New evidence on subjective well-being and the definition of unemployment in South Africa

Abstract

Access to new nationally representative, individual-level panel data from South Africa has allowed for the revalidation of Kingdon and Knight's discussion on the definition of unemployment. This paper investigates subjective well-being as a measure of comparison between labour-market statuses. It finds that on the grounds of subjective well-being the non-searching unemployed (or ‘discouraged’) are significantly worse-off than the not economically active. Moreover, evidence suggests that, with regard to the relationship between life satisfaction and labour-market status, the non-searching unemployed consistently are the worst-off. This is especially true of both the young and senior non-searching unemployed; however, the findings are largely driven by the African subsample. This paper does not advocate for a change in the official definition of unemployment but does advocate for the inclusion and recognition of the non-searching unemployed in policy relating to labour and development in South Africa.

Keywords:

• subjective well-being
• unemployment
• life satisfaction
• search state
• South Africa

1. Introduction

The definition of unemployment continues to be a debated topic; especially in countries, such as South Africa, where both strict and broad measures of unemployment continue to be high. South Africa recorded a strict unemployment rate of 25.2% and a broad unemployment rate of 36.7% in the first quarter of 2013 (StatsSA, 2013). For a long time, such unemployment rates appeared unrealistic, or even impossible, but since the financial crisis of 2008, many other countries are also experiencing such high levels of unemployment. Given the recent levels of unemployment in both the developing and developed world, and concerns over the impact of high youth unemployment (currently at 52.9% for ages 15 to 24), the debate surrounding who is and who is not unemployed has never been more pertinent.

A 2006 paper by Kingdon and Knight changed the approach taken in answering this question. They ask: are the searching and non-searching unemployed a similar or dissimilar group of people based on their relative happiness? This differs substantially from the prior research, which primarily focuses on probability of transition into employment. However, the South African data underpinning Kingdon and Knight's key findings is now 20 years old and South Africa has undergone considerable economic and social change during this time. This paper therefore revisits this argument. It attempts to reproduce the finding that the non-searching unemployed are at least as ‘unhappy’ as the searching unemployed, whilst adding to the literature in three ways. Firstly, Kingdon & Knight (2006) make use of the Southern Africa Labour and Development Research Unit (SALDRU) 1993 Household Survey, which is now 20 years old and unrepresentative of the modern-day South Africa. Data from the 2008 and 2010/11 National Income Dynamics Study (NIDS) are used in this study. Secondly, the longitudinal nature of the NIDS data adds to the study by allowing an examination of the direction of causality. Finally, the data provided by the NIDS are individual-level data – a vast improvement over the household-level life satisfaction data used by Kingdon and Knight in their original study.

However, one should note that this paper is not a replication or revision of Kingdon & Knight's (2006) work. The approach taken here differs in two important ways. Kingdon & Knight (2006) use subjective well-being as one of three arguments upon which to make a case for the broad measure of unemployment. In this paper we focus our attention entirely on subjective well-being, giving a more detailed analysis of the data and investigating issues of causality and model sensitivity. Furthermore, Kingdon and Knight's primary comparison is between the employed and the unemployed. In this paper we focus our attention primarily on differences between the non-searching unemployed and the not economically active, because the narrow definition of unemployment implicitly assumes that these unemployed individuals are not economically active. We show that these are two distinct groups, which reinforces Kingdon & Knight's (2006) claim that non-searching unemployment in South Africa is involuntary.

The paper begins with a revision of Kingdon & Knight's (2006) original argument and findings, as well as a brief overview of the surrounding literature. Thereafter, it discusses the key variables and data used in this paper, focusing on the life satisfaction variable. The analysis begins with a set of summary statistics and transition matrices that give an idea, both cross-sectionally and dynamically, of how the life satisfactions of each employment status compare. A dynamic perspective is crucial to observing the direction of causality in this relationship. The issue of causality is central to this debate and our position is reinforced by means of a probit model. Finally, the paper replicates the findings of Kingdon and Knight by means of an ordered-logit model. A sensitivity analysis expands on these results with comparisons across both age cohorts and race.

The findings reinforce the position of Kingdon & Knight (2006). We find that the non-searching unemployed are distinct from the not economically active with regard to subjective well-being. However, the results suggest that they may also be a distinct group from the searching unemployed. The non-searching umemployed are found to be significantly worse-off than both of these groups with regard to life satisfaction; hence warranting the title ‘discouraged’ and inclusion in the labour force. Furthermore, a sensitivity analysis by both age cohort and race confirms these results as well as highlights important groups, such as youth and senior non-searching unemployed, who are significantly worse-off.

2. Literature review

Prior to Kingdon & Knight's (2006) work on South African unemployment the international literature surrounding the unemployment debate centred around two main arguments. The first argument, the ‘discouraged worker’ hypothesis, suggests that the searching status of the unemployed is exogenously determined by economic conditions (see also Finegan, 1978). This line of reasoning attempts to separate the notions of endogenously and exogenously determined non-searching behaviour. Only if it can be shown that the non-searching unemployed halted their searching activities because of exogenously determined factors such as high search costs or a low perceived probability of success would one want to consider them unemployed, as opposed to not economically active (see also Dinkelman & Pirouz, 2002).

Kingdon & Knight (2006) separate this line of reasoning into two hypotheses, the ‘taste for unemployment’ hypothesis (endogenously determined non-searching behaviour) and the ‘discouraged worker’ hypothesis (exogenously determined non-searching behaviour). Evidence from Blundell et al. (1998), Tachibanaki (1991), Kuch & Sharir (1978) and Ondeck (1978) suggests that, in developed countries at least, adverse economic conditions significantly reduce the number of searching unemployed. This confirms the notion of a ‘discouraged worker’ effect. However, Finegan (1981:101) finds that ‘most persons whose labour-force participation is influenced by the discouraged worker affect (i.e., whose participation varies pro-cyclically) are never reported as discouraged workers, while many who are discouraged did not become so as a result of general economic conditions’.

