The temporal theory of regret (Gilovich and Medvec, 1994, 1995) is one of the most well-known regret theories and suggests that lifespan changes in regret intensity are driven by the nature of the regrettable decision itself (i.e., whether the regret relates to an action or an inaction). Specifically, the temporal theory of regret suggests that “[a]ctions produce greater regret in the short-term, whereas inactions generate more regret in the long run” (Gilovich and Medvec, 1994, p. 361). Gilovich and Medvec (1995) propose a variety of mechanisms that might explain why regret for inactions might increase over time compared to regret for actions, including the idea that outcomes resulting from mistaken actions can be ameliorated by opportunities afforded by the passage of time, whereas there may not be similar opportunities to correct the outcomes of regretted inactions.

Another key contemporary theory of regret is decision-justification theory (Connolly and Zeelenberg, 2002), which identifies the underlying decision rationale as key for regret generation. Similarly to other theories of regret, decision-justification theory proposes that one key influence on regret is the outcome of the decision: We feel more regret over decisions whose outcomes compare poorly to those of options foregone. However, decision-justification theory suggests furthermore that the intensity of regret felt also depends on the degree to which the individual identifies the decision as justifiable: A decision that results in a poor outcome will cause less regret if the decision was, in retrospect, still justified.

In contrast to both the temporal theory of regret and decision-justification theory, a third “belonging” theory of regret (Morrison et al., 2012) identifies the context of the regrettable decision as foundational in driving regret intensity. Specifically, Morrison et al. (2012) found that regrets in domains particularly relevant to our senses of social belonging (e.g., romantic or family domains) are more intensely felt than those in less social domains (e.g., work, education), suggesting that an individual experiences more regret when a poor decision is perceived to represent a threat to their sense of social belonging.

In sum, each of these three theories incorporates unique hypotheses concerning the key predictors of regret intensity; our study thus seeks to test these hypotheses. In addition, there is little evidence in the extant literature that the various predictors of regret specified in these three theories have been compared with each other in a single statistical model in order to assess their relative importance in generating regret intensity. Our study thus seeks to simultaneously assess the effects of these various predictors of regret intensity (whether a regret relates to an action or inaction, how justified the original decision was, and whether a regretted decision was in a social domain), and thereby determine which are the strongest drivers of regret intensity.

As noted above, one theory tested in this study is Connolly and Zeelenberg’s (2002) decision-justification theory. Connolly and Zeelenberg note, however, that little is known regarding the nature of justification generation. They suggest that one particularly effective form of justification is the presence of “a careful, competent decision” (Connolly and Zeelenberg, 2002, p. 215)”. However, justifications need not always be rational or explicit: Research also shows that justifications based on our implicit personal values or life-rules affect our experience of regret (Seta et al., 2001; Lönnqvist et al., 2006). This finding reflects core notions in theories of decision-making—like the theory of Regulatory Fit (Higgins, 2000; Camacho et al., 2003)—that maintain that consistency with our implicit values is an important factor in decision making. It is therefore feasible that decision justifications present themselves in two forms: explicit and implicit. To this end it seems likely that explicit axiom-based justifications, as described by Connolly and Zeelenberg (2002) shield one from regret, but that decisions lacking an explicit justification might still incur little regret if they were deemed consistent with an individual’s values or rules (i.e., a justification that is implicit in the sense that it stems from the consistency of an action or inaction with one’s individual life rules). We therefore hypothesized that regretted decisions that were consistent with participant’s personal life rules would be less intensely regretted.

An important issue with respect to the theories mentioned above is that while the theories discussed focus on explaining variation in the intensity of regret, much of the existing empirical research concerning these theories has actually assessed the frequency with which reported regrets fall into particular categories—rather than explicitly measuring regret intensity. For example, Gilovich and Medvec’s (1994) studies analyzed the frequency with which participants’ most greatly regretted decisions fell into particular categories (e.g., action/inaction), rather than directly measuring or analyzing regret intensity. More recent studies similarly have often focused on the frequency with which participants’ regrets fall into particular categories or life domains (e.g., Roese et al., 2009; Morrison and Roese, 2011; but c.f. Wilkinson et al., 2015). Emotional frequency and intensity are distinct processes underlying the experience of affect (Diener et al., 1985; Schimmack and Diener, 1997) and they are not necessarily highly correlated: Frequent emotions (e.g., happiness) may not necessarily be intense, while infrequent emotions (e.g., rage) may be very intense. Studies assessing regret frequency as a proxy for regret intensity are therefore potentially misleading, and this may explain why the few studies directly distinguishing regret frequency and intensity found inaction regrets to be more frequently reported but no more intense than action regrets (Feldman et al., 1999; Bonnefon and Zhang, 2008). This study will therefore focus on reported regret intensity (rather than frequency) in order to provide conclusions as to the key drivers of regret intensity across the lifespan.

