Health behavior and fluctuations of the business cycle:
Has the Great Recession influenced health behavior in Europe?
Katharina Koenz
1Supervised by dr. Gerard H. Kuper
2Master Thesis MSc Economics [EBM877A20]
8 June 2018
Abstract
Macroeconomic fluctuations affect large parts of the population simultaneously, evidence on how this impacts health behavior is surprisingly mixed. This paper estimates the effect macroeconomic fluctua-tions have on smoking, alcohol consumption and physical inactivity, utilizing a generalized estimating equation for pseudo-panel data based on a selection of European countries over the time period from 2004 to 2015. Results are significant, however, evidence on the effects is heterogeneous. The findings suggest that smoking and alcohol consumption vary countercyclical, whereas physical inactivity varies procyclical. I find that the 2007 recession has led to a significant change in physical inactivity but not in smoking and alcohol consumption. Effects are sensitive to sample selection and estimation method. Results indicate that the extensive margin of employment has a significant influence, while the intensive margin of work is not significant for health behavior. Macroeconomic conditions impact health behavior via the income- and the stress effect.
JEL classifications: C33, E24, E32, I12
Keywords: health economics, health behavior, macroeconomic fluctuation, recession, Europe, pseudo-panel, generalized estimating equation, cohort
1 S3488659 | K.Koenz@student.rug.nl
1
I. Introduction
“A rising tide lifts all boats”, John F. Kennedy popularized this phrase and numerous economists used it since to indicate that economic upswings benefit all economic agents, for example through lower unem-ployment and higher incomes. However, evidence on whether this statement also holds for health is mixed. Further, it is still unclear how exactly a falling tide, or a recession, affect individuals and their health. Health behaviors are actions taken by individuals that have consequences on health outcomes, examples are smoking, alcohol consumption or physical inactivity. This thesis provides insights into how macroeconomic fluctuations influence individual health behavior. Focusing on a selection of European countries during the period from 2004 to 2015, with special attention to the 2007 recession, this paper analyses the link from economic fluctuations to health behavior.
Based on the SHARE3 dataset and proxies for macroeconomic fluctuations, this paper reveals new
understandings whether, and how, the depth and the length of economic downturns translate into changes in individual behavior with regards to alcohol consumption, smoking, and physical inactivity. This paper adds to ongoing research, elaborating on various articles that have been published over the past decades.
Health constitutes a major factor for individual and collective welfare. A population in good health can boost economic growth by investing labor and capital effectively, resulting in higher gross domestic product (GDP) per capita (Wilkie & Young, 2009). Furthermore, health has an impact on equality. By na-ture, the health status is unequal among individuals. Some research finds that investments in the health status depend on socioeconomic status (Cutler, Lleras-Muney, & Vogl, 2012). To address inequalities re-sulting from this, reallocation of resources can take place via the health care system to increase equality and welfare, in both, monetary and non-monetary terms (Culyer & Newhouse, 2008; Waters, 2000). In addition, health is an input in the individual’s production function and simultaneously significant parts of health can be influenced by individual behavior. Health behavior is a mediator to health outcomes, how-ever, this research is solely concerned with impacts on health behavior. Moreover, on the cost side, health care expenditures present a significant part of GDP, and additionally, productivity lost due to illness can be harmful for the economy (Tompa, 2002). Continuously rising health care expenditures draw more at-tention to health-related topics. Overall, EU member countries experience increasing health care expend-itures since the 1970s. Measured in the percentage of gross domestic product (GDP), health care expenses have almost doubled for EU countries over the past four decades and account for approximately one-tenth of GDP in 2011, a significant part of each country’s spending4.
3 “Survey of Health, Ageing and Retirement in Europe (SHARE) is a multidisciplinary and cross-national panel database of micro data on health, socio-economic status and social and family networks of more than 120,000 individuals aged 50 or older” (http://www.share-project.org/ ac-cessed May 1, 2018)
2 Recessions are costly for an economy, leading to higher unemployment, loss of labor productivity, loss of individual income and consumption, fewer investments and lower GDP growth rates. If economic contractions have negative effects on health behavior, this could imply that recessions are even more costly. Economic fluctuations are a recurring phenomenon exposing households to systemic risk and af-fecting large parts of society simultaneously. Macroeconomic fluctuations might influence health behav-ior mainly via two channels, the income- and the stress effect. The income effect suggests a positive rela-tionship between macroeconomic conditions and health behavior and moves along the business cycle. In line with this argument, in economic upswings more income is available and thus more money is spent on consumption goods, including alcohol and cigarettes, and more physical inactivity. In contrast, the stress effect suggests a u-shaped relationship between macroeconomic fluctuations and health behavior, where stress is highest during peak and through. High (perceived) stress causes individuals to consume more stress-relieving substances such as alcohol, cigarettes, and be physically less inactive to cope with stress. This research aims to determine how macroeconomic fluctuations change individual health behavior and evaluate these changes in terms of the income- and stress effect. For this purpose, the paper utilizes the SHARE dataset and proxies for macroeconomic fluctuations. This study adds to current re-search by combining elements from modeling health with indicators for economic fluctuations on individual- and aggregate level. While many previous studies focus on the US, this paper is concerned with a selection of European countries during the period from 2004 to 2015. Due to cultural differences and also tremendously different institutions, results obtained from US studies cannot be generalized. Conse-quently, analyzing the effect for European countries is promising and important to gain a more profound understanding of the effects macroeconomic fluctuations have on health behaviors in Europeans.
There are several studies examining this effect, presenting conflicting results, thus underlining that in order to derive effective public policy measures to increase efficiency, realize saving opportunities, and achieve higher welfare, more research is needed. Moreover, gaining a deeper and more profound understanding of the link between macroeconomic fluctuations and health behavior is essential to com-prehend the underlying mechanisms. Interaction and interdependency between micro- and macroeco-nomics highlight the importance of joint approaches in order to provide an all-encompassing view.
3
II. Conceptual Framework
A. Definitions & Concepts
What is “health”?
A universal definition of health does not exist but rather a variety of approaches and understandings. This discussion is beyond the scope of this paper, for my purposes I will rely on the definition by Ruger (2010). Ruger defines health as the “confidence and ability to be effective in achieving optimal health given bio-logic and genetic disposition; intermediate and the broader social, political, and economic environment; and access to the public health and health care system” (Ruger 2010, p. 47). This definition indicates that the economic environment is crucial to health. While health has a physical and a mental component, I will focus on physical health and leave mental health for further research.
The health production function
Neoclassical models assume that households make decisions to optimally allocate their time and income. Households derive utility from consuming goods or, as added later, leisure time (Varian, 1992; Blundell & MaCurdy, 1999). Time is scarce and has opportunity costs. Households aim to maximize utility which can be best described by a strictly concave utility function. Becker (1965) and Michael and Becker (1973) add the household production function and underline the importance of human capital. Households can invest in themselves in order to become more productive. Each household may only consume what it produces. Production requires time and capital. Grossmann (1972) builds on the idea of human capital and expands the concept to health capital. According to Grossmann (1972), human capital is a stock of knowledge affecting productivity; health capital affects the time a household can spend producing capital or goods. Health is a consumption commodity, individuals derive disutility from being sick, and health is an invest-ment commodity, it influences the time available for work and leisure. The model will be discussed in more detail in chapter 3.