The second argument looks to differentiate the unemployed and not economically active on the basis of their probability of transition into employment. If an under-utilisation of the labour force is the primary concern with regard to unemployment, then an assessment of the probability of transition into employment would be an indication of such under-utilisation. The works of Clark & Summers (1979), Flinn & Heckman (1983), Gonul (1992), and Jones & Riddell (1999) fall in this tradition. However, Kingdon & Knight (2006) indicate that such tests often fail when unemployment is high. Moreover, this literature mainly consists of studies of North American and European economies and, at the time of Kingdon & Knight (2006), utilised longitudinal datasets for which there were no South African equivalents.³

Kingdon & Knight (2006) therefore adopt a similar approach to this last group in South Africa, but without longitudinal data. They compare the characteristics of the searching and non-searching unemployed, but avoid the probability of transition premise. Moreover, they go beyond a comparison of household and individual characteristics between labour-market statuses, as given by Dinkelman & Pirouz (2002). Instead they offer three new arguments on which to make the case that the non-searching unemployed are indistinguishable from the searching unemployed and therefore belong in the labour force. The first of these is a comparison of poverty rates; the second, a comparison of satisfaction levels; and the third, the relative impact on local wages.

Kingdon and Knight present evidence for both the first and third arguments in support of the ‘discouraged worker’ effect. They find that the non-searching unemployed are significantly more deprived than the searching unemployed; which, assuming an individual would not choose to be more deprived, sheds doubt on the ‘taste for unemployment’ hypothesis. For the third account they find that the number of non-searching unemployed has a significant impact on local wage determination. However, it is their second argument that forms the focus of this paper. Kingdon & Knight (2006) make use of self-reported life satisfaction data to compare the subjective well-being of the searching and non-searching unemployed. If the non-searching unemployed are found to be better off than the searching unemployed, then their decision to look for a job may be voluntary, which would warrant exclusion from the labour force. They find that on average the non-searching unemployed are as dissatisfied as the searching unemployed in comparison with the employed. On this basis they once again argue for the ‘discouraged worker’ hypothesis and the inclusion of the non-searching unemployed in the labour force.

Kingdon and Knight's analysis of the SALDRU 1993 life satisfaction data is not without its short-comings. The first of these, and possibly the most important, is that the life satisfaction variable they use is a household-level variable. The variable is derived in the response to the question: ‘taking everything into account, how satisfied is this household with the way it lives these days?’ Hence, it does not directly capture the relationship between life satisfaction and labour-market status of the relevant individual. Similarly, the household's searching and non-searching unemployment rate has to be used in the models as opposed to individual labour-market status. A second limiting factor in Kingdon & Knight's (2006) analysis is the use of cross-sectional data. Beyond the restrictions this puts on the econometrics techniques that can be used, it prevents a clear study of the direction of causality. Kingdon and Knight depend on the works of Winkelmann & Winkelmann (1998) and Clark (2003) to dismiss the concern over reverse causality. Winkelmann & Winkelmann (1998) show that those who enter a state of unemployment report a negative change in satisfaction and those who exit report a positive change in satisfaction, as one would expect if there was to be no concern for reverse causality. Finally, the SALDRU 1993 dataset, although extremely important, is now a fairly poor representation of the modern South Africa.

This study addresses each of the aforementioned short-comings. The NIDS longitudinal dataset, used here, currently consists of two waves from 2008 and 2010/11. It therefore allows for a more contemporary analysis of the topic, as well as an interesting look into pre-financial-crisis and post-financial-crisis South Africa. The longitudinal nature of the data allows for a more accurate look at the direction of causality within the South African context and, because the NIDS is an individual-level dataset, allows for a more accurate link between life satisfaction and labour-market status.

The primary argument of this paper differs somewhat from Kingdon & Knight (2006). We propose an argument on two levels. First, if it can be shown that non-searching unemployed are significantly worse-off, in terms of subjective well-being, than the rest of the not economically active (with whom they are grouped), then they should be treated as two distinct groups. In this case, the narrow definition of unemployment that defines the non-searching unemployed as not economically active would be incorrect. Second, assuming an individual would not voluntarily choose to be worse-off (or ‘unhappy’), if those who want to work, but are not actively searching for work, are significantly less satisfied than those who ‘voluntarily’ choose to be outside the labour force (the not economically active), then they cannot have voluntarily chosen to be in this position. Evidence of this nature would support the ‘discouraged worker’ hypothesis and inclusion of the non-searching unemployed in the labour force alongside the searching unemployed. However, one should note that this paper does not advocate for a change in the definition of unemployment, given that the current definition is agreed upon at a global level, but rather a deliberate inclusion of broad unemployment measures as well as attention given to the non-searching unemployed in South African policy debate.

3. Data and summary statistics

As previously mentioned, the data used in this analysis are those of the 2008 and 2010/11 waves of the NIDS, an individual-level longitudinal survey of South Africa (Brown et al., 2012). The sample used is that of working-aged adults (ages 15 to 64) who report a level of life satisfaction. This definition coincides with that reported by Stats SA in their Quarterly Labour Force Survey, which in turn upholds the International Labour Organization definition. The NIDS dataset includes 14 112 working-aged adults who completed the adult questionnaire in Wave 1 and 16 197 working-aged adults who were successfully interviewed in Wave 2 (this includes temporary sample members). This number excludes proxy members in the dataset because proxy interviewees were unable to answer life satisfaction questions. Of the balanced working-aged panel of 11 388 individuals, 9628 successfully answered the life satisfaction question in both waves. Of the Wave 1 working-aged sample, 77% were successfully re-interviewed; thus, there is a reasonably high attrition rate amongst this subpopulation. However, this rate corresponds closely to the total attrition rate of 19% – excluding individuals who died or moved ‘out of scope’ (Brown et al., 2012:22).

The primary variable of interest is that of perceived life satisfaction. The variable is derived from adults' answers to the following question: ‘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?’. This question follows sections in the adult questionnaire relating to labour-market status, education, health, and income; in fact, it is one of the last sections covered, followed only by biometric measurements and a numeracy test (only in 2008). Thus, the answers given should be equally biased by considerations of employment status, income and health. For a detailed discussion of the NIDS's satisfaction variable, including the various determinants of life satisfaction in the South African context, see Posel (2012). It is important to note that the variable of interest is a measure of life satisfaction and not happiness. For the purposes of comparison, it is the same definition used in Winkelmann & Winkelmann (1998). The majority of work done by Clark uses data from the British General Health Questionnaire and is constructed by combining a series of answers relating to mental health (see Clark & Oswald, 1994).⁴

Table 1 offers a brief set of summary statistics for the sample. The table reports mean satisfaction levels for both waves of data, each treated as a separate cross-sectional dataset. In addition, Table 1 reports the percentage of the sample that is ‘dissatisfied’, where ‘dissatisfied’ is defined as having a satisfaction level of four or less. It divides the sample into four labour statuses: employed, not economically active, searching unemployed, and non-searching unemployed (otherwise referred to as discouraged). The NIDS has defined these categories according to the conventions of the International Labour Organization (Brown et al., 2012). In particular, the non-searching unemployed are defined as those who are not employed, want to work, but have not actively looked for a job in the past four weeks. In addition, the table includes subsets of the searching and non-searching unemployed in an attempt to identify distinct groups amongst African-male cohorts and the youth labour force. The interest in the youth labour force relates to recent political and economic concerns regarding the impact and growth of youth unemployment in South Africa.

