Past and Present Subjective Well-being: the Role of Contrast and Memory

This study is an attempt to analyse the interrelationships between current, past and expected future subjective well-being (SWB) through the intermediating role of memory using the National Income Dynamics Study panel dataset for South Africa. The contribution of this study is in exploring the role of contrast (intertemporal and social) in determining (a) the current levels of SWB and (b) the drivers of recall bias, within a causal framework in the context of a country with low average SWB and high SWB inequality. The results indicate significant presence of hedonic adaptation and reveals past and future contrast as important correlates of current subjective well-being. In addition, a perceived improvement in current happiness from the past is associated with higher levels of current SWB, highlighting the intermediating role of memory. Memory, however, is observed to be biased with only 42% accuracy. Higher levels of current subjective well-being are found to enhance the probability of correct recall of the past. An overall improvement bias is evident among those in the lower segment of the SWB distribution. The results reinforce the hypothesised simultaneous relationships between current SWB, intertemporal contrast and recall behaviour. Furthermore, the differences in our findings from European studies emphasize the relevance of context in driving these relationships.


Introduction
The concept of subjective well-being (SWB) broadly refers to the level of life satisfaction in general that individuals self-report. 1 According to Tatarkiewicz (1966), present level of SWB is both retrospective and prospective. Therefore, SWB of an individual is expected to straddle over the lifespan of the individual (Gomez et al. 2013). One mechanism for this is through the confounding factor of individual personality. As the hedonic theory suggests, the inherent personality traits remain stable over the lifespan of an individual (Gutiérrez et al. 2005;Brickman and Campbell 1971).
The most commonly cited explanation for the hedonic treadmill is adaptation level theory which consists of two effects, namely contrast and habituation (Brickman et al. 1978). A contrast effect is when the quality (positive or negative) of an experience is dependent on its relation to a previous reference point (Tversky and Griffin 1991;Griffin and Gonzalez 2013). This highlights that happiness is also partially derived from being able to recall the memory of a major event. Memory is therefore an important conduit that links past and present satisfaction. As SWB is largely dependent on the individual's current satisfaction relative to their previous satisfaction, it would seem that the ability to accurately recall their circumstances related to their past satisfaction (through memory) is important for formulating current satisfaction.
Much of the literature assumes that people are able to access information regarding their long-term subjective well-being which is stored in their memory (Pavot and Diener 1993). However, attempts to understand how the past influences the determination of current life satisfaction, through the intermediating role of memory, have pointed to bias in memory recall (Kahneman et al. 1993;Adler 2013;Easterlin 2001;Prati and Senik 2020;Kaiser 2020). Kahneman et al. (1993) highlight the memory biases through their peak-end theory which argues that memory is crafted by the most intense moments and the way an experience ends, rather than the average (or sum) of the entire experience. Another type of bias in the assessment of recalled well-being is the influence of the current situation of the individual (Easterlin 2001). As people adapt to a new situation, they re-evaluate their view of the past in the light of their new benchmark (current situation). Therefore, a retrospective evaluation is likely to be distorted by the current situation and not entirely determined by the lived moments in the past.
In a similar vein, Adler (2013) pointed to the bias arising from heterogeneity in intrapersonal preference over time which can cloud effective comparison of individual well-being over time. Kaiser (2020) also emphasized that memory itself could be biased due to objective differences in the past and present circumstances, mood congruency, expectation or anticipation effect.
The recent empirical study of Prati and Senik (2020) has explored the interdependence of past and present SWB. Prior to that, the empirical attention on past satisfaction was limited either within the context of hedonic adaptation literature (Clark 1999;Wilson and Ross 2000;Van Praag and Ferrer-i Carbonell 2004;Prati and Senik 2020;Steffel and Oppenheimer 2009) or testing of reliability of life satisfaction measurement (Atkinson 1982;Michalos and Kahlke 2010). Prati and Senik (2020) found through a study of three national surveys across Britain, Germany and France that on average, people tend to overstate the improvement in their well-being over time and to understate their past happiness. But the study also finds asymmetry in perception; while happy people perceive the improvement over time to be better than it was, unhappy ones tend to exaggerate its worsening. The authors conclude that feeling happy today implies feeling better than yesterday.