In this study, we seek to test the temporal, decision-justification, and belonging theories of regret so as to determine which characteristics associated with a regretted decision are most strongly associated with the intensity of regret. Predictions implied by these theories have been tested in the past, but each theory has generally been tested in isolation. We also deliberately focus on studying the intensity of regretted decisions as opposed to simply the frequency with which regretted decisions of particular types are reported. Finally, we explore the effects on regret intensity of different forms of justification; namely, both explicit justification and implicit justification (i.e., whether the regrettable decision was consistent with the individual’s implicit life rules).

The following predictions were tested in relation to participants’ self-reported greatest lifetime regrets:

Three thousand individuals were randomly selected from the New Zealand electoral roll and invited to take part in a postal survey. The initial participant pool size of 3,000 was chosen on the basis that this number was considered sufficient to ensure the return of at least several 100 completed surveys for analysis.

A total of 677 surveys were returned. Of these, 177 did not contain responses to any of the items analyzed in this study, and were excluded from all analyses reported here. A large part of this apparent missing data was due to a screening item early in the questionnaire that asked: “Looking back on your life what level of regret would you say you have?,” with responses on a 5-item rating scale with endpoints of “No regrets” and “Many regrets.” The 126 participants who responded “No regrets” to this item were directed to skip through to a later portion of the questionnaire, bypassing the items examined here. These 126 surveys comprised the majority of the 177 without responses to any of the items analyzed in this study.

Within the resulting final sample of N = 500, participants ranged in age from 18 to 87 years, M = 47.3, SD = 15.2. Sixty-eight percent were female. In terms of ethnicity, the largest groups were Europeans/New Zealand Europeans (82%) and Mori (7%). There was great diversity in education level: 48% of the sample held only a high school qualification or below, with 24% having a trade certificate or diploma, and 28% a university degree or diploma.

Once the 3,000 participants had been randomly selected from the electoral roll, a three-stage postal survey process was used in order to increase response rates. Based on Dillman’s (2000) multi-stage tailored design method, the three stages in the current study consisted of (1) an initial letter of invitation to participate in a study described as looking “at the development of life regrets,” (2) a second introduction letter, the postal questionnaire and free-post return envelope, and (3) a reminder postcard.

This study was carried out in accordance with the recommendations of the Massey University code of ethical conduct for research, teaching and evaluations involving human participants with written informed consent from all subjects. All subjects gave written informed consent in accordance with the Declaration of Helsinki. The protocol was approved by the Massey University Human Ethics Committee.

Participants were asked to describe (in writing) their greatest life-time regret using the following specific prompt “Please describe in the space below the specific event or decision that is the single greatest regret from your entire lifetime.” They were provided with seven lines on an A4 page to write their description. Once they had completed this item, all subsequent questions were asked in relation to this specified regret. The provided written descriptions were generally fairly brief, M = 27 words, SD = 26 (excluding one outlier comprising a multi-page description).

Null hypothesis significance testing—the dominant data analysis framework in psychology—has many well-established problems (for reviews see Cohen, 1994; Falk and Greenbaum, 1995; Gigerenzer et al., 2004; Wagenmakers, 2007). In this study we use Bayesian estimation (see Kruschke, 2010 for an introduction). Bayesian estimation has three main advantages relevant to the current study: Firstly, unlike null hypothesis significance testing—or any frequentist method—it allows us to directly calculate and report the probability that a particular hypothesis is true (e.g., the probability that a true parameter is positive, given the data observed, P(β > 0 | Data)). Secondly, it also allows us to directly report the probability that a particular parameter falls in a specific region (traditional frequentist confidence intervals cannot do this; see Morey et al., 2015). In this study we report the 95% highest probability density (HPD) interval for each parameter. Finally, Bayesian estimation allows us to take into account prior information: For example, we know that in psychology, most effects are relatively small in size: Taking this pre-existing information into account helps us estimate parameters more accurately.