Table 1: Average life satisfaction and proportion dissatisfied amongst working-aged adults (ages 15 to 64) in 2008 and 2010/11

		2008	2010/11
Employed	Mean	5.793	5.499
		(0.082)	(0.118)
	% dissatisfied	26.9	33.5
		(0.015)	(0.021)
	N	5 250	4 042
Not economically active	Mean	5.505	4.760
		(0.105)	(0.107)
	% dissatisfied	35.9	48.7
		(0.017)	(0.019)
	N	4 542	6 530
Searching unemployed	Mean	5.159	4.410
		(0.137)	(0.113)
	% dissatisfied	3.95	5.30
		(0.022)	(0.027)
	N	1 657	1 187
African male	Mean	4.942	4.244
		(0.212)	(0.121)
	% dissatisfied	44.5	56.0
		(0.035)	(0.034)
	N	512	413
Ages 15 to 24	Mean	5.514	4.520
		(0.170)	(0.155)
	% dissatisfied	34.2	47.4
		(0.029)	(0.040)
	N	573	447
Ages 25 to 34	Mean	5.155	4.452
		(0.221)	(0.140)
	% dissatisfied	39.3	54.0
		(0.038)	(0.043)
	N	567	373
Non-searching unemployed	Mean	4.991	3.933
		(0.163)	(0.291)
	% dissatisfied	43.4	63.6
		(0.028)	(0.062)
	N	845	635
African male	Mean	4.378	3.445
		(0.396)	(0.303)
	% dissatisfied	54.4	74.5
		(0.066)	(0.075)
	N	160	198
Ages 15 to 24	Mean	5.175	3.935
		(0.291)	(0.344)
	% dissatisfied	41.5	62.4
		(0.047)	(0.078)
	N	289	237
Ages 25 to 34	Mean	4.568	3.768
		(0.264)	(0.298)
	% dissatisfied	50.2	65.9
		(0.040)	(0.069)
	N	228	157

Source: Authors’ own calculations using NIDS Wave 1 and Wave 2.

Notes: Wave-specific cross-sectional weights used for each wave, as well as corrections for survey design. Standard errors are reported in parentheses. Data in italics do not include controls for stratification because the sample included strata with only one cluster.

The first, and most immediate, observation from Table 1 is that average satisfaction amongst all unemployed (searching and non-searching) individuals is significantly lower than that of either employed or not economically active individuals in both waves. A second immediate observation is that all cohorts experienced a drop in average satisfaction between Waves 1 and 2. Posel (2012) suggests that this is largely driven by the African adults where just over one-half of subsample reported a lower satisfaction level in 2010/11. This may relate to changing economic conditions between 2008 and 2010/11, as a result of the 2008 Global Financial Crisis. Of interest, however, is the fact that even employed and not economically active individuals reported on average a decline in satisfaction during this period. Hence, the decline in satisfaction during this period need not only relate to cyclical unemployment, but to broader factors relating to employment conditions and remuneration. However, to restate, these figures represent two cross-sections and not a balanced panel.

Table 2 reports a transition matrix that tracks the labour-market status of the balanced panel between 2008 and 2010/11. Included in this are row percentages, number of observations, and the average change in satisfaction for each cell. The diagonal, which represents those whose labour-market status did not change between waves, confirms that even those who remained employed post 2008 reported a decline in satisfaction of –0.518. Of interest (although the sample is small), those who remained non-searching unemployed between waves reported an increase in average satisfaction of 0.12. Only those who entered a state of discouragement in 2010/11 report a large drop in satisfaction – in fact, the highest average drop across all statuses in 2010/11 is reported by individuals entering, for the first time, non-searching unemployment (given by column 2). In comparison, across 2008 labour-market statuses those who were originally not searching experienced a lower average decline in satisfaction between waves (given by row 2).

Table 2: Transition matrix of labour-market status with mean change in satisfaction

	Employment status in 2011
Employment status in 2008	Searching unemployed	Non-searching unemployed	Employed	Not economically active	Total
Searching unemployed	281	103	403	553	1 340
Row %	21.43	6.25	32.51	45.72	100
Mean change	–0.579	–0.798	–0.257	–0.671	–0.531
Non-searching unemployed	113	65	167	364	709
Row %	17.52	9.61	27.16	45.72	100
Mean change	–0.674	0.120	–0.334	–0.408	–0.327
Employed	272	163	2 479	1 069	3 983
Row %	6.79	3.44	68.90	20.87	100
Mean change	–1.555	–1.403	–0.518	–0.592	–0.626
Not economically active	376	216	520	2 722	3 834
Row %	12.18	5.44	15.04	67.34	100
Mean change	–0.773	–1.463	–0.236	–0.633	–0.629
Total	1 042	547	3 569	4 708	9 866
Row %	11.49	4.91	42.37	41.23	100
Mean change	–0.931	–1.130	–0.437	–0.616	–0.592

Source: Authors’ own calculations using NIDS Wave 1 and Wave 2.

Notes: Panel weights as well as corrections for survey design were used in the estimation of this table. Row percentages are reported as well as the mean change in satisfaction (only panel weights were used for the satisfaction estimates).

Table 1 demonstrated that on average the non-searching unemployed were worse off than even the searching unemployed. In addition, from Table 2 we see that the transition from searching to non-searching unemployment involved a drop in satisfaction, whilst those who remained non-searching experienced little change in satisfaction between waves. In fact, those who remain non-searching may even have experienced a slight increase in satisfaction. This suggests that with regard to labour-market status those individuals who would like to work but are not actively searching for a job are the worst-off with regard to life satisfaction.

