Biased health perceptions and risky health behaviors: Theory and evidence

Arni, Patrick; Dragone, Davide; Götte, Lorenz; Ziebarth, Nicolas R. Working Paper Biased health perceptions and risky health behaviors: Theory and evidence Quaderni - Working Paper DSE, No. 1146 Provided in Cooperation with: University of Bologna, Department of Economics Suggested Citation: Arni, Patrick; Dragone, Davide; Götte, Lorenz; Ziebarth, Nicolas R. (2020) : Biased health perceptions and risky health behaviors: Theory and evidence, Quaderni - Working Paper DSE, No. 1146, Alma Mater Studiorum - Università di Bologna, Dipartimento di Scienze Economiche (DSE), Bologna, https://doi.org/10.6092/unibo/amsacta/6374 This Version is available at: https://hdl.handle.net/10419/245888 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by-nc/3.0/ ISSN 2282-6483! ! ! ! ! ! Biased Health Perceptions and Risky Health Behaviors: Theory and Evidence Patrick Arni Davide Dragone Lorenz Goette Nicolas R. Ziebarth ! Quaderni - Working Paper DSE N°1146 ! ! Biased Health Perceptions and Risky Health Behaviors—Theory and Evidence Patrick Arni University of Bristol Davide Dragone University of Bologna Lorenz Goette University of Bonn Nicolas R. Ziebarth Cornell University ∗ April 23, 2020 Abstract This paper investigates the role of biased health perceptions as driving forces of risky health behavior. We define absolute and relative health perception biases, illustrate their measurement in surveys and provide evidence on their relevance. Next, we decompose the theoretical effect into its extensive and intensive margin: When the extensive margin dominates, people (wrongly) believe they are healthy enough to “afford” unhealthy behavior. Finally, using three population surveys, we provide robust empirical evidence that respondents who overestimate their health are less likely to exercise and sleep enough, but more likely to eat unhealthily and drink alcohol daily. Keywords: health bias, health perceptions, subjective beliefs, overconfidence, underconfidence, overoptimism, risky behavior, smoking, obesity, exercising, SF12, SAH, BASE-II JEL classification: C93, D03, D83, I12 ‡We would like to thank Teresa Bago d’Uva, Michele Belot, Kitt Carpenter, John Cawley, Davide Cesarini, Resul Cesur, Owen O’Donnell, Fabrice Etil´ e, Osea Giuntella, Glenn Harrison, Ben Hansen, Hendrik J¨ uerges, Jonathan Ketcham, Nadine Ketel, Gaurav Khanna, Audrey Laporte, Fabian Lange, Taryn Morrissey, Robert Nuscheler, Reto Odermatt, Ricardo Perez-Truglia, Gregor Pfeifer, Pia Pinger, Aldo Rustichini, Joe Sabia, Luis Santos Pinto, Tom Siedler, Rodrigo Soares, Pascal St-Amour, Alois Stutzer, Justin Sydnor, Harald Tauchmann, Erdal Tekin, Christian Traxler, Gerard van den Berg, Ben Vollaard, Justin White, V´ era Zabrodina. In particular we thank our discussants Matt Harris and Nathan Kettlewell for excellent comments and suggestions. Moreover, we thank conference participants at the the 2019 iHEA World Congress in Basel, the 2019 Workshop on the Economics of Risky Behavior in Bologna, the 2018 ASHEcon meetings in Atlanta, 2017 Bristol Workshop on Economic Policy Interventions and Behaviour, the 2017 Risky Health Behaviors Workshop in Hamburg, the 2014 iHEA/ECHE conference in Dublin as well as seminar participants at the University of Basel, University of Reno, the Berlin Network of Labor Market Research (BeNA), and The Institute on Health Economics, Health Behaviors and Disparities at Cornell University. We take responsibility for all remaining errors in and shortcomings of the article. This article uses data from the Innovation Panel of the Socio-Economic Panel Study (SOEP-IP) as well as the Berlin Aging Study II (BASE-II). BASE-II has been supported by the German Federal Ministry of Education and Research and Research under grant numbers 16SV5537/16SV55837/16SV5538/16SV5536K/01UW0808. We also would like to thank Peter Eibich and Katrin Schaar for excellent support with the BASE-II data, David Richter for his excellent support with the SOEP-IP data as well as Gert Wagner for his overall support of this research and with the BASE-II and SOEP-IP data. The research reported in this paper is not the result of a for-pay consulting relationship. The responsibility for the contents of this publication lies with its authors. Our employers do not have a financial interest in the topic of the paper that might constitute a conflict of interest. The project has been exempt from Cornell IRB review under ID # 1309004122. ∗Corresponding author: Cornell University, Department of Policy Analysis and Management (PAM), 426 Kennedy Hall, Ithaca, NY 14850, USA, phone: +1-(607)255-1180, e-mail: [email protected] Biased Health Perceptions and Risky Health Behaviors: Theory and Evidence by Patrick Arni, Davide Dragone, Lorenz Goette and Nicolas R. Ziebarth NON-TECHNICAL SUMMARY Can risky health behavior be “optimal”? Does overconfidence or optimism about our own health affect our decision to smoke, eat, or exercise? Do people behave differently when they think they are healthier than they actually are? Perhaps surprisingly, these questions have not been investigated thoroughly in the health economics literature. The economics literature has defined and measured “overconfidence” in many different ways. A common definition considers overconfidence as an overestimation of own performance or characteristics. Accordingly, it means that a person believes, for instance, to be smarter, more skilled, or more competent than she actually is. An alternative but related definition considers overconfidence as an individual assessment relative to other people. Accordingly, it means that a person overestimates her own ability, skills, or performance relative to a reference group. A third definition considers overconfidence in terms of the accuracy or precision of information. In the context of health, the first notion implies that a person believes to be healthier than she actually is; the second implies that a person believes to be healthier than her reference group, and the third implies that a person is overly certain that her perceived health is the true health. In this paper, we investigate theoretically and empirically the first and second notions, which we define as biases in absolute and relative health perceptions. We show that the two definitions are closely related and that one maps into the other under plausible assumptions. To measure absolute health perception biases , we use clinically diagnosed health conditions about high cholesterol and high blood pressure and how survey respondents’ perceptions about having such a condition deviate from the facts. Using a representative German survey, we find that 30% of the population has biased perceptions about their high cholesterol levels. To measure relative health perception biases, specifically for this research, we asked respondents to rank their health status relative to a reference group. Using two different high-quality surveys from Germany, we find that many respondents believe that, out of 100 people, there are (only) between 10 and 30 people in better health, a finding that is mathematically impos1 sible. We find that about a third of respondents believe that there are only 30 people in better health than them, whereas in reality there are 60 people. Next, we show theoretically that health perception biases affect health-related behavior through two possible channels. These operate in opposite directions. The first channel emphasizes the role of perceived health; the second channel emphasizes the role of the perceived costs of unhealthy behavior. The former implies that the more a person (wrongly) believes to be healthy, the more she engages in unhealthy behavior because she thinks she can “afford ”it. The latter, instead, implies an overestimation of the costs of unhealthy behavior. Accordingly, the higher the perception of her health, the less willing a person is to jeopardize her health; and therefore she engages in less unhealthy behavior. Which of the two channels matters more is an empirical question, which we tackle using three representative surveys. We find that individuals who overestimate their health are significantly more likely to not exercise, to eat unhealthy, to be overweight, and to sleep fewer hours. The statistical relationships are robust to correcting for socio-demographics, personality traits, cognitive skills, and risk aversion. They are also robust across our two notions of overconfidence. 