‘Fit to play’: Labour market behaviour in a cohort of public-sector antiretroviral treatment patients

ABSTRACT

Given the rapid scale-up of antiretroviral treatment (ART), it is necessary to explore the impact of ART on labour force participation, employment and labour productivity. This article investigates labour market outcomes in a prospective cohort of public-sector ART clients in the Free State province of South Africa. Empirical results suggest that labour force participation increased markedly as the proportion of those too ill to work declined, becoming indistinguishable from participation rates in the general population. Unemployment rates, however, remain above those reported for the general population. ART and its health-related benefits therefore translate into increases in labour force participation, but not employment. Employment status at HIV diagnosis strongly predicts absorption in the labour force. Public-sector ART clients should be referred to vocational rehabilitation and occupational therapy programmes, and to welfare-to-work programmes, and the unskilled to adult education and training and further education and training programmes.

KEYWORDS:

• Antiretroviral treatment
• labour markets
• South Africa

JEL CLASSIFICATIONS:

• I18
• J21

1. Introduction

The adverse microeconomic and macroeconomic impacts of the HIV and AIDS epidemic in South Africa and elsewhere is relatively well documented. Various household-level, firm-level and sector-level studies have shown how the epidemic stands to impact negatively on productivity, the size and composition of the workforce and the cost of doing business, translating into changes in demand and supply, and, in the case of small business, even resulting in business closures (Gaffeo, 2003; Fox et al., 2004; Liu et al., 2004; Rosen, 2004; Sendi et al., 2004; Bloom et al., 2007; Van Zyl & Lubisi, 2009; Abdulsalam, 2010; Abu et al., 2010; Oloo & Ojwang, 2010; Okezie et al., 2011). These supply-side and demand-side effects in turn filter through the macroeconomy to impact negatively on estimates of aggregate economic output, economic growth, unemployment and the income distribution (Cuddington, 1993; Cuddington & Hancock, 1995; Arndt & Lewis, 2000; Bonnel, 2000; Nicholls et al., 2000; Dixon et al., 2001; Ellis et al., 2003; Nattrass, 2003; Haacker, 2004; Burger & De Villiers, 2005; Corrigan et al., 2005; McDonald & Roberts, 2006).

With the rapid scale-up in the provision of subsidised or free antiretroviral treatment (ART) in developing countries such as South Africa, and its resultant short-term and sustained medium to long-term health-related and clinical benefits – which are well documented (Beard et al., 2009; El-Sadr et al., 2012) both in South Africa in general (Coetzee et al., 2004; Jelsma et al., 2005; Louwagie et al., 2007; Boulle et al., 2008, 2010) and in the Free State province in particular (Booysen et al., 2007; Fairall et al., 2008; Wouters et al., 2009) – the focus in research on the economics of HIV and AIDS has shifted towards determining how the expansion of ART may ameliorate these adverse microeconomic and macroeconomic consequences of HIV and AIDS (Charalambous et al., 2007; Jefferis et al., 2008; Rosen et al., 2008; Ventelou et al., 2008; Fultz & Francis, 2011). Modelling, moreover, confirms that scaled-up and universal antiretroviral (ARV) treatment is beneficial economically (Resch et al., 2011; Ventelou et al., 2012).

An important area of enquiry in this field is studies on the productivity, labour force participation and employment-related benefits of ART, which represents ‘a critical [economic,] social and health issue facing people living with HIV/AIDS’ (Worthington et al., 2012:231; labour force participants include the employed and searching unemployed, see footnote 4 for more detail). Empirical evidence in this area is mixed, with some earlier studies reporting that successful ARV treatment translates into return to work and increased productivity (Bernell & Shinogle, 2005; Lem et al., 2005; Razzano et al., 2006; Larson et al., 2008, 2009, 2013; Morineau et al., 2009; Habyarimana et al., 2010; Rosen et al., 2010), while other studies record improvements in labour participation or only ability to work, but not necessarily in employment (Ajithkumar et al., 2007; Coetzee 2007, Thirumurthy et al., 2008; Wagner et al., 2009). Another body of work investigated interventions in vocational counselling and occupational therapy that may enhance the work experiences of people living with HIV/AIDS (Salz, 2001; Berry & Hunt, 2005; Conyers, 2005; Escovitz & Donegan, 2005; Goldblum & Kohlenberg, 2005; Martin et al., 2005; McGinn et al., 2005; Bowyer et al., 2006; Braveman et al., 2006; Egan & Hoagland, 2006; Martin et al., 2006a, 2006b).

This article investigates labour market behaviour in the era of ART using data from a cohort study of patients enrolled in the Free State province’s public-sector ART programme during the initial phases of the ARV roll-out. More specifically, the article aims to explore five specific research factors: firstly, the article aims to document the health-related and clinical benefits of ART that are likely to facilitate changes in labour market behaviour among study participants. Secondly, the article documents changes in labour market outcomes across the ARV treatment career. In the third instance, the article compares the former trends with trends in corresponding labour market outcomes reported in national labour force statistics. Finally, the article determines the extent to which ART and its associated health-related and clinical benefits and employment status at HIV diagnosis represent important predictors of key labour market outcomes.