The implications of emotion on memory however can differ significantly between positive and negative experiences. Similar bias in memory is highlighted by studies that show that relatively unhappy people exhibit somewhat conflicting memorial tendencies vis-à-vis happiness, whereas very happy people show simpler, and less conflicting, tendencies (Liberman et al. 2009;Sharot 2008). On the other hand, other studies find that negative emotion can enhance memory recall (Kensinger 2007, Alexander et al. 2005Foa et al. 1989, Ochsner 2000. Through an experiment on Russian students, Balatsky and Diener (1993) showed that during predominantly high dissatisfaction levels, the mean recall of positive events was greater than the mean recall for negative events.
The above review of the relationship between memory recall and present SWB brings out the interdependence between the two. While the memory of the past may be contaminated by the present situation, the present SWB is also influenced by the memory of the past. Empirical literature reviewed above have addressed either of the simultaneous relationships, while ignoring the other, potentially creating biased estimates. Therefore, this study makes a contribution to literature by delving into the relationship between past and present satisfaction, through the conduit of memory, while accounting for the endogenous relationship. Lewbel's heteroscedasticity-based instrumental regression is used to estimate the endogenous relationship between SWB level and recall of past SWB level. This study further improves on Prati and Senik (2020) by considering the additional element of future expectations in driving recall behaviour, based on the contention in literature (Tatarkiewicz 1966;Gomez et al. 2013) that present level of SWB is both retrospective and prospective. Lastly, undertaking the analysis within the context of South Africa (a developing country ranking 106 in Happiness index (Helliwell et al. 2019)) provides additional insights from other recent studies focused on developed countries with higher average satisfaction levels (like Kaiser 2020; Prati and Senik 2020).
The proposed research will attempt to answer the following questions in the context of a country with low average satisfaction: Is past and future SWB related to present subjective well-being? How does intertemporal (past and future expectations) contrast and contemporary social contrast associate with current SWB? How accurate is recall of past subjective well-being? What are the correlates of bias in recalled subjective well-being?
These are fascinating questions that have been raised in psychology literature within the context of developed countries with high average levels of satisfaction, but less so within the context of a developing country with lower average satisfaction level. The context of the study is relevant as Prati and Senik (2020) argue that the nature of bias is dependent on the level of satisfaction, which varies significantly across countries of different developmental statuses. Based on this premise, overestimation bias is likely to dominate for societies with high satisfaction level (as developed countries), as happy people tend to overstate the time improvement in their life satisfaction. This would imply that in a society with low satisfaction, underestimation bias is likely to dominate as unhappy people tend to overstate the decline in their life satisfaction. However, average SWB level has been increasing and SWB inequality declining in South Africa in recent years, indicating that less happy people have been experiencing improved SWB. Future expectations therefore are likely to play a key role, and overestimation bias from less happy individuals is a possible outcome. The South African context with low but increasing average level of satisfaction and declining SWB inequality gives an ideal context to test this premise.

Data
In order to measure recall accuracy, panel data is required to compare the reported SWB of each person over at least two points in time, and compare it with recalled memory of the past at the latter time point. The National Income Dynamic Study (NIDS) data is perhaps the only available panel dataset in South Africa that has sufficient information regarding subjective well-being and questions that can be used to derive recall accuracy.
NIDS is a South African panel data study that has collected household survey data across the country for the same households, approximately in 2-year intervals across 5 waves (2008,2010,2012,2014,2017). Wave 1 contains data for more than 28,000 individuals over 7300 households which make up the initial nationally representative sample. The aim of NIDS is to collect information which can be used to analyse issues such as poverty, employment, education and health in the South African context. Interviews are conducted in person and proxy questionnaires are answered by a secondary member if the main member of the household is not available at the time of the interview. NIDS only interviews private households across South Africa, thereby excluding people who live in hostels, old age homes, hospitals, prisons or military accommodations from the sample.