However, this brings into question the direction of causality. Do the non-searching unemployed individuals have a tendency to give up searching because of their lower satisfaction levels and/or possible pre-disposition towards ‘giving up’; or are they less satisfied because outside factors have limited their employment prospects, thereby lowering the expected returns of search activities, and leaving them ‘discouraged’ and dissatisfied? This is another way of framing the endogenous–exogenous unemployment problem and will be discussed in detail in the following section.

4. Non-searching and satisfaction: in which direction does causality lie?

As previously mentioned, Kingdon & Knight (2006) are unable to give attention to the issue of causality in their paper because of the cross-sectional nature of the data at their disposal. Instead, they refer to the works of Clark (2003) and Winkelmann & Winkelmann (1998), whose work with German socio-economic panel data does address this issue. Given that we are working with the NIDS panel dataset, we are able to address such concerns. This is done both by means of simple descriptive statistics as well as a regression model. Table 3 replicates the transition matrix of Table 2, but in place of the average change in satisfaction and row percentage is average Wave 1 satisfaction.

Table 3 Transition matrix of labour-market status with Wave 1 satisfaction

	Employment status in 2011
Employment status in 2008	Searching unemployed	Non-searching unemployed	Employed	Not economically active	Total
Searching unemployed	281	103	403	553	1 340
Wave 1	4.918	4.305	5.193	5.194	5.082
Non-searching unemployed	113	65	167	364	709
Wave 1	5.051	4.994	4.873	5.407	5.056
Employed	272	163	2 479	1 069	3 983
Wave 1	5.811	4.781	5.313	6.037	5.825
Not economically active	376	216	520	2 722	3 834
Wave 1	5.377	5.563	5.552	5.438	5.508
Total	1 042	547	3 569	4 708	9 866
Wave 1	5.340	5.007	5.851	5.401	5.566

Source: Authors’ own calculations using NIDS Wave 1 and Wave 2.

Notes: Panel weights as well as corrections for survey design were used in the estimation of this table. Row percentages are reported as well as the mean change in satisfaction (only panel weights were used for the satisfaction estimates).

Table 3 appears to suggest that there may be a degree of reverse causality within the non-searching cohort. Those who are unemployed, but not searching in the Wave 2 (given by column 2) reported a lower average Wave 1 satisfaction level of 5.007. This is even lower for those who were either unemployed or employed in 2008 (where the averages are 4.305 and 4.781 respectively), but not for those who were not economically active in 2008 (given by 5.563). In contrast, the searching unemployed and not economically active in 2011 share a similar Wave 1 average satisfaction (given by columns 1 and 4 respectively). If ‘discouragement’ was an exogenous effect then one would expect the non-searching unemployed to have the same average Wave 1 satisfaction as these other cohorts. The distinction seems to suggest that there is a degree of reverse causality (or endogeneity). However, a transition matrix cannot establish this conclusion definitively. A regression analysis would be able to control for other factors that may be driving this apparent endogenous relationship.

Table 4 presents the results of a probit model, which estimates the significance of Wave 1 satisfaction on the probability of being unemployed, but not searching, in Wave 2. Life satisfaction is included in the model as a dummy variable due to the categorical nature of the variable. The dependent variable is a dummy variable of Wave 2 labour-market status where the non-searching unemployed are scored one and the other three labour market states are scored zero. The scale of the satisfaction variable is reduced from one to 10 to one to five as a means of simplifying the output (the satisfaction level three is excluded as the base case).

Table 4: Marginal effects for probit estimation of reverse causality, with a dummy variable of Wave 2 labour-market status as the dependent variable

	Unconditional	Conditional
Wave 1 satisfaction (= 1)	0.032***	0.015
	(0.009)	(0.010)
Wave 1 satisfaction (= 2)	0.019**	0.005
	(0.008)	(0.008)
Wave 1 satisfaction (= 4)	–0.013	–0.006
	(0.010)	(0.010)
Wave 1 satisfaction (= 5)	0.006	0.002
	(0.011)	(0.011)
Observations	8 800	8 657
Pseudo R-squared	0.0125	0.0824

Source: Authors’ own calculations using NIDS Wave 1 and Wave 2.

Notes: These estimations make use of the balanced panel together with the NIDS-prescribed panel weights. Robust standard errors are reported. Significance: *10% level, **5% level, and ***1% level.

The first column in Table 4 reports the results of an unconditional model, which shows that a reporting a satisfaction value of one or two (one to four using the original scale) increases the probability of being non-searching unemployed in Wave 2, when compared with those who report a satisfaction of three. However, reporting a satisfaction level higher than three (i.e. four or five) has no significant impact on the labour-market outcome. This partially supports the conclusions drawn from Table 3, as it suggests that those with a lower level of satisfaction have a higher probability of being non-searching unemployed. However, upon the inclusion of socio-economic and demographic controls (given by the second column) we find that this relationship falls away (see Appendix A for the full set of marginal effects). This suggests that there is a degree of unconditional reverse causality; however, there is no evidence of conditional reverse causality. It may be that the potential causal relationship observed in Table 3 is driven by endogenous factors such as geographical location, income and household characteristics. At the end of the day, there is little evidence to suggest that a lower satisfaction increases an individual's probability of not searching and causation lies in the direction initially assumed.

5. ‘Taste for unemployment’ versus ‘discouraged worker’ hypothesis: model and results

Having reaffirmed Winkelmann & Winkelmann's (1998) and Clark's (2003) prior conclusions with regard to causality, this paper moves on to examine more closely the relationship between employment status and satisfaction. It makes use of non-linear estimation techniques to compare the non-pecuniary costs associated with each employment status in an attempt to show that the non-searching unemployed are distinct from the not economically active and share similar characteristics to the searching unemployed with regard to subjective well-being. The dependent variable in each estimation is life satisfaction (or subjective well-being); however, because of the subjective and ordinal nature of the dependent variable, linear estimation techniques are not appropriate.