The household’s utility depends on health and all other consumption goods, as in the neoclassical approach. The household faces two restrictions; first, time is limited and second, the standard budget constraint applies. Macroeconomic fluctuations may influence the health production function in three ways: (1) relative cost of (health) goods or time changes, (2) changes in wage can shift the budget con-straint or (3) changes in the state of health, for example, the rate at which health deteriorates. Each of these changes might lead to changes in individual health behavior.
What is health behavior?
4 p.78). Those actions can promote or reduce health outcomes and include exercise, sleep, diet, cigarette, and alcohol consumption. Health behavior is not exogenous but may be influenced by underlying mecha-nisms, for example by future expectations or by stress. Individual decisions are often present-biased which implies that by hyperbolic discounting individuals value their current situation more than uncertain future states. This applies to health behavior as long-term consequences of behaviors are discounted stronger (Severens & Milne, 2004).
How to measure health
Measuring health is a challenge, there are over 100 health measurement instruments (McDowell, 2006). The most prominent ones are the mortality rate since death is the ultimate indicator of poor health and QALYs (Quality-Adjusted Life Years) which is an aggregate utility measure where the health status is cap-tured in a single number or score including both the quantity and the quality of life. Many studies evaluate health outcomes in terms of mortality rates, however, it is an oversimplification as it neglects all other health outcomes. For this paper, I will not evaluate health outcomes but examine the effect that macroe-conomic fluctuations have on health behavior.
Macroeconomic fluctuations
Macroeconomic fluctuations can be visualized by the business cycle describing the changes in an econ-omy’s GDP around its long-term growth trend. A recession is loosely defined as two-quarters of consecu-tive negaconsecu-tive GDP growth, while a depression is a period of prolonged recession. Economic downturns are costly for any economy due to higher unemployment, loss of output and unproductive resources. Macro-economic conditions, for example, employment levels and prices, change along the business cycle and can thus be used to proxy macroeconomic fluctuations. Most publications, discussed in the literature re-view, use changes in GDP growth or the unemployment rate as an indicator for recessions since it is a very accurate proxy to describe the ability to find a job and therefore a country’s labor market conditions. Proxies for macroeconomic conditions may also be on individual level, for example, the household in-come, employment status or the amount of hours worked.
Pro- & countercyclical variation
5 The time period from 2004 to 201 5
The dataset used in this analysis is focused on the time period from 2004 to 2015, which well suited to analyze the impact of macroeconomic fluctuation since Europe experienced major economic changes dur-ing that time period. The recession of 2007/8 is an example of a radical economic change in Europe. If there is a change in health behavior that is caused by economic downturns, this recession should be se-vere enough for these effects to be visible.
Figure 1 displays the economic development in the selection of the European countries for the time from 2004 until 2015, by using the average unweighted quarterly GDP growth rate. The graph highlights two major economic downturns; the first recession lasted from 2007 until 2009 with strong negative GDP growth. After a period of recovery, the second recession lasted from 2010 until 2012 and showed less severe negative GDP growth.
Figure 1: Average annual GDP growth in the countries subject to this study (in percent)
*Austria, Belgium, Switzerland Germany, Denmark, Spain, France, Greece, Italy, the Netherlands, Sweden
-2,5 -2 -1,5 -1 -0,5 0 0,5 1 1,5 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015
6
B. Literature Review
This section presents literature relevant to the research question, divided into the theoretical foundation and the empirical foundation. The theoretical part includes an overview of the theoretical concepts and an explanation of the transmission channels. The empirical part gives a short overview of the historical development in the field and presents the main publications in two tables. Table 1 contains papers inves-tigating multiple health behaviors, table 2 covers papers focusing on a single health behavior. The tables follow the same structure, naming the author and year, findings, the country and time period investigated, and which proxies were used for macroeconomic fluctuations and for health behavior, classified as either aggregate- or individual level. Finally, I highlight important findings and elaborate on each health behav-ior.
Theoretical Foundation
The idea that health interacts with economics and decision making is as old as economic thinking. Adam Smith (1776) was aware that during economic upswings, the option to maximize wages may incentivize individuals to work more than what is optimal for their health and they may “hurt their health by excessive labour‘’ (Smith, 1776, p. 82). More than a century later, Joseph Alois Schumpeter (1912) took an interest in the matter as well, he aimed to explain implications, consequences, and dynamics resulting from eco-nomic development for societies. He drew attention to the interdependencies of health and the ecoeco-nomic framework and how changes in a system influence individual behavior. Since then, the topic has been circling academic literature and public discussion.
Research shows that health behavior can vary in a pro- or countercyclical manner. From a theo-retical perspective, the relation between macroeconomic fluctuations and health behavior can be ex-plained by two effects, the income- and the stress effect, displayed in figure 2.
7 witness lower alcohol and cigarette consumption and less physical inactivity during economic down-swings. This is illustrated in figure 2, the income effect moves consistent with the business cycle, house-holds have a higher income during expansions and a lower income during recessions and adjust their health behavior.
The stress effect5 can argue for both, pro- and countercyclical reasoning, depending on the perception of stress (Schneiderman, Ironson, & Siegel, 2005). I assume a u-shaped relation between macroeconomic conditions and perceived stress. Both, peak and through are associated with high levels of stress for the individual as depicted in figure 2. The procyclical reasoning holds in phase one and two of the business cycle. First, when the economy moves away from the peak, stress also declines and second; when the economy recovers from a through, stress also decreases. The countercyclical reasoning holds in stage three and four of the business cycle. First, when the recession becomes more prolonged, stress increases as economic conditions keep declining, and second, when the economy moves out of depression and growth is positive, stress starts to decline.
In this line of argument, an economy both in a boom and in a depression leads to higher stress for the individual. In a flourishing economy, stress levels are high due to working overtime, rising require-ments or more efficiency and productivity. Consequently, consumption of stress-easing substances like alcohol and cigarettes increases. Similarly, reduced leisure time leads to higher physical inactivity. On the contrary, in an economy of enduring recession, stress may also be elevated due to high uncertainty, neg-ative future expectations, growing economic pressure caused by rising unemployment, decreasing wages or reduced working time, fear of job loss, foreclosure, or bankruptcy. In this line of argumentation, the
5 short for “psychophysiological stress response effect”
8 consumption of stress relieving substances, alcohol, cigarettes, and sports will be high. Stress is lower during economically calmer times, less alcohol, cigarettes, and sports are consumed to cope with stress.
Countercyclical variation of health behavior can only be explained by the stress effect. Conversely, procyclical variation can be explained by both, the income- and the stress effect. Thus, the underlying cause of procyclical variation is ambiguous. In the US, job loss will often lead to loss of health insurance, causing additional stress. However, this is not the case in Europe and therefore findings are expected to differ in this aspect.
For the analysis to follow, health behaviors are classified as relatively more or less time-intensive. The individual faces a budget- and a time-constraint, which is explained in more detail in Chapter 3. Hence, one should consider the time-intensity of health behavior and the time elasticity. Smoking is relatively less-intensive compared to other health behaviors and if smoking is limited during working hours, individ-uals may balance their cigarette consumption with increased smoking during leisure time. Smoking is rather time inelastic. When the marginal cost of time increases, individuals do not quit smoking, but rather adjust the intensity of smoking. This indicates that the time constraint will not have a major effect on smoking (Adda & Cornaglia, 2006). In the European Union, the price elasticity of cigarettes is estimated to be -0.4, indicating an inverse relation between cigarette price and consumption. Hence, smoking ought to be more affected by the budget than by the time constraint (Gallus, Schiaffino, La Vecchia, Townsend, & Fernandez, 2006). Alcohol consumption itself is little time consuming, however, since alcohol has an intoxicating effect, time-intensity rises with every additional drink consumed. An increase in the marginal cost of time will merely affect the consumption of the first alcoholic drink. Since this study does not meas-ure the intensity of drinking, but rather addresses the question whether alcohol is consumed, I classify alcohol consumption as less time-intensive. The price elasticity for alcohol is estimated to be -0.36 and thus alcohol consumption will be affected by the budget constraint rather than by the time constraint (Gallet, 2007). For most working individuals there are strict prohibitions to consume alcohol during work-ing hours, thus it takes place mainly durwork-ing free time hours.