Most studies on SWB are based on the assumption of interpersonal and intertemporal comparability of well-being responses. Kaiser (2020) calls this the assumption of 'common scale use' and explains that the sources of the heterogeneity may be due to a shift in reference distributions arising from altered intrapersonal or/and interpersonal comparisons. Intrapersonal comparisons can get altered due to revised expectations and aspirations, impacting the benchmark for the evaluation of life satisfaction. Also, the comparison group of the respondents itself may change according to current satisfaction level endogenously (Falk and Knell 2004). This has a subduing impact of altered circumstances on SWB (Wilson and Gilbert 2005).
Violation of the assumption of common scale use can bias the estimated effects of explanatory variable. The empirical results from Kaiser (2020) indicate that although, in principle, the violation of common scale use assumption could lead to sign reversals of coefficient estimates, in practice, it is unlikely and results in downward bias in the coefficient estimates. This study therefore proceeds with this caution.
For the purpose of this research, subjective well-being data is acquired from the adult questionnaires for wave 1 (2008), wave 2 (2010), wave 3 (2012), wave (2014) and wave 5 (2017). SWB or life satisfaction is measured using question M5, "M5: Using a scale of 1 to 10 where 1 means 'very dissatisfied' and 10 mean 'very satisfied,' how do you feel about your life as a whole right now?". Intertemporal contrast is obtained through self-reported change and observed change.
Self-reported change is obtained from question M6 from the adult questionnaire in wave 5 (2017), "M6: Are you happier, the same or less happy with life than you were 10 years ago?". While the language of questions M5 and M6 are not the same, to conduct this analysis, we make the assumption that question M6 can be synonymously interpreted as "Are you more satisfied, the same or less satisfied with life as a whole right now, than you were 10 years ago?" to make questions M5 and M6 comparable. This is a necessary assumption because in subjective well-being literature, happiness (affect) and life satisfaction are distinct components of subjective well-being. Given that M5 refers to satisfaction as 'as a whole right now' and M6 refers to 'happy with life', it is a reasonable assumption that both questions refer to satisfaction as a whole rather than the current emotional state. Furthermore, the two questions immediately follow one another in the NIDS questionnaire reinforcing the argument that the respondent's understanding of the two is likely to be related.
Observed actual change in SWB over the 10-year period is calculated as the difference between M5 question in wave 5 and wave 1. Positive differences are categorised as more satisfied, zero difference as same and negative difference as less satisfied. Recall accuracy is obtained comparing the self-reported change with actual observed change.
Social contrast is provided by the question (M3) "Please imagine a six-step ladder, where the poorest people in South Africa stand on the bottom (the first step) and the richest stand on the highest step (the sixth step)". Current social contrast is provided by the follow-up question (M3.2) "On which step are you today?". Future expectation is provided by M3.3 "On which step do you expect to be 2 years from now?". Detailed definitions of the other variables used in the analysis are provided in Table 10 in the Appendix.
For any panel dataset, the problem of attrition always needs to be addressed. While attrition is inevitable, it is more important that the attrition is non-random so that the sample will not become biased and not representative of the entire population. Brophy et al. (2018) show that attrition is relatively consistent across all NIDS waves. However, they also show for example that attrition rates are highest for White and Asian individuals, which may lead to a bias sample. To deal with the issue of attrition bias, the proposed research will use the NIDS panel weights which account for this bias and ensure that the sample remains representative of the whole population. Top up is deleted as data for first wave is not available.

Descriptive Statistics
The average level of life satisfaction is low in South Africa but Table 1 notes an increase in the life satisfaction over the 10-year period from an average of 5.05 (on a scale 1 to 10) over the first four waves to 5.56 in wave 5 (Table 1). The level of correct recall of life satisfaction contrast over the 10-year period is just over 42%. The black African population is over 80% in our sample in keeping with the South African demographics. Over 40% of our sample constitutes household heads. South Africa is a religious society with almost 90% of the individuals in the sample indicating religion to be important in their life. Furthermore, the very high level of economic inequality in South African society is represented through the extraordinarily high standard deviation of household income variable.
