Download Preview - American Economic Association

The Lazarus Drug
The impact of expansion of access to antiretroviral therapy for
HIV/AIDS on economic growth
Anna Tompsett∗
October 3, 2015
Abstract
How does health affect economic growth? This paper evaluates the short-run impacts of the rapid expansion in access to antiretroviral therapy for HIV/AIDS in low and middle-income countries that began in 2001.
The number of people receiving antiretroviral therapy in low and middle-income countries increased from
30,000 in 2001 to 9.7 million in 2012, precipitated by a rapid tenfold reduction in the price of antiretroviral
drugs faced by developing countries. Since the scale of a country’s antiretroviral therapy program may be
correlated with other factors that influence economic growth, I use predicted antiretroviral therapy coverage
as an instrument for observed coverage. I find that antiretroviral therapy coverage expansion led to increased
life expectancy and elevated growth rates in per capita GDP. The results suggest that the economic benefits
of antiretroviral therapy substantially outweigh the costs, and that the expansion of antiretroviral therapy
coverage is an important, but hitherto neglected, explanatory factor behind the African Growth Miracle.
∗ Email: [email protected]. Stockholm University, Department of Economics. Thanks to Heléne Lundqvist, Andreas
Madestam, Hanna Mühlrad, Cristian Pop-Eleches, Jakob Svensson, Jonas Vlachos, Brian Williams, Josh Graff Zivin and to
seminar participants at Stockholm University, Sussex University and the ASWEDE Conference on Development Economics
2014, for helpful comments and suggestions. Particular thanks to Paul Cahu and Falilou Fall for sharing data. Preliminary and
not for citation.
1
1
Introduction
Health improvements have intrinsic value in improving welfare. Studies of households and individuals
also confirm that relative improvements in health lead to relative increases in income and in investment in
human capital.1 However, a literature that dates back to Thomas Malthus raises the concern that health
improvements that increase the size of the workforce, or population, may not lead to aggregate growth
in per capita income, because of diminishing marginal returns to labour where there are fixed factors of
production.2 In particular, Acemoglu and Johnson’s influential study of the international epidemiological
transition3 concluded that, over a forty-year period, increases in population size outweighed increases in
total GDP, resulting in lower per capita GDP relative to prevalent trends.
Concerns about Malthusian mechanisms primarily relate to the medium- to long-term impacts of health
improvements. Many would counter that aggregate improvements in health are more likely to lead to
economic growth in the short run, when population is largely fixed. However, previous literature has provided
no robust causal evidence for these effects.4 The two principal empirical challenges are that other factors
may influence both health and income, and that increased incomes raise demand for goods that improve
health, such as clean water, leading to reverse causality. In this paper, I exploit the dramatic expansion
in access to antiretroviral (ARV) therapy for HIV/AIDS in low and middle-income countries, which began
in 2001, to provide the first robust causal evidence for positive short-run effects of health improvements on
growth in per-capita income.
Effective combinations of antiretroviral drugs were developed in the late 1990s, but prices were initially
prohibitively high, precluding widespread adoption in low and middle-income countries. In 2000, several
countries began producing generic copies of brand-name drugs. Pharmaceutical companies initially resisted
this competition aggressively but in 2001, facing mounting international opprobrium, they abruptly relaxed
their opposition to generic alternatives. The price of ARV drugs faced by low-income countries fell by a
factor of ten in under a year. Coverage of ARV therapy in low- and middle-income countries subsequently
rose from 30,000 in 2001 (Reich & Bery, 2005) to 9.7 million people at the end of 2012 (UNAIDS, 2013).
I exploit these dramatic global changes in ARV therapy coverage, combined with heterogeneity in HIV
infection rates at the time of the fall in prices, to construct a measure of predicted ARV therapy coverage
1 Reviews
include Strauss and Thomas (1998), Thomas and Frankenberg (2002), Bleakley (2010a).
as land or other natural resources such as forests or fisheries, or, in the short run, physical capital. Young (2005)
formalized these trade-offs in a model and simulation of the effect of the AIDS epidemic on economic growth, and concluded
that the epidemic would result in higher per-capita incomes in the long run, driven by the epidemic’s effect on population
growth.
3 Acemoglu and Johnson (2007).
4 Acemoglu and Johnson’s analysis focuses on the medium- to long-term effects of an improvement to health that took place
gradually over several decades.
2 Such
2
which I then use as an instrument for observed ARV therapy coverage.5 Predicted coverage reconstructs what
ARV therapy coverage would have looked like, if global changes in access to ARV therapy had been uniformly
distributed across all HIV positive individuals in low and middle-income countries. I thereby exploit only
the variation in ARV therapy coverage that is driven by global changes in drug price and availability. I find
that increasing ARV therapy coverage to an additional 1% of a country’s population is associated with a
5.3% increase in life expectancy (90% confidence interval: 2.6–7.9), and a corresponding 1.4 percentage point
increase in growth rates in per capita GDP (90% CI: 0.3–2.7).
HIV prevalence is associated with significant negative trends in measures of health and income growth
prior to ARV therapy coverage expansion. This pattern is consistent with the hypothesis that HIV/AIDS
itself has negative effects on both life expectancy and economic growth. These trends reverse at the onset of
ARV expansion, as shown in Figure 1.6 However, to reflect these trends, I include controls in all specifications
for pre-intervention (linear) trends, and my estimates reflect deviations from these pre-existing trends. In
this respect, my approach differs from previous studies that exploit similar strategies (Acemoglu & Johnson,
2007; Bleakley, 2007, 2010b; Cutler et al., 2010), which assume, and test to the extent possible, that preintervention disease prevalence has no direct effect on counterfactual trends.7
The empirical approach is motivated by the observation that access to a new health technology is partly
determined by the extent and success of a country’s implementation program, which may be correlated
with other variables that also influence growth. As a result, naı̈ve estimates of the effect of ARV therapy
on economic growth, which assume that ARV therapy coverage is not endogenous to growth, are probably
biased, although the sign of this bias is not ex ante obvious. Implementing the instrumental variables
strategy yields estimates that are larger than naı̈ve estimates, implying that the naı̈ve estimates are biased
downwards. One plausible explanation is that the size of a country’s ARV treatment program is correlated
with lagged economic growth, biasing downwards estimates of changes in growth rates resulting from the
ARV treatment program.
My analysis depends on the assumption that countries with high and low HIV prevalence did not experience any different simultaneous and unrelated trend changes when ARV therapy coverage expansion began.
This assumption would be violated if the expansion of ARV therapy coverage was itself caused by a reversal
in relative trends in life expectancy and growth in high HIV prevalence countries, but this seems highly unlikely, for two reasons. First, the drugs used to treat ARV were developed primarily for patients in wealthy
5 This approach was pioneered by Acemoglu and Johnson (2007), and later applied in other contexts with references to other
diseases including hookworm (Bleakley, 2007) and malaria (Bleakley, 2010b; Cutler, Fung, Kremer, Singhal, & Vogl, 2010).
6 Structural tests support the apparent correlation between the timing of the reversal in trends and the onset of ARV
expansion (Figure 6).
7 This identifying assumption has sometimes attracted criticism, including Bloom, Canning, and Fink (2009), in response
to Acemoglu and Johnson (2007).
3
countries, and second, the policy motivation for expansion of ARV coverage was falling life expectancy and
stagnant economic growth in countries with high HIV prevalence. Nonetheless, the results could still be
biased by unobservable factors that influence both health and growth, if their effects are correlated across
space with the HIV epidemic, and in time with the expansion of ARV therapy coverage. I examine several
potential alternative explanations — mineral, petroleum and export booms, and contemporaneous investment in fighting malaria — and conclude that these alternative explanations cannot account for the main
results. The results could also be biased upwards by regression to mean if high HIV prevalence in 2001 were
itself caused by poor economic growth in the preceding decade. I show that the results are robust to using
alternative instruments based either on observed HIV prevalence in 1990, or simulated HIV prevalence in
2001, using a simulation that exploits only the exogenous propagation dynamic of the HIV epidemic.8
The study exploits cross-sectional variation in the severity of the HIV epidemic, and what is essentially a
single global quasi-experiment in time. An inherent limitation of the study is the relatively small number of
countries affected. Only 15 countries have HIV prevalence greater than 5% in 2001, and all of these countries
are in sub-Saharan Africa.9 My main specification allows regions to experience different, flexible time trends,
and the results are similar when I limit the analysis to sub-Saharan Africa alone. I also show that the results
remain consistent when any one country is dropped from the sample, or when the three largest outliers
in HIV prevalence — which have the greatest leverage on the estimates — are together dropped from the
sample. Further, the results are robust to allowing aggregate trends to vary flexibly across space within
sub-Saharan Africa. The small number of affected countries ultimately restricts the precision with which
effects can be estimated, but this limitation must be weighed against the importance of the economic and
policy questions at stake.
