Methods: How can drivers of subjective well-being be analysed?
Data requirements, survey design principles and causality
Research aimed at better understanding the drivers of subjective well-being requires the inclusion of a wide range of co-variates in the analyses, including a number of standard demographic and control variables (described in Chapter 3), as well as measures of the drivers of interest, and their potential co-variates. Examples of key variables of interest are outlined in Diener, Diener and Diener (1995), Dolan, Peasgood and White (2007), Fleche, Smith and Sorsa (2011), and Boarini et al. (2012). Analyses of drivers require access to micro-level data, and may be undertaken by government analysts, researchers and organisations or institutes with an interest in informing government policy, academic enquiry and public discourse, as well as in organisational well-being (including business approaches to employee well-being).
Ideally, research into the drivers of subjective well-being should utilise data that enable some inferences about the causality of relationships between variables. Although the gold standard for determining causality involves experimental manipulations of hypothesised drivers under controlled conditions, this is close to impossible for most of the policy-relevant determinants of life evaluations and eudaimonia in particular. The model scenario for determining causality in real-life settings therefore tends to be randomised controlled trials (RCTs), which involve the random allocation of individuals into groups, each of which are assessed before and after receiving a different treatment (e.g. one group receives intervention A, one group intervention B, and a third group acts as the control, receiving no intervention). In practice, for many of the potential policy drivers of subjective well-being, RCTs are also rare, particularly in terms of ensuring perfect randomisation and double-blind or single-blind conditions.30
Quasi-experimental designs refer to “natural experiments” where a group of respondents exposed to a particular intervention can be matched with and compared to a similar group of respondents not exposed to that intervention. However, investigators tend to have little control over either the level of variation in the determinant of interest, or in the allocation of treatment groups, which is rarely completely random. It may also be difficult to obtain pre-intervention outcome measures - thus inferences may have to be based on measures collected only after the event has occurred. However, quasi-experimental designs do offer some advantages over RCTs, particularly where it would be unethical to randomise treatments, and/or where experimental designs can be challenged in terms of their real-world applicability.31 Quasi-experimental designs can also enable researchers to draw on much larger and more representative data sets, including national-level data.
Quasi-experimental designs typically require panel data (longitudinal surveys collecting repeated measures of individuals over time) or the collection of pre- and post-data for the populations of interest. These data offer the opportunity to explore whether a change in the level of a given determinant is associated with a subsequent change in subjective well-being over time.32 While panel data do not enable the researcher to experimentally manipulate the main variables of interest, and panels can suffer from attrition, this approach has the benefit of being able to utilise data sets from large and high-quality samples such as those obtained by national statistical agencies - thus enhancing the representativeness of the sample, and the generalisability of the findings.
Large sample sizes are particularly important for detecting the impact of minor drivers and/or drivers that typically affect only a small proportion of the overall population. In comparison to more experimental methods (such as RCTs), observational data also carry less risk of experimental demand characteristics (e.g. Hawthorne or placebo effects), where a respondent’s knowledge that he or she is part of a special treatment group may influence subjective well-being outcomes and/or how they are reported. The same is true of international comparisons, which offer another form of natural experiment, where a particular intervention has been applied in one country but not in another. However, it is very difficult to infer causality from international comparisons of cross-sectional (rather than longitudinal) data, given the variety of uncontrolled differences between countries in terms of both sample characteristics and other variables of interest.
At present, the majority of studies investigating the drivers of subjective well-being tend to rely on cross-sectional data, simply because these are the most widely-available data sets. Strictly speaking, these analyses are concerned with co-variates rather than drivers, although the term drivers will be used in the sections that follow to denote the underlying intention of the analyses described. Cross-sectional data do not enable causal inferences to be made directly, but can be interpreted alongside evidence about the direction of causality from other sources.