The majority reported being happier than 10 years ago (Table 2). While the difference between male and female is not significant, there is substantial gap between the black Africans and non-blacks observed in this regard. While almost 70% of nonblacks reported being happier than 10 years ago, the proportion of black Africans reporting the same was less than 60%.
Recall is considered accurate when the reported change in happiness over 10 years matches with the observed difference in life satisfaction between wave 5 (2017) and wave 1 (2008). Accurate recall is observed for 42% (sum of diagonals) of individuals in Table 3. Of the accurate recall, 71% is driven by those reporting improvements in SWB over the 10-year period.
We define over-reporting (of improvement of SWB) to be when individuals report an increase in SWB (more) but actual change is either zero (same) or negative (less); also where individuals report same level of SWB over the 10-year period, but in reality it has declined. Under-reporting is where individuals report decline (less) but the actual is either same or more; also where individuals report no change (same), but in reality, it has increased (more). In Table 3, 60% report improvement while only 44% observe it. On the contrary, 15% report decline, but 42% observe it. This indicates an overreporting in improvement and under-reporting of impairment. Overall, both contribute to an improvement bias of 43%. A further breakdown of recall patterns along the level of current SWB brings out interesting detail. Table 4 records the reported and actual differences in SWB for individuals with SWB below 4 on a scale 1-10. Recall is lowest among this group of individuals with 33.3% accuracy. Of the accurate recall, 77% are those reporting decline in SWB. There is evidence of under-reporting of decline, where 79% observe decline but only 32% report it. Recall is more accurate at higher levels of current SWB. On the other hand, there is over-reporting of improvement, with 34% reporting increased SWB while only 11% observe it. An improvement bias (70%) is hence evident from this among those along the left tail of the SWB distribution.
Those along the middle of the SWB distribution (with life satisfaction scores 4-7) have 37% correct recall. Of the accurate recall, 73% are those reporting improvement in SWB. Once again, improvement bias is noted with under-reporting of decline (12.5 vs 41.4%) and over-reporting of increase (63.4% vs 42%). Together an improvement bias is evident for those in the mid-segment of the SWB distribution. The improvement bias (53.4%) is however lower than that observed for those in the lowest segment.
The recall performance of those at the highest end of the SWB distribution with life satisfaction score of 8 to 10 is notably higher with 67% accuracy. Of the accurate recalls, 98% come from those reporting improvement in SWB over the 10-year period. Unlike the other segment of the distribution, there is no evidence of improvement bias observed in the right end of the SWB distribution. On the contrary, there is underreporting of improvement. While 79.6% reported an increase in SWB, actual improvement was observed by 82.4%. The improvement bias hence is −2.8%. Tables 3, 4, 5 and 6 indicate improvement bias to be higher among those with lower SWB levels. Figure 1 summarizes the correct recall and improvement bias observed across the three categories of SWB (low SWB defined as between 0 and 3, mid-SWB as between  Source: Author calculation from NIDS data 4 and 7 and high SWB as 8 and above). Correct recall is significantly higher for the high SWB category, whereas improvement bias is highest for the low SWB category. Comparing our findings with those of Prati and Senik (2020) reveals interesting similarities as well as differences. Both studies indicate the error in recall to be high at over 50%. Also, both studies find accuracy to be higher among higher SWB categories. However, the pattern of improvement bias (over-reporting of increase and under-reporting of decline) revealed through this study is different from that observed by Prati and Senik (2020). They found over-reporting to be higher for those with higher levels of SWB in their study of Britain, Germany and France. The differences in findings point to the relevance of study contexts. This questions the external validity of empirical findings developed within developed country contexts, especially for developing country contexts. South Africa, with lower average level of SWB and high SWB inequality, presents different contexts from the European countries with high average level of well-being and relatively lesser inequality.

