How Much Variance in Chronic Happiness Levels can be Explained by Genetic Factors?
It appears that Lyubomirsky et al. (2005) derived their figure of 50% for the heritability of well-being by combining, in a way that was not clearly specified, the results from three studies (Braungart et al. 1992; Lykken and Tellegen 1996; Tellegen et al. 1988). Lykken and Tellegen estimated that “the heritability of the stable component of subjective well-being approaches 80%” (p. 186). Tellegen et al. examined general personality traits in adult twins and determined that heritability ranged from 39 to 58% with an estimate of 48% for the personality trait labeled “Well-Being”. In contrast, Braungart et al. measured the influence of genetic factors on externally-reported developmental and emotional behaviors in children aged 1 or 2 and found heritability ranging from 35 to 57%. Even taking the lower bound of the estimated range of heritability from this study (which did not measure adult well-being), we arrive at an average figure of around 61%, whether or not the studies are weighted to take into account their sample sizes. About the best that can be said, then, for the figure of 50% for genetic factors is that it is surrounded by considerable uncertainty, and appears—based on the studies cited by Lyubomirsky et al.—likely to be a lower bound. Of course, any upward adjustment of this number would necessarily imply a downward adjustment of the percentage of variance attributable to volitional activities, if we accept the subtraction logic suggested by Lyubomirsky et al. Note that current overviews of the literature locate the heritability of overall (i.e., including momentary) happiness at 32–41%, but for the stable component of happiness (i.e., the chronic levels that are the outcome of interest in Lyubomirsky et al.’s model) heritability is reported in the 70–80% range (Nes and Røysamb 2017).
Origins of the Claim that Only 10% of the Variance in Well-Being is Due to Life Circumstances
Lyubomirsky (2007, p. 41) claimed that the figure of 10% for the amount of variance in happiness explained by life circumstances “represents an average from many past investigations, which reveal that all life circumstances and situations put together account for only about 10 percent in how happy different people are.” Lyubomirsky et al. (2005) cited two sources in support of the 10% claim: a book chapter by Argyle (1999) and an article by Diener et al. (1999). Argyle in fact gave two figures on his pp. 353–354: “less than 10 percent of the variance,” which he attributed to Andrews and Withey (1976), and 15%, which he reported as having been proposed by Diener (1984). The first of these figures appears to correspond to the following quote from Andrews and Withey (1976, p. 109): “The demographic variables, either singly or jointly, account for very little of the variance in perceptions of global well-being (less than 10 percent)”, while the second appears to refer to this: “individual demographic variables rarely account for more than a few percent of the variance in SWB, and taken together probably do not account for much more than 15% of the variance” (Diener 1984, p. 558). Notice that both of these quotes refer to demographic variables, which, crucially, are not the same as life circumstances, as we discuss below.
In the second article cited by Lyubomirsky et al. (2005) and Diener et al. (1999, pp. 278–279) reported three candidate percentages for the variance in happiness explained by circumstances: 20% according to Campbell et al. (1976), 8% according to Andrews and Withey (1976), and the same 15% mentioned by Argyle that was cited separately by Lyubomirsky et al. (which was, of course, Diener’s original informal estimate from 1984). Later in their article, Lyubomirsky et al. stated that “all circumstances combined account for only 8% to 15% of the variance in happiness levels” (p. 117), again with citations of Argyle (1999) and Diener et al. (1999); for some reason they omitted to mention the figure of 20%, derived from Campbell et al.’s study, which Diener et al. had cited in the sentence immediately preceding their discussion of the 8% and 15% figures. Lyubomirsky et al. did not report how they finally settled on their “headline” figure of 10%, but we tentatively assume that this was derived by some sort of informal interpolation between 8 and 15%. In summary, it seems that there is a considerable lack of clarity about how the (subsequently highly influential) figure of 10% was arrived at. Hence, we turned to the original sources to look for this information.