2 1 Introduction Can risky health behavior be “optimal”? While most non-economists would immediately refute this thought, economists would probably answer “it depends” and refer to one of their workhorse models, the Rational Addiction Model (Becker and Murphy,1988). Most economists fundamentally believe that rational forward looking agents would correctly assess the costs and benefits of their actions and then maximize their utility by consuming the optimal amount, such that the marginal costs equal the marginal benefits. Behavioral economists have challenged this traditional view by empirically and theoretically studying behavior that deviates from rational forward-looking decision-making (cf. Rabin,2013). In health economics, influential studies have shown that consumers pick dominated health plans and leave money on the table (e.g. Ketcham et al.,2012,2016;Abaluck and Gruber,2011,2016;Bhargava et al.,2017). Handel and Kolstad (2015) show that consumers lack insurance-related information and incorporate such frictions into welfare models. Other papers have cited behavioral phenomena as explanations for why people engage in “too much” risky health behavior (c.f. Cawley and Ruhm,2011). For example, theoretical papers model the role of hyperbolic discounting and time inconsistencies for smoking and overeating (Gruber and K˝ oszegi,2001;Cutler et al.,2003;Strulik and Trimborn,2018). In a seminal empirical paper, Della Vigna and Malmendier (2006) show that many gym members actually overpay as they exercise too little. They propose overconfidence about future self-control as one possible explanation for such observed behavior. To evaluate tools that nudge people to exercise regularly, follow-up studies have implemented field experiments. However, nudges appear to only be effective in the short-term (Royer et al.,2015;Carrera et al.,2018,2020).1This paper investigates the relationship between individual risky health behavior and health perception biases, about which only a small literature exists.2 As a first contribution, we formally introduce the concept of health perception biases into the (health) economics literature and document the existence of health perception biases in the population using three representative surveys. Specifically, we propose two individual mea1In one of the few lab experiments on this topic, Belot (2017) randomly updates students’ beliefs about their likelihood to contract diseases in the next 10 years and then study the impact on dietary choices. 2A notable exception is Harris (2017) who finds that people who overestimate their activity levels consume more calories. In earlier studies, Cawley and Philipson (1999) use life insurance data and the differential between perceived and predicted mortality risk to show that observed empirical patterns are inconsistent with standard insurance theory under perfect information. A possible explanation is that insurance mandates make low risk individuals worse off if many market participants are overconfident (Sandroni and Squintani,2007). 3 sures, absolute and relative health perception biases, whose empirical measurement we borrow from the growing literature on overconfidence (c.f. Tiefenbeck et al.,2018;Friehe and Pannenberg,2019). Absolute perception biases are biased perceptions of own health, whereas relative perception biases are biased perceptions relative to population health. We show that, under plausible assumptions, there exists a one-to-one positive mapping between the two health perception measures. To measure absolute health perception biases, we use objectively diagnosed health conditions (here: high cholesterol and high blood pressure) and how respondents’ perceptions about having such a condition deviate from the facts. In a representative German survey, we find that 30% of the population have biased perceptions about their high cholesterol levels. To measure relative health perception biases, specifically for this paper, we asked respondents to rank their health status relative to a reference group. The literature on overconfidence has used very similar survey measures. For the purpose of this paper, we included a health variant of this measures in two high-quality surveys from Germany (one representative survey and one interdisciplinary survey with rich measures on cognitive skills and risk aversion). Comparing the elicited subjective ranking to the objective ranking within the population health distribution (as measured by the SF12), produces our measure of relative health perception biases. Consistent with the absolute measure, we find that about 30% of the respondents overestimate their rank in the population health distribution by at least 30 ranks. That is, for instance, they believe that they rank at the 70th percentile when they actually only rank at the 40th percentile of the population health distribution. In particular, we find an excess mass of people who believe that they rank between the 70th and the 90th percentile of the population health distribution. We acknowledge that our definition of health perception biases may have different meanings and implications in different fields of the social sciences. Moreover, different fields— even subfields within the same discipline—have different empirical measures for phenomena such as overconfidence or overoptimism (cf. Camerer and Lovallo,1999;Barber and Odean, 2001;Benoˆ ıt and Dubra,2011;Merkle and Weber,2011;Burks et al.,2013;Spinnewijn,2015; Bago d’Uva et al.,2017;Cowan,2018). For example, Ortoleva and Snowberg (2015a) define overconfidence as the difference between respondents’ estimated inflation and unemployment rates and the true rates, net of respondents’ self-reported confidence about their estimates. Outside of economics, separate literatures on overoptimism biases (e.g. Weinstein,1980,1989; Sharot,2011) and self-esteem (e.g. Himmler and Koenig,2012) exist. In the literature on exper4 imental methodology, Harrison et al. (2015); Di Girolamo et al. (2015) and Harrison et al. (2017) show how to correctly elicit subjective beliefs in lab experiments. Regarding our definition of absolute health perception biases, there exists a related health economic literature on “reporting heterogeneity” or response errors in subjective health measures (Lindeboom and van Doorslaer,2004;J¨ urges,2007,2008;Bago d’Uva et al.,2008;Ziebarth, 2010) as well as objective health measures (Baker et al.,2004;Davillas and Pudney,2017;Choi and Cawley,2018). Moreover, absolute health perception biases could be interpreted as “unawareness.” Empirically, in the field, it is very challenging—if not impossible—to distinguish biased beliefs from reporting errors or unawareness. We entirely agree that such alternative labels and interpretations are feasible. As a consequence, we define perception biases broadly and deliberately allow for unawareness about one’s health condition. Precisely for that reason, we elicit and compare measures of absolute and relative perception biases. Similarly, to bound the impact of reporting errors in subjective health, we benchmark our relative health perception measure against both