In this study, we used a similar prior for all estimated effects (e.g., regression coefficients, mean differences, correlations). This was a normal distribution with mean zero and standard deviation of 0.31, but scaled to apply to the standardized effect size for each parameter (i.e., the correlation for a bivariate analysis or partial correlation for a multiple regression analysis). This prior is based on the finding of a meta-meta-analysis of over 25,000 studies in social psychology (Richard et al., 2003), which reported a mean absolute effect size of r = 0.21. The prior of N(0, 0.31²) effectively states that there is 50% prior probability that the standardized effect size will be of absolute magnitude of 0.21 or greater, corresponding with Richard et al.’s (2003) finding. This is an informative prior: All else being equal, it will tend to make our findings more conservative (shrinking our point and interval estimates slightly toward zero). The basic framework for using this approach to prior setting is described in Williams (2016). Non-informative priors were set on nuisance parameters (error variances and intercepts).

In terms of computation, data analysis was completed in R version 3.1.1 (R Core Team, 2014) using the MCMCpack package (Martin et al., 2011). With regard to item missingness, in addition to the 177 excluded surveys with no responses on the study variables, there were another 71 surveys with a response to at least one of the study variables, but with missing data (or an unclassifiable response) to one or more items. Multiple imputation (across five datasets) was used to impute the resulting missing data points in the mice package (Buuren and Groothuis-Oudshoorn, 2011). Each Bayesian analysis was conducted on each of the five imputed datasets, with inferential statistics (e.g., posterior means, highest posterior density intervals, etc.) calculated based on the pooled posterior distributions across all the imputed datasets.

Open data and analysis scripts for this study are available at https://github.com/mattnw/Towers-et-al-Regret.

Despite participants being specifically asked about their greatest lifetime regrets, there was considerable variation in regret intensity. The regret intensity ratings traversed the entirety of the possible scale range (1–9), with a mean of 5.97 (SD = 2.24). The time elapsed since participants’ greatest regrets also varied widely, with a mean of 21.77 years (SD = 15.09). The mean of 2.80 for the explicit justification variable (SD = 1.31) was near the midpoint of 3 (a regretted decision that was “somewhat” justified). In terms of the dichotomous explanatory variables used in this study, participants’ regrets were primarily in non-social (62%) rather than social domains, and the act or decision they regretted typically contradicted their personal life rules (57%). Slightly more of the regrets were for inactions (54%) as opposed to actions.

The following analyses describe the relationship between each of the individual explanatory variables in this study and the response variable (regret intensity). Firstly, the relationships between regret intensity and the three dichotomous explanatory variables were assessed. These dichotomous variables were whether the regret was for an action or for inaction; whether the life domain of the regret was social or non-social; and whether the regret contradicted the participant’s personal life rules. The results are displayed in Table 1. Interestingly, despite participants being more likely to report inaction regrets as their greatest lifetime regrets, the intensity of action regrets was higher. Participants also reported greater regret intensity when the regret was in a social life domain and when the regret contradicted their life rules.

Participants who reported higher levels of explicit justification reported lower regret intensity, r = -0.13, with strong evidence that the true correlation in the population was negative, 95% HPD [-0.22, -0.05], P(ρ < 0 | Data) = 0.999. On the other hand, time since the regret had virtually no relationship with intensity in the sample, r = 0.01, 95% HPD [-0.08, 0.09], with the direction of the true relationship remaining unclear, P(ρ > 0 | Data) = 0.564.

The analysis that follows assesses how the relationships between the explanatory variables and regret intensity change when their effects are assessed simultaneously. All explanatory variables were entered simultaneously in this model, the coefficients of which are shown in Table 2. It is immediately apparent that the effects of action/inaction and level of explicit justification are attenuated when the other explanatory variables are controlled for. The model suggests that when the other predictors are held at zero, the mean difference in regret intensity between an action and an inaction is just 0.08 units on the regret intensity scale (this scale ranging from 1 to 9). As in the bivariate analysis, it was actions that were associated with higher regret, although there was a great deal of uncertainty about this conclusion, P(β > 0 | Data) = 0.604.

The two explanatory variables with the largest effects were whether or not the regret was in a social domain, and whether it contradicted the participant’s life rules. On the other hand, explicit justification level had a small effect: An increase of one unit on the 1–5 explicit justification scale was associated with a reduction of just 0.16 units on the intensity scale, although there was sufficient evidence that the true parameter was negative, P(β < 0 | Data) = 0.984.

Importantly, the direction of the estimated coefficient for the interaction between action/inaction and time was not in accordance with the predictions of the temporal theory of regret. The point estimate of the coefficient was positive, suggesting that the “effect” of a regret being one of action was larger (and more positive) when more time had elapsed since the regretted decision. This said, the P(β > 0 | Data) value of 0.901 implies that while there is relatively high probability that the true direction of this effect is positive, there does remain some non-trivial uncertainty.