First, we examined Andrews and Withey’s (1976) book, which describes the results of a large, multiple-occasion study involving an overall total of 5142 US Americans. Of interest here are the surveys that were conducted in May 1972 (N = 1297), and April 1973 (N = 1433), which were the ones used by Andrews and Withey (1976, pp. 138–142) to illustrate their discussion of the limited value of “classification variables” in predicting overall life satisfaction. We found two tables (Exhibit 4.6 on p. 139 and Exhibit 4.7 on p. 141) showing that the percentage of variance explained by “six classification variables” was 8% in the May 1972 survey and 11% in the April 1973 survey. These “classification variables” are age, family income, education, race, sex, and “family life-cycle stage”, with the last of these being a categorical variable based on “the respondent’s own age and marital status, and the age of the youngest child living in the family” (Andrews and Withey 1976, p. 138). These variables, which Andrews and Withey (1976, p. 138) described as “demographic or social characteristics,” were used mainly as the basis of classifying their subjects into subgroups; it is clear, using Lyubomirsky et al.’s (2005) own definition (which we discuss below), that they do not constitute an exhaustive inventory of “life circumstances,” sufficient to allow conclusions to be drawn about the effect of such circumstances on an individual’s well-being (Fig. 2).
Turning to Campbell et al.’s (1976) book, which was based on a survey of 2147 US Americans conducted in 1971, we found a number of tables (e.g., those on pp. 230, 241, 253, 279, and 302) that explored the relation between satisfaction with an individual domain of life (e.g., “housing” or “life in the United States”) on the one hand, and participants’ subjective assessments of that domain and their “personal characteristics” on the other. The percentage of variance explained by the combination of subjective assessments and personal characteristics in these tables is typically in the range of 20–30%, with personal characteristics typically accounting for around 8–10%, but—as with Andrews and Withey’s (1976) “classification variables”—the lists of these characteristics were short (five or six items), with race and (where present) age and sex typically having the largest partial correlations with the relevant satisfaction measure. Indeed, Diener et al. made it clear that their estimates of the variance explained in these two books were based on “demographic factors,” not an extensive inventory of life circumstances:
Campbell, Converse, and Rodgers (1976) found that demographic factors (e.g., age, sex, income, race, education, and marital status) accounted for less than 20% of the variance in SWB. Andrews and Withey (1976) could only account for 8% by using these variables. (Diener et al. 1999, pp. 278–279)
In summary, although it is not clear exactly how Lyubomirsky et al. (2005) arrived at their estimate that only 10% of variance in well-being is explained by life circumstances, the available evidence suggests that this well-publicized figure may have been derived from an erroneous conflation of “classification variables” or “demographic factors” with the much wider category of “life circumstances,” in a few tables of results in two books describing studies that were conducted by researchers at the same institutionFootnote 5 (the Institute for Social Research, Ann Arbor, Michigan) in the early 1970s. Whether or not this constitutes “many past investigations,” as claimed by Lyubomirsky (2007) in the sentence quoted at the start of this section, is perhaps a matter of opinion. However, given that these sources provided the empirical evidence that was ultimately used by Lyubomirsky et al. to support their claims about the effect of life circumstances on well-being, it would seem interesting to examine the evidence provided by these studies, while not losing sight of the limitations just mentioned. We were able to obtain the data sets on which Andrews and Withey (1976) and Campbell et al. (1976) based their books, and reanalyze them to see what percentage of variance in these authors’ respective life satisfaction outcomes was explained by a set of 15 to 18 variables (depending on the exact questions asked in each survey) that were agreed by external evaluators to correspond to “life circumstances.” With the same statistical techniques that had been used by the original authors to arrive at the figure of 8–10% for demographic variables, we obtained values of between 18.13 and 26.47% for life circumstances. (A more modern approach using cross-validation results in a wide range of estimates, between 1.92 and 17.90%, highlighting how such estimates depend both on the statistical methods employed and on the variables available in the datasets.) Further details are provided in the Supplementary Information of the present article.
Measures of Well-Being are Specific to Given Populations and Periods of Time
Even if one were to assume that the variance decomposition suggested by Lyubomirsky et al. (2005) was valid (which we questioned above), it is not clear that these numbers still hold today (or, indeed, whether they did so when Lyubomirsky et al.’s article was published). Campbell et al. (1976) collected their data in 1971, while Andrews and Withey (1976) collected theirs in 1972–1973. This was a time when baby boomers were reaching adulthood and the Vietnam War was at its height. The relevance of these studies to the modern social landscape probably cannot be taken for granted. For example, the participants were between 87 and 90% White, with the rest being mostly reported as Black and no more than about 1% described as being of another race, while the demographic questions make it clear that the default expectation of the researchers was that cohabiting adults would be heterosexual and probably married. It is not clear whether the same results would be found were these studies to be repeated today, after 40 years of social change in the United States and with considerably greater diversity in the population and people’s lifestyles. Any estimate of the percentage of variance that can be explained by a certain combination of predictors is only meaningful with respect to the population from which the sample has been drawn. Of course, in the period spanning almost half a century since these studies were conducted, other researchers have collected data that show the relation between demographic variables and well-being (see Diener et al. 2018 for a review), but as we have already noted, such relations tell us little about the influence of variables that fall under the much wider category of life circumstances.