My results contribute to the literature on the impact of health on economic growth. While numerous
studies of households and individuals show that health has a direct effect on economic outcomes,10 these
studies measure the relative impacts of relative improvements in health, and are unable to capture the
impact of aggregate improvements in health. An extensive macro-economic literature includes health as
a variable in cross-country growth comparisons, with or without instruments for the health variables of
interest.11 However, the results of these studies are contentious, because exploiting only cross-sectional
variation renders the estimates more sensitive to violations of the identifying assumptions. The closest
macro-scale study is Acemoglu and Johnson (2007), who also exploit a global change in health technology,
8 Using
data kindly shared by Paul Cahu and Falilou Fall, see Cahu and Fall (2011).
countries (of a main sample of 86 low and middle-income countries who report HIV prevalence in 2001) have HIV
prevalence greater than 1% in 2001.
10 Examples include Bleakley (2007, 2010b), Cutler et al. (2010).
11 Examples include Bhargava, Jamison, Lau, and Murray (2001), Gallup and Sachs (2001), Bloom, Canning, and Sevilla
(2004).
9 40
4
to measure the causal impact of health improvements on medium to long-term economic growth.12
In contrast to Acemoglu and Johnson (2007), this study focuses on estimating the short-run impacts of
improvements to health. I also find an increase in the population growth rate (0.39 percentage points, 90% CI:
0.13–0.65), primarily driven by changes in death rates. Theory predicts that this increase in the population
growth rate will, over time, offset some of the increase in growth in income. A simple conceptual framework,
based on the neo-classical (Solow) growth model, motivates a comparison of the relative magnitudes of the
estimated effects on growth in per capita GDP and population. Subject to conservative assumptions about
the persistence of the estimated short-run effects, this analysis suggests that the long-run effect on GDP per
capita is likely to be positive. These findings differ markedly from those of Acemoglu and Johnson (2007).
The most likely explanation is that Acemoglu and Johnson focused on diseases that primarily influence infant
mortality — including malaria, measles, diptheria, scarlet fever and whooping cough — while HIV/AIDS
primarily affects adults of working age who may already be parents, and has significant morbidity impacts as
well as mortality impacts. As a result, we might ex ante expect the productivity and human capital channels
to dominate Malthusian effects.13 More generally, this paper forms part of a growing literature14 which uses
natural experiments to establish causal relationships in macroeconomic data.
The results also contribute to explaining the ‘growth miracle’ in sub-Saharan Africa in the first decade
of the new millennium.15 The ‘growth miracle’ remains largely unexplained in the literature, although
observers have posited a number of possible explanations including increases in Chinese trade and FDI, rises
in foreign aid more broadly, a boom in commodity prices, a recovery in rainfall, or changes in the quality
of institutions. An accounting exercise suggests that growth attributable to the expansion of ARV therapy
coverage corresponds in magnitude to around a third of observed growth in per-capita GDP in sub-Saharan
Africa, between 2004 and 2012.16 These results imply that ARV therapy coverage expansion, and the
apparent resulting reversal of the negative impacts of the HIV/AIDS epidemic, represents an important, and
hitherto neglected, explanatory factor behind the recent recovery of economic growth rates in sub-Saharan
Africa.
12 A relevant historical example is Hansen and Prescott (2002), who study the effect of the plague on wages and rents between
1275 and 1800, finding that sharp drops in population were accompanied by rises in wages and falls in rents, consistent with
a Malthusian model of land use. Since the population fall was sharp and sudden, these effects correspond to ‘medium-term’
impacts, after population has adjusted.
13 I also control for pre-trends in outcome variables associated with HIV prevalence, having shown these to be important and
significant, while the identifying assumption in Acemoglu and Johnson (2007) is that pre-trends are equivalent in countries with
high and low pre-intervention mortality. If this identifying assumption were not valid, this might also explain the difference in
the results. However, Acemoglu and Johnson (2014) deal at length with Bloom et al.’s (2009) critique of their assumption on
these grounds, and conclude that there is no evidence that trends associated with pre-intervention mortality can explain their
results. It is perhaps more likely that both assumptions and findings are valid, but that for the reasons discussed in the main
text, Malthusian effects did indeed dominate in the context studied.
14 See Fuchs-Schündeln and Hassan (2015) for a summary.
15 Documented in Young (2012), Pinkovskiy and Sala-i-Martin (2014), Rodrick (2014).
16 The counterfactual scenario implied by the data and model is one of negative growth in countries with high HIV prevalence,
had the HIV epidemic gone unchecked.
5
Overall, my results suggest that ARV therapy coverage expansion was responsible for an additional $US33
billion17 in GDP growth in sub-Saharan Africa in 2012. In comparison, a conservative estimate of the cost of
providing ARV therapy to the 7.5 million people in sub-Saharan Africa who received ARV therapy in 2012
is $US9.4 billion. This suggests that the monetized benefits of funding ARV coverage outweigh the costs
by a factor of approximately 3.5 to 1, even without taking into consideration the value of directly increased
welfare through reduced morbidity and premature deaths averted.
In summary, I find that the expansion in ARV therapy coverage has yielded substantial benefits in terms
of economic growth, which partly explain the sub-Saharan ‘growth miracle’, and I provide the first robust
causal evidence on the short-run effects of aggregate improvements in health on growth in per-capita GDP.
The paper proceeds as follows. Section 2 provides a conceptual framework which motivates the analysis.
Section 3 discusses the history of the HIV/AIDS epidemic and ARV therapy, and describes the data. Section
4 describes the empirical strategy, and section 5, the main results. Section 6 discusses estimates of the
overall benefit of antiretroviral therapy and the implications for long-run economic growth, summarizes, and
concludes.
2
Conceptual framework
Following Acemoglu and Johnson (2007), let us use the closed economy neoclassical growth (Solow)
growth model as a conceptual framework. Our particular concern here is to compare the short- and long-run
impacts of an improvement in aggregate health. I therefore assume that economy i has the constant returns
to scale aggregate production function:
α
1−α
Yit = (Ait Hit ) Kit
(1)
where α ≤ 1, Kit denotes capital and Hit is the effective units of labor given by Hit = hit Nit , where Nit is
the total population (and employment) and hit is human capital per person. Departing from Acemoglu and
Johnson (2007), I assume that Malthusian effects are primarily driven by the fact that the stock of physical
capital Kit is fixed in the short run, and that land does not enter the production function. This is because
my concern is with equilibriums with non-zero population growth.18
I also assume that health — and in particular, the fraction of the population who are infected by HIV,
17 In
contemporary US$.
that land enters a constant returns to scale production function yields a Malthusian growth trap whenever
population growth is above zero, without introducing further assumptions on technological change.
18 Assuming
6
φHIV it , and the fraction of the population who are treated with antiretroviral (ARV) therapy, φARV it 19 —
may increase output per capita through a variety of channels including an increase in the size of the effective
labour force, more rapid human capital accumulation or direct positive effects on total factor productivity
(TFP). To capture these effects in a reduced-form manner, I assume the following relationships:
γ
(2)
η
(3)
Ait = Āi (1 − φHIV it + φARV it )
hit = h̄i (1 − φHIV it + φARV it )
where Āi and h̄i designate baseline differences across countries These functional forms assume instant adjustment of Ait and hit to changes in φHIV it or φARV it . I will later on discuss the implications of relaxing
this assumption. Finally, health also naturally leads to greater population growth by reducing mortality and
possibly by increasing fertility, so I assume that
λ
nit = n̄i (1 − φHIV it + φARV it)
(4)
where nit is the population growth rate, and n̄i designates baseline differences in intrinsic population growth
rates across countries.20
Now imagine the effect of a change in access to ARV therapy from some baseline value φARV it0 at t0 to
a new value φARV it1 at time t1 . Suppose that while the aggregate level of health in the population changes
relatively quickly, the population and capital stock remain fixed in the short run at their original values N̄it0
and k̄it0 , where k̄it0 is the capita stock in per capita terms. In that case, I obtain the following relationship
between per capita income yit and ARV therapy coverage φARV it 21 :
ln yit = (1 − α) ln k̄it0 + α ln Āi + α ln h̄i − (α (γ + η)) φHIV it + (α (γ + η)) φARV it
(5)
for t = t0 , t1 . This equation shows that improvements to health obtained by increasing access to ARV
therapy should unambiguously raise income per capita in the short run.22
19 Where
φARV it ≤ φHIV it ≤ 1, reflecting the fact that HIV infection is a pre-condition for treatment with ARV therapy.
somewhat artificially constraints population growth to be positive, although since population growth is positive in
94% of observations in the main sample and 98% of observations in sub Saharan Africa, it is not particularly obvious that this
model should diverge too strongly from reality. Note that here I diverge again from Acemoglu and Johnson (2007) who assume
that population adjusts instantaneously to a new, constant level which is a deterministic function of life expectancy.
21 Noting that ln 1 + x ≈ x for small x.
22 Note that the above also assumes that φ
HIV it is fixed in the short run, which is probably reasonable, as since ARV therapy
20 This
7
In the long run, the supply of capital will adjust to reflect the changes to human capital and total
productivity, and to the increased population growth rate. To model the possible long-run outcomes, suppose
that each country i has a constant saving rate equal to si ∈ (0, 1),23 and assume that the depreciation rate
is negligible with respect to population growth so that the evolution of the capital stock in country i at
time t is given in per-capita terms by kit+1 = si yit + (1 − nit ) kit . After adjustment, the steady-state capital
stock will be ki =
si yi
nit .
