"من می ترسم خبر بد به شما بدهم ..." برآورد تاثیر اختلالات مختلف سلامت بر روی بهزیستن ذهنی

Abstract Bad health decreases individuals' happiness, but few studies measure the impact of specific illnesses. We apply matching estimators to examine how changes in different (objective) conditions of bad health affect subjective well-being for a sample of 100,265 observations from the British Household Panel Survey (BHPS) database (1996–2006). The strongest effect is for alcohol and drug abuse, followed by anxiety, depression and other mental illnesses, stroke and cancer. Adaptation to health impairments varies across health impairments. There is also a puzzling asymmetry: strong adverse reactions to deteriorations in health appear alongside weak increases in well-being after health improvements. In conclusion, our analysis offers a more detailed account of how bad health influences happiness than accounts focusing on how bad self-assessed health affects individual well-being.

Introduction Our health determines many facets of our life. It affects our employment opportunities and our incomes (Arrow, 1996), influences social activities (Gardner & Oswald, 2004) and impacts our mood and well-being more generally (Easterlin, 2003; Graham, 2008). Being in good health increases subjective well-being, just as illness or bad health decreases it (Graham, Higuera, & Lora, 2011; Veenhoven, 2008). An individual's subjective well-being (synonymously called “happiness” here) depends on a complex interacting web of factors, comprising economic (such as income, status or employment), situational (health, social relations), socio-demographic (gender, age, education), personal (personality and genes) and institutional factors (such as the extent of direct democratic participation), and the literature examining these relationships has vastly increased over the last few years (for overviews, see Dolan, Peasgood, & White, 2008; Easterlin, 2003; Frey & Stutzer, 2000). As one can consider subjective well-being to be a broad aspect of an individual's mental health, it is no wonder that many determinants of subjective well-being also determine health more generally (see, e.g., Contoyannis & Jones, 2004; Fuchs, 2004; Gardner & Oswald, 2004). In subjective well-being research, the relationship between subjective well-being and (mostly: self-assessed) health is well-researched and “studies consistently reveal a strong relationship between health and happiness” (Graham, 2008, p. 73). This is less surprising, for instance, for broad “mental well-being” measures (such as the GHQ-12) that incorporate some (mental) health aspects (Dolan et al., 2008, p. 100). But the positive relationship also holds when using life satisfaction as the dependent variable in regressions (Dolan & Kahneman, 2008; Dolan et al., 2008; Easterlin, 2003). It seems that causality runs in both directions: a high level of well-being certainly seems relevant for subsequent good health, with significant positive effects of well-being on health being observed two or three years later (Binder & Coad, 2010; Lyubomirsky, King, & Diener, 2005). The stronger relationship, however, seems to run from health to happiness. Numerous studies show that healthier individuals tend to be happier. Most studies here analyze the relationship between individuals' subjective health ratings and subjective well-being (Dolan et al., 2008; Easterlin, 2003) or the impact of disability on subjective well-being (Brickman, Coates, & Janoff-Bulman, 1978; Oswald & Powdthavee, 2008; Uppal, 2006), mostly for lack of more detailed data on objective health impairments. Very few studies also extend the analysis to more detailed health conditions (Dolan, 2011; Graham et al., 2011; Mukuria & Brazier, 2013; Shields & Wheatley Price, 2005). Even if large panel studies incorporate questions on individuals' health impairments, many of these illnesses are comparatively rare and typical multivariate regressions are ill-suited to deal with small numbers of observations in such cases (as well as lack of variation). In a cross-sectional analysis of Health Survey for England (HSE) data, Shields and Wheatley Price (2005) report significantly decreased psychological well-being for individuals with problems with muscular-arthritis-rheumatism, stomach problems and respiratory problems. For males, heart attack or stroke problems and migraine and epilepsy are associated with depressed psychological well-being, while hypertension and blood pressure problems seem associated with decreased psychological well-being in females (p. 529). Problems like cancer or diabetes are not related to psychological well-being in their sample. A similar cross-sectional