Multivariate Estimation Strategy
The first research question of how the past influences current SWB is tackled first using ordinary least squares regressions, as a benchmark estimation. Next, Lewbel's heteroscedasticity-based internal instrument approach (Lewbel 2012(Lewbel , 2018) is used to address the issue of possible endogenous regressor using the STATA module ivreg2h (Baum and Schaffer 2012). Identification is achieved by having regressors that are   Lewbel (2012), the Breusch and Pagan (1979) test is enough to test for heteroscedasticity. Therefore, following the OLS regression, the Breusch-Pagan/Godfrey/Cook-Weisberg test was conducted. Under the null hypothesis of homoscedasticity, the null hypothesis is rejected at the 1% level, therefore suggesting that heteroscedasticity is available to be exploited.
Lastly, the Hansen J test is used to test the identification restriction and test for the validity of the instruments. The post-estimation statistics for the Hansen J test are provided with the results. The Hansen J test results fail to reject the null hypothesis (at the 10% significance level) that the instruments are valid.
The dependent variable is y i life satisfaction for individual i measured using the NIDS question "M5: Using a scale of 1 to 10 where 1 means 'very dissatisfied' and 10 means 'very satisfied,' how do you feel about your life as a whole right now?" in wave 5. The influence of the past is embedded in three variables: (a) Average _ SWB i , average of M5 from waves 1 to 4 to give hedonic life satisfaction for individual i; (b) Reported _ Diff i , M6 from wave 5 to obtain the reported happiness compared to 10 years ago for individual i; (c) Actual _ Diff i ,actual difference in life satisfaction between the first and fifth waves calculated as the difference between life satisfaction (M5) reported in wave 5 and wave 1. Contemporary contrast is captured using the Social _ contrast variable. Lastly, expected income rank in 2-year time is used as a proxy for future expectations with the variable Future _ exp. X i is the vector of control variables incorporated based on literature on the covariates of SWB (Dolan et al. 2008;Kollamparambil 2020a;Booysen and Botha 2011;Posel and Casale 2011;Blaauw and Pretorius 2013). These include per capita household income, employment status, education, marital status, health, age, race, gender, religiosity and neighbourhood condition. The definition of variables included in the analysis is found in Table 10 in the Appendix.
Although the subjective well-being measure is technically ordinal, as highlighted by Ferrer-i-Carbonell and Frijters (2004), treating them as cardinal does not generally bias the results obtained. As a robustness check, the dependent variable is reclassified in 4 categories to run an ordinal logit model (Table 11 in the Appendix). For this purpose, the dependent variable of life satisfaction is reclassified such that it takes the value 1 if the reported life satisfaction ranges from 1 to 3, 2 if the reported life satisfaction ranges from 4 to 6, 3 if the reported life satisfaction ranges from 7 to 9 and 4 if the reported life satisfaction is 10.
In order to identify the drivers of recall (in)accuracy, we utilise a standard maximum likelihood logit model to estimate the probability of incorrect recall. The dependent variable is a binary variable for 'Incorrect recall', taking the value 1 (Y = 1) for an individual if one of the following 2 biases is observed: 1. Impairment bias: (a) Satisfaction in wave 5 (from the M5 question in wave 5) is greater than or equal to satisfaction in wave 1 (from the M5 question in wave 1), and individual reports that he/she is less happy with life than 10 years ago (from the M6 question in wave 5); (b) satisfaction in wave 5 is observed to be equal to satisfaction in wave 1 but is reported as being less happy with life now. 2. Improvement bias: (a) Satisfaction in wave 5 (from the M5 question in wave 5) is less than or equal to satisfaction in wave 1 (from the M5 question in wave 1), and individual reports that he/she is happier with life than 10 years ago (from the M6 question in wave 5); (b) satisfaction in wave 5 is observed to be equal to satisfaction in wave 1 but is reported as being more happy with life now.