the standard Self Assessed Health (SAH) measure and the 12-Item Short Form (SF12) health survey measure. By construction, the latter includes fewer systematic response biases. The empirical pattern and population distributions of two absolute and two relative bias measures across three different surveys yield important insights into the robustness and prevalence of such biases. As a second main contribution, we provide a theoretical framework that shows how biased health perceptions can affect risky health behavior. The framework is simple and flexible enough to explain nonlinear patterns between biased beliefs and risky behavior. Moreover, it highlights that biased health perceptions affect behavior through an extensive and an intensive margin, and that these margins operate in opposite directions. When the extensive margin dominates, risky health behavior and biased health perceptions are complements. This means that, the higher the perception of own health, the more an individual engages in unhealthy behavior, such as consuming fast food or not exercising. This is akin to saying: “Because I believe I am very healthy and can afford it, I eat more fast food and exercise less.” On the contrary, when the intensive margin of perceived health dominates, risky behavior and health biases are substitutes, hence a higher perceived health reduces risky behavior. This is akin to saying: “Because I perceive large health costs of risky behavior, I will eat less fast food and exercise more.” Whether the extensive or intensive margin dominates is an empirical question which we tackle using three representative surveys. 5 As a third contribution, we document a robust statistical link between health perception biases and risky health behaviors across all three surveys. Specifically, individuals who overestimate their health are significantly more likely to not exercise, to eat unhealthy, to be overweight and to sleep fewer hours. The statistical relationships are robust to controlling for sociodemographics, personality traits, cognitive skills, and risk aversion. They are also robust across our two notions of overconfidence. In the context of our model, these results are consistent with a dominant role of the extensive margin of health perception biases. Conversely, we do not find significant relationships for unbiased respondents and for those who are pessimistic about their health, a finding that is consistent with the extensive and intensive margin of health perception bias offsetting each other.3 Notably, smoking is not correlated with biased health perceptions. This result is consistent with Darden (2017), who finds that cardiovascular biomarker information at repeated health exams does not significantly alter smoking behavior. A possible explanation is that signals and information about own health, be it objective as in Darden (2017) or perceived as in this paper, are not powerful drivers of smoking behavior, possibly because its addictive nature prevents the proper evaluation of the health consequences of smoking. Finally, we would like to point out that this paper remains agnostic about the sources of biased beliefs. The origins of these biased beliefs are still poorly understood. They have been linked to image motivation (B´ enabou and Tirole,2002) and humans’ desire of being perceived positively by others (Burks et al.,2013;Goette et al.,2015;Charness et al.,2018), as opposed to managing a favorable self-image (Santos-Pinto and Sobel,2005;K˝ oszegi,2006;Weinberg, 2006) or self-serving biases (Babcock and Loewenstein,1997;Di Tella et al.,2015). Explanations based on models with rational agents have been proposed (Benoˆ ıt and Dubra,2011), but have not always been corroborated by the empirical evidence (Merkle and Weber,2011;Burks et al., 2013). 2 Defining Health Biases In this section, we define absolute and relative health biases. In the empirical section, we will investigate those biases. Our definitions are consistent with the definition of overconfidence in 3As it is common in this literature, our results cannot exclude that health perception biases and risky health behaviors are linked via unobservables, for instance genes (Linn´ er et al.,2019), or that the causality runs from risky behaviors to perception biases in the form of self-serving biases (B´ enabou and Tirole,2002). 6 were administered by psychologists but which are not the focus of this paper.5) The BASEII is representative of the elderly Berlin population up to age 89. As a supplement, BASEII also surveys a sample of younger Berlin residents aged 18 and above; the ratio between respondents above and below 60 is 3:1 (see Appendix, Table A2). Bertram et al. (2014) provide more information on the BASE-II. Our working sample consists of 1,780 respondents without missings on relevant variables. For this paper, we included a measure to elicit e riin the Socio-Economic Module of BASE-II, which was in the field between September and December 2012. Section 5.2 discusses the health bias measures in detail, also see Panel A of Table A2. Health Behavior. Panel B of Table A2 lists measures of risky health behavior: smoker,no sports,unhealthy diet,obese and BMI. As seen, 12% smoke, 36% do not exercise, 39% have an unhealthy diet, and 13% are obese. Socio-Demographics. Panel C of Table A2 lists socio-demographic variables. There are five main categories: (i) Demographics, (ii) Education, (iii) Employment, (iv) Behavioral Attitudes, and (v) Big Five. The average age of BASE-II respondents is 60; slightly more than half of them are female and married; a quarter are single. Fifty-six percent of the sample finished high school (13 school years) and almost half are still full-time employed. In our regression models, we also control for risk aversion and trust, which are important covariates when eliciting subjective beliefs (e.g. Harrison et al.,2015,2017). Accordingly, 15% of BASE-II respondents are risk loving (highest three categories of the standard 0 to 10 Likert risk aversion scale, see Dohmen et al., 2011). Almost forty percent say that they have “a lot of” or “quite some” trust in strangers. Finally, we also control for the Big-Five. The five dimensions are simple averages over three or four subscales which range from one to seven (Richter et al.,2013). Conscientiousness has the highest average of 5.6 and Agreeableness the lowest with 3.8 (Panel C, Table A2). 4.3 Socio-Economic Panel Study – Innovation Panel (SOEP-IP) Since 2012, the SOEP has been inviting researchers to submit proposals for innovative survey questions (Richter and Schupp,2015). Proposals are then reviewed by an expert committee. If accepted, the proposed questions become part of SOEP-IP, which is in the field annually from September to December. SOEP-IP respondents also answer the regular SOEP core questions 5Our findings are robust to controlling for cognitive measures. Detailed results are available upon request. 13 (Richter and Schupp,2017). In 2014, a total of 1,377 respondents answered the same health perception measure, e r, that we also included in BASE-II. Comparing relative health perceptions to true health allows us to construct relative health perception bias measures Ri(see Panel A of Table A3 and Section 5.2). Health Behavior. In 2014, the SOEP-IP did not ask about smoking, exercising, and respondents’ diet. However, the SOEP-IP elicited the average hours of sleep (Richter and Schupp, 2015). On average, Germans sleep 6.8 hours during the week and 7.6 hours on weekends (Panel B of Table A3). We use these information to generate sleep gap measures that indicate the difference to eight hours of sleep. Socio-Demographics. As above, Panel C of Table A3 lists socio-demographics. Because BASE-II and SOEP-IP both include SOEP’s socio-demographic core questions, we generate almost identical socio-demographic control variables. By design, representative SOEP-IP respondents are younger (51 vs. 60 years) but the shares of female and married respondents are very similar, slightly above fifty percent. In SOEP-IP, two thirds are full-time employed and the average monthly net income is e1,773. 