Using this value of the capital stock together with the earlier assumptions yields the
following long-run relationship between ARV therapy coverage and log income per capita
1−α
1−α
1−α
yit = ln Āi + ln h̄i +
ln si −
ln n̄i − (γ + η) −
λ φHIV it
α
α
α
1−α
+ (γ + η) −
λ φARV it
α
(6)
again for t = t0 , t1 . This equation shows that in contrast, improvements to health obtained by increasing
access to ARV therapy have an ambiguous effect in the long run, depending on whether or not the positive
effects on TFP, labour force participation and human capital (here, γ + η) outweigh the potential negative
effects resulting from the effect of the increase in population growth on the rate of capital accumulation
(here,
1−α
α λ).
Note also that the long-run impact on income via the human capital and TFP channels is
larger than the short-run impact, which follows from capital accumulation.
The empirical work I present here allows me to estimate λ, the effect of ARV therapy on population
growth, and and α (γ + η), the short-run effect of ARV therapy on growth through human capital and TFP.
A comparison between the relative magnitudes of these effects, adjusted by labour’s factor share to reflect
Equation 6, will then provide a first-order estimate as to whether the short-run increases in income per
capita are likely to translate into lasting positive effects. I will return to these comparisons in Section 6.
3
HIV/AIDS and antiretroviral therapy: Context and Data
3.1
The HIV/AIDS epidemic
The Human Immunodeficiency Virus (HIV) is a disease that attacks and weakens the immune system,
eventually rendering it susceptible to diseases or infections that healthy individuals would be able to fight off.
The disease has a relatively long latent period — two to fifteen years, depending on the individual — after
which the full blown form, Acquired Immunodeficiency Disease (AIDS), develops. AIDS is characterized by
is not a cure, a change in φARV it does not lead to any immediate change in φHIV it . However, the rate of change of φHIV it
responds to a change in φARV it via two offsetting effects: reducing the death rate of HIV positive patients tends to make the
rate of change of φHIV it increase, whereas reducing transmission rates tends to make the rate of change of φHIV it decrease.
23 We could also assume a savings rate s
it as a function of (1 − φHIV it + φARV it ), to reflect the fact that savings rates
respond to beliefs about life expectancy. This would lead to larger positive, or smaller negative, medium- and long-run effects
of ARV therapy coverage expansion on per capita income.
8
a number of specific opportunistic diseases, including pneumonia, wasting disease and various viral-induced
cancers, such as the otherwise rare Kaposi’s sarcoma.24
HIV is a retrovirus, a type of virus characterized by the procedure by which it replicates inside the cells
of its host. HIV is transmitted by the exchange of bodily fluids: blood, breast milk, semen, and vaginal
secretions. As a result, the primary transmission channels are unprotected sex, blood transmissions, and
sharing of contaminated needles or other medical equipment. HIV is also transmitted from mothers to their
children during pregnancy, labour or breastfeeding.
The first recognised cases of HIV/AIDS were in homosexual men in the United States, among whom an
unusually large number of cases of both Kaposi’s sarcoma and pneumonia were diagnosed. The term AIDS
was coined in 1982 (Kher, 2003) and the HIV virus itself was identified in 1983 (Barré-Sinoussi et al., 1983;
Popovic, Sarngadharan, Read, & Gallo, 1984). Medics in Uganda in turn identified a disease causing similar
symptoms, locally known as the ‘slim disease’ because of its wasting effect. There was initial skepticism as
to whether the two epidemics were related, given that the affected populations differed so widely in their
demographic characteristics, until tests proved that the virus was the same. It was later established that
HIV originates in Africa, in a family of viruses that is endemic to primate populations in West and Central
Africa (Keele et al., 2006). Genetic evidence shows that the disease was passed to humans via cross-species
transmission on several separate occasions, possibly during bushmeat hunting.
Most estimates date these transmission events to the first half of the 20th Century (Korber et al.,
2000). Researchers have offered multiple hypotheses to explain how these localized cases of animal-human
transmission developed into a full-blown pandemic, including social and medical changes that took place in
the 20th Century. Although HIV has been detected in stored blood samples from as early as 1959, it was
apparently rare at the time (Nahmias et al., 1986).25 By the mid-1990s, however, the disease had spread
across Western Africa, into East and Southern Africa, and overseas, with an estimated 13 million people
infected in sub-Saharan Africa out of a total of 20 million worldwide (Mertens & Low-Beer, 1996). Figure 2
shows the distribution of HIV prevalence in 2001.
There are multiple channels via which HIV/AIDS might influence economic growth, from direct effects on
patient productivity, to secondary productivity effects as a result of carers being forced to leave the workforce,
to the consequences for children of being orphaned as a result of both parents succumbing to AIDS. For
example, Evans and Miguel (2007) show that orphaned children were less likely to attend school, both
after and immediately before a parent’s death, suggesting that both parental morbidity and mortality affect
schooling participation. The risk of HIV/AIDS infection alone may also influence investment decisions,
24 A useful and well-referenced source of information on the HIV/AIDS epidemic and ARV therapy is www.avert.org, which
provided the background information for much of the following account.
25 One sample out of 818 collected in 1959 in Leopoldville (now Kinshasha) tested positive for the virus in 1986.
9
effectively leading to a higher discount rate. For example, Forson (2011) finds evidence that HIV/AIDS
prevalence has a negative impact on investments in human capital.
However, obtaining unbiased estimates of the aggregate impact of the epidemic is difficult. Studies that
compare countries with high HIV prevalence to those with low or no prevalence may be biased because the
factors that cause HIV to spread and develop, or that determine the extent and success of efforts to control
or halt the spread of the disease, may also be correlated with economic growth. Early estimates of the impact
of HIV/AIDS on growth in GDP per capita ranged from an increase of 0.2% to a decrease of 1.2% (United
Nations Department of Economic and Social Affairs, 2004).26 Two more recent, unpublished studies make
serious efforts to address the endogeneity problem but also reach essentially opposite conclusions: Ahuja,
Wendell, and Werker (2009) find no measurable impact on growth in GDP per capita27 while Cahu and Fall
(2011) find substantial negative impacts.28 In sum, the evidence to date on the impact of HIV/AIDS on
aggregate growth in per capita income is inconclusive.
3.2
Antiretroviral (ARV) therapy
The term antiretroviral (ARV) therapy describes a class of drugs that combat HIV by interrupting various
stages of the retrovirus’ life cycle. ARV therapy dramatically reduces, but does not eliminate, viral load.
ARV therapy does not therefore cure patients, but, in developed countries at least, it effectively converts
HIV from an imminent death sentence into a chronic latent condition. HIV mutates very rapidly, which
initially enabled it to development resistance to ARV therapy. Indeed, the first ARV drug was developed in
1987, but resistance severely and rapidly limited the drug’s effectiveness. In 1996, studies established that
ARV drugs are substantially more effective when prescribed in combination, primarily because this reduces
the likelihood that the virus can develop resistance (Hammer et al., 1997; Gulick et al., 1997). The cocktail
of three drugs that became standard reduced viral loads, incidence of opportunistic diseases, and mortality
by as much as 60-80%.29 ARV therapy appeared to bring back patients from the brink of death, leading to
the coining of the epithet the Lazarus drug.
Current WHO policy is to prescribe ARV therapy to patients above a threshold viral load in several waves.
First line therapy tries the cheapest, most readily available drugs, but second and even third line therapy
is available, should the first line drugs become ineffective in controlling a patient’s viral load. Because ARV
therapy reduces the viral load, it also dramatically reduces transmission of the virus (Palmer et al., 2008).
26 Estimated effects on total GDP were largely more negative (United Nations Department of Economic and Social Affairs,
2004).
27 Using male circumcision rates as an instrument for HIV prevalence.
28 Using instrumental variables to isolate the independent propagation dynamic of the epidemic and find substantial effects,
amounting to a long-run reduction of 12% in GDP per working age population in sub-Saharan Africa.
29 See e.g. Moore and Chaisson, 1999.
10
However, the price of combined ARV therapy was initially prohibitively high, on the order of $10,000 to
$15,000 per patient per year. As a result, expansion of coverage to the developing world was severely limited.
Between 2000 and 2001, a series of events took place which together combined to reduce the price of the
drugs faced by low and middle-income countries by an order of magnitude. India began production of generic
versions of ARV drugs at substantially lower prices. Big pharmaceutical companies initially resisted this
violation of intellectual property laws, but this resistance fell away in the face of overwhelming international
opprobrium. Two significant events signalled prevailing opinion. First, the Clinton administration decided
to ignore violations of American patent law in the case of sub-Saharan African countries providing ARV
therapy to their citizens. Second, pharmaceutical companies withdrew their attempt to prosecute the South
African government for allowing production and import of generic versions of ARV drugs. By the end of
2001, both brand and generic versions of the three-drug cocktail were available in low and middle-income
countries for less than US$1000 per patient per year (Médecins Sans Frontières, 2001). Figure 3 illustrates
this sharp fall in prices, which have since fallen further still; the price of first-line ARV drugs was US$115
per patient per year in 2013 (WHO, 2014).