2. Data

The Ethics committee of the Faculty of Humanities University of the Free State approved the study protocol. Study participants were sampled randomly from a list of patients who were eligible to commence ART during the first two months of the programme.¹ In each of the five health districts, 80 patients were sampled randomly from these lists proportional to the numbers at each health care facility who had commenced treatment or were waiting to initiate treatment.² As such, the study includes patients who had not yet initiated ARV treatment as well as patients on ART. Trained enumerators conducted structured, face-to-face baseline interviews with ART patients. Written, informed consent was obtained from study participants by the nursing personnel at the respective clinics, as well as by the enumerator.

Where cohort patients were lost to follow-up in subsequent survey rounds, replacements were sampled randomly from the original sampling frame. Enumerators obtained written, informed consent from study participants prior to each follow-up interview. Five rounds of follow-up interviews were conducted with patients at intervals of approximately six to nine months during the period January 2005–December 2008, although the actual timing between interviews does vary somewhat due to logistical and practical constraints (Table 1).

Table 1. Number of patient interviews, by survey round.

Survey round	Baseline	First follow-up	Second follow-up	Third follow-up	Fourth follow-up	Fifth follow-up	Total
A. Original cohort
Interviews conducted	371	307	268	242	219	202	1 609
Patients lost to follow-up		63	38	27	22	19	169
Attrition rate (%)		17.0	10.2	7.3	5.9	5.1	45.6
B. Cohort replacements
Patients recruited		8	29	44	2	0	83
Interviews conducted		8	29	73	64	61	235
Patients lost to follow-up			1	13	14	5	22
Attrition rate (%)			12.5	18.9	11.1	6.0	26.5
C. Follow-up intervals
Median months [interquartile range]		7.0 [5.9 to 8.2]	7.0 [5.8 to 8.2]	6.4 [6.0 to 7.2]	9.1 [7.5 to 10.0]	5.3 [4.4 to 6.3]	6.7 [5.6 to 8.4]

Note: Numbers interviewed and lost to attrition do not add up because of some follow-up interviews, as a result of practical and logistical considerations, being conducted not in the subsequent survey round but in the next round only (e.g. two patients were interviewed at baseline and then again in the third round, but not in round two). Attrition rates are calculated by dividing the number of patients lost to attrition by the number of patients ever interviewed up to and including the previous survey round.

In total, 1844 interviews were conducted with 452 individual patients who were eligible for and ready to commence ART. By the fifth follow-up, as per the note to Table 2, 169 patients had been lost, primarily due to mortality among study participants as well as refusal (Table 2). This translates into an aggregate attrition rate of 41.8%. A total of 83 study participants were recruited from the original sampling frame to replace subjects lost to attrition. In the subsequent pages, these data are employed to investigate the role of socio-demographics and ART dynamics in explaining observed labour market outcomes. First, an investigation is made into the possible existence of selection and attrition bias with regards to key socio-demographic characteristics, health status and the main study outcomes.

Table 2. Reasons for loss to follow-up.

Reason	Total
	n	%
Deceased	72	43.6
Refusal	28	17.0
Untraceable	21	12.7
Transferred	15	9.1
Migration	15	9.1
Withdrawn/defaulted	7	4.2
Health	3	1.8
Other	4	2.4
Total	165	100.0

Note: Reasons for loss to follow-up are not known for all patients lost to attrition. Percentages may not add up due to rounding.

The replacements do not differ significantly from the respondents interviewed at baseline in terms of sex, education and marital status (Table 3). The exception, however, is age. As expected, because of the cohort design of the study, replacements on average were interviewed at a higher age compared with respondents interviewed earlier, at baseline. As one would expect, replacements, who by that time would have been on treatment longer and hence are likely to have responded to treatment and be healthier, do report better health status. The differences are not statistically significant at the 5% level, however, and neither are the differences in employment status – although one does, as expected, see higher levels of employment and lower levels of illness, respectively. There is therefore no clear evidence of selection bias in the study.

Table 3. Socio-demographic characteristics of baseline respondents and replacements.

	Original baseline cohort		Replacements		p value
	%	n	%	n
Age
Mean	37.1	371	39.1	77	0.034
Median [interquartile range]	36 [31 to 43]	371	38 [33 to 45]	77	0.068
Female (%)	65.2	371	66.2	83	0.429
Education
None	4.1	15	2.4	2	0.180
Primary	30.9	114	21.7	18
Some secondary	46.6	172	47.0	39
Grade 12	14.6	54	24.1	20
Tertiary	3.8	14	4.8	4
Total	100.0	369	100.0	83
Marital status
Single	49.9	185	45.8	38	0.508
Not cohabiting	29.7	110	36.1	30
Cohabiting	20.5	76	18.1	15
Total	100.0	371	100.0	83
Health status
EQ-5D (0 to 1)	0.773	371	0.802	83	0.459
EQ-VAS (%)	65.2	371	65.8	83	0.816
Feeling ill (%)	30.9	371	21.6	83	0.092
Labour market outcomes
Employed	14.5	54	25.9	21	0.052
Unemployed	15.0	56	18.5	15
Discouraged	14.0	52	14.8	12
Too ill to work	45.8	170	35.8	29
Non-participant	10.5	39	4.9	4
Total	100.0	371	100.0	81

Note: Percentages may not add up due to rounding.