9
Empirical Foundation
A great variety of articles investigating the effect of macroeconomic fluctuations on health behavior have been published, empirical results are surprisingly conflicting. Taking a closer look at these studies is of importance since often results are used to justify policies and public spending.
Examining the link between economic activity and health is a long-running debate and of enduring concern since Brenner’s (1973; 1975; 1979) publications. Initial research was almost exclusively relying on ‘mortality’ as health variable, arguing that death is the ultimate indicator of poor health. However, this may present an oversimplification. With increasing medical treatments, at least in developed countries, death is no appropriate proxy for health outcomes since many health statuses which have implications on quality of life, on health care expenditures and on welfare, are lost. Brenner (1973; 1975; 1979) used aggregate time-series data, his results show countercyclical variation, for example, elevated mortality during economic downturns. However, his methods were criticized by many researchers for suffering from omitted variable bias, which is common in time series analysis, because underlying factors are not controlled for, for example, effects from medical progress, the invention of new antibiotics, leading to skepticism about his findings (Ruhm, 2000; 2005). In addition, these studies are sensitive to the selection of countries and time periods and therefore do not present a robust foundation to derive policy interven-tions.
Groundbreaking research6 on the topic was conducted by Ruhm (2000). He published his first
ar-ticle in a series of publications on the topic, provocatively titled “Are recessions good for your health?”. This fundamental publication gave grounds to many other empirical studies in the field. Ruhm (2000) examined the link between macroeconomic fluctuations and total mortality as an indicator of health, in aggregate data, finding mortality to vary procyclical. Procyclical fluctuation is also apparent in rising ciga-rette consumption and more physical inactivity. Ruhm’s (2000) work is especially highly recognized for solving the omitted variable bias problem inherent in Brenner’s (1973, 1975, 1979, 1983) work, primarily by using pooled macro data with location-specific fixed effects to control for variation between states while exploiting within-state changes and keeping time-invariant variables constant.
While Ruhm’s (2000) article is based on aggregate data, a later publication by Ruhm (2005) differs substantially as it used microdata from the Behavioral Risk Factor Surveillance System (BRFSS) for the time from 1987 to 2000, to underline that health behavior like smoking or physical inactivity decreases during recessions. Procyclicality can be explained to some extent by lower deaths by external sources, like acci-dents. Changes in lifestyle and behavior have an impact as well. Inspired by Ruhm’s (2000; 2005) findings, much research has been conducted since results are heterogeneous.
6 Cited: 1702 times (according to google scholar April 2nd, 2018)
10 Since the research question of this paper targets health behaviors, this literature review will focus on studies relating to health behavior rather than on mortality. In this literature review, I will provide two tables. Table 1 gives an overview of a selection of crucial empirical research on the link between macroe-conomic fluctuations and multiple health behaviors. Table 2 provides an overview of a selection of studies investigating how macroeconomic fluctuations affect single health behaviors.
11
Table 1: Selection of empirical studies investigating multiple health behaviors
Macroeconomic fluctuations Health behavior
Author/s (year) Findings Country Time period Variable Level Variable Level Data1
Ruhm (2005) procyclical US 1987-2000 unemployment rate, hours
worked, household income
both smoking, physical inactivity individual BRFSS, LAUS Macy, Chassin, & Presson
(2013)
procyclical US 2005-2011 hours worked, employment
sta-tus, financial strain
individual smoking, physical inactivity individual IUSS Asgeirsdottir, Corman,
Noonan, Olafsdottir, & Reich-man (2014)
procyclical Iceland 2007-2009 hours worked, real household in-come, loss in financial assets, mortgage debt
individual smoking,
alcohol consumption
individual Health & Wellbeing, PHII Jofre-Bonet, Serra-Sastre, &
Vandoros (2016)
procyclical England 2001-2013 unemployment rate aggregate smoking,
alcohol consumption
individual HSE Tekin, McClellan, & Minyard
(2013)
countercyclical US 1990-2014 unemployment rate aggregate smoking, alcohol
consump-tion, physical inactivity
individual BRFSS Kalousova, & Burgard (2014) countercyclical US 2008-2011 employment status, income individual smoking,
alcohol consumption
individual MRRS
Latif (2014) countercyclical Canada 1994-2009 unemployment rate aggregate smoking,
alcohol consumption
individual Canadian NPHS Cotti, Dunn, & Tefft (2015) countercyclical US 1984-2010 DJIA index, unemployment rate,
per capita income
aggregate smoking,
alcohol consumption
individual BRFSS, DJIA, NHCPD
Ruhm (2000) mixed US 1987-1995 unemployment rate,
employ-ment to population, personal in-come
aggregate smoking, alcohol consump-tion, physical inactivity
individual BRFSS
Charles and DeCicca (2008) mixed US 1997–2001 unemployment rate aggregate smoking, alcohol
consump-tion, physical inactivity
individual NHIS, BLS
Xu (2013) mixed US 1976-2005 hours worked, wage rate,
em-ployment status
both smoking, alcohol
consump-tion, physical inactivity
individual BRFSS, NHIS, CPS
Herzfeld, Huffman, and Rizov (2014)
mixed Russia 1994-2004 unemployment rate, prices,
Gross Regional Product (GRP) per capita, employment status, income
both smoking,
alcohol consumption
both RLMS, RSY
Colombo, Rotondi, and Stanca (2018)
mixed Italy 1993-2012 unemployment rate aggregate smoking, alcohol
consump-tion, physical inactivity
individual ISTAT, 2015
1Bureau of Labor Statistics (BLS), Behavioral Risk Factor Surveillance System (BRFSS), National Population Health Survey (NPHS), Current Population Survey (CPS), Dow Jones Industrial Average (DJIA), Health Survey
12
Table 2: Selection of empirical studies focusing on one health behavior
Macroeconomic fluctuations Health behavior
Author/s (year) Finding Country Time period Variable Level Variable Level Data1
Berniell, & Bietenbeck (2017) procyclical France 1998-2002 hours worked, income individual smoking individual ESPS
Kendzor et al. (2012) countercyclical US 2005-2007 unemployment, household income,
employment status
both smoking individual C.M., 2000
US Census Kaiser, Reutter, Sousa-Poza, and
Strohmaier (2017)
mixed Germany 1998-2014 unemployment, household income,
hours worked
both smoking individual SOEP
Colman and Dave (2013) countercyclical US 2003-2010 employment-to-population aggregate physical inactivity individual ATUS, PSID,
NLSY79
Colman and Dave (2014) countercyclical US 1998-2010 employment status individual physical inactivity individual PSID, NLSY79
Nie, Otterbach, & Sousa-Poza (2015)
procyclical China 1989-2011 hours worked individual physical inactivity individual CHNS
Ruhm (1995) procyclical US 1975-1988 unemployment rate, per capita
in-come, employment-to-population ratio
aggregate alcohol consumption aggregate AEDS
Freeman (1999) procyclical US 1970-1995 unemployment rate, per capita
in-come, employment-population ratio
aggregate alcohol consumption aggregate AEDS
Ruhm and Black (2002) procyclical US 1987-1999 unemployment rate, per capita income aggregate alcohol consumption individual BRFSS
Krüger and Svensson (2010) procyclical Sweden 1861-2000 GDP aggregate alcohol consumption aggregate CAN
Ólafsdóttir and Ásgeirsdóttir (2015)
procyclical Iceland 2007-2009 hours worked, income individual alcohol consumption individual DHI
Cotti, Dunn, and Tefft (2015a) procyclical US 2004-2010 unemployment rate, per capita income both alcohol consumption individual NHCPD Kossova, Kossova, and Sheluntcova
(2017)
procyclical Russia 2008-2012 unemployment rate, per capita in-come, consumption of electricity