study has been conducted by Graham et al. (2011) for a number of Latin American countries, where an EQ5D measure of health problems is related to health satisfaction and life satisfaction. Pain, anxiety and difficulties with usual activities are strongly negatively related to health satisfaction and, to a lesser degree, also to life satisfaction. Problems with mobility and self-care are not as clearly related to life satisfaction, which the authors interpret as evidence of a higher impact of acute and chronic mental illnesses over physical conditions (compare also Mukuria & Brazier, 2013). An explanation for this finding might include the uncertainty associated with some health problems, where the next anxiety or epilepsy attack cannot be anticipated (thus hindering adaptation). Similarly, Dolan (2011) finds that mental health has stronger effects on subjective well-being than physical health problems, while in preference elicitations, individuals value physical health more than mental health, probably due to focusing effects and faulty affective forecasting (Wilson & Gilbert, 2005). In the cases discussed, the cross-sectional data structure hinders investigation of self-selection, duration of the health condition, and the role of personality traits mediating the happiness–health relationship; so these estimates should be taken with care. While panel data regression techniques might offer valuable insights into the variation within individuals over time and thus help alleviate concerns about selection effects, as well as account for individual-specific (fixed) effects that capture the trait-like properties of subjective well-being (Diener & Lucas, 1999; Ferrer-i-Carbonell & Frijters, 2004), these techniques are ill-suited to deal with dummy variables that exhibit little variation, as in the case of specific illnesses. We therefore seek to obtain improved estimates of the causal impact of such illnesses on subjective well-being by applying matching estimators (Caliendo & Kopeinig, 2008; Imbens, 2004; Rubin, 1974). This allows us to address many of the above-mentioned shortcomings and estimate the impact of different health impairments on subjective well-being, at an improved level of detail. Similarly, we provide novel results concerning specific adaptation and recovery patterns for different health conditions. Indeed, the dynamics of illness conditions and their impact on subjective well-being need to be better understood, since it remains unclear to what extent subjective well-being can be permanently influenced by life events in general and health conditions in particular (Headey, 2010). This time dimension is also important in our context, as there is some evidence that individuals adapt differently to different health conditions. While some hedonic adaptation occurs, the level of adaptation seems far from complete: Oswald and Powdthavee (2008) find a rate of hedonic adaptation between 30% and 50% in their fixed-effects framework, depending on the degree of disability. As opposed to disability, patients who suffer from chronic diseases and chronic pain have difficulties adapting (Oswald & Powdthavee, 2008; Smith & Wallston, 1992). There are few studies in this field, and their results are complicated by the progressive nature of some of the diseases (Dolan & Kahneman, 2008, pp. 218–9). In sum, hedonic adaptation to adverse health conditions seems limited and domain-specific (Frederick and Loewenstein, 1999; Oswald & Powdthavee, 2008). The dynamic properties of subjective well-being and the extent of hedonic adaptation to adverse (but also to beneficial) life events motivates our later analysis of the causal effect of different health conditions on individuals' life satisfaction with different time lags. The paper is structured as follows: we present the dataset in Section 2. Our analysis is detailed in Section 3, and we conclude with a discussion in Section 4.

Results Fixed-effects regressions Table 3 presents a baseline model of the life satisfaction health relationship using standard fixed-effects (FE) regressions. Accounting for fixed-effects in subjective well-being regressions is the standard model choice (Ferrer-i-Carbonell & Frijters, 2004), since happiness is partly determined by genes and stable personality traits (Diener, Suh, Lucas, & Smith, 1999; Lykken & Tellegen, 1996). Table 3. Baseline regression analysis: Fixed-Effect (FE) regressions for the full sample as well as for subsamples by gender. (1) Life satisfaction (FE) (2) Life satisfaction (male) (3) Life satisfaction (female) Subj. health 0.1840*** (28.64) 0.1690*** (18.64) 0.1957*** (21.72) Doctor visits −0.0080 (−1.79) −0.0190** (−2.91) −0.0001 (−0.02) Accidents −0.0203* (−2.08) −0.0257* (−2.07) −0.0138 (−0.90) Log(hosp. days) −0.0123 (−1.62) −0.0274* (−2.26) −0.0016 (−0.16) Disabled −0.1520*** (−6.79) −0.1438*** (−4.69) −0.1565*** (−4.89) No. cigarettes −0.0025* (−2.13) −0.0027 (−1.80) −0.0023 (−1.21) Health condition dummies Arms −0.0307** (−2.69) −0.0129 (−0.82) −0.0477** (−2.91) Sight −0.0489* (−2.25) −0.0355 (−1.09) −0.0596* (−2.05) Hearing −0.0533** (−2.59) −0.0491 (−1.83) −0.0587 (−1.84) Allergy −0.0277 (−1.82) −0.0236 (−1.02) −0.0301 (−1.49) Chest 0.0033 (0.18) 0.0065 (0.26) 0.0004 (0.02) Heart −0.0014 (−0.09) −0.0054 (−0.25) 0.0042 (0.19) Stomach −0.0123 (−0.70) −0.0133 (−0.51) −0.0108 (−0.46) Diabetes 0.0276 (0.64) 0.0224 (0.40) 0.0340 (0.53) Anxiety −0.3890*** (−17.93) −0.4695*** (−12.43) −0.3515*** (−13.28) Drugs −0.1261 (−1.48) −0.1648 (−1.57) −0.0347 (−0.24) Epilepsy −0.0741 (−0.86) −0.1706 (−1.21) 0.0130 (0.13) Migraine −0.0611** (−3.13) −0.0392 (−1.13) −0.0688** (−2.92) Other −0.0496* (−2.48) −0.0537 (−1.60) −0.0483 (−1.95) Log(income) 0.0301*** (3.46) 0.0335** (2.89) 0.0282* (2.20) Age −0.0137 (−1.01) 0.0089 (0.48) −0.0310 (−1.70) (Age-mean age)2 −0.0001 (−1.33) 0.0001 (1.60) −0.0002** (−3.07) No. children −0.0023 (−0.27) 0.0097 (0.83) −0.0150 (−1.18) Education dummies Elementary 0.0378 (0.28) 0.0317 (0.18) 0.0880 (0.44) Basic vocational −0.0607 (−0.63) −0.2339 (−1.88) 0.1036 (0.75) Middle general 0.1985* (2.07) 0.0667 (0.56) 0.3477* (2.29) Middle vocational 0.3079* (2.17) 0.3549 (1.64) 0.3648 (1.88) High general 0.2475* (2.56) 0.1178 (0.97) 0.3976** (2.60) High vocational 0.1326 (1.27) 0.0302 (0.24) 0.2764 (1.62) Low tertiary 0.2483* (2.42) 0.1022 (0.84) 0.4170* (2.50) High tertiary 0.1788 (1.74) 0.0458 (0.35) 0.3249* (2.02) Marital status dummies Never married −0.0307 (−1.23) −0.0695 (−1.92) 0.0051 (0.15) Separated −0.1451*** (−3.56) −0.1863** (−3.20) −0.1120* (−2.00) Divorced −0.0081 (−0.23) −0.0017 (−0.03) −0.0134 (−0.28) Widowed −0.2344*** (−4.19) −0.1830* (−2.05) −0.2568*** (−3.59) Labor market status dummies Unemployed −0.3111*** (−10.99) −0.3428*** (−8.74) −0.2829*** (−6.93) Self-employed −0.0049 (−0.22) 0.0124 (0.46) −0.0387 (−0.96) Retired 0.0616* (2.50) 0.0744* (2.03) 0.0549 (1.66) Study, school 0.0472 (1.62) 0.0357 (0.82) 0.0574 (1.46) Maternity leave 0.2865*** (6.61) 0.3459 (1.43) 0.2734*** (6.11) Long-term sick −0.2869*** (−7.52) −0.3319*** (−5.67) −0.2434*** (−4.85) Family-care −0.0523* (−2.20) −0.2179* (−2.25) −0.0406 (−1.58) Other −0.0162 (−0.30) −0.1271 (−1.58) 0.0707 (0.98) Observations 100,265 46,850 53,415 R2 0.047 0.050 0.047 t statistics in parentheses. Key to significance levels: *p < 0.05, **p<0.01, ***p<0.001. Table options Regarding our health variables, the FE models exhibit strong positive effects of good subjective health status on life satisfaction and strong negative effects from disability, long-term sickness, as well as health conditions such as anxiety. Ceteris paribus effect sizes (the coefficient magnitudes) are small for most objective health impairments, and insignificant for many of these problems. With reference to our subsequent matching analysis, it can be conjectured that this is not due to absence of effects, but rather an artifact resulting from small numbers of observations for these conditions, as well as their slow-changing nature (circumstances under which FE models underperform). The two health conditions referring to having a stroke or cancer are not shown in Table 3, as they have been only elicited halfway into our sample horizon. When we included both conditions, our sample size fell from 100,265 observations to 61,143. In this model, neither health problem was significantly related to life satisfaction (stroke: b = −0.0891, t-stat −1.27; cancer: b = 0.0342, t-stat 0.59) and due to the nearly halved sample size, some other (health) coefficients were also insignificant. We find typical results in our model regarding other variables. Income has a significant effect on life satisfaction, which seems to be driven by males. Being unemployed has a strong negative impact on life satisfaction, irrespective of gender. We find no relationship between self-employment and life satisfaction in the unmatched sample, as do most studies (Dolan et al., 2008, but see Binder & Coad, in press). We also find a significant negative relationship for separation and widowhood, while being divorced has no significant impact (as expected, if divorce finalizes a long decline in quality of marriage). In the gender disaggregation, education is unrelated to life satisfaction for males, but it influences female life satisfaction more strongly. However, the relationship between education and subjective well-being appears