There likewise is no evidence of attrition bias in the main socio-demographic characteristics and labour market status. As expected, health status is considerably and significantly poorer among subjects lost to follow-up, as implied by mortality being the most important reason for loss to follow-up (Table 4). Where these subjects had remained in the study, one could possibly expect continued poor health and hence poorer labour market outcomes, which implies that the improvements in labour market outcomes reported in this study may be overstated.

Table 4. Attrition analysis.

	Balanced cohort		Attrition		p value
	%	n	%	n
Age
Mean	38.1	202	36.9	169	0.184
Median [interquartile range]	37 [32 to 44]	202	36 [31 to 43]	169	0.495
Female (%)	68.8	202	60.9	169	0.113
Education
None	4.0	8	4.2	7	0.286
Primary	30.0	60	31.9	54
Some secondary	50.0	100	42.6	72
Grade 12	14.0	28	15.3	26
Tertiary	2.0	4	5.9	10
Total	100.0	200	100.0	169
Marital status
Single	48.5	98	51.4	87	0.643
Not cohabiting	31.6	64	27.2	46
Cohabiting	19.8	40	21.3	36
Total	100.0	202	100.0	169
Health status:
EQ-5D (0 to 1)	0.837	202	0.697	169	<0.001
EQ-VAS (%)	68.0	202	62.0	169	0.004
Feeling ill (%)	25.2	202	37.8	169	0.008
Labour market outcomes
Employed	14.3	29	14.7	25	0.408
Unemployed	16.8	34	13.0	22
Discouraged	15.3	31	12.5	21
Too ill to work	41.5	84	50.8	86
Non-participant	11.8	24	8.8	15
Total	100.0	202	100.0	169

Note: Percentages may not add up due to rounding.

3. Method

In the bivariate analysis in Sections 4 and 5 of the article, t tests, chi-square tests and one-way analyses of variance were employed to determine whether trends in treatment outcomes and labour market outcomes differ statistically significantly across treatment duration. The specification of the multivariate regression models reported in Section 6 was informed by conceptual frameworks and overviews published in the relevant literature (Ferrier & Lavis, 2003; Hergenrather et al., 2004; Maguire et al., 2008; Worthington et al., 2012). Various proxies of treatment and treatment benefits and independent, explanatory variables representing key socio-demographic characteristics and other controls were regressed on labour market outcomes. The following model was estimated using panel fixed effects (FE) logistic regressions:where:

• is a vector of (binary) labour market status variables: too ill to work (1 = yes), labour force participation (1 = yes) and employment/absorption (1 = yes).

• x_it is a vector of socio-economic and other control variables: age, age squared, sex, race, education, type of dwelling, marital status, household size, household dependency ratio, disability grant status, number of employed household members (excluding the patient), need for secrecy about HIV status and district.

• ARV_Treatment_it is a vector of ARV treatment/health status variables: being on treatment (treatment status, 1 = yes/on treatment), treatment duration (categorical: pre-ART, 0 to 3 months, 3 to 12 months, 12 to 24 months, 24 to 36 months, >36 months), treatment duration (number of months on treatment) and employment status at HIV diagnosis (1 = employed). Furthermore, treatment status was interacted with self-reported health-related quality of life (EQ-VAS and EQ-5D) and measures of immunologic (CD4 mm³) and virologic (RNA ml) response to ART.

• is the logistic cumulative distribution function (cdf), with e^z/1−e² (Cameron & Trivedi, 2005:795), while and γ are vectors of model parameters, and α_i are individual FE.

The FE model represents an appropriate econometric strategy for assessing the causal effect of the ARV_Treatment variables on the dependent variables (Angrist & Pischke, 2009:223), as it addresses endogeneity driven by the correlation between treatment and time-invariant unobservables and allows for the consistent estimation of the marginal effects of these explanatory variables on the labour market outcome variables (Cameron & Trivedi, 2009:237).³ Labour market outcomes, the dependent variables, were defined in line with the South African Labour Force Survey (LFS), based on responses to a set of survey questions.⁴ The Hausman specification test is employed to determine whether the null hypothesis that the random effects regression model is appropriate can be rejected in favour of the FE regression model (Cameron & Trivedi, 2009:266–7). In both cases the FE model will yield consistent estimates; if the null hypothesis is not rejected, however, the random effects estimator would be more efficient.

4. Antiretroviral treatment and health outcomes

ART most probably impacts on labour market outcomes via its impact on ARV patients’ health status. Table 5 presents the significant improvements over time in study participants’ health status with respect to the two key bio-medical markers of treatment response, namely CD4 and viral load, which are used to measure immunologic and virologic responses to treatment, respectively.

Table 5 Subjective health status and clinical outcomes, by treatment duration.