5 Measuring Health Perception Biases This section shows how we operationalize health perception biases using survey data. 5.1 Measuring Absolute Health Perception Biases To measure absolute health perception biases, we use the German National Health Survey EastWest 1991 (GNHSEW91). This dataset contains information on individual blood pressure and cholesterol levels, collected by the Institute for Prevention and Public Health in Berlin, Germany. We use these measures as proxies for the individual objective health status Hi. We dichotomize these continuous objective measures depending on whether they are above or below the medically defined threshold to indicate specific clinical conditions. Specifically, for blood pressure, we define that a respondent has high blood pressure, and BPiequals one, if the systolic value is larger than 160 mmHg and/or the diastolic value exceeds 95 mmHg; BPiequals zero otherwise. Analogously, for high cholesterol levels, we define a dummy Choliequal to one for values larger than 6.2 mmol/l, and zero otherwise. Because GNHSEW91 also elicits 14 perceptions about these conditions, we then compare these objective clinical outcomes BPiand Choliwith respondents’ perceived high blood pressure, g BPi, and high cholesterol levels, g Choli, to obtain an assessment of absolute health overconfidence. Each surveyed individual provides self-assessed binary measures of g BPiand g Choli.6 Panel B of Table A1 (Appendix) shows a mean cholesterol level of 6.1 millimole per liter (mmol/l) and that 44% of all Germans have high cholesterol levels (BPi=1). Panel B of Table A1 also shows mean systolic blood pressure levels of 135 millimetres of mercury (mmHg) and mean diastolic blood pressure levels of 83 mmHg.7Following the official WHO definition at the time, 21% of Germans had hypertension (Choli=1).8Panel C shows that 21% of respondents knew that they suffered of hypertension (g BPi=1) and that 25% knew that they had high cholesterol ( g Choli=1). [Insert Figure 1about here] As good health corresponds to BPi=0 or Choli=0, absolute overconfidence results if g BPi<BPior g Choli<Choli. Figure 1a illustrates the four possible combinations of the binary measures of objective and perceived health for cholesterol. Individuals in the bottom-right corner display absolute overconfidence. This corresponds to 30% of respondent who actually have high cholesterol levels, but are not aware of it. The bottom-left corner shows that 50% of all respondents do not have high cholesterol levels, and consistently report that they do not have high cholesterol. As shown in the top-right corner, 14% correctly state that they have high blood cholesterol levels.9 Figure 1b has the same setup and shows the analogous distribution for high blood pressure. Accordingly, 66% of all respondents correctly state that they do not have high blood pressure (bottom-left corner), and 12% correctly state that they do have high blood pressure (top-right corner). Nine percent indicate that they never had high blood pressure although the clini6Note that there is a small literature on misreporting of clinical diagnoses (Baker et al.,2004;Davillas and Pudney,2017;Choi and Cawley,2018). It is distinct from the literature on overconfidence. It is certainly up to scientific debate on how to define this phenomenon. People are either unaware and have biased health perceptions (our interpretation) or they are aware of their health condition but deliberately misreport it, for example, due to a desirability bias. 7Each measure was taken three times from each respondent; we use data from the second measure. 8In the meantime, the official definitions have been downgraded. In November 2017, the American Heart Association and the American College of Cardiology redefined the thresholds to 130/80 (American Heart Association, 2017). 9We ignore the 6.5% in the top-left corner who had no high cholesterol at the time of the survey, but who claim that they had been diagnosed before (as the statement may or may not be true). 15 cal measures show the opposite (bottom-right corner).10 The smaller perception bias for high blood pressure as compared to high cholesterol is consistent with the notion that it is easier to check for high blood pressure than high cholesterol levels outside of clinical settings. 5.2 Measuring Relative Health Perception Biases Ri To measure relative health perception biases Ri=e ri−ri, we make use of the BASE-II and the SOEP-IP. To measure Hi(which is needed to construct ri), both the BASE-II and the SOEP-IP contain the standard SAH measure as well as the SF12 measure. Both measures have been routinely used by health economists and public health scientists. SAH asks about the overall health status; respondents self-categorize as being in excellent, very good, good, fair, or poor health. However, although widely available and easy to collect, the literature has documented systematic SAH response biases with respect to age and gender (Lindeboom and van Doorslaer,2004; J¨ urges,2008;Bago d’Uva et al.,2008;Ziebarth,2010;Spitzer and Weber,2019), which we control for in our setting. To minimize concerns about reporting biases, we employ the generic and continuous SF12 as an alternative Himeasure (Andersen et al.,2007). The SF12 belongs to the “health-related quality of life measures.” Using a specific algorithm, the SF12 weights and aggregates the answers to twelve health questions into a physical health (pcs) and a mental health (mcs) summary scale. Compared to SAH, the SF12 is a “more” objective health measure and was developed to minimize reporting biases. It “can be used to compare the health of different groups, for example, the young and the old or the sick and the well” (RAND,1995). Both subscales of the SF12, pcs and mcs, have continuous values between 0 and 100, mean 50, and a standard deviation of 10. We use equal weights of 0.5 to generate the overall continuous SF12 measure. Figures A1 and Figure A2 in the Appendix show the distributions of SAH and SF12 in our BASE-II (Figure A1) and SOEP-IP sample (Figure A2). The left panels refer to SAH and the right panels refer to SF12. The health distributions appear very similar, both across measures and across databases. The SAH and SF12 allow us to infer Hias well as the population health distribution F(Hi). Hence it is straightforward to calculate the individual rank riin the health distribution. For 10Note that the prevalence of the medical condition also determines the prevalence of absolute overconfidence. However, correcting responses by the prevalence rate is outside the scope of this paper. 16 SAH, each respondent self-categorizes into one of the five SAH categories. Then, we assign every respondent the upper cdf threshold of the category chosen in the SAH distribution. For example, 9% of all respondents self-categorize to be in the highest category “excellent” health. Hence we assign ri,SAH =91 to all respondents in the second highest category “very good” and, using the same principle, we do the same for the other categories. Because SF12 is continuous, ranges from 0 to 100 and has mean 50, it directly yields riwithout further manipulation. To measure e ri, we added the following question to BASE-II and SOEP-IP: “Imagine one would randomly select 100 people in your age, what do you think: How many of those 100 people would be in better health than you?”