This fall in prices precipitated a concerted global effort to increase access to ARV drugs in low and middleincome countries. The price of ARV therapy represented one critical barrier to expansion of treatment,
but effective expansion of treatment coverage faced other important barriers, such as the absence of trained
staff, and the need to establish HIV testing programs and reliable supply chains for ARV drugs. Importantly,
these barriers likely differed between countries, for institutional and other reasons that may also have affected
economic growth. The WHO have repeatedly issued targets for scaling up access to ARV therapy. The Global
Fund for AIDS, TB and Malaria, founded in 2002, estimates that programs sponsored by the fund were
responsible for bringing access to ARV therapy to 6.6 million people (by the middle of 2014), approximately
two thirds of those with access to ARV therapy in low and middle-income countries. By 2012, a total of
9.6 million people in low and middle income countries were receiving ARV therapy. Figure 4 shows ARV
therapy coverage as a percentage of people with HIV infection in 2012.
Clinical studies show that combination ARV therapy dramatically reduces both morbidity and mortality
(Moore & Chaisson, 1999). Studies in contemporary developing country settings indicate that ARVs have
corresponding impacts in real-world settings. For example, Bor, Herbst, Newell, and Bärnighausen (2013)
showed that scale-up of ARV therapy in a population with high HIV prevalence led to large increases in life
expectancy; increasing coverage to 7% of the population raised population life expectancy from 49.2 years
to 60.5 years.
A number of studies provide estimates of the positive impact of ARV therapy on individual productivity
(Thirumurthy, Zivin, & Goldstein, 2008; Bor, Tanser, Newell, & Bärnighausen, 2012; Habyarimana, Mbakile,
11
& Pop-Eleches, 2010). Studies that take the individual as the unit of analysis may underestimate the
aggregate impact of ARV therapy on labour market outcomes, since HIV/AIDS may also affect carers’
participation in the labour force, or may affect other decisions about investment or labour force participation
via beliefs about risk. For example, a recent study by Baranov, Bennett, and Kohler (2012) finds far-reaching
impacts of ARV therapy on productivity, even among the HIV negative, supporting the case for mechanisms
beyond labour market participation. Further, the scale-up of ARV therapy may have a direct impact on
local economies as a direct result of local employment and procurement associated with the provision of
ARV therapy. This paper provides what is, to my knowledge, the first robust estimates of the impacts of
ARV therapy programs on growth in per-capita income at country scale.
3.3
Data
I use publicly available yearly country aggregate data from the World Bank and UNAIDS. Key variables
include GDP per capita in constant 2005 US$, population, life expectancy, HIV prevalence30 and the number
of people receiving ARV therapy.31
Data on GDP per capita and population is available for 168 countries.32 HIV prevalence data is only
available from 1990, and is reported in 2001 for 91 countries, of which 5 are high-income countries, 23 are
upper middle-income countries, 38 are lower middle-income countries and 25 are low-income countries. The
sample of 86 low and middle income countries for which HIV prevalence is reported in 2001 constitutes the
main sample for the analysis.33
Life expectancy from birth measures the expected lifetime of a hypothetical newborn who experiences
prevailing age-specific mortality rates throughout their lifetime.34 Life expectancy calculations require agespecific mortality rates, but data on adult mortality are not available for all countries, particularly in subSaharan Africa. In the absence of full data on mortality, life expectancy calculations use data on infant
mortality, but also reply on appropriate model life tables. For countries with HIV prevalence higher than 2%
of the population, the estimates are adjusted to explicitly model the effect of HIV/AIDS, drawing on data
collected from UNAIDS, including treatment data (UNDESA, 2014). The results I present on life expectancy
— which I include to demonstrate that ARV therapy has a first order effect on health — must therefore be
30 All
obtained via the World Bank Databank. HIV prevalence for the population of people aged 15-49 drawn from UNAIDS
estimates.
31 Obtained from UNAIDS Global AIDS Response Progress Reporting.
32 Including 52 high-income countries, 45 upper middle-income countries, 44 lower middle-income countries and 27 low-income
countries.
33 In robustness checks, I also test whether the results change when I include other low and middle income countries for which
HIV prevalence is not reported, assuming that HIV prevalence is zero in these countries.
34 Data on life expectancy provided by the World Bank is compiled from other sources, including the United Nations Population Division World Population Prospects; United Nations Statistical Division Population and Vital Statistics Report; and
Census reports and other statistical publications from national statistical offices.
12
interpreted with caution, given that I cannot rule out the possibility that I may be partially recovering a
modelled impact. For comparison, I also present results on crude death rates and mortality broken down by
age group and gender. However, some of these estimates may also be adjusted, so the degree to which they
provide an independent measure of the impact of ARV therapy on health is limited.
I calculate the overall global proportion of HIV positive patients receiving ARV therapy using data from
reports by UNAIDS and the WHO (WHO, 2011, 2013). Data on the number of people receiving ARV
therapy in a given country is reported on an annual basis since 2004 for between 94 and 111 countries. For
the largest set of countries reporting, 9 are high income countries, 36 are upper middle income countries, 40
are lower middle income countries and 26 are low income countries. My estimates of global coverage suggest
that prior to 2004, less than 1% of HIV positive patients in low and middle-income countries received ARV
therapy. Lacking better information, I therefore treat ARV therapy coverage as zero in low and middleincome countries before 2004.35
4
Empirical Strategy
The empirical analysis is motivated by Equation 5, which describes the short-run impact of increasing
access to antiretroviral (ARV) therapy, and thereby improving health, on income. As such, the right hand
side variable of interest is the proportion of the population in country i that receives antiretroviral therapy
at a time t, which I will denote ARVit . This is the natural way to construct the variable of interest — rather
than regarding the variable of interest as the proportion of HIV positive patients receiving antiretroviral
therapy — because the fraction of the population receiving ARV therapy captures the effect on the effective
labour force. Figure 5 shows how this measure of ARV therapy coverage varies across countries in 2012.
If we were prepared to make the very strong assumption that all short-run impacts happen in the same
year, then following Equation 5, the natural specification to examine might be:
∆yit = π∆ARVit + ζi + µrt + it
(7)
where π is the coefficient of interest, ∆yit is the change in log per-capita GDP in country i in year t, ∆ARVit
is the increase in ARV therapy coverage, and ζi and µrt are country and region-year fixed effects, respectively.
The assumption that income adjusts immediately seems implausibly strong, given any reasonable level of
labour market frictions. Neglecting lagged effects would likely bias upwards the results, as there is positive
35 To
the extent that this fails to capture the small increases in coverage in 2003 and 2003, this introduces additional
measurement error, which might further bias OLS estimates of the effect of ARV therapy on outcome variables towards zero.
13
serial correlation in ∆ARVi,t .36
Adjusting the equation above to reflect these lag effects yields the distributed lag equation:
∆yit =
k
X
πs ∆ARVit−s + ζi + µrt + it
(8)
s=0
where k is some suitable time period after which there are no persistent (lagged) effects on growth rates, and
the coefficients πs , s ∈ (0, k) are the coefficients of interest, which presumably differ given that lagged effects
are likely to differ from contemporaneous effects. However, it proves unfeasible to estimate this equation
using the instrumental variables strategy, for reasons that will become clear when I discuss the instrument
in more detail. As a result, I impose the further assumptions that i) k >= 1137 and ii) πs = π ∀ s ∈ (0, k),
or equivalently, that the coefficients on lagged and contemporaneous effects be equal. In practice, the result
of making this assumption implies that I estimate the mean effect of a change in ARV therapy over the
following years. This yields the following main equation:
∆yit = πARVit + ζi + µrt + θ (HIVi2001 × t) + it
(9)
I include the term θ (HIVi2001 × t) because countries with high HIV-prevalence experienced different time
trends in outcome variables prior to the onset of ARV therapy scale-up, as clearly shown in Figure 1.38 The
inclusion of this control will be essential for the validity of the exclusion restriction.
Our fundamental concern with this equation is that even conditional on country and region-year fixed
effects, ARVit may still be correlated with other, possibly unobservable factors that influence the outcome
variables through other channels. In formal terms, that cov (ARVit , it ) 6= 0. This would result in biased
estimates of πs . Ex ante, the sign of bias is not clear: the presence of HIV is a necessary condition for the
presence of an ARV therapy programs, and HIV prevalence might influence the estimates of πs even after
including the country fixed effects and linear trends that vary as a function of HIV prevalence. External
support for ARV therapy programs might be concentrated in countries with low expectations of economic
growth, which ceterus paribus would tend to bias the estimates downwards. However, countries that are
successful in providing ARV therapy to their citizens might also be successful on other metrics; for example,
36 Estimating equations analogous to those in the main results sections with similar specifications to Equation 7 indeed yield
much larger and more statistically significant coefficients.
37 Noting that ARV is effectively zero for all years before 2001.
it
38 In preliminary analysis, I confirm that these differences in trends are statistically significant (Table 1). A more general, but
less parsimonious approach, would be to include country-specific linear trends in the regressions; I also implement this strategy
and show that the results change very little.
14
they might have improving institutions more generally. This would tend to bias the estimates upwards.