aggregate alcohol consumption aggregate FSSSR
Dee (2001) countercyclical US 1984-1995 unemployment rate, per capita income aggregate alcohol consumption individual BRFSS
Dávalos, Fang, and French (2012) countercyclical US 2001-2005 unemployment rate, personal income, employment status
both alcohol consumption both NESARC
Frijters, Johnston, Lordan, & Shields (2013)
countercyclical US 2004-2011 unemployment rate, insured unem-ployment rate
aggregate alcohol consumption aggregate Google Insights
Ettner (1997) mixed US 1988 employment status aggregate alcohol consumption individual NHIS
Johansson, Böckerman, Prättälä, & Uutela (2006)
mixed Finland 1975-2001 unemployment rate, GDP aggregate alcohol consumption both THL
1Alcohol Epidemiologic Data System (AEDS), American Time Use Survey (ATUS), Behavioral Risk Factor Surveillance System (BRFSS), China Health and Nutrition Survey (CHNS), Enquete sur la Sant e et la Protection
13
Empirical studies investigating multiple health behaviors
Table 1 suggests that empirical studies find opposite results. Procyclical variation is confirmed by some authors, one important contribution by Ruhm (2005) showed that smoking decreases during tempo-rary economic downturns and leisure-time physical inactivity declines. To proxy macroeconomic con-ditions he uses both, individual- and aggregate level data. Procyclicality is confirmed by Macy, Chassin, & Presson (2013), however, their study provides evidence that on individual level neither the weekly working time nor a change in employment is crucial but the financial strain on households. In addition, Asgeirsdottir, Corman, Noonan, Olafsdottir, & Reichman (2014) confirm procyclical variation for smok-ing and drinksmok-ing when investigatsmok-ing the 2008 economic downturn in Iceland. They find that health behaviors are mainly affected mainly via the price mechanism. Jofre-Bonet, Serra-Sastre, & Vandoros (2018) conclude that unemployment does not explain all variation in health behavior.
Countercyclical variation is supported by some papers for instance by Cotti, Dunn, & Tefft (2015b), who use data from Dow Jones Industrial Average (DJIA) and stock market crashes as indicators to proxy macroeconomic fluctuations. Higher smoking and more alcohol consumption occur in periods of crashes or great negative monthly DIJA returns, thus supporting countercyclical variation. They con-clude that the stock market drives future expectations. Tekin, McClellan, & Minyard (2013) focus on the time period around the Great Recession. Their research supports countercyclical variation, how-ever, they underline that the strength of the link between economic conditions and health behavior has weakened considerably, and many times is not significant any longer. Kalousova & Burgard (2014) support this finding. Even though they also find evidence for countercyclical variation, they stress that the importance of distinguishing between different types of recessions, and a perceived versus an ac-tual economic decline. Latif (2014) finds that the countercyclical effect of the unemployment rate is stronger for men than for women.
14 This paper focuses on three health behaviors; alcohol consumption, smoking and physical in-activity. Each health behavior will be described in more detail in the following parts. For every health behavior, selected empirical studies will be discussed, further detail can be obtained from table 2. I abstain from other health behaviors, such as diet or calorie intake, obesity and drug abuse. Research in this direction is beyond the scope of this thesis and is left for future research.
Smoking
Smoking is one behavior that relates to health and may be affected by macroeconomic changes. To-bacco consumption is relevant to health policy since globally 1 in 5 people smoke, resulting in negative health outcomes and high costs for healthcare systems and society. Smoking can be addressed because it is a modifiable behavior, with health-compromising effects only. Additionally, a rising number of smokers report their willingness to quit smoking.
On a worldwide scale, smoking was the second leading cause of health burden, contributing to approximately 6.1 million (preliminary) deaths in 2013, marking it a major cause of ill health (Forouzanfar et al., 2015). The WHO estimates that worldwide tobacco can be held accountable for 12% of deaths in adults over 30. On a European scale, cigarette consumption is even more pronounced than at the global level, 16% of deaths are caused by tobacco7. Tobacco does not only lead to
prema-ture deaths but also leads to increased risk for other conditions, for example, cancer, cardiovascular and respiratory diseases (Courtney, 2015). Thus, smoking is a considerable strain on EU’s health care expenditures and the largest avoidable risk for people’s health. Overall, data from the ECHI (European Core Health Indicators) shows a downward sloping trend in the countries of this research, displayed in figure E1. While in 2004 on average 23% of the population older than 15 years were regular daily smokers, ten years later this number has decreased to 19%. I control for this time trend by the using a country-time specific fixed-effect.
Empirical evidence for smoking
Empirical evidence focusing on the relationship between macroeconomic fluctuations and smoking is mixed. Berniell & Bietenbeck (2017) find that smoking varies procyclical with worked hours, utilizing the reduction of working hours to 35 in France, to investigate the effect. In sharp contrast, Kendzor et al.’s (2012) study supports countercyclical variation of smoking by individual- and area-level unem-ployment. The most recent study, Kaiser, Reutter, Sousa-Poza, & Strohmaier (2017) finds mixed evi-dence. Their study indicates countercyclical movement in the tendency to start smoking, however, procyclical variation intensity of consumed cigarettes by smokers.
15
Alcohol consumption
Alcohol consumption is one behavior that relates to health. The WHO distinguishes between different stages of drinking, ranging from mild, moderate, social, heavy drinking, problem drinking, binge drink-ing to heavy episodic drinkdrink-ing. Forouzanfar et al. (2016) find alcohol abuse to be a significant contrib-utor to diseases in their study on the root of the causes of burden of disease. In high-income countries, such as subject of this paper, alcohol consumption is evaluated as the sixth highest cause of poor health and early deaths. The WHO finds alcohol use to be a central risk factor for disease, related to the cause of more than 200 health conditions (Shield, Rylett, & Rehm, 2016). Drinking leads to mainly negative health outcomes, both acute – hypoglycemia, nausea– and chronic –cancer, cirrhosis, cardiovascular disease– for the individual but also for society, through traffic accidents or increased health care ex-penditures (Gutierrez, 2016; Rehm, & Parry, 2009). The social cost of alcohol includes direct costs such as unemployment and loss of productivity and indirect costs, such as higher criminality and social costs. Anderson & Baumberg (2008) estimate the direct cost to amount to €125 billion and indirect costs to amount to €270 billion in 2003, making it an economically relevant factor. These costs increase with the amount consumed. In the 2014 global health report, the WHO finds that 70% of Europeans drink alcoholic beverages and the amount of pure alcohol consumed annually per capita is on average 10.7 liters (WHO, 2014). In contrast to that, the global average of liters of pure alcohol consumed annually per capita is 6.13 liters and 38.3% drink alcohol (WHO, 2014). Over the time period from 2004 to 2014, pure alcohol consumption measured in liters per capita for individuals aged fifteen or higher, varied slightly within the eleven countries analyzed in this paper. Overall, the average long-term trend of alcohol consumption of pure alcohol in liter per capita among individuals age fifteen and older is de-creasing as shown in figure E2. This trend is again captured by the fixed-effect.