.11 From the raw untransformed response to this question, e bi, we compute e ri=100 −e bi. In both surveys, we obtain a high response rates of above 90% for our e bimeasure. Even among the elderly in BASE-II, only 10 respondents (<1%) are coded “don’t know” and 129 respondents (6%) are coded “does not apply.” The high response rates may be a function of the natural reference group—100 people in the same age group. This framing allows meaningful comparisons without being too restrictive or too complex. Note that, despite avoiding many of the methodological criticisms of earlier studies (Benoˆ ıt and Dubra,2011), our question does not elicit entire belief distributions (Di Girolamo et al.,2015). Moreover, we did not specifically incentivize respondents (Harrison and Rutstr¨ om,2006). Eliciting entire belief distributions in an incentive-compatible environment is typically feasible in lab experiments (Harrison,2015), which is costly and can only be implemented in large samples under specific conditions. In addition to the advantages above, maybe the main advantage of our measure of e biis its simplicity and cost-effectiveness. Using one simple question, our proposed question has the power to elicit subjective relative beliefs in representative population surveys. [Insert Figure 2about here] Figure 2shows the raw untransformed distributions of e bifor BASE-II and SOEP-IP. Under the assumption of 100 random people being orderly ranked from 1 to 100 and under full rationality and common priors, e biwould be uniformly distributed between 0 and 100 with a mean of 50 (Goette et al.,2015). However, as seen in Figure 2, few respondents say that more than 50 respondents in their age would be in better health and the distribution is clearly skewed to the 11This survey question has been successfully tested in other contexts. For example, using the same format, respondents in the Swiss “Amphiro” study were asked about their income position, their water use, and their knowledge of energy conservation (Tiefenbeck et al.,2018;Friehe and Pannenberg,2019). 17 left. It is worthwhile to emphasize the similarity of the e bidistributions in BASE-II and SOEPIP; the mass of the distributions lies between 10 and 30. In other words, a significant share of respondents believe that (only) 10-30 out of 100 people are in better health (e bi∈(10; 30)) implying that they rank themselves in the 70th to 90th percentile of the population health distribution, e ri∈(70; 90). This yields first evidence for the existence of relative health perception biases at the population level. [Insert Figure 3about here] More evidence for the existence of relative health perception biases is illustrated by Figure 3, which plots the bins of e rion the x-axis and the average values for rion the y-axis. The scatters, whose size indicate the share of respondents falling into each bin, would be lined up along the 45-degree line if e ri=ri. However, as seen, while the scattered line has a slightly positive slope, it is clearly flatter than the 45-degree line. Using riand e ri, we can now calculate Rifor each respondent in BASE-II and SOEP-IP. Because BASE-II is representative of the elderly Berlin population, whereas SOEP-IP is representative of the entire German population, comparing the results of both surveys will inform us about the generalizability of the empirical findings. Calculating Ri,SF12 =e ri−ri,SF12 is straightforward because both the SF12 and e riare continuous. When calculating Ri,SAH =e ri−ri,SAH, recall that e riis continuous, but SAH has five categories. However, because e riis continuous and varies within SAH categories, Ri,SAH is continuous as well, as shown by Figures 4a (SOEP-IP) and A3a (BASE-II). Both figures also demonstrate that Ri,SAH looks very similar in BASE-II and SOEP-IP and that the gender differences are negligible. [Insert Figure 4about here] Figures 4b (SOEP-IP) and A3b (BASE-II) show the distributions of Ri,SF12. As seen in Panels A of Tables A2 and A3, the mean Ri,SF12 values are 14 (SOEP-IP) and 20 (BASE-II), that is, clearly positive and implying that respondents overestimate their rank on average by 14 and 20 positions. Conditional on being overconfident, respondents overestimate their health rank by 23 (BASE-II) and 19 (SOEP-IP) positions. Again, the Ri,SF12 distributions are very similar for BASE-II and SOEP-IP. When studying determinants of Ri,SF12 (details available upon request), one finds that the bias decreases in age and with the number of siblings. Moreover, it is highly correlated with 18 education and decreases in the number of school years. Similarly, blue collar workers and lower-income respondents tend to be more biased, as are people with a lower self-reported trust level. Interestingly, Ri,SF12 is not significantly correlated with the risk tolerance, being religious or the nationality. 6 Health Perception Biases and Risky Health Behaviors In this section, we first study the empirical link between absolute health perception biases and risky health behavior using the GNHSEW91. Then we study the link between relative health perception biases and risky health behavior using the BASE-II and SOEP-IP. We will provide non-parametric evidence and evidence from multivariate regression models. 6.1 Absolute Health Perception Biases and Risky Health Behavior Figure 5tests whether respondents who have biased perceptions about their blood cholesterol levels are more likely to (a) not exercise, (b) have higher BMIs, (c) drink alcohol daily, (d) smoke. Figure 6tests the same relationships for respondents who have biased perceptions about their blood pressure levels, see Section 5for details about how we generated the perception bias measures. Each of the figures shows four bar diagrams along with 95% confidence intervals. [Insert Figures 5and 6about here] Figure 5a shows that respondents who state that they do not have high cholesterol but who, in fact, do have high cholesterol are a highly significant 11 percentage points more likely (43% vs. 54%) to not exercise at all. The BMI differential is also significant (Figure 5b). Similarly, respondents with absolute health perception biases are significantly more likely to drink alcohol daily—the share of daily drinkers is almost 50% higher among this group (21% vs. 14%, Figure 5c). Figure 6d, however, does not provide much evidence that smoking is significantly linked to biased perceptions about high cholesterol levels. Comparing Figure 6to Figure 5, the similarity and robustness of the link between both absolute health perception bias measures and four risky health behavior measures is worthwhile to point out. Not only do all statistical links have identical signs and significance levels, but the risky behavior differentials and their sizes are also very similar. This is even more surprising, given the low correlation between the two perception bias measures of only 0.11. 19 In conclusion, there is robust evidence that absolute health perception biases (“health overconfidence”) are significantly linked to three out of four risky behaviors. According to Proposition 3, this implies that the extensive margin effect dominates the intensive margin effect meaning that overconfident people engage in more unhealthy behavior because they (wrongly) believe that they can “afford” it. One exception appears to be smoking, where the effect on the intensive margin appears to be stronger. This intensive margin effect reduces the inclination to engage in risky behavior because the bias increases the marginal costs of risky behavior, see Section 3for more details. 