rather unstable in the literature (Binder & Coad, 2011; Dolan et al., 2008). We also find evidence of gender differences in subjective well-being, which can be conjectured to interact with other variables of interest, such as job status (in our case) or age (Plagnol & Easterlin, 2008).1 Matching estimates While FE models are preferable to simple pooled models for panel data, we may be “overcontrolling” – removing some slow-changing variables of interest. Moreover, fixed-effect regression suffers from other drawbacks of regression models (in particular, potential lack of a common support for treatment and control groups). Both points are germane to our context. Consider the dummies for different illness conditions. These would exhibit little variation if individuals mostly transition into a health problem and stay there, due to chronic illness. For those categories that refer mostly to conditions with chronic or progressive disease characteristics, our dummies will not capture much variation, making FE estimates unreliable. Moreover, the above FE regressions with the prima facie high number of observations obscures a crucial fact regarding illness conditions, namely the comparatively few cases available in the sample. By disaggregating descriptive statistics into different illness conditions ( Table 1), one can clearly see that a sickness condition can include as few as 533 observations (in the case of drugs), which in consequence leads to non-significant results in the FE regressions. Despite an overall high number of observations, coefficients in such cases are derived from smaller numbers of observations. Matching estimates do not obscure this fact, as the smaller numbers of observations in Tables 4 and 5 reveal. Table 4. Matching estimates: propensity score matching (PSM), nearest-neighbor matching (NN), and transitions into the sickness condition. Lower part of the table refers to changes in subjective health status. Condition t + 1 t + 2 Transitions PSM: ATT NN: SATE SE SE t-stat z-stat Obs Obs PSM: ATT NN: SATE SE SE t-stat z-stat Obs Obs Sick t + 1 Healthy t + 1 Sick t + 2 Healthy t + 2 Arms −0.4067*** 0.0288 −14.1176 15,491 −0.5331*** 0.0504 −10.5675 9379 5357 20,782 1970 13,746 −0.3012*** 0.0267 −11.2854 15,529 −0.3894*** 0.0480 −8.1170 9617 Sight −0.6319*** 0.0619 −10.2099 13,025 −0.6988*** 0.1066 −6.5543 6920 1906 20,782 491 13,746 −0.4372*** 0.0570 −7.6741 13,398 −0.4100*** 0.1196 −3.4271 8725 Hearing −0.5309*** 0.0574 −9.2487 13,126 −0.4053*** 0.0761 −5.3258 7679 1831 20,782 707 13,746 −0.3743*** 0.0565 −6.6269 13,440 −0.1833* 0.0881 −2.0806 8877 Allergy −0.4736*** 0.0406 −11.6533 13,898 −0.3769*** 0.0622 −6.0573 8916 2921 20,782 948 13,746 −0.3330*** 0.0362 −9.2038 14,050 −0.1740** 0.0582 −2.9896 9004 Chest −0.5915** 0.0528 −11.2005 13,544 −0.4700*** 0.0904 −5.1988 8629 2373 20,782 799 13,746 −0.4872*** 0.0456 −10.6913 13,682 −0.4173*** 0.0753 −5.5443 8905 Heart −0.5852*** 0.0465 −12.5942 14,089 −0.5493*** 0.0648 −8.4821 9150 3147 20,782 1449 13,746 −0.4501*** 0.0412 −10.9140 14,201 −0.4621*** 0.0605 −7.6404 9316 Stomach −0.6113*** 0.0572 −10.6946 13,647 −0.6598*** 0.0824 −8.0031 8767 2531 20,782 761 13,746 −0.5280*** 0.0435 −12.1323 13,837 −0.5468*** 0.0779 −7.0231 8895 Diabetes −0.5858*** 0.1048 −5.5886 11,883 −0.6941*** 0.1366 −5.0800 7100 431 20,782 275 13,746 −0.7552*** 0.1399 −5.3969 12,594 −0.7855*** 0.1800 −4.3625 8582 Anxiety −1.1024*** 0.0526 −20.9616 13,444 −1.2213*** 0.0932 −13.1027 8453 2385 20,782 712 13,746 −1.0980*** 0.0500 −21.9534 13,707 −1.0426*** 0.0898 −11.6128 8862 Drugs −1.3839*** 0.1692 −8.1802 8006 −1.0139*** 0.2745 −3.6933 958 171 20,782 33 13,746 −1.3846*** 0.2220 −6.2354 12,424 −1.0824* 0.4385 −2.4684 8428 Epilepsy −0.3906 0.2110 −1.8512 6360 −1.1355** 0.4382 −2.5913 702 85 20,782 34 13,746 −0.2013 0.3250 −0.6194 12,371 −0.5242 0.4468 −1.1733 8426 Migraine −0.5820*** 0.0475 −12.2527 13,299 −0.7317*** 0.0859 −8.5217 8446 2034 20,782 584 13,746 −0.4683*** 0.0441 −10.6136 13,465 −0.5920*** 0.0812 −7.2934 8750 Other −0.5678*** 0.0487 −11.6662 13,603 −0.5615*** 0.0978 −5.7399 8217 2237 20,782 445 13,746 −0.4338*** 0.0438 −9.9033 13,694 −0.4354*** 0.1014 −4.2953 8690 Cancer −0.7364*** 0.1176 −6.2605 8687 −0.5552** 0.1703 −3.2604 5737 364 20,782 152 13,746 −0.6012*** 0.1404 −4.2827 12.574 0.0200 0.2058 0.0970 8520 Stroke −0.6851*** 0.1539 −4.4524 6084 −0.9297*** 0.2689 −3.4578 4372 284 20,782 105 13,746 −0.1259 0.2354 −0.5348 12.514 −0.1206 0.3774 −0.3196 8477 Δhealth >+1 −0.0970* 0.0415 −2.3357 24,226 −0.0443 0.0634 −0.6988 19,644 21,015 46,142 593 31,077 0.0020 0.0467 0.0434 24,532 0.0386 0.0749 0.5156 19,825 Δhealth +1 −0.0440** 0.0165 −2.6678 30,896 0.0067 0.0237 0.2819 22,516 12,694 