The dependent variable, satisfaction, is technically an ordinal, rather than a cardinal, measure of an individual life satisfaction; hence, the use of linear (ordinary least-squares) estimations is not appropriate. Nonetheless, a large share of the literature appears to ignore this complicating factor and treats satisfaction as a continuous cardinal variable (for further discussion, see Ferrer-i-Carbonell & Frijters, 2004).⁵ Two alternative methodologies exist, which do not require the assumption of cardinality or continuity (Winkelmann & Winkelmann, 1998). Firstly, the satisfaction variable can be reduced to a binary variable of ‘satisfied’ and ‘dissatisfied’ (where ‘satisfied’ may translate to a satisfaction level of five or greater) upon which a probit/logit model can be used. Alternatively, assuming that satisfaction is ‘interpersonally ordinal’, one can make use of an ordered-probit/logit model as in Clark & Oswald (1994; Ferrer-i-Carbonell & Frijters, 2004). In this case, collapsing the satisfaction variable from a scale of one to 10 to a scale of one to five simplifies the estimation. This paper makes use of an ordered-logit model, with a satisfaction variable of scale one to five. The employment status variable is reported as a set of dummy variables, with not economically active as the base case; this for ease of comparison between the non-searching unemployed and the not economically active. Below are reported the marginal effects for each employment dummy. Given the use of the ordered-logit model, the results are reported for each possible dependent variable outcome; in this case, the values one through five. These marginal effects are estimated using a pooled dataset of both the Wave 1 and Wave 2 cross-sections, excluding Wave 2 temporary sample members.

Included in the model are individual and household level socio-economic and demographic controls (see Appendix B for the full set of marginal effects). The income variable used is the log of real per-capita household income. Other studies, such as Winkelmann & Winkelmann (1998), use total household income; however, because of the vast range of household sizes in the NIDS the use of per-capita household income was deemed more appropriate. Per-capita income was estimated simply as household income divided by household size. In addition, the income variable was deflated to constant September 2008 prices. The household-level variables include number of children and elderly in the household, as well as household size. The individual characteristics controlled for are age, age-squared, gender, subjective health (see Ferrer-i-Carbonell & Frijters, 2004), a dummy variable for religious involvement, education dummies, population group, as well as provincial and geographical dummy variables (e.g. urban informal). The inclusion of both age and age-squared variables corresponds to the findings of Clark et al. (1996) regarding a U-shaped relationship between satisfaction and age. Moreover, age ‘is used as a proxy for cohort effects or unobserved social status and health deterioration’ (Ferrer-i-Carbonell & Frijters, 2004:646).

Table 5 demonstrates that there is no significant difference between the life satisfaction probability structure of the employed and the not economically active, who are excluded as the base case in this estimation. Individuals in these two states seem to report similar levels of satisfaction. On the other hand, the results suggest that there is a significant difference between the subjective well-being of the non-searching unemployed and the not economically active as well as for the searching unemployed and the not economically active. In fact, the difference is even stronger for the non-searching unemployed than for the searching unemployed. For example, at the mean of the independent variables, a non-searching unemployed individual has a 4.4% higher probability of reporting a satisfaction level of one. As one would expect, the probability of such an individual reporting a satisfaction of five is 3% lower than any not economically active individual. Neither of these probabilities is as strong in the case of the searching unemployed, although they are significant at a 10% level. Appendix C presents separate marginal effects for each wave. The significantly lower satisfaction of the non-searching unemployed, given above, is observed in both waves (although it is stronger in Wave 2), whilst the searching unemployed appear to be significantly worse-off only in Wave 1. In Wave 2 the employed are better-off than the not economically active.

Table 5: Marginal effects of labour-market status on life satisfaction

	Satisfaction value
	(1)	(2)	(3)	(4)	(5)
Searching unemployed	0.013*	0.009*	–0.002*	–0.010*	–0.009*
	(0.007)	(0.005)	(0.001)	(0.005)	(0.005)
Non-searching unemployed	0.044***	0.028***	–0.009***	–0.033***	–0.030***
	(0.010)	(0.007)	(0.002)	(0.008)	(0.007)
Employed	–0.005	–0.003	0.001	0.004	0.004
	(0.006)	(0.004)	(0.001)	(0.005)	(0.004)
Observations	26 884
Pseudo R-squared	0.0715

Source: Authors’ own calculations using NIDS Wave 1 and Wave 2.

Notes: Cross-sectional post-stratified weights were used in all estimations. Standard errors were adjusted for serial correlation and heteroskedasticity, by clustering at the individual level. Significance: *10% level, **5% level, and ***1% level.

However, treating the not economically active as a homogeneous base category for our analysis may be problematic as reasons for labour-force non-participation tend to vary by age and this may distort the comparison we are trying to make. For example, the youth (ages 15 to 19) should still be in school. Two approaches were taken to test the potential impact of this heterogeneity on our results; in particular, the inclusion of youth (ages 15 to 19) within the not economically active. First, the model was replicated within a sample that excluded youth; and second, a separate labour-market status variable for not economically active youths was added to the original model. This last approach, rather unsurprisingly, found that the not economically active youth are significantly better-off than their senior counterparts. However, the key result of the model – that the non-searching unemployed are distinct from the not economically active in terms of subjective well-being and that they are the worst-off amongst the labour-market statuses – did not change in either of these approaches.

An alternative view of the output from the ordered-logit model is a probability matrix in which each cell reports the probability of attaining a specific y-variable outcome (Long & Freese, 2006). By setting the value of each labour-market dummy to either zero or one, one is able to attain the probability of achieving each satisfaction outcome given a specific labour-market status (including the not economically active), conditional on the other independent variables. All other independent variables are assumed to hold their mean value. Table 6 reports these probabilities. As expected, they tell a similar story to Table 5. The probability of an individual reporting a satisfaction level of three (which represents five or six on the original scale) is about 30% for all labour-market statuses. However, the non-searching unemployed have a higher probability of reporting a value of one, and lower probability of reporting a five. This relationship holds, but is weaker, for the searching unemployed.

Table 6: Probability of reporting each satisfaction level

	Satisfaction value
	(1)	(2)	(3)	(4)	(5)
Not economically active	15.88	22.93	30.64	19.84	10.72
Searching unemployed	17.25	23.80	30.32	18.82	9.82
Non-searching unemployed	20.65	25.54	29.25	16.55	8.01
Employed	15.36	22.57	30.73	20.24	11.09

Source: Authors’ own calculations using NIDS Wave 1 and Wave 2.

This observation suggests that there is a distinct difference between the subjective well-being of the non-searching unemployed and the not economically active, holding other factors constant. These results therefore agree with ‘discouraged worker effect’ put forward by Kingdon & Knight (2006). Moreover, they agree that on the grounds of subjective well-being the non-searching unemployed are on average worse-off than the not economically active and form a distinct group from the not economically active. These results go further to suggest that the non-pecuniary costs associated with being a non-searching unemployed individual outweigh those associated with any other state, including being an active searcher. The non-searching unemployed appear to be distinct from (and less satisfied) than their searching counterparts on the basis of subjective well-being.