A further caveat to note here is that when estimating the amount of variance in an outcome that can be attributed to certain predictors, analyses are necessarily constrained by the amount of variability in those predictors (as already explained for the case of volitional activities above). For example, broader economic circumstances are often constant within a survey at a given time point in a given sample. However, changes in these circumstances can have remarkable effects on well-being, as indicated by the World Happiness Report (United Nations 2015), which is based on the Gallup World Poll. This report periodically measures the self-reported well-being of the inhabitants of most of the countries of the world on a “ladder” from 0 to 10 (Cantril 1965). The 2015 report showed that between the measurement periods of 2005–2007 and 2012–2014, self-reported well-being in Greece fell from 6.327 to 4.857, a drop that represents almost one and a half steps down the ladder from the best possible life towards the worst. This reduction of 1.470 points also corresponds to 31.0% of the range of country-average well-being scores in the 2015 report from the highest (Switzerland, at 7.587) to the lowest (Togo, at 2.839). It seems unlikely that more than a trivial part of this decline in well-being over the better part of a decade can be accounted for by a change in the genetic characteristics of the population of Greece, or a decision by Greek citizens to eschew en masse the forms of intentional activity that might contribute to their individual happiness. Rather, the drastic change in their (economic) circumstances forced on the majority of Greeks by the ongoing financial crisis affecting their country would seem to be by far the most plausible candidate to explain this drop in national well-being (cf. Rhodes Hatzimalonas 2017). Of course, one could argue that these changes are in some sense transitory, and ought to fade over the course of the years as the Greek economy recovers. However, if such a multi-year criterion is to be used to distinguish between mere fluctuations in circumstance and true changes to people’s chronic level of well-being, it should logically also be applied to intervention studies claiming to produce sustainable changes in happiness.
The wide range of scores between the top and bottom countries on the World Happiness Report list is another indication that there might be many circumstantial factors that affect well-being beyond those that were (or, indeed, could have been) included by Andrews and Withey (1976) and Campbell et al. (1976) in their surveys of US Americans. On a continuum from the smaller democracies of northern Europe at one end through to the poorest countries of sub-Saharan Africa at the other, there are considerable variations in political freedom or repression, institutional transparency or corruption, and the availability or lack of basic services such as electricity and clean drinking water, to name just a few examples of factors that might be expected to influence well-being, but would not be expected to vary much among any sample drawn from just one country. Indeed, Lyubomirsky et al. (2005, p. 117) acknowledged that “Happiness-relevant circumstances may include the national, geographical, and cultural region in which a person resides,” but these authors then failed to limit their own claims to the specific population from which the data were drawn (i.e., US Americans from the 1970s). We suggest that future research into the effects of life circumstances on well-being should take into account potential cross-cultural concerns and acknowledges the limitations of the samples being studied. For example, it has recently been suggested that researchers should explicitly limit conclusions to specific target populations to avoid unjustified generalizations (Simons et al. 2017).
Likewise, researchers should carefully reconsider the usage of terms such as “volitional activities”. What might be a matter of choice for some people can be a question of circumstances for others. For example, middle-class Western parents with a comfortable household income can exercise a wide range of choice over the diet that they and their children consume; they can choose to drive their car to the farmer’s market instead of taking the bus to the discount grocery store, and afford to place organic blueberries rather than canned, syrup-laden fruit cocktail in their shopping cart. They can also enjoy a wide range of recreational activities, such as vacations to awe-inspiring or culturally important destinations, enabling them to enjoy the benefits of feeling that they are part of a meaningful world (Gilovich and Kumar 2015). Indeed, Lyubomirsky et al. (2005) themselves described the types of volitional activities that they were considering in decidedly middle-class terms: “rather than running on a track, a fitness-seeking wilderness lover might instead choose to run on a trail through the woods… rather than learning classical pieces, a jazz-loving piano student might instead choose to work on jazz standards” (p. 122). A subsistence farmer working from dawn to dusk in a developing nation, a worker who spends 14 h per day assembling electronic devices in a south-east Asian factory, or a single Western parent living in less fortunate circumstances than the family just mentioned, may not be in a position to make such choices. As with many phenomena studied by primarily Western psychologists (cf. Henrich et al. 2010), the list of volitional activities that are available to improve one’s well-being may be less than universal, not just for cultural reasons but also precisely because of one’s life circumstances.
