The Impact of Company-Level ART Provision to a Mining Workforce in South Africa: A Cost–Benefit Analysis

Background HIV impacts heavily on the operating costs of companies in sub-Saharan Africa, with many companies now providing antiretroviral therapy (ART) programmes in the workplace. A full cost–benefit analysis of workplace ART provision has not been conducted using primary data. We developed a dynamic health-state transition model to estimate the economic impact of HIV and the cost–benefit of ART provision in a mining company in South Africa between 2003 and 2022. Methods and Findings A dynamic health-state transition model, called the Workplace Impact Model (WIM), was parameterised with workplace data on workforce size, composition, turnover, HIV incidence, and CD4 cell count development. Bottom-up cost analyses from the employer perspective supplied data on inpatient and outpatient resource utilisation and the costs of absenteeism and replacement of sick workers. The model was fitted to workforce HIV prevalence and separation data while incorporating parameter uncertainty; univariate sensitivity analyses were used to assess the robustness of the model findings. As ART coverage increases from 10% to 97% of eligible employees, increases in survival and retention of HIV-positive employees and associated reductions in absenteeism and benefit payments lead to cost savings compared to a scenario of no treatment provision, with the annual cost of HIV to the company decreasing by 5% (90% credibility interval [CrI] 2%–8%) and the mean cost per HIV-positive employee decreasing by 14% (90% CrI 7%–19%) by 2022. This translates into an average saving of US$950,215 (90% CrI US$220,879–US$1.6 million) per year; 80% of these cost savings are due to reductions in benefit payments and inpatient care costs. Although findings are sensitive to assumptions regarding incidence and absenteeism, ART is cost-saving under considerable parameter uncertainty and in all tested scenarios, including when prevalence is reduced to 1%—except when no benefits were paid out to employees leaving the workforce and when absenteeism rates were half of what data suggested. Scaling up ART further through a universal test and treat strategy doubles savings; incorporating ART for family members reduces savings but is still marginally cost-saving compared to no treatment. Our analysis was limited to the direct cost of HIV to companies and did not examine the impact of HIV prevention policies on the miners or their families, and a few model inputs were based on limited data, though in sensitivity analysis our results were found to be robust to changes to these inputs along plausible ranges. Conclusions Workplace ART provision can be cost-saving for companies in high HIV prevalence settings due to reductions in healthcare costs, absenteeism, and staff turnover. Company-sponsored HIV counselling and voluntary testing with ensuing treatment of all HIV-positive employees and family members should be implemented universally at workplaces in countries with high HIV prevalence.


Choice of parameters, distributions and shape parameters for probabilistic sensitivity analysis
Beta distributions were assigned to binomial events such as the proportion of individuals that experience either a HIV-related or non HIV-related separation, treatment failure or loss to ART retention in a specific time period, as well as the proportion of recruits that are newly HIV-infected in a specific year. Normal distributions were assigned to the number of absenteeism days experienced by individuals on or off ART for different CD4 cell count categories. Dirichlet distributions were assigned to the upward and downward transition probabilities between CD4 cell-count defined health states (further details below table). Lastly, because of the over dispersed nature of cost data, gamma distributions were assigned to the costs of inand outpatient care for individuals on or off ART and to the cost of ART (which includes the cost of drugs, labs, other medical supplies, staff time, site programme cost and central management cost). Specific details on the parameter distributions used for each parameter and the justification for those assumptions are given in Table S2. Notation for the Gamma, normal and beta distribution is standard: Beta(α,β) gives a continuous approximation to a binomial distribution with α successes and β failures; N(µ,s 2 ) denotes a normal distribution with mean µ and standard deviation s; and G(a,b) denotes a gamma distribution with scale a and shape b. The distribution for all separation rates was estimated directly from the data used to derive it based on the number that separationed out of each CD4 health state (α in Beta(α,β)) from the total sample (α+β) for specific CD4 health state The distribution for all separation rates was estimated directly from the data used to derive it based on the number that separationed out of each job grade strata (α As is standard practice, gamma distributions were assumed for all cost parameters because of their usual over dispersion. The shape (a) and scale (b) parameters for each distribution G(a,b) were derived using the method of moments where a=(mean/SE) 2 and b=SE 2 /mean, with mean being the mean cost from the sample and SE being the standard error around the mean of the cost estimates in the sample. This was done for the cost data collected for in-patient as well as out-patient costs with each being stratified by the CD4 cell count of the individual. This was also done for the cost of ART for each individual. This was not stratified by CD4 count because the cost of ART did not vary by this variable within our data.