To isolate variation in ARVit , that is otherwise uncorrelated with factors influencing the outcome variables
of interest, I exploit a strategy first implemented in Acemoglu and Johnson (2007) with respect to the
international epidemiological transition and later replicated in other contexts for other diseases (e.g. Bleakley,
2007; Bleakley, 2010b; Cutler et al., 2010). This strategy uses predicted coverage of a health intervention as
an instrument for observed coverage of the health intervention. In this case, I predict ARV therapy coverage
using the interaction between HIV prevalence in 2001 (at the time when prices fell precipitously) and global
coverage of ARV therapy in low and middle income countries in the following first stage equation. Predicted
coverage essentially reconstructs what ARV therapy coverage would look like, if global advances in coverage
were evenly distributed across all HIV positive individuals in the world.
ARVit = β HIVi2001 × ARV t + ζ̃i + µ̃rt + θ̃ (HIVi2001 × t) + νit
(10)
where ARV t is the proportion of HIV positive individuals receiving ARV therapy across all low and middleincome countries at time t. I also include the term θ̃ (HIVi2001 × t) for consistency, although the inclusion
of this control matters little for the first stage estimates. Exactly equivalent first stage equations can be
constructed to predict ∆ARVit using HIVi2001 and ∆ARV t .
For predicted ARV coverage to be a valid instrument for observed ARV coverage, it must first predict
observed ARV coverage in a sufficiently strong ‘first stage’. Table 2, Panel A, Column 3, shows the results
from this first stage. The F-statistic associated with the coefficient on the instrument is 84, indicating that
the instrument (predicted ARV coverage) is a strong predictor of the endogenous variable (observed ARV
coverage). The instrument must also satisfy monotonicity: predicted ARV coverage should always have a
positive impact on true ARV coverage. It is difficult to construct a case where countries with high HIV
prevalence should have lower total coverage of ARV therapy as a fraction of the population, since HIV
prevalence is a prerequisite for prescribing antiretroviral drugs.
I then use predicted ARV coverage, given by HIVi2001 ×ARV t , to instrument for observed ARV coverage,
ARVit , in Equation 9. The reason that I cannot use this approach to estimate Equation 8, the distributed lag
specification, should now be clear. The instrument only exploits one dimension of cross-sectional variation,
and one dimension of temporal variation. As a result, lagged values of the instrument are near to collinear,
resulting in a weak instrument problem if I were to use contemporaneous and lagged variables of predicted
changes in ARV coverage to instrument for the corresponding set of contemporaneous and lagged changes
in observed ARV coverage.
15
I construct similar equations to Equation 9 to estimate effects on other outcome variables including life
expectancy (in logs) and differences in log population. I present results on life expectancy to illustrate that
the expansion of ARV therapies indeed had an impact on health, a necessary pre-condition to measuring how
improvements in health affect aggregate economic growth. However, the results on life expectancy may be
affected by the way in which life expectancy is calculated in countries with limited data on adult mortality,
and in particular, may be partially recovering a modelled effect of the HIV/AIDS epidemic, rather than a
measured observation.
The final condition for instrument validity is that the instrument (predicted ARV therapy coverage) must
satisfy the exclusion restriction i.e. conditional on the fixed effects included, must not be correlated with the
outcome variables (health, economic growth) via any other channel than the endogenous variable (observed
ARV therapy coverage). One major concern is that ARV therapy coverage is mechanically correlated with
HIV prevalence. Although including country fixed effects partially accounts for this, the HIV epidemic
is dynamic, and not static. For this reason I also allow countries to experience different linear trends in
growth that vary as a function of HIV prevalence. Including the differential time trend term means that the
measured effect is on changes in trends. The identifying assumption is therefore that changes in trends in
outcome variables following the expansion of ARV therapy coverage are uncorrelated with HIV prevalence
in 2001 through any other channel. In more formal terms, the exclusion restriction is that:
cov HIVi2001 × ARV t , it = 0
(11)
This assumption seems plausible, but could be violated if there are non-linear trends associated with HIV
prevalence, and in particular, if there are unobservable factors that are correlated across space with the HIV
epidemic and across time with the scaling-up of antiretroviral therapies. I will test four specific plausible
alternative explanations. Noting that Chinese influence — whether measured by aid, trade or FDI — follows
a similar pattern in time to the scale-up of ARV therapy, I allow countries with high relative levels of mineral
and oil resources, or high levels of exports, to have different flexible time trends. Given that spending on
malaria control via the Global Fund for Aids, Tuberculosis and Malaria also increased over a similar time
frame, I also allow countries with different malaria ecologies to have different flexible time trends. None of
these additional controls substantially alters the main results, ruling out these alternative explanations for
the main results.
My implementation of the approach of using predicted access to a health intervention to instrument for
observed access differs from prior applications (Acemoglu & Johnson, 2007; Bleakley, 2007, 2010b) where
16
the identifying assumption — tested to the extent possible — is the equivalence of pre-trends in countries
with high and low disease prevalence prior to the intervention. In this case, I clearly reject the possibility
that pre-trends are equivalent, and therefore adjust the regression specification accordingly.
Both dependent variables and treatment variables are likely to exhibit both spatial and serial correlation.
The presence of positive serial and/or spatial correlation tends to lead to over-rejection of null hypotheses,
when observed differences are in fact a product of chance. Tests indicate that, conditional on the modelled
fixed effects, serial correlation persists in this context for 2-3 years.39 I therefore adopt the conservative
approach of clustering standard errors at the country level, which allows for the correct statistical inference
in the context of any type of serial correlation (Wooldridge, 2001; Bertrand, Duflo, & Mullainathan, 2004).
Constructing Newey West standard errors with a lag of 5 years yields considerably more precise estimates,
and could also be justified given the observed pattern of serial correlation. However, I err on the conservative
side and cluster at the country level.
I also test for spatial correlation by regressing the residual in one country on the mean residual in its
neighbouring countries.40 These tests suggest that spatial correlation is primarily associated with variations
in time trends by region; spatial correlation disappears either in specifications which include region-year
fixed effects, or when considering sub-Saharan Africa alone. As I include region-year fixed effects in my main
specification, I do not further adjust the standard errors for spatial correlation.
5
Results
In this section, I first present preliminary analysis that demonstrates that a trend break exists around
the time of the fall in prices of ARV drugs, in 2001. Before 2001, HIV prevalence is associated with falling
life expectancy and falling growth rates in GDP per capita. After 2001, the reverse is true. Structural tests
confirm that the location of the break-point is close to 2001. This reduced form evidence motivates the rest
of the analysis; any alternative explanation for the results must explain this reversal of trends associated
with HIV prevalence.
I then present the main results, showing that covering an additional 1% of the population with ARV
therapy leads to a 5.3% increase in life expectancy and a 1.4 percentage point increase in the growth rate of
GDP per capita. I show that these results are robust to a series of specification checks, and that collapsing
the data to a cross section yields similar reduced form estimates. I then examine in detail four alternative
plausible mechanisms which might explain the results, and in each case find no evidence to support these
alternative mechanisms.
39 See
40 See
Appendix Table A1, Panel A.
Appendix Table A1, Panel B.
17
Throughout, I present results on life expectancy to demonstrate that scaling up ARVs has a first order
effect on health. However, the results on life expectancy should be interpreted with caution, given that life
expectancy is calculated using modelled data where statistical data is missing or unreliable, and in particular,
incorporates explicit modelling of the HIV/AIDs epidemic in countries with significant HIV prevalence. As a
result, I cannot rule out the possibility that the analysis of effects on health may partially recover modelled
effects.
Finally, I also present results on population growth, showing that covering an additional 1% of the
population with antiretroviral therapy is associated with a 0.39 percentage point increase in the population
growth rate, driven primarily by mortality and migration. In the conclusion, I discuss the implications of
these findings for the prospects for long-run economic growth.
5.1
Preliminary analysis
In this section, I show that countries with high HIV prevalence experience different trends — relative to
countries with low HIV prevalence — before and after the precipitous fall in ARV prices in 2001. Figure 1
illustrates these trend breaks. Countries with high HIV prevalence have lower life expectancies than other
low- and middle-income countries in 1990. These differences widen until just after 2001, when the trend
reverses and the gap in life expectancies begins to close. Countries with high HIV prevalence have similar
annual growth rates in GDP per capita in 1990, with growth rates actually marginally but insignificantly
higher in high HIV prevalence countries than in other low- and middle-income countries. Growth rates in
high HIV prevalence countries decline, relative to other low- and middle-income countries, over the following
decade, but this trend also reverses after 2001. The take-off of ARV therapy coverage is slow in the first
years after 2001, but visual comparison suggests that the shape of the response after 2001 is consistent with
the shape of the ARV scale-up curve. However, the large and unexpected drop in prices in 2001 provides a
natural motivation to focus on 2001 as the timing of the discontinuity.