Moving Forward: Between-Subjects Versus Within-Subject Designs and Causal Inference
We noted earlier that there appeared to be a mismatch between the evidence (such as it was) provided by Lyubomirsky et al. (2005) for their model of variance decomposition—based on between-subjects studies—and the suggestion that individuals can improve their happiness with volitional activities, which is inherently a within-subject effect. Such limitations of positive psychology research have been noted numerous times in the past (e.g., Lazarus 2003; Nickerson 2007, 2014), and calls for more longitudinal, within-subject research are not uncommon even in the discussion sections of studies based on cross-sectional, between-subjects data. In fact, it appears—at least from an informal examination—that within-subject designs may be becoming more popular within positive psychology. In early December 2018, we examined the 50 most recently published articles in the Journal of Happiness Studies that reported empirical data, and compared these with a random selection of 50 articles published in the same journal approximately 10 years earlier. Whereas in 2007–2008 only seven out of 50 articles used within-subjects designs, by 2017–2018 this number had risen to 16,Footnote 6 with a variety of methods being used. For example, on a shorter timescale, data collected with the Experience Sampling Method (Csikszentmihalyi and Hunter 2003) were used to investigate how detachment from workplace stresses affect the quality of interactions in romantic relationships (Debrot et al. 2018), while on a longer timescale, data collected as part of a panel study of 1500 households in Nepal were used to investigate the cost of coping on different forms of well-being (Chindarkar et al. 2018).
Part of the appeal of within-subject designs is that they can, if properly analyzed, control stable inter-individual differences, rendering a causal interpretation of associations more plausible (although such designs do not automatically warrant causal conclusions, as time-varying confounders can offer alternative explanations). Causal inference might not always be the explicitly stated goal of empirical investigations of subjective well-being, but it often seems to be what authors are most interested in; indeed, Lyubomirsky et al. (2005, p. 116) referred to the genetic set point, life circumstances, and intentional activity as the “three primary types of factors that we believe causally [emphasis added] affect the chronic happiness level.” For example, if there is a between-subjects association between a certain activity and happiness (i.e., on average, people who do X are happier), but no within-subject association (i.e., a person who starts doing X does not become happier), we would be quick to dismiss the between-subjects association as spurious and refrain from advising people to do more X to become happier.
However, this advantage becomes a complication when the aim is to give a holistic picture of all of the determinants that explain happiness (whether this is measured over an extended period of time or as short-term fluctuations). Many inter-individual differences, such as personality or socio-economic status, hardly vary within individuals, making it difficult to establish any within-subject association with the outcome. Even research on fairly common life events such as marriage or childbirth makes use of the massive sample sizes of ongoing panel studies to ensure that sufficient individuals have actually experienced the event in question while they were part of the study.
Of course, the above comments merely underscore the self-evident truth that no single type of research design can answer all questions. Within-subject designs may pave the way for interesting new research questions with a different scope, such as whether individuals reliably differ in the volatility of their well-being, or in their reactivity to stressors, and which factors shape these inter-individual differences in intra-individual variation. However, the complex nature of human experience is such that data collected over a range of timescales, from a few minutes to many decades, will likely be needed to fully identify the determinants of well-being as individuals experience daily existence, major life milestones, and the general aging process (as well as any psychological interventions that we might subject them to). To both enhance and complement these efforts, it could be fruitful if researchers working on subjective well-being considered more explicitly whether they are trying to draw causal conclusions (the answer might be “yes” more often than not), and, if so, made use of systematic and established approaches from other fields of research. For example, counterfactuals (e.g., Morgan and Winship 2015; Neyman 1923/1990), directed acyclic graphs (e.g., Kuipers et al. 2018; Pearl 2000), and various types of natural experiments (e.g., Dunning 2008; Rutter 2007) might have been under-utilized in psychological well-being research (likely because they are not routinely taught to psychologists). A more careful consideration of the admittedly complex issue of causality might lead to the sobering conclusion that a complete decomposition of happiness into genes, circumstances, and intentional activities is not only an elusive research goal, but not even a well-defined one.