Since the mid-1990s, there is an increased interest in the relationship between macroeconomic fluctuations and alcohol consumption. The evidence is extensive, however, mixed with regards to whether the variation is pro- or countercyclical.
Empirical evidence for alcohol consumption
Empirical studies focusing on the effects of macroeconomic conditions on alcohol consumption are extensive. The majority finds procyclical variation, but some research also suggests that variation is countercyclical or that evidence is mixed.
16 (2010) focuses on Sweden for a long time period and uses aggregate level data for health behavior, namely on alcohol sales. Cotti, Dunn, & Tefft (2015a) follow a similar approach, their study is also based on alcohol sales and they conclude as well that variation is procyclical. Their study emphasizes the importance of dynamics and short- versus long-run variation. Kossova, Kossova, & Sheluntcova (2017) show that in Russia the procyclical relationship holds for vodka and beer, but not for wine.
Countercyclical variation is supported by some empirical studies. Dee (2001) finds binge drink-ing to vary countercyclical, risdrink-ing durdrink-ing economic downturns. In line with Dee (2001), the study by Dávalos, Fang, & French (2012) also finds excessive alcohol consumption to increase during economic downturns. An innovative approach was used by Frijters, Johnston, Lordan, and Shields (2013), sug-gesting a countercyclical relationship based on Google searches. An internet search may provide real-time information and solves the measurement error caused by reporting bias in self-evaluations. Data is analyzed for the US on state-level for the years 2004-2011. The authors find that during the reces-sion, having controlled for state and time-effects, a 5% increase in unemployment is followed by roughly 15% rise in alcoholism-related searches over the next year. Results have to be interpreted with caution since this process of data collection is still in its early stages.
Mixed evidence is found by Ettner (1997) and Johansson, Böckerman, Prättälä, & Uutela (2006). Ettner (1997) raised awareness that reverse causality can impact outcomes, as consumption of alcohol can affect employment. Hence, she used a two-stage instrumental variable method to address endogeneity and showed that economic downswings raise overall alcohol consumption but lead to a decrease in dependence for individuals. Johansson et. al (2006) combine aggregate and individual level data for alcohol consumption. Their study suggests countercyclical variations on macro-level, but mi-cro-level data reveals procyclicality; more alcohol is consumed during expansions without having an impact on the probability of being a drinker. Moreover, they find that the greatest recession in their sample constitutes an exception, in the most severe depression drinking decreased.
Alcohol consumption has implications for the health of individuals. Evidence on whether, and via which channels, macroeconomic fluctuations influence drinking, is mixed. Effects change with the data set, time period, geographic region, the specification of the model, and the variables used for macroeconomic fluctuations and alcohol consumption.
Physical Inactivity
17 Multiple studies have proven adverse impacts of inactivity on health. Physical inactivity results in the rising probability to suffer from chronic diseases such as coronary and heart diseases (Warburton et. al, 2010) and type 2 diabetes (Miller, & Dunstan, 2004). The WHO estimates physical inactivity to play a major role for certain diseases, causing over 1 in 5 cases of breast or colon cancer, and more than a quarter of diabetes conditions and around one-third of ischemic heart disease. Häußler (2014) underlines the importance of physical activity, his study shows that the effect of an active lifestyle dominates the effect of obesity on health care cost. The study by Abu-Omar, Rtten, & Robine (2004) underlines this finding, their empirical study also indicates a positive relationship between physical activity and self-rated health. Smith, Ng, and Popkin (2014) link physical inactivity to health outcomes such as cardio-metabolic consequences. They draw as a conclusion that regular or multiple physical activities have health-promoting benefits, more than a single activity has. Humphreys, McLeod, & Rus-eski (2011) confirm these findings. Based on the Canadian Community Health Survey their study high-lights that physical activity reduces the likelihood of suffering from some diseases, among these are heart diseases, high blood pressure, diabetes, asthma, and arthritis. The marginal effect on health is, however, diminishing in terms of frequency and intensity of physical activity above some moderate level.
The level of physical inactivity differs among the European countries. Figure E3 displays the weekly time in minutes people spend on non-work-related, cardiovascular exercise, classified into four groups. Country-fixed effects are used in the model to control for these disparities.
Empirical evidence for physical inactivity
The literature on linking macroeconomic fluctuations to physical inactivity is less extensive than for other health behaviors, findings are mixed.
19
III. Methodology
Chapter 3 presents the methodology, starting with an explanation of the hypotheses of this research, before discussing the theoretical and empirical model, and the dataset, and finishing with a description of the variables.
A. Hypotheses
The research question of this paper is: How do macroeconomic fluctuations influence health behavior? How have macroeconomic fluctuations from 2004 to 2015 influenced alcohol consumption, smoking and physical inactivity in a selection of eleven European countries? Insights from the theoretical con-cepts and previously mentioned research provide the foundation for the hypotheses to be tested:
- Hypothesis 1: Smoking varies countercyclical
Smoking is relatively little time-intense, it does not lead to a lasting intoxication and thus is not strongly impacted by the time constraint. Prices for cigarettes are high, therefore smoking may be impacted via the budget constraint. Some previous studies, following a similar method, have found countercyclical variation, at least for the extensive margin for smoking. Despite the price of cigarettes, I suppose that smoking is a major response to stress. Hence, the stress effect could dominate the income effect. This implies that even though households have a lower disposable income during downswings, they might smoke more to cope with stress caused by the recession. I expect macroeconomic conditions to nega-tively impact smoking behavior.
- Hypothesis 2: Alcohol consumption varies procyclical
Alcohol consumption is relatively little time-intense, as explained earlier. Therefore, the time re-striction is less influential than the budget rere-striction. The majority of empirical studies have found procyclical variation for alcohol consumption. Procyclical variation can be explained by the income- and partially the stress effect. During an upswing, a higher disposable income allows households to spend more money on alcohol. The increasing stress of the upswing intensifies procyclicality when individuals consume more alcohol to cope with the stress. Therefore, I expect macroeconomic condi-tions to positively impact alcohol consumption.
- Hypothesis 3: Physical inactivity varies procyclical
20 and the stress effect. Here, the income effect refers to the opportunity cost of time. During upswings, individuals work more and leisure time becomes relatively costlier. If physical activity helps to reduce the stress experienced due to economic conditions, the psychological stress effect and the income effect work in the same direction. Hence, and in line with results from previous studies, I expect phys-ical inactivity to vary procyclphys-ical.
- Hypothesis 4: Health behaviors before and after the 2007 crisis differ significantly
The great recession of 2007 constitutes a marked macroeconomic change. Therefore, the relation be-tween macroeconomic conditions and health behaviors should become apparent in a direct compari-son of health behaviors before and after the crisis. For alcohol consumption and physical inactivity, I expect macroeconomic conditions to have a positive relation and thus to see a decrease in alcohol consumption and physical inactivity as consequences of the crisis. For smoking, I expect the opposite, a negative effect of macroeconomic conditions on smoking and thus an increase in response to the crisis. I will use a sub-specification of the model to compare health behaviors before and after the crisis.