The results shown in Table 1 confirm that HIV prevalence in 2001 is associated with negative trends in life
expectancy and growth rates in GDP per capita before 2001, and that these trends reverse thereafter. These
relationships are statistically significant, and similar both across the larger sample of low and middle-income
countries, and within sub-Saharan Africa. To be specific, I estimate the following equation:
yit =θ1 (HIVi2001 × t) + θ2 (HIVi2001 × P OSTt )
(12)
+ θ3 (HIVi2001 × P OSTt × t) + ζi + µt + it
where HIVi2001 is HIV prevalence in 2001; t is event time (relative to 2001) and P OSTt is an indicator
18
variable which takes the value 1 after 2001, and 0 in the years up to and including 2001. The coefficient
θ1 captures pre-ARV trends associated with high HIV prevalence. The coefficient θ3 captures the change in
these trends after expansion of ARV coverage begins. Rejecting the null hypothesis that θ3 = 0 implies that
there are statistically different time trends in countries with high HIV prevalence before and after 2001. I
include the interaction term associated with the coefficient θ2 for completeness. As in all specifications, I
cluster standard errors by country.
HIV prevalence is associated with significant negative trends in life expectancy (columns 1 and 2) and
growth in per-capita GDP (columns 3 and 4) before 2001. HIV prevalence is also associated with significant
positive changes in these trends after 2001. These patterns are consistent in both the main sample of all
low- and middle-income countries and within sub-Saharan Africa alone. The coefficients on the changes in
trend (θ3 ) are larger in magnitude than the initial negative trends (θ1 ), suggesting that the negative trend
more than reverses, also consistent with Figure 1.41 The interaction between the post-2001 dummy variable
and HIV prevalence in 2001, included for completeness, is generally small relative to the trend coefficients,
amounting to between 0 and 3 years of growth at pre-ARV trends, consistent with a trend break — rather
than a change in levels — in 2001.
The above analysis shows that trends associated with HIV prevalence differ before and after 2001, the
year in which prices of ARV drugs fell precipitously. However, these tests assume that the year of the trend
break is known. The above analysis therefore suggests a second set of tests, which treat the year of the trend
break as unknown, to establish whether the timing of the trend break is consistent with the timing of onset
of expansion of ARV therapy coverage. I present results from structural tests that confirm 2001 as the most
likely timing of the trend break with respect to growth in per capita GDP in Figure 6.
The results shown in Figure 6 represent three tests for the timing of a trend break, following Hansen
(2001). First, I run a series of regressions with equations similar to Equation 12, but with possible break
years between 1996 and 2006. Figure 6a shows the results of F-tests of the joint significance (Chow tests)
of either coefficients equivalent to θ2 and θ3 , or to θ3 alone. One approach to identifying the location of a
trend break is to pick the break year which gives the maximum value of the Chow test. These two tests
select the years 2001, and 2000, respectively. However, Hansen’s preferred test is to select the break year
that minimises the residual variance or sum of squares. This test selects 2002 as the break year.
Taken together, the results support a structural trend break around 2001, at which point the relationship
between trends in health and economic growth, and HIV prevalence, reverses. This preliminary analysis
motivates the rest of the paper; an alternative mechanism must explain why trends associated with HIV
41 However,
I generally cannot reject the hypothesis that the coefficients θ1 and θ3 sum to zero, except for the full sample
with log life expectancy as an outcome variable.
19
prevalence reverse at the specific time that ARV therapy coverage expansion began. Further, this preliminary
analysis emphasises the importance of measuring impacts relative to pre-existing trends, because countries
with high and low HIV prevalence clearly experience different trends prior to 2001. In the main analysis,
I achieve this in one of two ways: either by including an interaction term between HIV prevalence in 2001
and a linear time trend, or by including country-specific linear time trends. Either of these two approaches
yields very similar results.
5.2
Main results
Table 2 shows the main results. I find that increasing coverage of ARV therapy leads to substantial
increases in life expectancy and growth rates in per capita GDP.
Panel A shows the first stage regression, which demonstrates that the instrument (predicted ARV coverage) is a strong predictor of true ARV coverage, conditional on country and region-year fixed effects, and
linear time trends associated with HIV prevalence. The F-statistic for the coefficient on the instrument is
84.
Panels B and C show the results from a series of specifications for the main outcome variables: life
expectancy and growth in per capita GDP, respectively. In column 1), I show the raw correlations between
the outcome variables and observed ARV therapy coverage in the country. These correlations do not, of
course, reflect causal relationships. Countries with higher observed ARV therapy coverage mechanically
also have high HIV prevalence, and have correspondingly lower life expectancies. Countries with higher
observed ARV therapy coverage have slightly higher mean growth rates in per capita GDP, although it is
not particularly clear which sources of bias drive this correlation.
In column 2), I show how these correlations alter when I include controls to remove overall global time
trends, modelled by region-year fixed effects, mean differences between countries over time reflecting timeinvariant country characteristics, and linear trends in the outcome variables associated with HIV prevalence.
These regressions correspond to naı̈ve estimates of Equation 9, treating observed ARV therapy coverage
as exogenous conditional on overall time trends and country fixed effects. After including these controls,
the sign of the coefficient on life expectancy reverses (Panel B, column 2), implying that life expectancy
increases, relative to previous trends, after countries expand antiretroviral therapy coverage. The coefficient
on growth (Panel C, column 2) doubles in magnitude, but is not statistically significant.
Column 3) shows the reduced form results, in which I regress the outcome variable directly on the
instrument and controls from Equation 9. These regressions show significant relationships, in the expected
directions. Countries with high HIV prevalence show increasing life expectancy and increasing growth rates
20
in per capita GDP that are strongly correlated in time with global mean coverage of ARV therapy.
Column 4) shows the results from implementing the instrumental variables strategy, in which I use
predicted ARV therapy coverage as an instrument for observed ARV therapy coverage. The coefficients
resulting from the reduced form and IV estimates are consistent in sign with the OLS regressions that
include controls, but are substantially larger than the OLS relationships, and are statistically significant.
This pattern of results suggests that observed ARV therapy coverage is correlated with other factors that are
negatively associated with changes in trends in life expectancy and economic growth. At this time, the most
plausible explanation seems to be that the scale of a country’s ARV therapy program, conditional on HIV
prevalence, is correlated with higher than average growth in the years preceding widespread availability of
antiretroviral drugs. As a result, the naı̈ve OLS estimates of the change in growth trends after ARV scale-up
are biased downwards.
Column 5) shows results obtained using a similar approach to that of column 4), but including countryspecific linear trends instead of including the term which captures differential linear pre-intervention time
trends associated with HIV-prevalence. These country-specific linear trends capture any linear trend in
growth rates over time. Including country-specific linear trends makes very little change to either the
coefficient or standard errors, which is to be expected given that the instrument is only correlated with
country-specific linear trends through the linear relationship with HIV prevalence in 2001. These results
assuage any remaining concern that the results could be driven by differences in mean trends in growth rates
between countries with high and low HIV prevalence. However, since including country-specific linear trends
makes very little difference to the results, I proceed with the specification that controls for differential linear
trends as a function of HIV prevalence, as the more parsimonious specification.
I present results on life expectancy to demonstrate that the scale-up of ARV therapy has a direct effect
on health. The main concern with these results is that they may partially recover a modelled effect, given
that aggregate statistics provided by the United Nations Population Division are adjusted to reflect the
effect of HIV/AIDS. This concern is partly, but not completely, assuaged by the fact that the IV results are
stronger than the OLS results, because it is not clear why the effect should be biased downwards in the OLS
results if the effect recovered is purely modelled. A further question raised by these findings concerns the
direction of causality. Does the effect on life expectancy reflect primarily the direct causal effect of ARV
therapy on health? If there are improvements in health that are not the direct causal effect of expansion of
ARV therapy coverage, these might be the product of positive feedback effects, via which improved health
leads to increased income or reduced congestion in other health services which in turn further improves
health. However, improvements in health that are not related to ARV therapy might also raise concerns
about violations of the exclusion restriction as a result of omitted variables that influence both economic
21
growth and health improvements.
However, while it is hard to definitely rule out alternative channels, the evidence is largely consistent with
a direct impact of ARV therapy on health. Life expectancy reflects an aggregated index of all-age mortality.
An analysis of mortality by age and gender42 shows that the effects on life expectancy are consistent with
those that the health literature would associate with a causal effect of ARV therapy. Results are stronger
for adult female mortality than for adult male mortality43 and while effects on infant and under-5 mortality
are strong,44 there are no significant effects on neonatal mortality.45 However, the same principal concern
applies to all these estimates, in that all mortality estimates in the datasets used are adjusted to reflect
the simulated effect of HIV/AIDS.46 As such, the results of these comparisons only add limited further
information regarding the effect of ARV therapy on health.
5.3
Robustness checks
Table 3 shows that the main results are largely robust to changes in the specification or sample. Column
1) replicates the main results from Table 2, column 4) for comparison. Panels A shows results for life
expectancy, while panel B shows results for growth in per capita GDP.
Since all of the countries with high HIV prevalence47 are in sub-Saharan Africa, I include region-year
fixed effects in the main specification, to absorb differences between regions and to avoid conflating other
differences in overall trends between sub-Saharan Africa and other low and middle-income countries with
the causal impact of ARV therapy coverage expansion.48 Column 2) shows that when overall time trends
are modelled as year fixed effects, the coefficients are larger and somewhat more strongly significant. These
results confirm that including region-year fixed effects is the more conservative specification, at the cost of
absorbing variation of potential interest in comparing growth rates in sub-Saharan Africa with the rest of
the world.49 In column 3), I restrict the sample to sub-Saharan Africa alone. This considerably reduces
42 Results
shown in Appendix Table A2.
women are infected with HIV in low and middle-income countries than men, particularly in sub-Saharan Africa, and
pregnant women were one of the first groups targeted by ARV therapy expansion campaigns.