46,142 4950 31,077 0.0270 0.0160 1.6831 30,900 0.0611* 0.0244 2.5061 22,530 Δhealth −1 −0.2116*** 0.0172 −12.3089 30,880 −0.1433*** 0.0262 −5.4591 22,067 12,637 46,142 4250 31,077 −0.1486*** 0.0165 −9.0065 30,887 −0.0507 0.0269 −1.8866 22,102 Δhealth >−1 −0.4890*** 0.0433 −11.2935 24,550 −0.5948*** 0.1084 −5.4868 18,622 2188 46,142 330 31,077 −0.3556*** 0.0454 −7.8339 24,710 −0.5362*** 0.1275 −4.2072 19,648 Notes: Sample Average Treatment Effects (SATEs) and Average Treatment effects for the Treated (ATTs), with z-statistics in parentheses. Key to significance levels: *p < 0.05, **p < 0.01, ***p < 0.001. Table options Table 5. Matching estimates: recovery. Propensity score matching (PSM), nearest-neighbor matching (NN), and transitions into the sickness condition. Condition t + 1 t + 2 Transitions PSM: ATT NN: SATE SE SE t-stat z-stat Obs Obs PSM: ATT NN: SATE SE SE t-stat z-stat Obs Obs Sick t + 1 Healthy t + 1 Sick t + 2 Healthy t + 2 Arms 0.1568*** 0.0303 5.1792 11,538 0.2145*** 0.0401 5.3499 7882 4854 19,964 2357 17,424 0.2626*** 0.0314 8.3647 11,605 0.3242*** 0.0445 7.2880 8756 Sight 0.1120 0.0726 1.5429 1984 0.1992* 0.0966 2.0624 1222 1781 2749 1042 2109 0.0969 0.0684 1.4167 2005 0.1669 0.0868 1.9230 1323 Hearing −0.1002 0.0549 −1.8242 3420 −0.0586 0.0736 −0.7963 2288 1559 6040 723 5269 −0.0227 0.0572 −0.3970 3445 0.0317 0.0789 0.4017 2563 Allergy 0.0241 0.0383 0.6292 4871 −0.0400 0.0517 −0.7744 3035 2958 7695 1599 6147 0.0590 0.0386 1.5274 4895 0.0234 0.0533 0.4393 3415 Chest −0.0256 0.0465 −0.5518 5168 −0.0268 0.0626 −0.4274 3651 2256 9724 1154 8572 0.0070 0.0484 0.1451 5195 0.0871 0.0650 1.3403 4041 Heart −0.0090 0.0429 −0.2111 7391 −0.0308 0.0630 −0.4891 5321 2538 12,583 1112 11,227 0.0341 0.0447 0.7635 7421 −0.0039 0.0688 −0.0567 5823 Stomach 0.1541** 0.0528 2.9205 3215 0.2181** 0.0680 3.2078 1981 2363 4684 1360 3598 0.2384*** 0.0517 4.6101 3243 0.2737*** 0.0672 4.0752 2158 Diabetes −0.0177 0.1594 −0.1111 1620 −0.0473 0.2280 −0.2073 1136 179 3091 71 2980 0.1682 0.1792 0.9387 1706 −0.0206 0.3154 −0.0654 1515 Anxiety 0.5621*** 0.0589 9.5412 3196 0.7146*** 0.0825 8.6575 2021 2228 4938 1231 3909 0.7326*** 0.0572 12.8009 3212 0.8526*** 0.0738 11.5499 2229 Drugs 0.0069 0.3445 0.0199 165 −0.0182 0.4436 −0.0410 63 155 323 90 250 0.5941* 0.2461 2.4142 202 0.5966* 0.2847 2.0952 145 Epilepsy 0.1489 0.3769 0.3949 199 −0.1982 0.6252 −0.3171 37 81 638 40 595 −0.0029 0.3536 −0.0083 341 0.2995 0.3837 0.7804 288 Migraine 0.0908 0.0516 1.7596 3174 0.1300 0.0682 1.9061 1915 2181 4972 1160 3956 0.1538** 0.0497 3.0975 3179 0.2622*** 0.0670 3.9164 2212 Other 0.0067 0.0700 0.0954 1974 −0.2100* 0.0976 −2.1502 1235 2018 2138 1344 1419 −0.0338 0.0654 −0.5173 1996 −0.0484 0.0803 −0.6027 1358 Cancer 0.2127 0.1350 1.5758 490 0.2927 0.1893 1.5461 324 274 27,039 190 26,920 0.1949 0.1680 1.1603 4947 0.1691 0.1911 0.8845 4738 Stroke 0.3600* 0.1814 1.9845 361 0.3733 0.2628 1.4205 230 223 26,946 119 26,851 −0.2189 0.2286 −0.9574 4835 0.2054 0.3025 0.6793 4652 Notes: Sample Average Treatment Effects (SATEs) and Average Treatment effects for the Treated (ATTs), with z-statistics in parentheses. Key to significance levels: *p < 0.05, **p < 0.01, ***p < 0.001. Table options In order to obtain improved estimates of the causal impact of different health conditions on life satisfaction, we turn to our matching estimates (Caliendo & Kopeinig, 2008; Imbens, 2004; Rubin, 1974). Matching is an econometric technique that bears similarities to an experimental setup in medical research, with two groups of randomly-selected participants, whereby one is the control and the other the treatment group, which is subjected to medical treatment. Unlike clinical trials, however, matching estimators can be applied to observational data, to ensure that “treatment” and “control” groups are closely comparable in terms of observed characteristics. This means no actual trial is conducted, where randomly-chosen individuals are subjected to some “illness conditions” to identify the effects of these “treatments” on the participants' subjective well-being. Matching techniques applied to observational data can recreate a control group that is comparable to the treatment group in terms of observed variables, although we cannot entirely rule out differences between the control and treatment groups in terms of unobserved variables. To deal with this latter point, a conditional independence assumption (CIA) is invoked, which holds that the potential outcome (subjective well-being) and “treatment” participation (experience of the bad health condition) are independent for individuals with the same exogenous characteristics. CIA may be a strong assumption, and moreover it cannot be