Outcome	Treatment duration												Sample (n)
	Pre-ART		0 to 3 months		3 to 12 months		12 to 24 months		24 to 36 months		>36 months
CD4 (copies/mm^3)	143	(17.0)	165	(11.0)	254	(10.3)	354	(10.1)	402	(13.3)	449	(20.8)***	1 153
RNA (copies/ml)	249 652	(51 685)	204 747	(33 201)	24 692	(8 121)	3 811	(1 373)	11 909	(4 031)	9 660	(4,918)***	1 137
CD4 > 350 and RNA < 500 (%)	7.6	(4.3)	4.0	(2.0)	17.6	(2.3)	42.1	(2.7)	51.7	(3.3)	60.6	(4.9)***	1 050
Feeling ill (%)	39.3	(4.3)	26.9	(3.2)	26.9	(2.3)	19.7	(1.8)	19.5	(2.1)	15.3	(10.4)***
EQ-5D (0 to 1)	0.706	(0.034)	0.818	(0.020)	0.816	(0.014)	0.833	(0.012)	0.854	(0.013)	0.895	(0.013)***	1 714
EQ-VAS (%)	63.1	(2.0)	66.5	(1.4)	67.7	(1.0)	70.0	(0.9)	71.4	(1.1)	74.9	(1.3)***	1 710
Side effects (%)	50.0	(15.0)	52.6	(4.0)	31.4	(2.5)	10.0	(1.4)	8.1	(1.4)	6.6	(1.7)***	1 541
Severity of side effects (%)
No side effects	50.0		47.7		68.5		89.9		92.8		93.9	***	1 535
Non-disruptive side effects	8.3		4.0		2.3		0.7		0.0		0.0
Minor side effects	16.7		10.6		10.5		2.6		1.6		3.1
Severe side effects	25.0		37.8		18.7		6.8		5.6		3.1
Hospitalised (%)	11.2	(2.8)	8.9	(2.0)	13.2	(1.8)	6.5	(1.1)	6.8	(1.2)	9.0	(2.0)**	1 705

Notes: Standard errors reported in parentheses. Results for side effects only include patients on ARV treatment at the time. Results exclude those patients known to have interrupted their ARV treatment at some time or other during the study (n = 27; N = 130).

***Differences statistically significant at the 1% level. **Differences statistically significant at the 5% level. Medians of continuous variables, however, not reported here due to constraints of space, also differ statistically significantly across treatment duration (p < 0.001).

A similar trend is observed in more subjective, self-reported measures of health status. Table 5 shows all such outcomes to have improved significantly across the treatment career for ART clients observed in each survey round. The question, however, is to what extent these observed improvements in health status have translated into changes in labour market behaviour among patients on ART.

5. Labour market outcomes across the ARV treatment career

In the subsequent pages, four main labour market outcomes (i.e. being too ill to work, labour force participation, unemployment and labour force absorption) are compared across treatment duration. First, however, labour force status is compared across treatment duration (Table 6).

Table 6. Labour force status, by treatment duration.

	Treatment duration							Total
	HIV-positive diagnosis	Pre-ART	0 to 3 months	3 to 12 months	12 to 24 months	24 to 36 months	>36 months
Employed	18.4	12.3	13.2	16.2	20.9	26.8	26.8	20.4
Unemployed	8.6	14.6	11.6	29.5	25.9	24.1	24.7	23.7
Discouraged	3.2	15.4	15.9	13.3	16.9	13.8	17.0	15.3
Too ill to work	67.6	46.9	46.6	33.7	27.5	25.5	20.1	31.1
None-participants	2.2	10.8	12.7	7.4	8.8	9.8	11.3	9.6
Total	100.0	100.0	100.0	100.0	100.0	100.0	100.0	100.0
Sample (n)	315	130	189	353	455	377	194	1 698

Note: HIV diagnosis refers to employment status at first HIV-positive test result. Total excludes figures for ‘HIV-positive diagnosis’. Percentages may not add up due to rounding. Results exclude patients known to have interrupted their ARV treatment at some time or other during the study (n = 27; N = 130). Differences are statistically significant at the 1% level (χ² = 92.37).

There is a statistically significant difference in labour force status across treatment duration (p < 0.001). Prior to treatment initiation, almost half of the study participants reportedly were too ill to work. This percentage was even higher at first HIV diagnosis, where two in three respondents indicated that they were too ill to work, which reflects the general trend that patients only present for testing when relatively ill (Schwarcz et al., 2011; Dickson et al., 2012; Dowson et al., 2012; Jürgenson et al., 2012). This percentage declined markedly and consistently to one in five respondents as treatment duration increased. Concurrently, the proportion of employed study participants declined between HIV diagnosis and the early part of ART initiation but increased thereafter, rising to one in four study participants at ≥2 years of ARV treatment, which is well below the approximate 50% reported in studies in developed countries (Worthington et al., 2012). Dray-Spira et al. (2006) likewise found employment losses to be relatively common in the earlier, pre-ART stages of HIV disease. The proportion of unemployed and discouraged respondents remained relatively high, however, and even increased over the ARV treatment career, as individuals too ill to work joined the labour force, but did not find employment. In the subsequent sections, trends in individual labour market outcomes are assessed relative to calendar time, which allows a comparison with corresponding labour market trends observed in the general population.