The 'Incorrect recall' binary variable will equal to 1 (Y = 1) if either of the biases are observed. The logit model takes the following form: where Pr(Y = 1 | X) is the probability of 'Incorrect recall' and; Although Eq.
(3) cannot claim to be an accurately specified model given that human behaviour is intrinsically difficult to model, its contribution to identify the covariates of incorrect recall, while accounting for endogeneity (using the Lewbel regression), is significant. The above analysis does not differentiate between the two types of incorrect recall, viz. improvement bias and impairment bias. Prati and Senik (2020), in their study of Germany, France and the UK, found that happy people tend to overstate the time improvement in their life satisfaction (dynamic overestimation or improvement bias) and unhappy people tend to overstate the decline in their life satisfaction (dynamic underestimation or impairment bias). However, this study argues that this need not be a universal truth. Rather, recall bias is dependent on the SWB dynamics of a specific context as it is not just past and present but also future expectation that influence assessment of the present and past (Tatarkiewicz 1966). Therefore, the analysis of the relationship between SWB level and memory cannot be devoid of the context. Logit and Lewbel estimates are undertaken for improvement bias to get a more nuance understanding of the recall bias.

Contrast Effects
The first two research questions set out in the "Introduction" are answered through the ordinary least squares and Lewbel heteroscedasticity-based instrumental variable estimations presented in Table 7. Past average levels of subjective well-being (Average SWB) are positively and significantly correlated with current levels of SWB pointing to the hedonistic treadmill argument that individuals tend to maintain a consistent level of SWB. However, the coefficient estimates although positive and significant are not unitary, indicating that factors other than permanent personality trait also contribute to current SWB level.
The role of temporal contrast is brought out by the significance of the recall measures. The reported difference in SWB (Reported_Diff) is positive and significant at 99% confidence level. This can be interpreted as perceived improvement from the past driving current SWB. The actual difference (Actual_Diff) between the present and past is also positively correlated with current levels of SWB. Comparing the OLS and Lewbel coefficients, it is clear that neglecting the feedback relationship between satisfaction level and difference from the past (observed and reported) leads to an underestimation of their effects. This underscores the existence of simultaneity and brings to question the past estimations in literature that do not account for endogeneity.
Furthermore, the higher coefficient of Actual Diff (compared to Reported_Diff) lends credence to the questions raised in literature on the ability of individuals to undertake Standard errors in parentheses ***p < 0.01, **p < 0.05, *p < 0.1 intertemporal comparisons accurately due to objective differences in the past and present circumstances. Nevertheless, the perceived difference (even if not entirely accurate) still is an important driver of current SWB level. The contemporary contrast variable of relative income is also positive and significant. The result reinforces the hypothesised relevance of contemporary (social) reference points in determining current SWB levels. The recall of satisfaction is not just dependent on their memory of their own life situation but also dependent on their social comparison in the present. Also, the satisfaction due to what someone has achieved over the last 10 years may be dependent on what other people have achieved in the same time period. Brandstätter (2000) argues that social comparison can be seen as a contemporary contrast effect. Even if circumstances are not ideal, one may be relatively happy if the average experience in society is worse. As such, local-and communitybased comparison as well as national-level comparison would affect how they recall their happiness relative to now. Luttmer (2005) shows that having neighbours who earn relatively higher incomes is associated with lower reported happiness. Future expectations are also seen as significant drivers of current SWB.
Along with the significance of the hedonistic, memory and contrast variables, other key variables continue to be significant. Black African has a significant negative coefficient in line with other studies in the South African context. Per capita household income has significant positive association with SWB except for the non-black households, which have higher average household incomes compared to black African households. On the other hand, neighbourhood characteristics are important across the board, with violence having a significant and negative effect on SWB. Religion as well as self-reported health is also important for all.