6. Sensitivity analysis

The summary statistics reported in Table 1 suggest that there may be important life-satisfaction differences between the unemployed of different age groups and races. For this reason, a sensitivity analysis by age and race should both confirm and expand on the findings discussed above.

6.1 Age variations in the relationship between search state and subjective well-being

Table 1 demonstrated that the relationship between subjective well-being and search state, as discussed above, may vary across age groups. The following analysis tests this hypothesis by differentiating across both search state and age cohort. The dummy variables for the searching and non-searching unemployed are separated across three age groups: youth (ages 15 to 29), middle-aged (30 to 49), and senior (50 to 64). These age groups are used elsewhere in the literature on subjective well-being and unemployment (see Winkelmann & Winkelmann, 1998). They also reflect three important stages in any individual's working life where employment holds a very different meaning. For example, an individual who loses their job after the age of 50 would be concerned about the probability of re-employment and their savings for retirement, whilst a young individual is concerned about making an entry into the workforce. The individual-level and household-level controls used in the previous models remain the same. The marginal effects for this analysis are reported in Table 7.

Table 7: Marginal effects of searching state and age cohort on life satisfaction

	Satisfaction value
	(1)	(2)	(3)	(4)	(5)
Non-searching unemployed	0.064***	0.042***	–0.013***	–0.049***	–0.045***
(aged 15 to 29)	(0.013)	(0.008)	(0.003)	(0.010)	(0.009)
Non-searching unemployed	0.008	0.005	–0.002	–0.006	–0.006
(aged 30 to 49)	(0.017)	(0.011)	(0.003)	(0.012)	(0.011)
Non-searching unemployed	0.056*	0.036*	–0.011*	–0.042*	–0.039*
(aged 50 to 64)	(0.031)	(0.020)	(0.006)	(0.023)	(0.021
Searching unemployed	0.010	0.006	–0.002	–0.007	–0.007
(aged 15 to 29)	(0.008)	(0.006)	(0.002)	(0.006)	(0.006)
Searching unemployed	0.009	0.006	–0.002	–0.007	–0.007
(aged 30 to 49)	(0.012)	(0.008)	(0.002)	(0.009)	(0.008)
Searching unemployed	0.068***	0.044***	–0.013***	–0.051***	–0.047***
(aged 50 to 64)	(0.021)	(0.014)	(0.004)	(0.016)	(0.015)
Employed	–0.006	–0.004	0.001	0.005	0.004
	(0.006)	(0.004)	(0.001)	(0.005)	(0.004)
Observations	26 884
Pseudo R-squared	0.0718

Source: Authors’ own calculations using NIDS Wave 1 and Wave 2.

Notes: These estimations make use of a pooled dataset of both Wave 1 and Wave 2 cross-sectional datasets (includes Wave 2 temporary sample members). Cross-sectional post-stratified weights were used in all estimations. Standard errors were adjusted for serial correlation and heteroskedasticity, by clustering at the individual level. Significance: *10% level, **5% level, and ***1% level.

The results suggest that the relationship found in Tables 5 and 6 is not the same across each of these age groups. For the non-searching unemployed, both the youth and senior unemployed report significantly lower levels of subjective well-being. Those youth who want to work but are not searching have 6.4% higher probability of reporting a satisfaction level of one (one and two on the original scale). Similarly, their senior counterparts have a 5.6% higher probability. For the senior unemployed, the relationship appears significant regardless of search state. However, amongst the youth only the non-searching unemployed report a lower satisfaction. This may suggest that the determinants of a lower satisfaction are different for these two groups. For the senior unemployed, any form of unemployment is a cause for concern, because the issues associated with unemployment are the same. However, youth who are actively searching for employment are not dissatisfied, whilst those who potentially have ‘given up’ searching are. This suggests for the youth that the non-searching unemployed may best be described as ‘discouraged’.

Amongst the searching unemployed, only the senior unemployed share a significantly lower satisfaction. Being unemployed between the ages of 50 and 64 increases the likelihood of reporting a satisfaction level of one by 6.8% but decreases the likelihood of reporting a satisfaction of five by 4.7%. In comparison, there appears to be no distinction between the youth and middle-aged unemployed and the not economically active. These non-pecuniary costs at an older age are expected given an older individual's concerns over retirement, savings, and the probability of future employment.

Another way of viewing these results is to predict the probability of each satisfaction outcome for a set of profiles as was done earlier in Table 6. In Table 8 we report the probability of achieving each of the five satisfaction outcomes for three different profiles: a 25-year-old non-searching unemployed youth with median income, a 55-year-old non-searching unemployed individual with median income, and a 55-year-old searching unemployed individual with median income.

Table 8: Probability of reporting each satisfaction level by age cohort

	Satisfaction value
	(1)	(2)	(3)	(4)	(5)
Non-searching youth	22.82	28.08	29.48	14.00	5.62
Non-searching senior	20.99	27.31	30.35	15.11	6.23
Searching senior	22.55	27.98	29.61	14.16	5.70

Source: Authors’ own calculations using NIDS Wave 1 and Wave 2.

The probability matrix for all three of these profiles is comparable with that of the non-searching unemployed in Table 6, which in comparison with the other labour-market statuses were considerably worse-off. Therefore, the youth and senior age cohorts amongst the non-searching unemployed are considerably worse-off in terms of subjective well-being relative to both the employed and the not economically active. Remaining with a strict definition of unemployment would remove these two groups from the labour force despite their clear dissatisfaction with their state.

6.2 Variations by race

The previous sensitivity analysis was executed by means of interacted dummy variables. To test for racial differences, separate regressions are run for the African and non-African (white, coloured and Asian/Indian) subsamples. It should be noted that the African subsample represents about 80% of the entire sample, but close to 90% of the searching unemployed subsample in both waves. Of the non-searching unemployed subsample, 80% is African in Wave 1 and 87% in Wave 2. Thus, any relationship found within the African subsample not only represents the majority of the population, but also the majority of the unemployed population. Hence, the African subsample carries a greater weight in this analysis since their share unemployed sample is greater than their share of the total sample.