5.1. Being too ill to work

Figure 1 illustrates how the proportion of study participants too ill to work declined over time and by early 2008, although still somewhat higher, was not significantly different from the proportion of the working-age population in South Africa who are disabled or too ill to work.

Figure 1. Trends in illness/disability among the workforce (%).

Note: Outcomes for ART patients represent results for a balanced panel only (i.e. patients observed in all six survey rounds). Provincial estimates of the numbers of ill/disabled persons in the workforce are not available for all Labour Force Survey (LFS) years. Estimates are for unadjusted LFS data series (Statistics South Africa, 2007–08). LFS estimates are weighted in accordance with the race and sex composition of the CP6 cohort.

5.2. Labour force participation

Early in the treatment career, in 2004, the labour force participation rate among ARV patients was significantly below the estimates from the LFS. Subsequently, however, labour force participation rates among public-sector ART clients remain indistinguishable from participation rates observed among the general population, tracking the general upward trend reported over the period (Figure 2).

Figure 2. Trends in labour force participation rates (%).

Note: Outcomes for ART patients represent results for a balanced panel only (i.e. patients observed in all six survey rounds). Estimates are for adjusted Labour Force Survey (LFS) data series (Statistics South Africa, 2007–08). LFS estimates are weighted in accordance with the race and sex composition of the CP6 cohort.

5.3. Unemployment

The (narrow) unemployment rate, although increasing in the early treatment career and having declined statistically significantly by mid 2006 to end 2007, in the longer term increased to levels observed early in the ARV treatment career (Figure 3). As expected, given the fact that public-sector ART clients are likely to represent relatively poor, uneducated individuals (refer to Table 3), the unemployment rate among study participants, with the exception of the first period under treatment, remains significantly higher throughout than among the general population.

Figure 3. Trends in unemployment rates (%).

Note: Outcomes for ART patients represent results for a balanced panel only (i.e. patients observed in all six survey rounds). Estimates are for adjusted Labour Force Survey (LFS) data series (Statistics South Africa, 2007–08). LFS estimates are weighted in accordance with the race and sex composition of the CP6 cohort.

5.4. Labour force absorption

The labour force absorption rate among study participants exhibits a sluggish but upward trend. The absorption rate among public-sector ART clients had significantly increased by 2006/07, well beyond rates observed early in the ARV treatment career. Only in 2005, however, did the rate exceed the provincial rate. Beyond 2005, the absorption rate is indistinguishable from the provincial rate (Figure 4). The difference between the study cohort and the general population therefore persists only for the unemployment rate, with patients facing higher unemployment rates throughout their treatment career.

Figure 4. Trends in labour force absorption rates (%).

Note: Outcomes for ART patients represent results for a balanced panel only (i.e. patients observed in all six survey rounds). Estimates are for adjusted Labour Force Survey (LFS) data series (Statistics South Africa, 2007–08). LFS estimates are weighted in accordance with the race and sex composition of the CP6 cohort.

6. ARV treatment as a predictor of labour market outcomes

Labour force participation increases markedly as the proportion of public-sector ART clients who are too ill to work declines. Participation for many does not translate into employment. Despite absorption rates being equivalent in the two populations, higher unemployment rates remain endemic. The question addressed in the subsequent regression analysis, moreover, is how treatment status and duration as well as the health-related and clinical benefits of ART may independently predict observed labour market outcomes.⁵

Receiving ARV treatment and being on ART for longer than one year is associated with statistically significantly smaller odds of being too ill to work and with statistically significantly larger odds of labour force participation (Table 7). Treatment duration in months is not significantly associated with either being too ill to work, participating in the labour force or being absorbed in the labour force.

Table 7. ARV treatment as predictors of labour market outcomes in public-sector ART clients.

	Too ill to work			Labour force participation			Absorption
	OR	95% CI		OR	95% CI		OR	95% CI
Treated (yes/no)	0.329	[0.133 to 0.813]	p = 0.016	3.257	[1.263 to 8.398]	p = 0.015	1.260	[0.181 to 8.771]	p = 0.814
	F = 2.82 (p < 0.001)	Hausman = 80.15 (p < 0.001)	n = 192 N = 975	F = 1.59 (p = 0.062)	Hausman = 56.40 (p < 0.001)	n = 183 N = 805	F = 2.77 (p < 0.001)	Hausman = 11.32 (p = 0.078)	n = 112 N = 500
Treatment duration (comparison = pre-ART)
0 to 3 months	0.512	[0.181 to 1.444]	p = 0.205	1.019	[0.333 to 3.118]	p = 0.973	1.044	[0.115 to 9.471]	p = 0.969
3 to 12 months	0.236	[0.091 to 0.608]	p = 0.003	4.671	[1.717 to 12.708]	p = 0.003	1.141	[0.130 to 9.986]	p = 0.904
12 to 24 months	0.174	[0.052 to 0.583]	p = 0.005	5.714	[1.571 to 20.774]	p = 0.008	4.098	[0.375 to 44.772]	p = 0.245
24 to 36 months	0.161	[0.033 to 0.778]	p = 0.023	4.830	[0.882 to 26.430]	p = 0.069	5.101	[0.329 to 78.882]	p = 0.241
> 36 months	0.090	[0.013 to 0.595]	p = 0.013	5.853	[0.797 to 42.972]	p = 0.082	3.833	[0.170 to 86.184]	p = 0.394
	F = 2.84 (p < 0.001)	Hausman = 71.72 (p < 0.001)	n = 192 N = 975	F = 2.28 (p = 0.001)	Hausman = 48.39 (p < 0.001)	n = 183 N = 805	F = 2.23 (p = 0.003)	Hausman = 18.33 (p = 0.074)	n = 112 N = 500
Treatment duration (months)	0.960	[0.909 to 1.014]	p = 0.144	1.030	[0.971 to 1.093]	p = 0.314	1.008	[0.903 to 1.126]	p = 0.879
	F = 2.87 (p < 0.001)	Hausman = 72.84 (p < 0.001)	n = 192 N = 975	F = 1.33 (p = 0.169)	Hausman = 68.64 (p < 0.001)	n = 183 N = 805	F = 2.51 (p = 0.001)	Hausman = 41.88 (p < 0.001)	n = 112 N = 500