6.2 Relative Health Perception Biases and Risky Health Behavior Figure 7non-parametrically links Rto xacross the entire Rdistribution using kernel-weighted local polynomial smoothing plots. Table 1provides the equivalent parametric multivariate regressions using a rich set of controls. [Insert Figure 7about here] No physical exercise. Figure 7a shows a monotonically increasing relationship between relative health overconfidence, R>0, and not exercising. On average, 30% of those who accurately assess, or who underestimate their health, do not exercise at all. This share monotonically increases to 50% for respondents who overestimate their rank in the population distribution by 50, that is, who exhibit strong health overconfidence. Next, we run the following parametric regression model controlling for a rich set of sociodemographics: xi=β0+β1R+ i+β2R− i+Ziβ3+ρt+ei(11) where xirepresents risky health behavior and Ristands for our relative health bias measure. Specifically, we will replace the continuous Rimeasure with two measures R+ i{Ri|Ri∈ (0; 100)and R− i{Ri|Ri∈(−100; 0).R+ iis truncated from below and measures the degree of health overconfidence. R− iis truncated from above and measures the degree of health underconfidence. 20 Zicontains socio-demographic controls as listed in Table A2. Note that we also control for risk aversion, trust, and cognitive abilities. We also include interview month fixed effects, ρt;ei is the error term. [Insert Table 1about here] Table 1shows the results of four regression models as in equation (11). The models in columns (1) and (3) do not control for personality traits and behavioral attributes (such as trust, measures or risk aversion or cognitive abilities), whereas the models in columns (2) and (4) do. The first two columns use Rimeasures based on the SF12 benchmark, and the last two columns use Rimeasures based on the SAH benchmark (Section 5). Broadly speaking, the results from the four models confirm the non-parametric findings in Figure 7: there is no evidence that a negative health bias is significantly linked to not exercising. By contrast, similar to the evidence for absolute health overconfidence in Figures 5a and 6a, we find a highly significant link between Ri>0 and not exercising: an increase in Riby 10 ranks is associated with a 1.5ppt higher likelihood to not exercise (column (2)). The size of the association is larger when using RSAH (columns (3) and (4)) but overall robust across columns. [Insert Tables 2and 3about here] BMI, Obesity and Unhealthy Diet. Figure 7b and Table 2show the equivalent findings for BMI, while Figure 7c and Table 3show the findings for having an unhealthy diet. Both figures and tables reinforce our previous findings. Figure 7b shows a non-linear relationship between Ri>0 and BMI which is very similar to Figure 7a: Respondents who accurately assess their health or who underestimate their health do not have higher BMIs. However, the average BMI increases monotonically in the size of the health bias for Ri>0. In other words, the more people overestimate their health, the heavier they are. To test this link parametrically, we generate a binary obesity indicator as dependent variable12 and run regressions similar to equation (11). The results in Table 2confirm the nonparametric visual evidence: An increase in the health bias Riby 10 ranks is associated with a 1.5 percentage point (about 12%) higher probability to be obese in the first two columns; it is associated with a 2.6 percentage point higher probability in the last two columns. As above, 12Using the continuous BMI measure yields robust results. 21 the effect size is larger for RSAH but generally robust. The inclusion of rich sets of individual covariates, such as detailed measures of cognitive skills, do not matter for the strength of the empirical relationship. Figure 7c and Table 3corroborate the findings above. Whereas no link exists for respondents with Ri≤0, that is, unbiased or underconfident people, we find a clear and positive statistical link between being overconfident about one’s health and unhealthy eating. An increase in the bias by 10 ranks increases the likelihood of an unhealthy diet by almost 2 percentage points (or about 4.5%). This estimate is very robust across the health bias definitions and all four columns. [Insert Table 4about here] Smoking. Figure 7d and Table 4show the results for being a smoker. The graphical evidence provides no evidence for a statistical link between health perception biases and smoking status. This is confirmed by the parametric regressions in Table 4—none of the eight health bias measures is significantly linked to smoking status and the effect sizes are very small. It is worthwhile to emphasize that this finding is exactly in line with the findings from above in Figures 5and 6, where we found no association between absolute health overconfidence and smoking. According to our model, this implies that the health perception biases operate through a strong intensive margin effect, which emphasizes the costs of smoking (Section 3). This finding is consistent with Darden (2017), who reports that updated (and objective) cardiovascular biomarker information has not altered smoking behavior in the population of the Framingham Heart Study—Offspring Cohort. [Insert Figure 8and Table 5about here] Sleep. Finally, we use the representative SOEP to investigate the empirical relationship between relative health perception biases and sleep. Our outcome measures indicate the ”sleep gap” between the actual hours of sleep and eight hours. Figure 8a shows again the characteristic pattern from above. The sleep gap is monotonically increasing in the size of the health bias—but only for Ri>0, that is, for overconfident people. For the underconfident and those who do not exhibit biases, no clear association between Riand hiexists. This holds for sleep during the week (Figure 8a) as well as sleep on weekends (Figure 8b). Table 5shows the results of regression models as in equation (11), using the sleep gap during the week as dependent variable (results for the weekend are similar and available upon 22 K˝ oszegi, B. (2006). Ego utility, overconfidence, and task choice. Journal of the European Economic Association 4(4), 673–707. Lindeboom, M. and E. van Doorslaer (2004). Cut-point shift and index shift in self-reported health. Journal of Health Economics 23(6), 1083–1099. Linn´ er, R. K., P. Biroli, E. Kong, S. F. W. Meddens, R. Wedow, M. A. Fontana, M. Lebreton, A. Abdellaoui, A. R. Hammerschlag, M. G. Nivard, A. Okbay, C. A. Rietveld, P. N. Timshel, S. P. Tino, M. Trzaskowski, R. d. Vlaming, C. L. Z¨ und, Y. Bao, L. Buzdugan, A. H. Caplin, C.- Y. Chen, P. Eibich, P. Fontanillas, J. R. Gonzalez, P. K. Joshi, V. Karhunen, A. Kleinman, R. Z. Levin, C. M. Lill, G. A. Meddens, G. Muntan´ e, S. Sanchez-Roige, F. J. v. Rooij, E. Taskesen, Y. Wu, F. Zhang, A. Auton, J. D. Boardman, D. W. Clark, A. Conlin, C. C. Dolan, U. Fischbacher, P. J. F. Groenen, K. M. Harris, G. Hasler, A. Hofman, M. A. Ikram, S. Jain, R. Karlsson, R. C. Kessler, M. Kooyman, J. MacKillop, M. M¨ annikk¨ o, C. Morcillo-Suarez, M. B. McQueen, K. M. Schmidt, M. C. Smart, M. Sutter, A. R. Thurik, A. G. Uitterlinden, J. White, H. d. Wit, J. Yang, L. Bertram, D. Boomsma, T. Esko, E. Fehr, D. A. Hinds, M. Johannesson, M. Kumari, D. Laibson, P. K. E. Magnusson, M. N. Meyer, A. Navarro, A. A. Palmer, T. H. Pers, D. Posthuma, D. Schunk, M. B. Stein, R. Svento, H. Tiemeier, P. R. H. J. Timmers, P. Turley, R. J. Ursano, G. G. Wagner, J. F. Wilson, J. Gratten, J. J. Lee, D. Cesarini, D. J. Benjamin, P. D. Koellinger, and J. P. Beauchamp (2019). Genome-wide association analyses of risk tolerance and risky behaviors in over 1 million individuals identify hundreds of loci and shared genetic influences. Nature Genetics 51, 245–257. Merkle, C. and M. Weber (2011). True overconfidence: The inability of rational information processing to account for apparent overconfidence. Organizational Behavior and Human Decision Processes 116(2), 262–271. Moore, D. A. and P. J. Healy (2008). The trouble with overconfidence. Psychological review 