44 The literature on the life expectancy of children born with HIV makes for depressing reading. UNAIDS (2014) estimates
that without treatment, about one third of children born with HIV die in their first year of life, and half will not reach their
second birthday. Expanding ARV therapy coverage reduces rates of child mortality both directly, through treatment of children
born with HIV, and indirectly, by decreasing mother-to-child transmission and thereby reducing the number of children born
with HIV.
45 Neonatal mortality rates remain high in children born to HIV positive women, even those receiving ARV therapy (Ades
et al., 2013) and some ARV therapy combinations have been shown to increase neonatal mortality rates (Fowler et al., 2015),
although these are not the combinations most widely used.
46 This is because data from surveys on, for example, infant mortality, is biased because of elevated mortality rates among
young mothers.
47 e.g. HIV prevalence above 5% of the population
48 The regions included in the main sample are East Asia and Pacific (6 countries); Europe and Central Asia (9 countries);
Latin American and the Caribbean (17 countries); Middle East and North Africa (6 countries); South Asia (6 countries);
sub-Saharan Africa (39 countries).
49 The main sample consists of countries which report HIV prevalence in 2001. In Appendix Table A3, column 2), I show
that the results are similar if I extend the sample to all low and middle-income countries that report the outcome variables,
43 More
22
the sample size but yields very similar results, which is to be expected given that most of the variation of
interest is within sub-Saharan Africa.
There are a relatively small number of countries with high levels of HIV prevalence, and the distribution
of HIV prevalence is highly skewed.50 This raises the concern that influential observations drive the results.
To examine the influence of individual countries on the estimates, I run a series of leave-one-out regressions,
identical to the main specification but in each case dropping one country. The resulting estimates of the
effect of antiretroviral therapy on the main outcome variables (coefficients and p-values) are plotted in Figure
7. This figure confirms that, as we expect, the countries with highest HIV prevalence in 2001 — Botswana,
Swaziland and Zimbabwe — have the largest influence on the coefficients, but that this influence is largely
balanced. Excluding Zimbabwe tends to decrease the magnitude of the estimated impacts; excluding either
Botswana or Swaziland tends to increase the magnitude of the estimated impacts. For the main results, the
estimated impact nonetheless remains consistent when any individual country is dropped from the regression.
For growth in GDP per capita, the result is not statistically significant when some individual countries are
dropped from the regression, although this partly reflects the conservative approach to specification choice:
without region-year fixed effects, all of these regressions remain statistically significant, and with Newey West
standard errors, all but one are statistically significant. Given the influential nature of the observations from
Botswana, Swaziland and Zimbabwe — the countries with highest HIV prevalence — I also test whether the
results are robust to the exclusion of all three of these countries. Column 4) of Table 3 shows these results;
the effect on life expectancy is larger, while the effect on growth in GDP per capita is smaller, but both
remain statistically significant.51
A further concern is that the results could be driven by spatial correlation in trends within sub-Saharan
Africa, given that HIV prevalence is higher in East and Southern Africa, than in West Africa. Columns 5)
and 6) show that the magnitude of the coefficients are increased when I allow time trends to vary flexibly
with linear (column 5) and quadratic (column 6) functions of longitude and latitude.52 The effect on growth
in GDP per capita becomes marginally insignificant in column 5) (p-value 0.111) but is more significant in
column 6) (p-value 0.054). Taken together, these checks suggest that the results cannot be explained by
broad East-West and North-South variation in trends within sub-Saharan Africa.
I attribute the correlation between HIV prevalence in 2001 and decreasing growth rates in the preceding
assuming that where HIV prevalence is not reported it is zero, and therefore ARV coverage is also zero.
50 Appendix Table A4 lists all countries with HIV prevalence above 1% in 2001, in rank order.
51 In Appendix Table A3, columns 3 and 4), I show that the results are similarly consistent when I exclude observations with
extreme values of the dependent variable, by excluding the three countries associated with the largest absolute values of the
dependent variable in each regression (column 3) and winsorizing the top and bottom 1.5% of observations (column 4). The
results are marginally insignificant, but again this reflects conservative specification choice: the results are significant either
with year fixed effects instead of region-year fixed effects, or with Newey West standard errors with 5 lags.
52 Specifically, I include latitude-year and longitude-year fixed effects in column 5), and additional terms in latitude and
longitude squared and the interaction between latitude and longitude in column 6).
23
decade to a causal effect of high HIV prevalence on trends in growth rates in per capita GDP and life expectancy. However, I cannot rule out causality in the opposite direction: decreasing growth rates throughout
the 1990s may have caused HIV prevalence in 2001 to be higher than would otherwise have been the case; or
an unobserved omitted variable may have caused both increasing HIV prevalence and falling growth rates. If
this were the case, then the results might be biased upwards by regression to mean. My preferred approach
to addressing this concern is to test robustness to alternative instruments and, in particular, to alternative
measures of HIV prevalence.
I first demonstrate that the results are robust (and in fact considerably larger in magnitude) when I
use HIV prevalence in 1990, the earliest year for which HIV prevalence is estimated, in the construction of
predicted ARV therapy coverage. These results are shown in column 7) of Table 3. HIV prevalence in 1990
is more strongly associated with decreasing growth rates in the 1990s than HIV prevalence in 2001. This
assuages the concern that HIV prevalence in 2001 is associated with decreasing growth rates in the 1990s via
reversed causality i.e. falling growth rates causing higher HIV prevalence. Second, I construct an alternative
instrument using simulated values of HIV prevalence in 2001, which isolate the independent propagation
dynamic of the HIV epidemic using instrumental variables including distance from the Democratic Republic
of Congo, circumcision rates, religion and road networks.53 These data are only available for sub-Saharan
Africa. Comparing the results in columns 3) and 8) demonstrates that using the simulated values of HIV
prevalence makes only very small changes to the estimated coefficients, slightly decreasing the estimated
effect on life expectancy, and fractionally decreasing the estimated effect on growth in GDP per capita.
The precision of the estimates are also naturally somewhat reduced, given that the instrument is slightly
weaker, and the p-value on the estimated effect on growth in sub-Saharan Africa falls just outside statistical
significance (p-value 0.132).54 However, the small changes in the observed coefficients, particularly for growth
in GDP per capita, overall lend support to the claim that the estimated effects are not driven by regression
to mean.55
53 I thank Cahu and Fall for generously sharing this data with me. See Cahu and Fall (2011) for full details of simulation
procedure.
54 Cahu and Fall’s data provides estimates for simulated HIV prevalence in two countries for which UNAIDS data are
missing. Excluding these countries partly reduces the difference in results between those obtained using the observed values of
HIV prevalence in 2001 and those obtained using the simulated values.
55 An alternative approach that some have suggested is to create a matched control group. However, it is not clear whether this
approach makes sense. First, the meaning of a matched control group is unclear in a context where there is variation in treatment
intensity within the treated group. In this context, the association between negative pre-trends and HIV prevalence exists within
the group of countries with significant HIV prevalence, as well as between countries with and without high HIV prevalence.
Creating a ‘control’ group matched on pre-trends therefore attenuates, but cannot remove completely, the relationship between
HIV prevalence and negative trends in outcome variables prior to the expansion of access to ARV therapy. Second, pro-actively
selecting ‘control’ countries with lower growth rates in the 1990s for reasons unrelated to HIV prevalence may also have an
attenuating effect, through regression to mean in this ‘control’ group. This exercise cannot therefore distinguish between i) main
estimates that are biased upwards because of regression to mean in the ‘treated’ group and ii) adjusted estimates that are biased
downwards because of regression to mean in the matched ‘control’ group. For the sake of completeness I nonetheless carry out
this exercise using three different matching strategies based on growth trends prior to 2001. The results, shown in columns 5)
to 7) of Appendix Table A3, are very similar to the main analysis. Full details on the matching strategy employed are available
on request. This discussion also motivates a further test, which is to exclude the ‘control’ group completely and focus only on
24
5.4
Collapsed reduced form regression
Many empirical growth studies adopt a long difference approach, rather than a panel. The long difference
approach more naturally reflects our interest in growth over longer periods than a single year. Further, the
long difference approach reduces concerns about inference driven by serial correlation in growth between
years and possible measurement error in changes in GDP over shorter time scales. I therefore collapse the
data to a single cross section and demonstrate that the reduced form relationships remain compelling in
both visual scatterplots (Figure 8) and cross-sectional regressions (Table 4). The scatterplot presentation
further helps to allay concerns that the results are driven by unusual outliers, as the countries with high HIV
prevalence are even distributed around the best fit line. The reduced form estimates presented here do not
directly compare to the instrumental variable estimates.56 However, they are consistent in magnitude with
the preliminary reduced form analysis.