Note: Results are for multivariate fixed effects logistic regression models. Estimates are adjusted for age, age squared, sex, race, education, dwelling, marital status, household size, dependency ratio, disability grant status, number of employed persons in household (excluding patient), need for secrecy of HIV status, and district – full results are available from the authors on request. p values for coefficients of explanatory variables are derived from robust, Jackknife standard errors. Where regression models failed to converge, the number of iterations in the estimation procedure was set to 100. Test statistics for the overall fit of the regression model and the Hausman specification test are reported for each estimate together with sample sizes (N/n). CI = confidence interval; OR = odds ratio.

The improved health-related quality of life for those on ARV treatment, however, translates into statistically significantly lower odds of being too ill to work and statistically significantly larger odds of not only participating in the labour force but also being absorbed into the labour force (Table 8). Among the clinical outcomes, only immunologic response predicts labour market outcomes, more specifically being too ill to work and participating in the labour force. Apart from self-reported health status, the health-related benefits of ARV treatment do not translate into improved odds of being absorbed into the labour force.

Table 8. Health status and clinical outcomes as predictors of labour market outcomes in public-sector ART clients.

	Too ill to work			Labour force participation			Absorption
	OR	95% CI		OR	95% CI		OR	95% CI
Treated * EQ-VAS	0.974	[0.966 to 0.983]	p < 0.001	1.017	[1.008 to 1.026]	p < 0.001	1.017	[1.001 to 1.033]	p = 0.028
	F = 4.32 (p < 0.001)	Hausman = 72.72 (p < 0.001)	n = 192 N = 975	F = 2.00 (p = 0.010)	Hausman = 52.29 (p < 0.001)	n = 183 N = 805	F = 2.41 (p = 0.002)	Hausman = 19.43 (p = 0.021)	n = 112 N = 499
Treated * EQ-5D	0.149	[0.078 to 0.283]	p < 0.001	5.268	[2.574 to 10.783]	p < 0.001	3.332	[0.891 to 12.455]	p = 0.073
	F = 3.87 (p < 0.001)	Hausman = 80.34 (p < 0.001)	n = 192 N = 975	F = 2.19 (p = 0.004)	Hausman = 59.24 (p < 0.001)	n = 183 N = 805	F = 2.50 (p = 0.001)	Hausman = 26.89 (p = 0.002)	n = 112 N = 500
Treated * CD4 > 350 mm^3	1.072	[0.566 to 2.029]	p = 0.829	0.967	[0.540 to 1.734]	p = 0.912	1.155	[0.606 to 2.200]	p = 0.657
	F = 1.62 (p = 0.058)	Hausman = 54.11 (p < 0.001)	n = 133 N = 559	F = 1.89 (p = 0.020)	Hausman = 27.62 (p = 0.003)	n = 121 N = 465	F = 1.81 (p = 0.034)	Hausman = 25.60 (p = 0.012)	n = 84 N = 324
Treated * RNA < 500 ml	0.516	[0.282 to 0.942]	p = 0.032	3.620	[1.695 to 7.732]	p = 0.001	0.770	[0.336 to 1.768]	p = 0.534
	F = 1.69 (p = 0.044)	Hausman = 48.46 (p < 0.001)	n = 137 N = 561	F = 2.21 (p = 0.005)	Hausman = 27.48 (p = 0.003)	n = 124 N = 468	F = 5.80 (p < 0.001)	Hausman = 55.45 (p < 0.001)	n = 80 N = 307
Treated * CD4 > 350 mm^3 and RNA < 500 ml	1.081	[0.562 to 2.079]	p = 0.812	1.138	[0.631 to 2.053]	p = 0.663	1.230	[0.610 to 2.478]	p = 0.557
	F = 1.22 (p = 0.252)	Hausman = 49.80 (p < 0.001)	n = 124 N = 490	F = 1.56 (p = 0.079)	Hausman = 26.42 (p = 0.022)	n = 111 N = 412	F = 4.49 (p < 0.001)	Hausman = 58.26 (p < 0.001)	n = 76 N = 284

Note: Results are for multivariate fixed effects logistic regression models. Estimates are adjusted for age, age squared, sex, race, education, dwelling, marital status, household size, dependency ratio, disability grant status, number of employed persons in household (excluding patient), need for secrecy of HIV status, and district – full results are available from the authors on request. p values for coefficients of explanatory variables are derived from robust, Jackknife standard errors. Where regression models failed to converge, the number of iterations in the estimation procedure was set to 100. Test statistics for the overall fit of the regression model and the Hausman specification test are reported for each estimate together with sample sizes (N/n). CI = confidence interval; OR = odds ratio.

