules may reflect the varying responses of alcohol drinkers from different age groups when exposed to new health information.
Figure 3. Hypertension diagnoses and alcohol consumption by age groups.
Second, individuals’ drinking intensity also varies with age. Figure 3, panel B, shows that while Chinese adults’ wine and Baijiu consumption do not vary significantly with age, their beer consumption does. In particular, their beer consumption is much higher than their wine and Baijiu consumption but drops significantly after age 45; it becomes lower than Baijiu consumption after age 65.
These two age-related patterns suggest one possible explanation for the differential impacts of DBP-based and SBP-based hypertension diagnoses found above. Chinese adults are more responsive to DBP-based diagnosis than SBP-based diagnosis when they are relatively younger because the former is more common among younger and middle-aged adults (Figure 3, panel A). However, they become less responsive to hypertension diagnoses (either DBP-based or SBP-based) when they become older because their alcohol consumption (especially beer consumption and total alcohol intake) drops rapidly with age (Figure 3, panel B).
To test this explanation, we ran separate RD regressions for sample individuals under 60 and those above 60 based on both diagnostic rules. The results, reported in Table 4, provide evidence that supports this explanation: individuals under 60 are responsive to hypertension diagnoses (but only to DBP-based diagnoses), while those aged 60 and above are not significantly affected by hypertension diagnoses based on either the SBP or the DBP rule.
Table 4. Non-parametric RD estimates of the effects of first-ever hypertension diagnoses on alcohol consumption by age
Data source: Author’s analysis using data from the China Health and Nutrition Survey (1993-2015).
Notes: Non-parametric RD estimates with the triangular kernel and the associated optimal bandwidths are reported. The analytical sample size is determined by the optimal bandwidth chosen by minimizing the mean squared error at the threshold (Calonico, Cattaneo and Titiunik, Reference Calonico, Cattaneo and Titiunik2014).
Standard errors in parentheses, clustered at the community level.
*** p<0.01, ** p<0.05, *p<0.1.
This exploration reveals another informative finding: for individuals under age 60, while their beer and Baijiu consumption drop in response to DBP-based hypertension diagnoses, their wine consumption increases by 40.4 ml/week, an effect that is significant at the 0.079 level (Table 4, column 2).Footnote 17 This finding suggests that we may find a more pronounced pattern when focusing on a subsample that is more responsive to (DBP-based) hypertension diagnoses. To test this possibility, we divide the sample of individuals under age 60 by their residential area (urban versus rural), as urban residents consume more alcohol than rural ones and are more likely to be sensitive to hypertension diagnoses.Footnote 18 The results reported in Table 5 confirm this expectation. Despite its much smaller size, the urban sample yields point estimates of hypertension-diagnosis effects that are, in general, larger than those found in the rural sample (with the only exception of beer consumption). In particular, the impact of hypertension diagnosis on wine consumption is much larger and statistically more significant for the urban sample than in the rural sample.
Table 5. Non-parametric RD estimates of the effects of first-ever DBP-based hypertension diagnoses on alcohol consumption among individuals aged under 60 (rural vs. urban)
Data source: Author’s analysis using data from the China Health and Nutrition Survey (1993-2015).
Notes: Non-parametric RD estimates with the triangular kernel and the associated optimal bandwidths are reported. The analytical sample size is determined by the optimal bandwidth chosen by minimizing the mean squared error at the threshold (Calonico, Cattaneo and Titiunik, Reference Calonico, Cattaneo and Titiunik2014).
Standard errors in parentheses, clustered at the community level.
*** p<0.01, **p<0.05, *p<0.1.
The increase in wine consumption, together with the significant decline in beer and Baijiu consumption, suggests a substitution effect upon (DBP-based) hypertension diagnosis. It is likely that Chinese adults substitute wine consumption for beer and Baijiu consumption upon (DBP-based) hypertension diagnosis because they believe wine has more health benefits compared with beer and Baijiu. In fact, red wine offers various health benefits that have been scientifically documented. For example, resveratrol, presenting in the skin and tannins of red grapes, has been extensively researched and demonstrated to induce vasorelaxation, elicit anti-inflammatory responses, and scavenge reactive oxygen species (ROS), thus helping to protect heart function (Das et al., Reference Das, Mukherjee and Ray2010, Reference Das, Sato, Ray, Maulik, Engelman, Bertelli and Bertelli1999; Wang et al., Reference Wang, Zou, Cao, Hsieh, Huang and Wu2005). In addition, resveratrol has been found to exhibit antitumor activity and neuroprotective abilities (Castello and Tessitore, Reference Castello and Tessitore2005; Zhuang et al., Reference Zhuang, Kim, Koehler and Doré2003). Knowing these benefits, some (young urban) hypertensive individuals who enjoy alcohol drinking may decide to switch from beer and Baijiu consumption to wine consumption while having to cut back on total alcohol intake.