Recall Accuracy
The "Descriptive Statistics" section has revealed only 42% recall accuracy within our sample. In this section, we explore the variables that are closely associated with recall inaccuracy. The results of binary logit regression indicate that higher levels of current SWB are negatively associated with incorrect recall (Table 8). The upward bias brought about through the simultaneity in relationship between level of satisfaction and recall accuracy is evident from the reduced magnitude of the coefficient in the Lewbel estimation. Nevertheless, the sign of the coefficient is retained, albeit at a lower confidence level. These findings are aligned with psychology literature (Liberman et al. 2009) that predicts happier individuals to have more ability to recall accurately. However, our results emphasize the need to consider endogeneity bias in the estimation of the relationship between recall accuracy and current satisfaction.
A non-linear effect of current SWB is observed in models 3 and 4, with the effect of current SWB being negative and stronger at higher levels of SWB (Table 8). The black African population as well as the head of the household is observed to have significantly lower probability of inaccurate recall. Future expectations are seen to have significant negative association with incorrect recall. However, social contrast is observed to contribute positively to incorrect recall. This can be explained better by looking deeper into the nature of bias (Table 9).
Further to the binary logit for incorrect recall regression, we next undertake estimation to uncover the nature of bias within incorrect recall. The inaccuracy can emanate either from reported change in SWB being higher than it actually is observed (improvement bias) or from reported change in SWB being lower than actually observed (impairment bias). Table 9 presents the results of the logit and Lewbel estimation of improvement bias. A negative relationship is found between the level of SWB and improvement bias. The relationship is however not linear. At lower levels of SWB, increase in SWB increased improvement bias. On the contrary, a negative relationship is found with improvement bias at higher levels of SWB. The signs of the coefficients are positive and significant at the lowest levels of SWB, and negative and significant at the highest levels of SWB. The positive sign of coefficients associated with SWB1 and SWB2 means that, everything else being equal, respondents with low levels of SWB are more likely to over-estimate improvement and more likely to over-report improved SWB than someone with SWB = 5. The coefficients are negative and significant for SWB≥7, indicating that respondents with high SWB are less likely to over-report SWB. These results are different to the findings of Prati and Senik (2020) who found happy people tend to overstate the improvement in their life satisfaction and less happy people to underreport improvement in their life satisfaction.
The findings of this study need to be understood in the South African context of low average SWB and high SWB inequality. Nevertheless, what is remarkable is that SWB inequality has been falling in recent times despite increasing income inequality (Kollamparambil 2020a). Kollamparambil and Mathentamo (2020)   SWB inequality to the non-income factors that drive SWB like increased access to public amenities among the previously deprived population and the political empowerment of the majority (with low SWB levels due to decades of apartheid discrimination) brought about by the end of apartheid and shift to majority government. While these factors have contributed to an increase in SWB among the less happy, the improvement bias is driven by the enhanced optimism and hope for future even though the material and economic challenges for the vast majority remain resulting in low current SWB level. On the other hand, the privileged minorities, which have high level of SWB attributed to higher economic status, find themselves politically side-lined under the democratic black African majority government. The perceived fall in SWB can be attributed to increasing anxiety and pessimism (Hook 2020), despite the continued economic privilege that retains the high average current SWB (Kollamparambil (2020b). The mirage effect therefore is seen to operate differently at different ends of the SWB spectrum, fed by the stark economic and SWB inequality which is increasingly divorced from political empowerment and hope for future.

Study Limitations
While the results clearly point to hedonic adaptation and the relevance of memory through contrast effects, our analysis is not without limitations. The cardinal validity of subjective well-being reports has been called into question given that the individualspecific nature of formulating one's level of SWB would make measuring and interpreting subjective well-being difficult between people and over time. The current study is based on the strength of Kaiser (2020) results that indicates that although, in principle, the violation of the assumption of common scale use could lead to sign reversals of coefficient estimates, in practice, it is unlikely and results in downward bias in the coefficient estimates. While the primary objective is to find whether people accurately recall SWB or not, it may be of additional value to distinguish between complete recall and other 'levels' of recall. For example, reporting higher SWB as a response to level of happiness compared to 10 years ago is documented as correct recall as long as wave 5 SWB level is higher than wave 1 SWB level. Technically, the difference however could take any value between 9 and 1 but is not differentiated in the analysis. This is an area that presents itself for future research.