Table 9. Employment status at HIV-positive diagnosis as predictor of labour market outcomes in public-sector ART clients.

	Employed at first HIV+ test (yes/no)
	OR	95% CI
Too ill to work	0.682	[0.409 to 1.138]	p = 0.143
	F = 6.17 ( < 0.001)	LR = 65.85 (<0.001)	n = 305; N = 1 519
Labour force participation	2.396	[1.373 to 4.179]	p = 0.002
	F = 5.67 (<0.001)	LR = 78.47 (<0.001)	n = 302; N = 1 277
Absorption	9.941	[4.600 to 21.484]	p < 0.001
	F = 4.16 (<0.001)	LR = 106.90 (<0.001)	n = 302; N = 1 277

Note: Results are for multivariate random-effects (RE) logistic regression models. Estimates are adjusted for age, age squared, sex, race, education, dwelling, marital status, household size, dependency ratio, disability grant status, health*EQ-VAS, number of employed persons in household (excluding patient), need for secrecy of HIV status, and district – full results are available from the authors on request. p-values for coefficients of explanatory variables are derived from robust, jackknife standard errors (SE). Where regression models failed to converge, the number of iterations in the estimation procedure was set to 100. Test statistics for the overall fit of the regression model and the likelihood ratio (LR) specification test are reported for each estimate together with sample sizes (N/n). CI = confidence interval; OR = odds ratio.

The impact on labour market outcomes of employment status at HIV diagnosis is significantly larger in magnitude compared with the impact of treatment duration and health-related benefits. Being employed at HIV diagnosis increased the odds of labour force participation more than two-fold and the odds of absorption in the labour force almost 10-fold. Thus, respondents not employed when first having learnt their HIV-positive status, other things being equal, are likely to remain locked out of the labour market (Table 9).

7. Limitations

The presented findings, their interpretation and the following resultant conclusions should be interpreted with caution due to various methodological and econometric limitations. This study focuses mainly on paid employment and therefore underestimates the potential labour-related benefits of ARV treatment with respect to unpaid unemployment, particularly production within the household. Clinical markers (recorded at facility visits) and labour market outcomes (reported at interview dates) were observed at different points in time. As a result, it is methodologically impossible to determine the extent to which clinical treatment response predicts labour market outcomes.

In addition, true counterfactuals of labour market outcomes remain unknown due to the absence of corresponding data from comparative samples of known HIV-negative persons and known HIV-positive persons not on ART, which requires a case–control study design. The said comparisons with estimates from the LFS are used as an indicative yardstick rather than as a counter-factual.

Attrition bias is present in the study, but the bias could not be corrected for due to econometric considerations. A set of attrition probit regressions was estimated using the variables in Table 4 as predictors of study attrition. In all three regressions, the coefficients on the health variables were found to consistently differ significantly from zero. As expected, being in poorer health was positively associated with attrition in all three probit attrition regressions. However, the overall fit of all three attrition probit regressions were poor: the overall Wald chi-square statistic was only statistically significantly different from zero at the 5% level for the attrition probits in which EQ-VAS and EQ-5D were the health indicators, while the pseudo R² value for the three probit regressions ranged between 5.3 and 6.9%. Using these attrition probits to construct inverse probability weights to address attrition bias is inappropriate insofar as the low predictive power of these attrition models (as evidenced by poor overall model fit statistics) leads to the violation of the ignorability assumption, which is crucial for identification (Kerr & Teal, 2015). Importantly, moreover, the main study outcome – employment status – does not differ statistically significantly by attrition.

Labour market outcomes, moreover, were observed at one point in time only – namely at the interview date – with no information being collected on the timing and duration of these events. As a result, there are limitations in using these data to fully explore the dynamics of labour market behaviour among study participants. Presnell (2006) and Worthington et al. (2012) emphasise that quantitative analysis such as that presented here only reflects part of the larger narrative of how HIV-positive individuals fair in the labour market. Further qualitative research is therefore required to present a full picture of labour market behaviour and related issues as experienced by people living with HIV/AIDS.

8. Conclusion

Results from descriptive analyses show that both clinical and subjective markers of health status have improved markedly with treatment duration. These analyses also show that labour force participation increased markedly as the proportion of public-sector ART clients that are too ill to work declined, becoming indistinguishable from participation rates in the general population. Labour force participation for most study participants, however, did not translate into employment, with unemployment rates remaining relatively high, markedly above rates reported for the general population. Treatment and its health-related benefits predominantly translate into increases in labour force participation. Sustainable, effective ARV treatment therefore has the positive spin-off of growing the country’s army of potential workers. For this reason, it remains important to further scale-up access to ART by addressing existing barriers to uptake of HIV testing (Deblonde et al., 2010) through tailor-made interventions aimed at specific populations (Creel & Rimal, 2011), not only to maximise the prevention benefits of ART (Granich et al., 2010), but to fully exploit the wider direct and indirect economic benefits of ARV treatment.