The marginal effects are reported in Table 9. It is immediately clear that the aforementioned relationship between subjective well-being and employment status only holds for the African subsample. As in Table 5, we see that the marginal effect of not searching on the probability of reporting a satisfaction of one is three times higher than that of the searching unemployed; being an African non-searching unemployed individual increases your probability of reporting a one by 6.2%.

Table 9: Marginal effects of searching state for the African and non-African subsamples

	Satisfaction value
	(1)	(2)	(3)	(4)	(5)
Africans
Searching unemployed	0.016*	0.009*	–0.006*	–0.010*	–0.008*
	(0.008)	(0.005)	(0.003)	(0.005)	(0.004)
Non-searching unemployed	0.062***	0.034***	–0.024***	–0.039***	–0.032***
	(0.012)	(0.007)	(0.005)	(0.008)	(0.007)
Employed	–0.008	–0.004	0.003	0.005	0.004
	(0.007)	(0.004)	(0.003)	(0.005)	(0.004)
Observations	21875
Pseudo R-squared	0.0458
Non-Africans
Searching unemployed	0.012	0.020	0.034	–0.022	–0.045
	(0.009)	(0.016)	(0.027)	(0.017)	(0.035)
Non-searching unemployed	–0.006	–0.010	–0.017	0.011	0.023
	(0.011)	(0.019)	(0.032)	(0.020)	(0.042)
Employed	0.002	0.004	0.006	–0.004	–0.008
	(0.005)	(0.010)	(0.016)	(0.010)	(0.021)
Observations	5009
Pseudo R-squared	0.0535

Source: Authors’ own calculations using NIDS Wave 1 and Wave 2.

Notes: Cross-sectional post-stratified weights were used in all estimations. Robust standard errors are reported. Significance: *10% level, **5% level, and ***1% level.

However, none of the employment status variables are significant in the non-African subsample. This may be because of the vastly different employment rates amongst the different samples. The African sample used in the above analysis had a broad unemployment rate of about 36% in Wave 1, whilst the non-African sample's rate was 21%. Although there is probably a whole set of factors, for which race is merely a proxy, which explains this relationship (for a more detailed discussion on differences in life satisfaction between South Africa's racial groups, see Ebrahim et al., 2013).

7. Conclusion

Kingdon & Knight (2006) offer three new arguments on which to make the case that the non-searching unemployed belong in the labour force. This paper re-examined the second of these arguments – that the non-searching unemployed are just as ‘unhappy’ as the searching unemployed – using data from the NIDS. The results confirm Kingdon & Knight's (2006) ‘discouraged worker’ hypothesis by showing that there are significant non-pecuniary costs associated with being unemployed but not actively searching for a job. Moreover, the panel data allowed for an analysis of the direction of causation, which confirmed that conditional on socio-economic status a lower satisfaction does not increase the probability of not searching, but rather employment status is in fact a significant determinant of subjective well-being. This suggests that the ‘discouragement’ is not an endogenously determined phenomenon. If the assumption holds that no one would choose to be worse-off (or ‘unhappier’), then the evidence presented here strongly supports the ‘discouraged worker’ hypothesis. Indeed, the non-searching unemployed do appear to be distinct from the not economically active on the basis of subjective well-being.

The evidence goes further to suggest that this group is distinct from the searching unemployed. The non-pecuniary costs associated with being unemployed but not searching appear to be higher than for those who are searching. Indeed, at a cross-sectional level the non-searching state appears to be the worst with regard to subjective well-being. In sharp contrast, the life satisfaction of the not economically active who are outside the labour force appears to be similar to the employed. This implies a world of difference between the not economically active and the non-searching unemployed.

A more detailed comparison of different age cohorts reveals that the young non-searching unemployed are significantly worse-off than their searching counterparts. This is particularly important in light of the high level of youth unemployment in South Africa. The senior non-searching unemployed appear to share a similarly low satisfaction to their searching counterparts. An additional sensitivity analysis by race reveals that the non-pecuniary costs to unemployment are even greater amongst the African subsample (by far the majority of the population), but the relationship is not significant amongst the non-African subsample. However, given the greater weight of the African subsample, these racial differences do not negate the importance of the overall findings.

As in 1993, it remains true in 2010 that the non-searching unemployed should be included in the labour force. The fact that, in 2008 and 2010, this group is worse off even than the searching unemployed is worthy of further exploration elsewhere. Here, it serves to accentuate just how miserable life is for South Africa's non-searching unemployed. We are not advocating for a change in the official definition of unemployment. This definition is agreed upon at a global level and South Africa has to report internationally in a consistent fashion. However, we are making a case for the deliberate inclusion of the non-searching unemployed in South African policy debate. We have shown that this is an especially vulnerable group and attention to this group needs to complement discussion of changes in the formal unemployment rate. The fact that Stats SA produces and release statistics on the non-searching unemployed makes it possible to do this in policy targeting and assessment.

Acknowledgements

Neil Lloyd acknowledges the financial support of the National Research Foundation and the NIDS. Murray Leibbrandt acknowledges the Research Chairs Initiative of the Department of Science and Technology and National Research Foundation for funding his work as the Research Chair in Poverty and Inequality.

Notes

³Dinkelman (2004) is one study on the definition of unemployment in South African that makes use of longitudinal data (although the data are not nationally representative). The findings serve mainly to reinforce the link between search activities and labour-market outcomes and the household determinants of search activities.

⁴One could argue that the variable derived from the NIDS survey is a more obvious proxy for individual utility as the question relates directly to satisfaction and is not derived from questions relating to mental health and state of mind (Posel, 2012).

⁵Linear (ordinary least-squares) estimation techniques generally report the same set of significant independent variables; however, the interpretation of ordinary least-squares coefficients is not appropriate given the ordinal and categorical nature of the satisfaction variable. Thus, despite their appeal and ease of interpretation, ordinary least-squares estimations are not used.

