Interpreting drivers of subjective well-being
Two key questions for interpreting analyses of the drivers of subjective well-being are what size of impact can be expected? and how can the impacts of different drivers be compared?. The first question needs to be considered with reference to the overall sensitivity of subjective well-being measures, as well as the time frames for the analysis. The second question raises issues about the interpretation of regression coefficients - including problems of correlations among independent variables, the effects of unmeasured or omitted variables, the possibility of shared measurement error (or method variance) between variables and outcomes, the possible presence of reverse or two-way causality between variables and outcomes, and the generalisability of results.
What size of impact can be expected?
There are a priori reasons not to expect large movements in subjective well-being data as a result of single drivers. Many of the interpretive issues described in the first section of this chapter (looking at basic patterns of change over time and differences between groups) also apply to analyses involving the drivers of subjective well-being.35 Factors such as the initial distribution of responses, the proportion of the sample affected, the number of significant drivers in the model, frame-of-reference effects and adaptation can all limit the impact on subjective well-being that one might expect as a result of any one individual driver, as well as the size of sample required in order to detect that impact.
Senik (2011) notes that the typical model R2 value of an ordinary least squares estimate of life evaluation varies between 3% and 15%, depending on the control variables and drivers included in the model and the sample size. Drawing on two waves of Gallup World Poll data (2009 and 2010) from all 34 OECD countries and including measures for a large number of key known drivers of subjective well-being, Boarini et al. (2012) obtained R2 values of 0.35 in the case of life evaluations, and 0.19 in the case of affect balance. Fleche, Smith and Sorsa (2011) report cross-country comparisons of key life satisfaction drivers over two to three waves of data (from the World Values Survey 1994-2008) for 32 different countries and find R2 squared values ranging from 0.40 in New Zealand to 0.14 in Turkey, with an OECD average of 0.22. In Helliwell, Barrington-Leigh, Harris and Huang’s (2010) analysis of data from a global sample of between 50 000 and 140 000 respondents in 125 countries, income plus a range of social and cultural variables explained between 30 and 45% of the individual-level variance in life evaluations.
Given the very large number of potential drivers of subjective well-being, these R2 values suggest that the proportion of variance explained by any one driver is likely to be small. Furthermore, if the initial amount of variability in a given driver is itself limited (for example, because only a small proportion of the sample is affected by it), then the proportion of variability in the subjective well-being outcome it can explain will also be limited. The key statistic of interest in the analysis of drivers, however, is not the overall model R2, but the size and significance of the individual regression coefficients associated with each driver - which (in the absence of correlations among independent variables) indicate how much the dependent variable is expected to increase or decrease when the independent variable increases by one unit.
As noted in Section 1, a mean change of 0.3 or 0.5 scale points on a 0-10 life evaluation scale may represent a very sizeable result that one might expect only in response to major life events at the individual level. Between countries, differences may be larger due to the cumulative impact of differences across a wide range of subjective well-being drivers.
Using an appropriate time horizon for analysis is another key consideration when interpreting effect sizes. As with any attempt to evaluate the impact of a driver, it will be important to consider the mechanisms through which that driver is assumed to operate - and how long it may take for these effects to emerge. Due to psychological resilience and adaptation over time, the immediate impact on subjective well-being for some life events and interventions may be greater than the impact several years down the line. As noted previously, the variables that influence the rate and extent of adaptation to life events over time may be of key interest to policy-makers, and thus adaptation should be considered a feature rather than a “bug” in subjective well-being data. Nonetheless, the process of psychological adaptation means that close attention to time horizons is warranted, in particular to avoid misinterpreting the effects of exogenous events.