The more recent literature documents evidence in support of the positive impact of ARV treatment, not only on labour force participation (Thirumurthy & Zivin, 2012; Venkataramani et al., 2014), as we found here, but on employment status, both in the short and longer term (El-Sadr et al., 2012). In a Ugandan study, the likelihood of working was found to have increased at six and 12 months on treatment (Linnemayr et al., 2013). In India, rapid and sustained improvements in employment status had been achieved two years following ART initiation (Thirumurthy et al., 2013). In South Africa, Bor et al. (2012) found that employment levels by the fourth year of treatment had recovered to 90% of levels observed three to five years prior to treatment initiation. Rosen et al. (2014) moreover report improvements in employment outcomes through years three to five of an ARV treatment programme in South Africa. Employment status has also been found to be positively correlated with CD4 counts in a study from Uganda (Thirumurthy et al. 2013).

There are various reasons as to why in this study we do not observe corresponding improvements in employment prospects. Subjects in this population are likely to be uneducated and potentially unemployable, partly as a result of the pro-poor rationing of ARV treatment during the early phases of the roll-out of treatment in South Africa. HIV-positive people also face discrimination in the labour market and therefore could find it difficult to find employment. The studies, moreover, are not directly comparable insofar as the cohorts are from different countries and settings and from a range of different treatment programmes, each with its peculiar characteristics.

In order to inform policy recommendations, further research is required to better understand why public-sector ART clients are unemployed. Potential policy recommendations include the following: insofar as ARV treatment in this study does not guarantee employment per se, but rather prior employment status – in this case, more specifically, being employed when first testing HIV-positive – ARV clients should be referred to vocational rehabilitation and occupational therapy programmes aimed at facilitating return to work. For unskilled, unemployable individuals, who are likely to make up a large share of public-sector ART clients, referral to and enrolment in adult education and training and in further education and training programmes could be beneficial. In the absence, moreover, of employment opportunities for public-sector ART clients and income support for unemployed adults, the disability grant is likely to remain an important safety net for most patients and their households. The risk of patients adhering sub-optimally to treatment to retain their social welfare grants and of patients becoming dependent on social assistance, however, remains real and calls for (referrals to) an integrated welfare-to-work programme in the public sector, to maximise the sustained economic benefits of ARV treatment.

Acknowledgements

The authors thank the study participants and the fieldwork staff as well as the Free State Department of Health and the National Health Laboratory Service.

Additional information

Funding

The study was funded by the following institutions: Australian Agency for International Development (AUSAID), Canadian International Development Agency (CIDA), Development Cooperation Ireland (DCI), Department for International Development (DfID), Canadian International Development Research Centre (IDRC), Joint Economics, Aids and Poverty Programme (JEAPP), United Nations Development Programme (UNDP), United States Agency for International Development (USAID), International Development Research Centre, Medical Research Council, National Research Foundation, Research-Foundation Flanders, UK Department for International Development DfID. The findings, interpretations and conclusions in the article are those of the authors and do not reflect the views of funding agencies.

Notes

1 The public-sector ART programme did not commence simultaneously in all five health districts. The original sampling frame excluded patients eligible for ART (i.e. CD4 ≤ 200 and/or WHO AIDS stage 4) but who were not certified as ready to commence with treatment by a physician, in many cases due to patients having to first complete their tuberculosis treatment. The results thus cannot be generalised to all patients eligible for ART, but rather to public-sector patients ready to commence ART.

2 In Xhariep district, the list included less than 80 patients: as a result, all treatment and non-treatment cases were included in the study.

3 Theoretically, the data also allow an investigation of predictors of labour market transitions, such as becoming a labour force participant, gaining employment or being newly absorbed in the labour force. Unfortunately, the numbers of transitions in the dataset are relatively small, which substantially reduces the statistical power of the ensuing FE multivariate regression analyses, given the ‘expensive [nature of FE] in terms of degrees of freedom’ (Gujarati, 2003:647) as a result of losing those observations where y_it =0 for all t or y_it =1 for all t (Cameron & Trivedi, 2009:614).

4 The labour market indicators and definitions used in the LFS comparisons correspond to those used by Statistics South Africa in its LFSs and Quarterly Labour Force Surveys. Unemployment refers to all individuals aged 15 to 64 years who were not employed in the seven days preceding their interviews; who were willing and able to work during this time, while having taken active steps to find employment in the four weeks prior to the interview. Labour force participants are either employed or unemployed, while the labour force participation rate is obtained by dividing the number of labour force participants by number of people aged 15 to 64 years. The labour force absorption rate is obtained by dividing the number of employed individuals by the number of individuals aged 15 to 64 years.

5 Admittedly, the general economic upswing experienced in South Africa during 2006/07 may play an important part in explaining the observed trends in labour market outcomes during this period. However, it is not possible to determine its role in explaining these trends in the absence of comparable data for a representative panel of economically active individuals in the Free State province.