verified directly; only with reference to theoretical considerations of what drives treatment and outcome (in order to assure that CIA holds, we have selected our matching variables with respect to findings from the subjective well-being literature first, see Dolan et al., 2008, Graham et al., 2011, Helliwell and Wang, 2012, Veenhoven, 2008. Secondly our variables conform to different criterion schemes presented by the UN on factors driving bad health, see Lechner and Vazquez-Alvarez, 2011, pp. 395–396). The CIA is more reasonable for some conditions (e.g. “arms”) than for others, where decreases in well-being may for instance be endogenous to decisions to turn to alcohol or drugs (Veenhoven, 2008, p. 461). The second matching assumption is known as “overlap”, or the “common support condition.” This assumption ensures that individuals with the same characteristics have a positive probability of being either “participants” (i.e. becoming sick) or “nonparticipants” (not becoming sick). In further analysis we find considerable support for the common support condition (the methodological background of matching is further discussed in Binder & Coad, in press; Caliendo & Kopeinig, 2008; Oakes & Kaufman, 2006). We match our treatment group (those experiencing a change in health; more specifically, entry into a certain health impairment category) with a control group with unchanged health, at time t. We then track these individuals over time and observe differences between the treatment and control groups. Our analysis applies two different types of matching – propensity-score matching as well as multidimensional nearest-neighbor matching. Nearest-neighbor matching finds a match in many dimensions simultaneously, while propensity score matching collapses all covariates into one composite variable (the so-called “propensity score”). Our numbers of observations and variables allow us to use the same covariates for both matching estimators. Our covariates are: previous change in life satisfaction, log(income), gender, age, a quadratic age term, number of children, education, personality trait scores, dummies for being disabled, never married, separated, divorced or widowed, as well as being unemployed, retired, still studying or in school, on maternity leave or on family care, or self-employed; ethnicity dummies, year dummies, and regional dummies for former Metropolitan counties and Inner/Outer London. Table 4 shows the estimates. While the two matching algorithms produce similar results, we focus our interpretation mostly on propensity-score estimates.2 First, note the decreased number of observations. These are partly because matching estimators are clearer about which observations are used for the estimates (only the few occurrences of a sickness condition are presented here, as opposed to being obfuscated in a FE regression with many observations). Second, we only consider cases where individuals report sickness conditions for at least two periods, so that transitions from healthy to sick can be observed (for the second lag specification, this is extended to individuals reporting health conditions for three consecutive periods). Third, the matching algorithm allows us to discard observations where off-support inference would take place (individuals that are very different in terms of matching covariates are not compared with each other, in order to avoid “comparing apples with oranges”). These properties render matching estimates more informative than standard FE regressions. Note finally that it would be inappropriate to directly compare coefficient sizes between matching estimators and FE regressions (not because of the different sample sizes but) because matching estimates refer to total effects on life satisfaction while regression coefficients are ceteris paribus effect sizes, holding other variables of interest constant (Oakes & Kaufman, 2006, p. 382). By making treatment and control groups more comparable (finding the “perfect twin”, Almus & Czarnitzki, 2003, p. 231), we achieve significant results for most sickness conditions with the same dataset used for FE regressions. Carefully establishing a well-suited control group by discarding the “evil twins” thus increases precision and significance of our estimates. The lower part of Table 4 shows that the causal impact of a two-category (or more) decrease in subjective health assessment is highly significant (−0.49) and even increases after two years (−0.59). A smaller decrease in health (by one category) still affects subjective well-being quite strongly (−0.21 in t + 1, −0.14 in t + 2). With the above-mentioned caveat in mind, it is instructive to compare these effect sizes