B. Model
This section is split in two parts. The first part describes in an economic model how macroeconomic conditions can be linked to individual health behavior. Based on this, the second part explains how this translates into an econometric model, the fundament of the analysis to follow.
An economic approach to macroeconomic fluctuations and health behavior
Based on the Grossman model (1972, 2000) described previously, macroeconomic fluctuations and health behavior are linked through the individual production function. Xu (2013) derives that the indi-vidual’s utility U depends on the health state, H, consumption of health commodities, and consump-tion of all other goods, N, and individual characteristics, Z. To capture a possible time effect, health commodities are split into time-intense goods, A, e.g. physical activity, and less time-intense goods, C, e.g. alcohol consumption and smoking. The utility function can be expressed as:
𝑈 = 𝑈(𝐻, 𝐶, 𝐴, 𝑁, 𝑍)
The static model abstracts from dynamic investments in health. This is a serious drawback since individual-specific information on habits or additions are not measured8. The individual produces
health by the health production function. Health behaviors, A and C, are inputs to the individual health production function. Environmental circumstances, E, and time spent working 𝑇𝑤 are inputs as well.
21 Other influences, such as genetic predispositions are captured by ε. The health production function can be expresses as:
𝐻 = 𝐻(𝐶, 𝐴, 𝑇𝑤, 𝑍, 𝐸, 𝜀)
Further, the individual faces a budget and time constraint. The assumption that individuals can maximize their utility by freely altering their working hours does not hold, during recessions, individu-als involuntarily face short time or unemployment. Accordingly, they cannot freely set their working hours in the short run. The time constraint is given by:
𝑇 − 𝑇𝑤− 𝑡𝐶𝐶 − 𝑡𝐴𝐴 − 𝑡𝑁𝑁 = 0
where T is the total time, 𝑇𝑤 is time spent working, and 𝑡𝑖𝑖 is the time spent on consumption of a good with i= C, A, N and by assumption: 𝑡𝐴 > 𝑡𝐶 since A is the time-intense health behavior.
The budget constraint can be described by:
𝑤𝑇𝑤+ 𝑝𝐶𝐶 − 𝑝𝐴𝐴 − 𝑝𝑁𝑁 = 0
where w is the wage rate per hour, and 𝑝𝑖 are the prices of goods. The Lagrange function is thus given by:
𝐿 = 𝑈[𝐻(𝐶, 𝐴, 𝑇𝑤, 𝑍, 𝐸, 𝜀), 𝐶, 𝐴, 𝑁, 𝑍] + 𝜆(𝑇 − 𝑇𝑤− 𝑡𝐶𝐶 − 𝑡𝐴𝐴 − 𝑡𝑁𝑁) + 𝛾 (𝑤𝑇𝑤+ 𝑝𝐶𝐶 − 𝑝𝐴𝐴 − 𝑝𝑁𝑁)
Maximum utility is achieved when individuals choose the optimal combination of goods under the two constraints. Solving for first-order conditions yields the reduced-form demand function:
𝑌𝑖 = 𝐷𝑖 (𝑃𝐶, 𝑃𝐴, 𝑃𝑋, 𝑡𝐶, 𝑡𝐴, 𝑡𝑁, 𝑤, 𝑇𝑊, 𝑍, 𝐸, 𝜀)
Where Y is demand and 𝑖 = 𝐴, 𝐶, 𝑁. The demand for health commodities depends on prices, time, income, environment, personal traits, and the initial health state. Utility is directly affected by health goods and indirectly affected via the health production function, influencing the health status or the marginal cost of consumption. Any health behavior that is positively related to health lowers the mar-ginal cost of the consumption of this good.
22 The econometric model
In order to determine the effect macroeconomic fluctuations have on health behavior, an econometric regression model is estimated, which is presented first. As shown above, the demand for health com-modities depends on prices, time, income, environment, and personal characteristics.
In the regression model, I expect changes in prices of goods to have no or a small influence since the prices of alcohol and cigarettes are sluggish and therefore will not change in the short-run if the economy experiences a downturn. Environmental factors and personal characteristic will change only marginally or stay constant. The two parameters from the demand function that are essential for the empirical specification to determine the effect of macroeconomic variation on health behavior are time and income.
Proxies for macroeconomic conditions can be used to determine the influence of business cycle fluctuations on the individual’s demand for health commodities. Proxies can be on individual level or on aggregate level. On aggregate level, proxies can be the short-term and the long-term unemploy-ment rate, as well as the GDP growth rate. In an economic downswing, unemployunemploy-ment will increase which influences the time the household has available as the relative price of time decreases. Higher unemployment generally results in lower income. When the economy moves towards a recession, ag-gregate demand decreases, companies might lay-off workers or close down altogether, thus the risk to be affected by unemployment rises. Even those individuals who stay employed are affected indi-rectly, they are less likely to receive a bonus, a promotion or achieve a higher income from working extra hours. During a recession, workers cannot work overtime and there may be reduced working hours, leading to a reduction in the relative price of time and to a lower income for households.
23 Following the current literature, the regression models used to identify if changes in macroeco-nomic fluctuations explain changes in health behavior are specified as:
𝐵𝑐𝑗𝑡 = 𝛼𝑗+ 𝛽𝑋𝑐𝑗𝑡+ 𝛾𝑀𝑗𝑡+ 𝜆𝑡+ 𝜁𝑗𝑡+ (𝑢𝑐 + 𝜀𝑗𝑡) (1) 𝐵𝑐𝑗𝑡 = 𝛼𝑗+ 𝛽𝑋𝑐𝑗𝑡+ 𝛾𝑀𝑗𝑡+ 𝜆𝑡+ 𝛿𝐷𝑡+ 𝜇𝐷𝑡∗ 𝑀𝑗𝑡+ 𝜂𝐷𝑡∗ 𝑋𝑐𝑗𝑡+ (𝑢𝑐 + 𝜀𝑗𝑡) (2)
𝐵𝑐𝑗𝑡: health behavior
𝑋𝑐𝑗𝑡: vector of independent variables
𝑀𝑗,𝑡: vector of measures macroeconomic fluctuations 𝛼𝑗 : country fixed-effect
𝜆𝑡 : period fixed-effect
𝜁𝑗𝑡: trend (country-time) fixed-effect 𝑢𝑐: cohort effect
𝜀𝑗𝑡 : regression error term
𝐷𝑡: dummy variable (0: 2007 and before, 1: after 2007) 𝛽, 𝛾 and 𝜂 are vectors of unknown parameters
the subscript c denotes the cohort, j the country and t the time
Equation (1) provides the baseline model, equation (2) provides an extension used to test whether health behavior has changed in European countries after the economic crisis. Observations are spread equally over the groups if the dummy variable takes the value 0, the equation reduces to equation (1). Multiple sub-specifications of the baseline model are tested and explained subsequently.
24
C. Data
This section elaborates on the data used for the empirical analysis. First, more detail regarding the indicators for macroeconomic fluctuations is provided. Second, I elaborate on the SHARE dataset and recodings that were necessary for the analysis. For the main specification of the model, I created co-horts thus the panel was transformed into a pseudo-panel, which will be explained in greater detail, as well as the generalized estimating equation before commenting on stationarity, structural break, and multicollinearity.