115(2), 502. Ortoleva, P. and E. Snowberg (2015a). Overconfidence in political behavior. American Economic Review 105(2), 504–35. Ortoleva, P. and E. Snowberg (2015b). Overconfidence in political behavior. The American Economic Review 105(2), 504–535. Rabin, M. (2013, June). Incorporating limited rationality into economics. Journal of Economic Literature 51(2), 528–43. RAND (1995). User’s manual for the medical outcomes study (mos) core measures of healthrelated quality of life. Technical report. Richter, D., M. Metzing, M. Weinhardt, and J. Schupp (2013). SOEP Scales Manual— Erhebungsinstrumente Berliner Altersstudie II. SOEP Survey Papers: Series C - Data Documentations (Datendokumentationen) 138. panel.gsoep.de/soep-docs/ surveypapers/diw_ssp0138.pdf, retrieved January 17, 2014. Richter, D. and J. Schupp (2015). The SOEP Innovation Sample (SOEP IS). Schmollers Jahrbuch 135(3), 389–399. Richter, D. and J. Schupp (2017). Questionnaire for the soep innovation sample (boost sample). SOEP Survey Papers 454, DIW Berlin, The German Socio-Economic Panel (SOEP). Robert Koch Institut (2012). German National Health Survey East-West 1991 (GNHSEW91), Public Use File OW91. Robert Koch Institut. http://www.rki.de/DE/Content/ Gesundheitsmonitoring/PublicUseFiles/informationen_datensaetze/ info_datensaetze_node.html, last accessed on June 11, 2016. 29 Royer, H., M. Stehr, and J. Sydnor (2015). Incentives, commitments, and habit formation in exercise: Evidence from a field experiment with workers at a fortune-500 company. American Economic Journal: Applied Economics 7(3), 51–84. Sandroni, A. and F. Squintani (2007). Overconfidence, insurance, and paternalism. The American Economic Review 97(5), 1994–2004. Santos-Pinto, L. and J. Sobel (2005). A model of positive self-image in subjective assessments. American Economic Review 95(5), 1386–1402. Sharot, T. (2011). The optimism bias. Current biology 21(23), R941–R945. Spinnewijn, J. (2015). Unemployed but optimistic: Optimal insurance design with biased beliefs. Journal of the European Economic Association 13(1), 130–167. Spitzer, S. and D. Weber (2019). Who is telling the truth? biases in self-reported physical and cognitive health status of older Europeans. PLoS ONE 10(14), e0223526. Strulik, H. and T. Trimborn (2018). Hyperbolic discounting can be good for your health. Center for European, Governance and Economic Development Research Discussion Papers 335, University of Goettingen, Department of Economics. Tiefenbeck, V., L. Goette, K. Degen, R. Lalive, and T. Staake (2018). Overcoming salience bias: How real-time feedback fosters resource conservation. Management Science 64(3), 1458–1476. Viscusi, W. K. (1990). Do smokers underestimate risks? Journal of Political Economy 98(6), 1253– 1269. Wagner, G. G., J. R. Frick, and J. Schupp (2007). The German Socio-Economic Panel Study (SOEP)–scope, evolution and enhancements. Schmollers Jahrbuch : Journal of Applied Social Science Studies / Zeitschrift f¨ur Wirtschaftsund Sozialwissenschaften 127(1), 139–169. Weinberg, B. A. (2006). A model of overconfidence. Working paper, Ohio State University. Weinstein, N. (1989). Optimistic biases about personal risks. Science 246(4935), 1232–1233. Weinstein, N. D. (1980). Unrealistic optimism about future life events. Journal of Personality and Social Psychology 39(5), 806. Ziebarth, N. R. (2010). Measurement of health, health inequality, and reporting heterogeneity. Social Science & Medicine 71, 116–124. 30 Figures Figure 1: Absolute Health Perception Bias about High Cholesterol Levels (a) High Cholesterol Levels (b) High Blood Pressure Source: GNHSEW91. See Section 4.3 and 5for more details. Figure 2: Perceived Population Share in Better Health (e bi) 0 .05 .1 .15 .2 .25 Fraction 0 10 20 30 40 50 60 70 80 90 100 Number of respondents with better health status 0 .05 .1 .15 .2 .25 Fraction 0 10 20 30 40 50 60 70 80 90 100 Number of respondents with better health status Sources: BASE-II (left panel), SOEP-IP (right panel). Responses to the question are plotted: “Imagine one would randomly select 100 people in your age. How many of those 100 people would be in better health than you?” People answering 0 believe nobody is in better health; people answering 99 believe everybody is healthier than them.” 31 Figure 3: Actual (ri) and Perceived ( ˜ ri) Ranking in Population Health Distribution BASE-II SOEP-IP 0 10 20 30 40 50 60 70 80 90 100 ri 0 10 20 30 40 50 60 70 80 90 100 ri ~ e Ri<0 e Ri>0 0 10 20 30 40 50 60 70 80 90 100 ri 0 10 20 30 40 50 60 70 80 90 100 ri ~ Ri<0 Ri>0 Source: BASE-II. Relative health perception biases are defined as e Ri=e ri−ri, with ˜ ri= 1 −e bi, see Figure 2and main text. The true rank in the population health distribution, ri, is based on the SF12 indicator. Figure 4: Distribution of RiBased on (a) SAH, (b) SF12 in SOEP-IP 0 .005 .01 .015 .02 Density -100 -50 0 50 100 Overconfidence measure SAH (rank differences) males females kernel = epanechnikov, bandwidth = 9.8000 0 .005 .01 .015 Density -100 -50 0 50 100 Overconfidence measure SF12 (rank differences) males females kernel = epanechnikov, bandwidth = 9.8000 Source: SOEP-IP. Figure displays distributions of e Ri=e ri−riin the representative SOEP-IP, with ˜ ri= 1 −e bi, see Figure 2and main text. Subfigure (a) uses SAH and subfigure (b) uses the SF12 as ri. 32 Figure 5: Absolute Health Perception Bias (High Cholesterol) and Risky Health Behavior .4 .45 .5 .55 .6 mean of NoSport 0 1 26 26.5 27 27.5 28 mean of bmi 0 1 .12 .14 .16 .18 .2 .22 mean of AlcoholDaily 0 1 .3 .31 .32 .33 .34 mean of smoker 0 1 Source: GNHSEW91, own calculations, own illustration. Bar diagrams show, along with 95% confidence intervals, the (a) share of respondents who do not exercise, (b) mean BMI, as well as the share of respondents who (c) drink alcohol daily and (d) smoke. 33 Figure 6: Absolute Health Perception Bias (High Blood Pressure) and Risky Health Behavior .45 .5 .55 .6 mean of NoSport 0 1 26.5 27 27.5 28 28.5 mean of bmi 0 1 .15 .2 .25 .3 .35 mean of AlcoholDaily 0 1 .26 .28 .3 .32 .34 mean of smoker 0 1 Source: GNHSEW91, own calculations, own illustration. Bar diagrams show, along with 95% confidence intervals, the (a) share of respondents who do not exercise, (b) mean BMI, as well as the share of respondents who (c) drink alcohol daily and (d) smoke. 34 Figure 7: Relative Health Perception Bias and Risky Health Behavior .2 .3 .4 .5 .6 No sports -25-20-15-10 -5 0 5 10 15 20 25 30 35 40 45 50 Confidence measure (percentage point deviation) 95% CI lpoly smooth kernel = epanechnikov, degree = 0, bandwidth = 5.47, pwidth = 8.21 24.5 25 25.5 26 26.5 BMI -25-20-15-10 -5 0 5 10 15 20 25 30 35 40 45 50 Confidence measure (percentage point deviation) 95% CI lpoly smooth kernel = epanechnikov, degree = 0, bandwidth = 9.37, pwidth = 14.05 .3 .35 .4 .45 .5 No healthy diet -25-20-15-10 -5 0 5 10 15 20 25 30 35 40 45 50 Confidence measure (percentage point deviation) 95% CI lpoly smooth kernel = epanechnikov, degree = 0, bandwidth = 5.19, pwidth = 7.78 .05 .1 .15 .2 Smoker -25-20-15-10 -5 0 5 10 15 20 25 30 35 40 45 50 Confidence measure (percentage point deviation) 95% CI lpoly smooth kernel = epanechnikov, degree = 0, bandwidth = 8.33, pwidth = 12.5 Source: BASE-II, own calculations, own illustration. Figure shows non-parametric kernel-weighted local polynomial smoothing plots. The y-axis shows (a) the likelihood that respondents do not exercise, (b) their BMI, as well as the likelihood that respondents (c) have an unhealthy diet, (d) smoke. 