to those of the few other studies using matching estimators: the impact of self-employment on subjective well-being has been found to be around 0.11 ( Binder & Coad, in press), the effect of sustained volunteering at 0.11 ( Binder & Freytag, 2013), and getting unemployed decreases well-being by −0.72 (unpublished results by the authors). This shows that the effects we report here are large. Surprisingly, we cannot find a reciprocal effect of increased subjective health rating — the effect is negative in the first lag and sometimes insignificant. Further research could explore why hedonic adaptation to increases in health should be so strong. For our specific health impairments, a number of conditions display significant negative effects on subjective well-being. The strongest treatment effect concerns alcohol and drug abuse (−1.38), followed by anxiety, depression and other mental illnesses (−1.10), cancer (−0.74) and stroke (−0.69). Sight (−0.63), stomach (−0.61), chest (−0.59), heart (−0.59), migraines (−0.58), diabetes (−0.59) and the catch-all “other” condition (−0.57) also depress subjective well-being. Smaller causal effects appear for “arms” (that is, arms, legs, hands, feet, back, or neck problems; the estimate is −0.41) and for hearing (−0.53) and allergy problems (−0.47). Comparatively severe health impairments such as epilepsy (in the first lag) yield no significant results, however, but a strong negative impact in the second lag, despite a minuscule sample size of only 34 individuals who transitioned into the condition and remained there for two years (see the last columns of Table 4). Our results can be related to the few studies' results that also addressed the impact of objective health conditions on subjective well-being. Shields and Wheatley Price (2005) found a strong negative (cross-sectional) association between mental well-being and migraines, heart-conditions-and-stroke, and epilepsy. Graham et al. (2011) found strong negative impacts of anxiety and strong pain for Latin American countries (also cross-sectional). Opposed to severe adverse physical conditions, extreme pain and anxiety in their study remained significantly associated with unhappiness even after including an optimism personality variable (in an attempt to control for individual fixed-effects). These independent findings support the conclusion that adaptation is easier for physical conditions than chronic pain, or psychological conditions such as anxieties (Dolan, 2011). Even if personality traits mediate problems of bad health and their impact on individual life satisfaction, this is less true for the above-mentioned health conditions. Our study goes beyond both studies in establishing that, in many objective health conditions, there is a strong significantly negative effect on life satisfaction (even after matching individuals by personality traits). We can corroborate the one consistent finding from the aforementioned studies, that mental health matters greatly for subjective well-being, and adaptation to it is not easy (the effect increases in the second lag, see below). However, our findings extend beyond the few studies tackling objective health conditions in that we can establish clear negative impacts of other physical ailments that (substantially) decrease subjective well-being even when taking personality into account. The high negative impact of drug abuse on subjective well-being is a case in point, providing further evidence against theories of rational addiction. But also cancer and stroke are physical conditions that severely impact individuals' subjective well-being (in the case of stroke, even increasing over time). In this respect, our findings show the limits of the current interpretation in the literature that physical impairments are less relevant to subjective well-being. What seems more plausible is that “concrete” ailments play an important role, and physical conditions relating to arm problems or allergies might indeed have less hedonic impact than mental problems, but that severe physical impairments (e.g. stroke, cancer) approach the impact that anxieties or migraines can have on the individual. Clearly, further research should delve deeper into these differences in health conditions. It also should be noted that our estimates are conservative in the sense that they might underestimate the impact of these health conditions on life satisfaction because of attrition – if an illness is so severe that it hinders the individual from responding, the existing sample might represent the comparatively less severe cases of bad health conditions. We cannot completely rule out this source of downward bias, but in general, a