This project investigates (1) how changes in macroeconomic conditions relate to changes in health behavior and (2) how health behavior has changed in European countries after the economic crisis of 2007. To determine if macroeconomic fluctuations have a significant effect on health behavior, the model is based on a pseudo-panel9. Data for the empirical analysis is based on SHARE data and proxies
for macroeconomic fluctuations; the unemployment rate, the long-term unemployment rate, and the GDP growth rate obtained from Eurostat.
Proxies for macroeconomic fluctuations
Macroeconomic fluctuations are proxied by aggregate-level and individual-level data. On the aggre-gate level, I use the short-term unemployment rate, the long-term unemployment rate, and the GDP growth rate from Eurostat. Both, the short and long-term unemployment rates are quarterly averages of the percentage of the active population (people aged 15 to 74). GDP is the gross domestic product at market prices (chain-linked volumes), seasonally and calendar adjusted in percentage change of the previous period. Macro-level conditions have a direct influence on individuals, a rising unemployment rate suggests that some of the individuals in the sample will be affected by job loss as well. In addition, there is an indirect effect from the overall macroeconomic vibe that can cause a change in expectations or lead to higher stress. On the individual level, I use the household’s net income, the total hours worked per week in the individual’s main job and the classification of the individual’s main job. Changes in these variables have a more direct effect on the individual.
The SHARE dataset
SHARE is a multidisciplinary and cross-national panel database of micro data on self-reported, individ-ual characteristics. Among these are indicators for health, socio-economic status and social and family networks. The easySHARE dataset provides an unbalanced panel of data of individuals from multiple countries over a time period. SHARE started in 2004 with eleven European countries. Even though it currently covers 27 European countries plus Israel, this paper will focus on the subgroup of eleven countries, for which data is available prior and post the economic crisis. From 2004 to 2015, data is
25 available for Austria, Belgium, Switzerland, Germany, Denmark, Spain, France, Greece, Italy, the Neth-erlands, and Sweden. The SHARE survey contains data of individuals aged 50 or older, discussed later as one of the limitations of the paper.
The total number of observations in the data set is 288736 however, for my purposes, I dropped observations on countries that are not part of my sample (80631 observations), following common practice (Dávalos, Fang, & French, 2012). I dropped wave three because it does not include the variables of interest (23863), I dropped individuals who are retired (94626) or provided no infor-mation on their employment status (2438) and those observations that did not provide a response to the health behavior questions (22827). Individuals who are retired are not of interest for my research question, for them the income effect and the stress effect do not work to the same extent as for the active population. Retirees have more time and a compared to the active part, a relatively fixed in-come, in form of a pension. I assume that missing values are randomly distributed. Moreover, I drop observations from all individuals who were not aged 14-74 in the respective wave (3733), since this matches the active population. The adjusted dataset results in a large sample of 51773 observations from 31467 individuals, spread among countries and waves. Overall, the number of observations per country and wave is sufficiently large.
Pseudo-panel data
This research is based on panel data, which allows to measure effects of explanatory variables over time. In this dataset, observations on individual level are collected over multiple time periods, however, the panel is unbalanced. This implies that the panel in incomplete in the sense that not all individuals responded to every wave. For survey data this is not uncommon, reasons can be, among others, attrition, nonresponse or late entry. On average, each individual is observed in 1.6 time peri-ods, which allows only very limited conclusions about changes in behavior. Since the research objective is to investigate how changes in macroeconomic conditions affect health behavior, it is necessary to observe individuals over multiple time periods. From 1.6 time periods, one could not detect sufficient macroeconomic fluctuations. To address this shortcoming, I create cohorts based on the individual level data. Using cohorts for pseudo-panel analysis dates back to Deaton (1985). As common practice, cohorts are created based on individual characteristics that do not change over time, in this case, the year of birth and the country of residence.10 Observing cohorts instead of individuals thus allows for
drawing more conclusions, since every cohort is observed in multiple time periods, time availability is higher and attrition lower (Verbeek, 2008; Guillerm, 2017). Furthermore, cohort analysis reduces the bias that occurs from participating in a survey. Zwane et al. (2011) showed that taking part in a survey could skew results as it may lead to changes in behavior because of higher awareness. This holds es-pecially for health-related research, participants of a study are for example more likely to purchase
26 health insurance. Creating cohorts reduces this effects because mean over groups of individuals are calculated (Guillerm, 2017).
Creating cohorts implies facing a bias‑variance tradeoff. Large cohorts have a smaller sampling bias11, however, this implies that the total number of cohorts is smaller and thus the variance will be
higher and precision is lost (Verbeek and Nijman, 1992). The average cohort size in my data set is 62.5, the total number of cohorts is 503 with, on average, 11.3 observations per cohort. Hence, the panel is still unbalanced. By creating cohorts, some detail is lost, however, on the other side, cohorts are now present over more time periods and allow for stronger conclusions. This trade-off is typical for individ-ual panel data since attrition rates are high (Guillerm, 2017). In an alternative specification, I estimated the model on individual level, results are presented in table D8 of the appendix.
Generalized estimating equation: c ohort analysis, unbalanced pseudo-panel, frac-tional logistic model, two-way-fixed effects
Creating cohorts leads to a transformation of the dependent variables from binary or ordinally scaled variables to fractionally scaled variables. On the cohort level, smoking, alcohol consumption and phys-ical inactivity indicate the likelihood of the cohort to display the health behavior. Since the dependent variables are fractional by creating cohorts, neither the linear fixed-effects nor the logistic conditional maximum likelihood estimation no longer yield consistent estimates. Hence, a fractional logistic model should be used (Fitzmaurice, Laird, & Ware, 2011). I use a Generalized Estimating Equations approach (GEE) to fit a population-average model and control for time, country and country-time fixed-effects. Weights adjust for different cohort sizes, however, adjustment is imperfect such that results have to be interpreted with caution.
The GLM is a generalization of the ordinary least square (OLS) model that allows for a non-normal error distribution. The generalized estimating equations (GEE) method, proposed by Liang and Zeger (1986) is an extension of the generalized linear model (GLM) based on a quasi-likelihood ap-proach. The GEE approach extends the GLM by allowing the dependent variable to be exponentially distributed (Horton & Lipsit, 1999). The distribution has to be specified from the family of exponential distributions, as well as the specification of the link is required. In the model at hand, all dependent variables are fractional, therefore a binomial logistic regression is used by choosing a binomial distri-bution with logic link. This ensures that the estimated coefficients are within the zero to one range.
The goal of estimating GEE is to obtain estimates and inferences about the population, taking into account the within-subject correlation. The estimates and inferences obtained by GEE are asymp-totically correct. Since GEE is a population-averaged approach, the marginal effects are averaged across individuals sharing the same predictors across all panels. The specification of the correlation
11 Verbeek, & Nijman (1992) suggests to create cohorts of 100-200 individuals to significantly reduce the risk of bias.
27 matrix accounts for potential correlation in the data. Therefore, GEE is suitable for longitudinal or clus-tered data. Even if the specification of the correlation matrix was wrong, the estimates still converge to the true value when N approaches infinity. In addition, GEE can handle unbalanced panels and miss-ing data as long as these are missmiss-ing at random (MAR).