Appendix A

Table A1: Full marginal effects for probit estimation of reverse causality, with a dummy variable of Wave 2 labour-market status as the dependent variable

	Unconditional	Conditional
Wave 1 satisfaction (= 1)	0.032***	0.015
	(0.009)	(0.010)
Wave 1 satisfaction (= 2)	0.019**	0.005
	(0.008)	(0.008)
Wave 1 satisfaction (= 4)	–0.013	–0.006
	(0.010)	(0.010)
Wave 1 satisfaction (= 5)	0.006	0.002
	(0.011)	(0.011)
Log of real per-capita household income^a		–0.068***
		(0.022)
Age		0.0004
		(0.001)
Age squared		–0.00002
		(0.0002)
Male		–0.005
		(0.007)
Poor health		0.006
		(0.009)
Religious		0.026**
		(0.010)
Number of children in household		–0.006**
		(0.003)
Number of elderly in household		–0.005
		(0.006)
Household size		0.001
		(0.002)
African		0.051**
		(0.023)
Coloured		0.048*
		(0.025)
Asian/Indian		0.069**
		(0.033)
Relationship status		No
Education		No
Province		Yes
Geographical location		Yes
Observations	8 800	8 657
Pseudo R-squared	0.0125	0.0824

Source: Authors’ own calculations using NIDS Wave 1 and Wave 2.

Notes: These estimations make use of the balanced panel together with the NIDS-prescribed panel weights. Robust standard errors are reported. Significance: *10% level, **5% level, and ***1% level. ^aThe marginal effect for income is reported as a semi-elasticity.

Appendix B

Table B1: Marginal effects of labour-market status on life satisfaction, including all controls

	Satisfaction value
	(1)	(2)	(3)	(4)	(5)
Wave 2	0.059***	0.038***	–0.012***	–0.044***	–0.041***
	(0.004)	(0.003)	(0.001)	(0.003)	(0.003)
Searching unemployed	0.013*	0.009*	–0.002*	–0.010*	–0.009*
	(0.007)	(0.005)	(0.001)	(0.005)	(0.005)
Non-searching unemployed	0.044***	0.028***	–0.009***	–0.033***	–0.030***
	(0.010)	(0.007)	(0.002)	(0.008)	(0.007)
Employed	–0.005	–0.003	0.001	0.004	0.004
	(0.006)	(0.004)	(0.001)	(0.005)	(0.004)
Log of per-capita real income^a	–0.243***	–0.176***	0.027***	0.192***	0.200***
	(0.017)	(0.013)	(0.004)	(0.014)	(0.015)
Age	0.010***	0.007***	–0.002***	–0.008***	–0.007***
	(0.001)	(0.0007)	(0.0002)	(0.0008)	(0.0008)
Age squared	–0.0001***	–0.00008***	0.00003***	0.00009***	0.00009***
	(0.00001)	(0.00001)	(0.000005)	(0.00001)	(0.00001)
Male	0.006	0.004	–0.001	–0.004	–0.004
	(0.005)	(0.003)	(0.0009)	(0.004)	(0.003)
Poor health	0.053***	0.034***	–0.011***	–0.040***	–0.036***
	(0.007)	(0.005)	(0.002)	(0.005)	(0.005)
Religious	–0.064***	–0.041***	0.013***	0.048***	0.044***
	(0.007)	(0.005)	(0.002)	(0.006)	(0.005)
Number of children in household	–0.003	–0.002	0.0005	0.002	0.002
	(0.002)	(0.002)	(0.0005)	(0.002)	(0.002)
Number of elderly in household	0.003	0.002	0.0005	–0.002	–0.002
	(0.004)	(0.002)	(0.0008)	(0.003)	(0.003)
Household size	–0.001	–0.0009	0.0003	0.001	0.001
	(0.002)	(0.001)	(0.0003)	(0.001)	(0.001)
African	0.092***	0.060***	–0.019***	–0.069***	–0.064***
	(0.011)	(0.007)	(0.003)	(0.008)	(0.008)
Coloured	–0.013	–0.008	0.003	0.010	0.009
	(0.014)	(0.009)	(0.003)	(0.010)	(0.010)
Asian/Indian	–0.083***	–0.054***	0.017***	0.063***	0.058***
	(0.022)	(0.014)	(0.005)	(0.017)	(0.015)
Relationship status	Yes	Yes	Yes	Yes	Yes
Education	Yes	Yes	Yes	Yes	Yes
Province	Yes	Yes	Yes	Yes	Yes
Geographical location	Yes	Yes	Yes	Yes	Yes
Observations	26 884
Pseudo R-squared	0.0715

Source: Authors’ own calculations using NIDS Wave 1 and Wave 2.

Notes: These estimations make use of a pooled dataset of both Wave 1 and Wave 2 cross-sectional datasets (includes Wave 2 temporary sample members). Cross-sectional post-stratified weights were used in all estimations. Standard errors were adjusted for serial correlation and heteroskedasticity, by clustering at the individual level. Significance: *10% level, **5% level, and ***1% level. ^aMarginal effects for the long of per-capita real income are reported as semi-elasticities.

Appendix C

Table C1: Re-estimation of Table 5 for Waves 1 and 2 separately

	Satisfaction Value
	(1)	(2)	(3)	(4)	(5)
Wave 1: 2008
Searching unemployed	0.017*	0.013*	0.0007	–0.015*	–0.015*
	(0.010)	(0.008)	(0.0005)	(0.009)	(0.009)
Non-searching unemployed	0.027**	0.021**	0.001	–0.024**	–0.025**
	(0.013)	(0.010)	(0.001)	(0.011)	(0.012)
Employed	0.0009	0.0007	0.000	–0.001	–0.001
	(0.008)	(0.006)	(0.0003)	(0.007)	(0.007)
Log of per-capita household income	–0.204***	–0.168***	–0.030***	0.185***	0.217***
	(0.023)	(0.019)	(0.006)	(0.021)	(0.025)
Observations	12 159
Pseudo R-Squared	0.056
Wave 2: 2010/11
Searching unemployed	0.012	0.006	–0.004	–0.007	–0.006
	(0.010)	(0.005)	(0.004)	(0.006)	(0.005)
Non-searching unemployed	0.051***	0.027***	–0.018***	–0.032***	–0.027***
	(0.016)	(0.008)	(0.006)	(0.010)	(0.008)
Employed	–0.017*	–0.009*	0.006*	0.011*	0.009*
	(0.010)	(0.005)	(0.003)	(0.006)	(0.005)
Log of per-capita household income	–0.268***	–0.163***	0.076***	0.182***	0.173***
	(0.025)	(0.016)	(0.008)	(0.018)	(0.018)
Observations	14 725
Pseudo R-squared	0.093

Source: Authors’ own calculations using NIDS Wave 1 and Wave 2.

Notes: Cross-sectional post-stratified weights were used in all estimations. Robust standard errors are reported. Significance: *10% level, **5% level, and ***1% level.