decreasing health condition has been shown not to affect response rates in the BHPS (Uhrig, 2008, p. 28). We investigated attrition bias in several ways (available on request): we re-estimated Table 4 focusing on transitions into a sickness condition for these individuals who get sick in t + 1, yet recover in t + 2. These individuals represent the (arguably) less severe or short-term cases in the respective sickness categories. For cancer, anxiety, diabetes and sight problems, the lighter cases lead to noticeable reductions in coefficient size; for the other cases, the reduction is modest (−0.10 or less). As we are interested in the dynamic aspects of well-being, we also investigate lagged effects of these health conditions. A robust finding in happiness research is that individuals often adapt to changes in life circumstances. Hedonic adaptation, the hedonic dulling of repeated or constant affective stimuli (Frederick and Loewenstein, 1999) is highly domain-specific, varying with the concrete stimulus (e.g. hedonic adaptation to marriage is faster and more complete than hedonic adaptation to repeated unemployment, see Clark, Diener, Georgellis, & Lucas, 2008; Dolan & Kahneman, 2008). These relationships are complicated by different effects taking place, in the case of unemployment, for example, a “saddening effect” of unemployment can be mitigated by a “time composition effect” when the unemployed can pursue more leisure activities (Knabe, Rätzel, Schöb, & Weimann, 2010, p. 868). Our panel dataset allows us to include a second year to check for hedonic adaptation. In three cases, the effect remains comparable (sight, stomach, and the “other” category). For other conditions, we find a few cases with significant changes in life satisfaction two years after the individual became ill. In many cases, the impact of the health problem becomes smaller (cancer, hearing, allergy, chest, heart and drug abuse). In other cases, however, the point estimates increase at the second lag (arms, diabetes, anxiety, epilepsy, migraine, and stroke), which means that the health impairment's negative effect increases with time. We attribute this increasing impact to a gradual worsening of the health conditions (e.g. progressive diseases/health impairments) in some cases. The strong deterioration in well-being caused by epilepsy is particularly striking in the second year. These findings underline how specific the phenomenon of hedonic adaptation is in the health domain (Dolan & Kahneman, 2008, pp. 218–9). Note that dynamic effects vary only in a small number of health conditions when considering the nearest-neighbor-matching estimates, suggesting our estimates are robust. Finally, we examine to what extent individuals recover their lost life satisfaction after recovering from health impairments (Table 5). In line with the asymmetric finding regarding positive (subjectively-assessed) health changes, it is striking to observe that transitioning out of different health conditions does not usually lead to significantly higher life satisfaction in subsequent years (with the exception of conditions such as anxiety, stomach, arm problems but also migraines and strokes). Overall it seems that “objective” physical conditions (problems with arms, sight etc.) have smaller negative impacts, and that subsequent recovery brings less noticeable improvements in life satisfaction. Mental conditions, on the other hand, apparently lead to much stronger decreases in life satisfaction and exhibit more pronounced recovery patterns. Graham et al. (2011) conjecture that it might be easier to adapt to such “objective” physical conditions than to mental problems such as anxiety, which would explain our findings. Due to our dataset's lag structure, however, we cannot say whether the positive effect on life satisfaction after recovery never occurs, or whether it occurs within a year and the individual has adapted after just one year. It could also occur after two years, when recovery takes longer to fully set in; this alternative explanation seems unlikely, however – it is not clear why there should be a delay of more than two years between absence of the sickness condition and any resulting increase in subjective well-being. Pain or negative health impairments do have — by their biological origin and purpose — a higher behavioral relevance and it seems that nature has endowed individuals with the corresponding mechanism that we might call a “psychological immune system” (Dolan & Kahneman, 2008, p. 222): going into states of ill-health decreases well-being much more strongly than subsequent recovery, presumably to motivate the individual to modify behavior accordingly.