The GEE is semiparametric, this implies that it combines both, parametric and nonparametric components. A parametric component is finite dimensional vector, for example, the distribution of the error, where the density function is known but includes a nuisance parameter. This nuisance parame-ter will converge to the true parameparame-ter as N approaches infinity. The nonparametric component is infinite-dimensional, the properties depend upon the structural function and can be from an unknown distribution. In a semiparametric model, the component of interest is parametric. Nonparametric com-ponents can be added, but are usually not of core interest. The GEE relies on the first two moments, mean and variance. The structure of the variance can be specified using the quasi-likelihood infor-mation criterion (QIC) (Cui, 2007). For the dataset of this research, the QIC is lowest for an independent correlation structure.
GEE is often used in the context of cohort studies since the structure of the variance allows for unmeasured dependence and it estimates the population-average effect. In addition, GEE is unrestricting regarding the distribution of residuals than mixed models requiring normality. GEE re-quires a specification of the working correlation matrix. If correctly specified, GEE will be the most efficient estimator. The robust, or sandwich variance estimator can be included so that the model pro-vides consistent estimates of the variance of the unknown parameters even if the working correlation matrix is not equal to the true variance. This allows controlling for bias arising from possible heteroske-dasticity and autocorrelation by using clustered robust errors (Kauermann & Carroll 1999). Kelly Ven-ezuela, Botter & Sandoval (2007) provide proof that if the sandwich variance estimator is used even though the working correlation matrix was correctly specified, the estimator will reduce to the model-based variance estimator. Hence, I include the sandwich variance estimator in the model to control for possible heteroskedasticity and autocorrelation.
Stationarity
The stationarity assumption states that the data sample is drawn from a stationary process implying that the mean, the variance, and the autocorrelation structure are time-independent. The stationarity assumption addresses the time dimension of the panel and is essential for the reliability of the estima-tors as 𝑇 → ∞. The panel at hand is unbalanced and 𝑇 < 𝑁 such that 𝑇/𝑁 → 0. Hence, the asymp-totic convergence is not achieved, a unit-root test has very low power, and test results would not be sufficiently reliable. Therefore, I do not test for stationarity, I partly account for possible non-station-arity by using time- and trend fixed-effects in the model12 (Baltagi, 2005).
28 No structural break
The data should not encounter an unexpected shift in the coefficients also referred to as structural break. In presence of a structural break, estimates are less reliable. In this research, a structural break may occur due to the 2007 crisis. The Great Recession constituted a shock of considerable magnitude. Therefore, the Chow test can be employed to assess whether the coefficients in the model are constant over all periods. Splitting the dataset into two groups, categorized by time (2004-2007 and 2008-2015), allows testing the null hypothesis of no structural stability. For smoking, the null hypothesis is rejected at the 5% significance level, for alcohol consumption, one fails to reject the null hypothesis and for physical activity, the null hypothesis is rejected at the 1% significance level. Details are presented in table B3 of the appendix. Since there is no statistically significant structural break for all variables, I decide against subdividing the dataset. A possible extension could be to spilt the dataset to compare if coefficients change.
No multicollinearity
29
D. Variables
This section describes the variables in more detail. First, all dependent variables are explained before providing more information on the explanatory variables and the control variables. Finally, I evaluate the specification of the model. Summary statistics are presented in table A3, more detail on the varia-bles and sources can be obtained from table A2.
Dependent variables
The vector of dependent variables includes variables from the module “Behavioural Risks (BR)” for alcohol consumption, smoking, and physical inactivity.
Smoking is a binary response variable, where 0 indicates that the individual is not smoking at the present time and 1 indicates the individual smokes at the present time. Since the data is averaged to cohort level, smoking becomes a fractional variable.13
Alcohol consumption14 is presented on an ordinal scale, indicating how many days per week
the respondent consumed alcohol last 6 months; ranging from “not at all”, “less than once a month”, “once or twice a month”, “once or twice a week”, “three or four days a week”, “five or six days a week” to “almost every day”. For the empirical analysis, the variable is coded into a continuous one, where for example “alcohol consumption 5 or 6 times per week” is coded to “[(5 + 6)/2]/7 = 0,7857”. The continuous variable for alcohol consumption is fractional, it can take any value from, and including, zero to one.
The variable for physical inactivity is also ordinally scaled, indicating how often individuals do sports or activities that are vigorous or a job that involves physical labor, ranging from “more than once per week”, “once per week”, “one to three times a month” to “hardly ever, or never”. By com-puting a continuous variable, the output shows the likelihood to be physically active per day. The con-tinuous variable for physical inactivity is also fractional, it can take any value from, and including, zero to one.
Explanatory variables
The vector of independent explanatory variables includes proxies for macroeconomic fluctuations which can be divided into two groups; first, proxies based on aggregate-level data and second, proxies based on individual-level data. The first group contains three aggregate-level proxies, namely the short-term unemployment rate, the long-term unemployment rate and the GDP growth rate15. All
three variables are rounded to one decimal place. The second group within the independent variables contains individual-level data from the SHARE survey from the section ‘work and money’ and makes
13 Coding it back to a binary variables loses much detail, results are presented in table D9 in the appendix 14 “During the last six months, how often have you drunk any alcoholic beverages, like beer, cider, wine, spirits
or cocktails?“
30 use of changes in employment status, total hours worked per week, and the imputed household net income. The total hours worked per week refer to the individual’s main job. The variable for the house-hold’s net income is based on the income from the previous year in the local currency, which is Euro for all countries in my sample with the exception of Switzerland. The variable for the weekly worked hours contains 4990 observations, there is a substantial number of missing observations. These indi-viduals were filtered because they had selected earlier in the questionnaire that they do not work, either because they are unemployed, permanently sick or disabled, or homemaker. For that reason, their total hours worked per week contain missing values16.
I used variables on the employment status to create a dummy variable that distinguishes be-tween having a job and not having a job. This dummy variable takes on the value 1 for “having a job” if the individual participates in the labor market, including all individuals who are employed, civil serv-ants, self-employed or other. The dummy variable takes the value 0 for “not having a job” if the indi-vidual does not participate in the labor market, including those who are permanently sick or disabled, unemployed or homemaker. The reasons not to participate in the labor market differ, homemaker are a country’s hidden reserve and could be available to the labor market. Their decision to have no job is mostly a voluntary choice. Individuals, who are unemployed are involuntarily out of the labor market but available, while individuals who are permanently sick or disabled are involuntary out of the labor market and not available to the labor market. Creating a dummy variable leads to a loss of detail, nevertheless, since the scale of employment is not ordinal, creating a dummy allows to draw conclu-sions from the employment level on a higher level. For the purpose of this analysis, I accept to lose some degree of detail to, in turn, be able to measure the effect of having a job on health behavior. Control variables
The vector of independent control variables consists of information on demographics (gender, age, household size, years of education), and health and health behavior (Body mass index (bmi), and whether the individual has trouble sleeping). Gender is a dummy variable where 0 indicates male and 1 indicates female. The bmi is the individual’s weight divided by the square of her height, rounded to seven digits. The bmi is an attempt to generalize information on body mass to be able to draw conclu-sions if the individual is over- or underweight. Yet, since the bmi is a generalized concept, it does not accurately classify all individuals correctly but can help to see general trends. The dummy variable sleep indicates whether an individual has trouble sleeping, where 0 means having no trouble sleeping and 1 indicates having trouble to sleep. I chose sleep as a control variable since trouble sleeping is often associated with stress (Akay, Martinsson, & Ralsmark, 2017) and stress is one of the channels how macroeconomic conditions influence health behavior. The other variables are self-explanatory, more information can be obtained from table A1 in the appendix.