Inflation adjustment
Cost are given in 2010 constant USD. Even though inflating to the most recent year for which a Consumer Price Index (CPI) value is available is standard methodology, in the case of the healthcare cost in South Africa, the use of the general CPI for adjusting for inflation the expected value of a past cost analysis has its limit. Healthcare costs do not follow the general CPI, since salaries are subject to separate negotiations and drug prices (especially for antiretrovirals) have undergone dramatic downward developments since 2010. We think that inflating costs over a total of eight years (from 2006 to 2014) would have exaggerated them and render the final cost figures close to useless.

Model calculations
Standard methodology (Drummond 2005) suggests that the choice of health states in a health-state transition model be reflective of important differences in disease progression, or healthcare utilization and cost, or both, in order to best represent survival and cost associated with the disease or an intervention against it. In analysing the workplace data to decide on the number and definition of health states, we found differences in separations (morbidity and mortality) and promotion rates between job grades and age groups in all employees; and differences in incidence and separations between job grades, genders and age groups in HIV positive employees, as well as in absenteeism and cost between HIV-positive  8 A small proportion (default value: 0.1%) of Infecteds in first-line treatment failure are assumed to move back to successful first-line treatment as a result of their viral load being re-suppressed after intensified adherence counselling and an improvement in adherence 9 Despite s being included as a subscript to all transition probabilities, the transition probabilities are specific to the type of care 10 If the workforce is set to be reduced during one year, the resulting number of recruits will be negative, signifying the number of people who will be retrenched, rather than recruited, during this year.

a) All recruits/ retrenchees
Recruits/ retrenchees in cycle t+1(in specific job grade, age and gender group) = (Workforce required in year y + all separations in cycle t + all retirements in cycle t -current workforce in cycle t) * proportion of required workforce in year y that needed in job grade * proportion of workforce in cycle t that is of this gender, age, and job grade out of all workforce in this job grade in cycle t Rt+1(a,g,j) 11 = (Ny + D1…6t + Et -Nt) * Ny(j) / Ny * Nt(a,g,j) / Nt(j) (Equation 1.1)

Susceptibles (S)
Susceptibles in cycle t+1 = Susceptibles in cycle t -HIV incident cases + HIV-ve recruits -non-HIV related separations -promotions to higher job grade -losses to older age group + gains from younger age group + promotions from lower job grade (all in cycle t; all for the relevant age-, job grade-and gender-specific cohort) -t+1(a,g,j)-St(a,g,j) * d1,2,3(j)

-St(a,g,j) * (pr(j,y) + ar(a))/4 + St(a-1,g,j) * ar(a-1)/4 + St(a,g,j-1) * pr(j-1,y)/4 (Equation 2)
3a. Untested infected (Iu) 11 For each parameter, variables in brackets denote the categories that the parameter was stratified by. Parameters without variables in brackets denote the total population or rate across all categories. 12 If the number of recruits is positive in a year, prevalence in recruits is based on our analysis of workforce prevalence data (by year, gender and job grade); if it is negative (ie, retrenchees are being calculated), prevalence in recruits is based on general workforce prevalence in the model in the same year (by year, age, gender and job grade) Untested infected in cycle t+1 = Untested infected in cycle t + (HIV+ve recruits + incident cases in cycle t) * (1 -testing coverage) -untested infected in cycle t * testing coverage -HIV related and unrelated separations -promotions to higher job grade -losses to older age group + gains from younger age group + promotions from lower job grade -transitions to lower health states 13 + transitions from higher health states (all in cycle t; all for the relevant health state-, age-, job grade-and gender-specific cohort) -Int(s,a,g,j) * (pr(j,y) + ar(a))/4 + Int(s,a-1,g,j) * ar(a-1)/4 + Int(s,a,g,j-1) * pr(j-1,y)/4 -Int(s,a,g,j) * all x≥1 tpn,s,s-x + all x≥1Int(s+x,a,g,j) * tpn,s+x,s