Conclusion
This study is an attempt to analyse the interrelationships between current, past and expected future SWB through the intermediating role of memory. The contribution of this study is in exploring the role of contrast (intertemporal and social) in determining (a) the current levels of SWB and (b) the drivers of recall bias, within a causal framework in the context of a country with low average SWB and high SWB inequality. While psychology literature has made attempts in understanding the intermediating role of memory through which past influences the present level of SWB, these studies for the most part ignore the simultaneity in the relationship between recall of past SWB and current SWB level. Methodologically, the study contributes to literature by using the Lewbel heteroscedasticity-based instrumental variable estimation to account for possible bias arising out of endogeneity.
The key objective of understanding the association between past and present SWB is achieved by exploring hedonic adaptation, memory and contrast variables in the analysis. The contrast variables included, apart from the conventional contemporary social contrast, intertemporal contrast (past as well as future expectations). Both the ordinary least squares and the Lewbel regression are consistent in their findings that (a) past levels of SWB are positively and significantly correlated with current levels of SWB, offering support to the hedonic adaptation argument that individuals tend to maintain a consistent level of SWB; (b) a perceived improvement in current happiness from the past is associated with higher levels of current SWB, highlighting the intermediating role of memory; (c) the contemporary (social) contrast variable of relative income is a positive predictor of current SWB in line with Brandstätter (2000) and Luttmer (2005); and (d) lastly, the study also finds future expectations to be a significant driver of current SWB. These findings lend support to Tatarkiewicz (1966) that current levels of SWB reflect past levels, as well as future expectations.
These initial findings underline the importance of memory and raise the question of recall accuracy. It is therefore noteworthy that only 42% record accurate comparison between present and past SWB. Along the lines of Prati and Senik (2020), this study records an overall improvement bias and identifies incorrect recall to be negatively associated with current level of life satisfaction. The latter finding is also in keeping with psychology literature (Liberman et al. 2009) that predicts happier individuals to have more ability to recall accurately.
The above similarities with Prati and Senik (2020) aside, there are key differences to note as well. This study found a negative relationship between the current level of SWB and improvement bias. The relationship is however not linear. Current SWB is observed to have a positive association with improvement bias only at lower levels of SWB. On the contrary, current SWB has a negative association with improvement bias at higher levels of subjective well-being. These findings differ from Prati and Senik (2020) who found improvement bias to be prominent among happy individuals. The different outcome observed in the South African context is explained by the crossover effects between present SWB and future expectations. Within the context of declining SWB inequality and improved future expectations of previously disadvantaged (less happy) individuals under the democratic government, it is not surprising that improvement bias is higher among those on the left end of the SWB distribution. Similar conclusions were arrived at by Møller (1996) in a study of college students in South Africa, where the black African students with lower mean levels of SWB had higher future expectations compared to others. The difference in the finding from Prati and Senik (2020) underscores the relevance of dynamics within the lived realities of individuals in driving recall bias.
Overall, this study finds that while past and future shape the present, memory recall and bias are also shaped by the present and future expectations. The results reinforce the hypothesised simultaneous relationships between current SWB, intertemporal contrast and recall behaviour.  1 if the reported level of happiness in comparison to 10 years ago matches with the observed differences in the reported level of life satisfaction in wave 5 and wave, 0 otherwise.

Binary
Improvement bias Waves 1 and 5 Satisfaction in wave 5 is less than or equal to satisfaction in wave 1, but is reported as more happy with life than they were 10 years ago Binary Impairment bias Waves 1 and 5 Satisfaction in wave 5 is greater than or equal to satisfaction in wave, but is reported as less happy with life than they were 10 years ago  Consent to Participate NA. No primary data was collected for this study.

Conflict of Interest
The authors declare no competing interests.