35 Figure 8: Relative Health Perception Biases and the Sleep Gap to 8 Hours (SOEP-IP) .8 1 1.2 1.4 1.6 sleepweekBench -25-20-15-10 -5 0 5 10 15 20 25 30 35 40 45 50 Confidence measure (percentage point deviation from rational benchmark) 95% CI lpoly smooth kernel = epanechnikov, degree = 0, bandwidth = 5.5, pwidth = 8.25 0 .2 .4 .6 .8 sleepweekendBench -25-20-15-10 -5 0 5 10 15 20 25 30 35 40 45 50 Confidence measure (percentage point deviation from rational benchmark) 95% CI lpoly smooth kernel = epanechnikov, degree = 0, bandwidth = 5.77, pwidth = 8.65 Source: SOEP-IP, own calculations, own illustration. Figure shows non-parametric kernel-weighted local polynomial smoothing plots. The y-axis shows the difference between 8 hours of sleep and actual hours of sleep (a) during the week, (b) on weekends. 36 Table 1: Relative Perception Bias and Likelihood to Not Exercise (1) (2) (3) (4) No Sports (SF12) (SF12) (SAH) (SAH) Overconfidence (Ri>0) 0.0022*** 0.0022*** 0.0031*** 0.0031*** (0.0006) (0.0006) (0.0007) (0.0007) Underconfidence (Ri<0) 0.0001 0.0005 0.0011 0.0012 (0.0014) (0.0014) (0.0011) (0.0012) Age 0.0017 0.0034 0.0019 0.0034 (0.0067) (0.0068) (0.0067) (0.0068) Age2 -0.0000 -0.0000 -0.0000 -0.0000 (0.0001) (0.0001) (0.0001) (0.0001) socio-demographics & education yes yes yes yes employment char. & income no yes no yes month FE yes yes yes yes R20.0257 0.0301 0.0289 0.0329 Source: Berlin Aging Study II (BASE-II), own calculation and illustration; * p<0.1, ** p<0.05, *** p<0.01; standard errors in parentheses. The descriptive statistics are in the Appendix (Table A2). The model is estimated by OLS and has 1,868 observations; the binary dependent variable measures the likelihood that a respondent does not exercise at all. Health overconfidence (Ri>0) and underconfidence (Ri<0) are continuous health bias measures. For more information, see Section 4.2 and 5. 37 Table 2: Relative Perception Bias and Likelihood to be Obese (1) (2) (3) (4) Obesity (SF12) (SF12) (SAH) (SAH) Overconfidence (Ri>0) 0.0010*** 0.0011*** 0.0022*** 0.0022*** (0.0004) (0.0004) (0.0005) (0.0005) Underconfidence (Ri<0) 0.0008 0.0008 -0.0002 -0.0002 (0.0010) (0.0010) (0.0008) (0.0008) Age 0.0097** 0.0109** 0.0102** 0.0113*** (0.0043) (0.0044) (0.0043) (0.0043) Age2 -0.0001* -0.0001** -0.0001** -0.0001** (0.0000) (0.0000) (0.0000) (0.0000) socio-demographics & education yes yes yes yes employment char. & income no yes no yes month FE yes yes yes yes R20.0271 0.0306 0.0360 0.0391 Source: Berlin Aging Study II (BASE-II), own calculation and illustration; * p<0.1, ** p<0.05, *** p<0.01; standard errors in parentheses. The descriptive statistics are in the Appendix (Table A2). The model is estimated by OLS and has 1,868 observations; the binary dependent variable measures the likelihood that a respondent is obese (BMI>30). Health overconfidence (Ri>0) and underconfidence (Ri<0) are continuous health bias measures. More information on the variables, see Section 4.2 and 5. 38 Figure A3: Distribution of RiBased on (a) SAH, (b) SF12 in BASE-II 0 .005 .01 .015 .02 Density -100 -50 0 50 100 Overconfidence measure SAH (rank differences) males females kernel = epanechnikov, bandwidth = 9.8000 0 .005 .01 .015 Density -100 -50 0 50 100 Overconfidence measure SF12 (rank differences) males females kernel = epanechnikov, bandwidth = 9.8000 Source: BASE-II. Figure displays distributions of Ri=e ri−ri, with ˜ ri= 1 −e bi, see Figure 2and main text. Subfigure (a) uses SAH and subfigure (b) uses the SF12 as ri. Table A1: German National Health Survey East-West 1991 Variable Mean Std. Dev. Min. Max. N A. Health Bias Measures Absolute Health Bias Cholesterol, Ai>1 0.2976 0.4572 0 1 6429 Absolute Health Bias Blood Pressure, Ai>1 0.0935 0.2911 0 1 6429 B. Objective Health Measures (Hi) Total blood cholesterol [mmol/l] 6.128 1.2292 2.33 12.9 6429 High total blood cholesterol [>6.2 mmol/l] 0.4394 0.4964 0 1 6429 Systole, 2. measure [mmHg] 134.6558 20.2632 88 256 6429 Diastole, 2. measure [mmHg] 83.4033 12.1276 34 158 6429 Hypertension 0.2095 0.407 0 1 6429 C. Subjective Health Assessment ( ˜ Hi) High Cholesterol 0.2072 0.4053 0 1 6429 High Blood Pressure 0.2518 0.4341 0 1 6429 D. Health Behavior Alcohol Daily 0.1618 0.3683 0 1 6429 Current Smoker 0.3282 0.4696 0 1 6429 Body-mass-index [kg per m2] 26.6467 4.6113 15.02 75.467 6429 Obese (BMI>30) 0.2019 0.4014 0 1 6429 No sports 0.4652 0.4988 0 1 6429 Sources: GNHSEW91, own illustration.[mmol/l] stands for millimole per liter. [mmHg] stands for millimetres of mercury. [kg per m2] stands for kilogram per square meter. 45 Table A2: Descriptive Statistics BASE-II Variable Mean Std. Dev. Min. Max. N A. Health Bias Measures Relative Health Bias SAH, Ri3.8770 26.2816 -98 95.7662 1751 Ri>0, SAH 11.9119 17.6585 0 95.7662 1751 Ri<0, SAH 2.2780 10.0479 0 98 1751 Relative Health Bias SF12, Ri19.9118 26.2824 -72.5012 94.11023 1751 Ri>0, SF12 23.2566 21.3903 0 94.1102 1751 Ri<0, SF12 3.3443 8.8070 0 72.5012 1751 B. Health Behavior BMI 25.5181 4.3132 13.7143 64.0923 1780 Obese 0.1264 0.3324 0 1 1780 Smoker 0.1169 0.3213 0 1 1780 No sports 0.3567 0.4792 0 1 1780 Unhealthy diet 0.3866 0.4871 0 1 1780 C. Covariates Demographics Age 60.0596 16.7442 18 89 1780 Female 0.5242 0.4996 0 1 1780 Married 0.5674 0.4956 0 1 1780 Single 0.2528 0.4347 0 1 1780 Partner in Household 0.6528 0.4762 0 1 1780 # kids 1.3118 1.1355 0 5 1780 # daughters 0.6573 0.8284 0 4 1780 No kids 0.3101 0.4627 0 1 1780 German 1.0124 0.1105 1 2 1780 Education 8 school years 0.1236 0.3292 0 1 1780 10 school years 0.2539 0.4354 0 1 1780 13 school years 0.5584 0.4967 0 1 1780 Employment Blue collar worker 0.0275 0.1637 0 1 1780 White collar worker 0.1893 0.3919 0 1 1780 Civil servant 0.0185 0.1349 0 1 1780 Full-time employed 0.4697 0.4992 0 1 1780 Part-time employed 0.1404 0.3476 0 1 1780 Gross labor earnings 549.7449 1300.6576 0 20,000 1780 Net labor earnings (last month) 378.4079 831.9664 0 10,000 1780 Total income (last month) 1565.3478 1295.8649 0 20,950 1780 Behavioral Attitudes Trust (in strangers) 0.3966 0.4893 0 1 1780 Trust (general) 0.7545 0.4305 0 1 1780 Risk aversion (scale) 5.077 2.2269 0 10 1780 Risk averse 0.2624 0.44 0 1 1780 Risk loving 0.1528 0.3599 0 1 1780 No religion 0.6348 0.4816 0 1 1780 Big Five Openness 4.9872 1.1524 1.3333 7 1780 Conscientiousness 5.6126 0.9755 1.6667 7 1780 Extraversion 4.7374 1.1743 1 7 1780 Neuroticism 3.7723 1.2778 1 7 1780 Agreeableness 5.2368 0.9820 1.3333 7 1780 Sources: Berlin Aging Study II (BASE-II), own illustration. 46 Table A3: Descriptive Statistics SOEP-IP Variable Mean Std. Dev. Min. Max. N A. Health Bias Measures Relative Health Bias SAH, Ri-0.2969 27.69 -100 96.0114 1377 Ri>0, SAH 10.234 16.9727 0 96.0114 1377 Ri<0, SAH 4.256 14.1312 0 100 1377 Relative Health Bias SF12, Ri13.8467 27.4498 -96.2857 99 1377 Ri>0, SF12 18.917 20.6165 0 99 1377 Ri<0, SF12 5.0687 11.6833 0 96.2857 1377 B. Health Behavior Sleep in hours, weekday 6.8221 1.3144 2 13 1377 Sleep deficit, week 1.1641 1.3151 -5 6 1377 Sleep in hours, weekend 7.5737 1.567 2 14 1377 Sleep deficit, weekend 0.4242 1.5551 -6 6 1377 C. Covariates Demographics Age 51.1438 18.4188 17.011 93.2301 1377 Female 0.5178 0.4999 0 1 1377 Married 0.5272 0.4994 0 1 1377 Single 0.2585 0.438 0 1 1377 # kids 0.5810 0.9436 0 5 1377 German 0.9390 0.2394 0 1 1377 Education 8 school years 0.6572 0.4748 0 1 1377 10 school years 0.2012 0.401 0 1 1377 13 school years 0.1264 0.3324 0 1 1377 Employment Blue collar worker 0.0741 0.262 0 1 1377 White collar worker 0.3471 0.4762 0 1 1377 Civil servant 0.0312 0.174 0 1 1377 Full-time employed 0.3573 0.4794 0 1 1377 Part-time employed 0.1046 0.3061 0 1 1377 Gross labor earnings 1293.5229 1836.4051 0 12,540 1377 Net labor earnings (last month) 879.6253 1167.7447 0 8,000 1377 Total income (last month) 1772.902 1736.4825 0 14,200 1377 Behavioral Attitudes Risk averse 0.3086 0.4621 0 1 1377 Risk loving 0.1503 0.3575 0 1 1377 Risk Loving Health 0.0763 0.2655 0 1 1377 Risk Averse Health 0.5534 0.4973 0 1 1377 Risk Loving Trust 0.0821 0.2746 0 1 1377 Risk Averse Trust 0.4379 0.4963 0 1 1377 Sources: SOEP-IP, own illustration. 47 