VI. Conclusion
Exploiting the fact that hypertension is diagnosed when a person’s blood pressure reading exceeds a medically determined threshold, this paper estimates the effects of a first-ever hypertension diagnosis on Chinese adults’ alcohol consumption behavior based on a two-dimensional RD design. Analyzing a longitudinal dataset (CHNS) involving 10,787 Chinese adults (aged 18 or above), our analysis reveals that first-ever hypertension diagnoses based on DBP readings exert some desirable effects, significantly reducing hypertensive adults’ drinking frequency, their beer and Baijiu (Chinese spirits) consumption, and the incidence of excessive drinking about three years later. Further analyses reveal that young and middle-aged adults (under age 60) are more responsive to DBP-based hypertension diagnosis than older adults (age 60 and above). Interestingly, younger urban adults were found to substitute wine consumption for beer and Baijiu consumption upon (DBP-based) hypertension diagnosis. In contrast, no significant effects of SBP-based hypertension diagnoses on alcohol consumption were found.
Two policy implications can be derived from these findings. First, the alcohol consumption curbing effects of (DBP-based) hypertension diagnosis underscore the necessity of regular hypertension screening. With the rising incidence of hypertension among younger populations, routine health checkups enable early detection and timely treatment of hypertension and related illnesses. Healthier behavioral changes may also be triggered Second, the insignificant impact of hypertension diagnoses among older adults highlights the need for special attention paid to the alcohol-drinking habits of this specific population. Encouraging them to adopt a healthier lifestyle requires more than just routine health checkups. For example, an alcohol policy with strong legislation and comprehensive health promotion may be considered.
Acknowledgments
We thank the Editor, Karl Storchmann, an anonymous reviewer, and participants of the 2023 Annual Meeting of the Agricultural and Applied Economics Association in Washington, D.C., for their helpful comments on earlier versions of this paper. This work was financially supported by the National Natural Science Foundation of China [grant number 71973134] and the 2115 Talent Development Program of China Agricultural University. This research was conducted based upon the data from the China Health and Nutrition Survey (CHNS). We thank the National Institute of Nutrition and Food Safety, China Center for Disease Control and Prevention; the Carolina Population Center, the University of North Carolina at Chapel Hill; the National Institutes of Health (NIH; R01-HD30880, DK056350, and R01-HD38700); and the Fogarty International Center, NIH, for financial support for the CHNS data collection and analysis files since 1989.
Competing interests
The authors declare no competing interests.
Appendix: Supplementary and Figures Tables
Table A1. Non-parametric RD estimates of the effects of hypertension diagnoses on alcohol consumption
Data source: China Health and Nutrition Survey (1993-2015).
Notes: Non-parametric RD estimates with the triangular kernel and the associated optimal bandwidth are reported. The sample size (N) is the same for all panels in each column; the analytical sample size is determined by the optimal bandwidth chosen by minimizing the mean squared error at the threshold (Calonico, Cattaneo and Titiunik, Reference Calonico, Cattaneo and Titiunik2014).
Standard errors in parentheses, clustered at the community level.
*** p<0.01, **p<0.05, *p<0.1.
Figure A1. Continuity in pre-determined characteristics at hypertension diagnostic thresholds.
Figure A2. Densities of systolic and diastolic blood pressure readings.
Table A2. Non-parametric RD estimates with different bandwidths
Source: Author’s analysis using data from the China Health and Nutrition Survey (1993-2015).
Notes: Non-parametric RD estimates with the triangular kernel and the associated optimal bandwidths are reported. The analytical sample size is determined by the optimal bandwidth chosen by minimizing the mean squared error at the threshold (Calonico, Cattaneo and Titiunik, Reference Calonico, Cattaneo and Titiunik2014).
Standard errors in parentheses, clustered at the community level.
*** p<0.01, **p<0.05, *p<0.1.
Open access