With respect to life expectancy, I collapse the data to the difference in 10-year growth rates in life
expectancy before and after ARV therapy coverage expansion begins in 2001 i.e. (log LEi2011 − log LEi2001 )−
(log LEi2001 − log LEi1991 ). Collapsing the data in this way removes differences across countries due to
different baseline levels of life expectancy and to different trends in life expectancy, thus approximating the
main reduced form specification. Plotting this measure of medium-run changes in life expectancy against
HIV prevalence in 2001 (Figure 8) shows that HIV prevalence in 2001 is positively associated with this
measure of increasing life expectancy. A simple cross-sectional regression confirms that this relationship is
statistically significant both within the main sample and between sub-Saharan African countries only (Table
4, columns 1 and 2). The coefficients imply that an additional 1% of the population with HIV in 2001 is
associated with a change in life expectancy that is 1.4 percentage points higher between 2001 and 2011 than
between 1991 and 2001,57 or 1.3 percentage points higher in sub-Saharan Africa alone.
With respect to growth in GDP per capita, I collapse the data to differences in differences in 5 year
growth rates i.e. ((yi2011 − yi2006 ) − (yi2006 − yi2001 )) − ((yi2001 − yi1996 ) − (yi1996 − yi1991 )). Collapsing the
data in this way removes differences between countries due to inherent differences in long-run average growth
rates, and due to different linear trends in growth rates. Again, this approximates the main reduced form
specification. Plotting this measure of changes in growth trends against HIV prevalence in 2001 (Figure 8)
shows that HIV prevalence is positively associated with this measure of increasing growth rates, compared to
previous trends. Cross-sectional regressions also confirm that the relationship is statistically significant both
variation within countries with HIV prevalence above 1% in 2001. This specification yields slightly larger estimates for GDP
per capita. Results are shown in column 8) of Appendix Table A3.
56 The instrument interacts HIV prevalence with global ARV therapy coverage, and ARV therapy is selectively provided to
the subset of the HIV positive who have the highest viral loads.
57 Decomposing this effect, the change in life expectancy was 1.1 percentage points lower for each additional 1% of the
population with HIV between 1991 and 2001, and 0.3 higher between 2001 and 2011.
25
for all low and middle-income countries that report HIV prevalence in 2001, and for sub-Saharan African
countries only (Table 4, columns 3 and 4). These coefficients imply that growth in per capita GDP increased
by 1.2 percentage points more in the decade following the onset of expansion of ARV therapy coverage than
in the preceding decade. Once again, the effect sizes are not directly comparable to the panel estimates, but
are consistent with the preliminary analysis.
5.5
Alternative explanations
The main results, presented in Table 2, could be biased by omitted variables that influence the outcomes,
if these omitted variables were correlated in space with HIV prevalence, and in time with the global expansion
of ARV therapy coverage. In this section, I discuss four plausible alternative explanations, but show that
none can explain the main results. I present the results for growth in GDP per capita in Table 5.58
First, Chinese overseas influence follow a similar pattern in time to the expansion of ARV therapy
coverage. The stock of Chinese FDI investments in sub-Saharan Africa increased by an order of magnitude
between 2005 and 2012, and official records show similar patterns in both exports and imports between China
and Africa (Information Office of the State Council, 2013; Chinese Ministry of Commerce, 2009; Heritage
Foundation, 2004). Chinese aid flows are also likely to be correlated in time and space with other measures
of Chinese influence. If factors correlated with HIV prevalence were also correlated with factors that predict
Chinese investment, trade or aid; and Chinese investment, trade or aid also influence the outcome variables,
then the main results would be biased.
One plausible channel is via the presence of mineral resources. Chinese overseas investment is widely
believed to be correlated with the local availability of commodities such as minerals, and mines attract large
populations of transient workers which have historically been associated with high rates of HIV prevalence
(see e.g. Corno and de Walque, 2012). In Panel A of Table 5, I show how the main results (restated for
reference in columns 1 and 4) change when I alter the specification to allow countries with high dependence
on mineral rents — proxied by the fraction of GDP coming from mineral rents in 2001 — to have different
time trends before and after 2001. For this and the following analyses, I test robustness to both differential
linear trends before and after 2001 (here, the fraction of GDP from mineral rents interacted with time, a post
2001 dummy, and their interaction term) and differential flexible time trends (here, fraction of GDP from
mineral rents interacted with year dummies). I report results both for the main sample and for sub-Saharan
Africa only.
The estimated effect of ARV therapy on growth in GDP per capita increases when I allow countries
with high mineral dependency to have different time trends, suggesting that Chinese investment in mineral
58 The
corresponding results for life expectancy are included in Appendix Table A5.
26
rich countries cannot explain the results on antiretroviral therapy (Panel A, columns 2, 3, 5 and 6, Table
5). Countries with high mineral dependency do have systematically different trends in growth in GDP per
capita, with growth rates falling relative to other countries before 2001. These negative trends do recover
somewhat after 2001, but the recovery is an order of magnitude smaller than the negative pre-trend, and is
only significantly different from zero for the main sample (Panel A, columns 2 and 5, Table 5).
Second, Chinese investment may also be correlated with the presence of petroleum resources. If there
is any correlation between oil resources and HIV prevalence — perhaps through lower local pump prices,
increased mobility and hence increased transmission of HIV — this could also bias the results. In Panel B
of Table 5, I show how the main results change when I alter the specification to allow countries with high
dependence on petroleum rents — proxied by the fraction of their GDP coming from this source — to have
different time trends before and after 2001.
Countries with high petroleum dependence do experience relatively higher growth in GDP per capita
after 2001 (Panel B, columns 2 and 5, Table 5) although this is a change in levels, rather than the trend
break observed in Figure 1 and measured in Table 1.59 Including linear or flexible controls for petroleum
dependence results in a slightly reduced coefficient on ARV therapy, reducing from 1.36 to 1.27 for the
main sample and 1.34 to 1.26 for sub-Saharan Africa. The coefficient on ARV therapy becomes marginally
insignificant (p-values between 0.109 and 0.112). However, the small change in the point estimate, the
relative lack of explanatory power, and the inconsistency with the observed trend break together suggest
that this mechanism is unlikely to explain the main results.
Third, the increase in trade with China, and to some extent the rest of the world, may have differentially
benefited countries with higher exports as a fraction of GDP. If there are higher levels of mobility in these
countries, this may be associated with a greater spread of HIV prevalence.60 In Panel C of Table 5, I show
how the main results change when I alter the specification to allow countries with large export shares of
GDP to have different time trends before and after 2001. Countries with high dependence on exports do
not experience any significant relative trends, either before 2001 or after 2001 (Panel C, columns 2 and 5,
Table 5). Further, while including these additional time trends reduces the estimated effect of ARV therapy
on growth in per capita GDP for the main sample (Panel C, column 2, Table 5), their inclusion increases
the estimated effect for sub-Saharan Africa alone (Panel C, column 5, Table 5). Including these superfluous
controls also decreases precision and renders the coefficient in this specification marginally insignificant (pvalues between 0.103 and 0.156). However, the lack of explanatory power and the increase in the point
estimates in sub-Saharan Africa suggest that this mechanism is also unlikely to explain the main results.
59 Data on petroleum exports as share of GDP was missing for 27 countries in the main sample in 2001, but cross-referencing
with other data suggest that it is reasonable to assume that missing data implies zero or negligible petroleum exports.
60 See e.g. Oster, 2012.
27
Finally, global investment in fighting malaria also increased over a relatively similar time period, after
the founding of the Global Fund for Aids, Tuberculosis and Malaria. Malaria prevalence may be correlated
with HIV prevalence through climate and ecology, given that malaria thrives in warm, wet climates, and HIV
originated in primate populations with forest habitats. If this is the case, this might lead me to overestimate
the impact of ARV therapy alone, confounding it with other interventions in malaria control which took
place at the same time.
Since the extent of a country’s malaria control program is likely endogenous to other health and economic
factors, I focus instead on geographical variation in climate and ecology that influences malaria transmission.
I use the malaria ecology index developed in Kiszewski et al. (2004) aggregated to a population-weighted
country dataset by McCord (n.d.).61 To facilitate interpretation of the coefficient, I normalize the variable
by its maximum across the main sample. The raw coefficient on malaria ecology can then be interpreted as
the change in the outcome variable resulting from going from zero malaria ecology to its highest value in the
sample.
Malaria ecology has little predictive power for trends in growth in per-capita GDP (Panel D, columns 2
and 5, Table 5). Including the malaria ecology interaction terms also increases the coefficient on antiretroviral
therapy in both the main sample (Panel D, columns 2 and 3, Table 5), and sub-Saharan Africa alone (Panel D,
columns 5 and 6, Table 5). The coefficient becomes marginally statistically insignificant in the specification
that allows flexible trends (p-value 0.102) within sub-Saharan Africa, as the standard errors also increase.
However, taken together, these results suggest that confounding investment in fighting malaria is unlikely to
explain the main results.
One important note is that I am unable to replicate this analysis for tuberculosis (TB) infection rates,
as the TB epidemic is so closely correlated with the HIV epidemic. The WHO estimates that one third of
people living with HIV are also infected with TB, and treatment programs for TB are also delivered alongside
ARV treatment programs. Where improvements to treatment for TB are delivered at the same time and in
the same places as ARV therapy coverage expansion, my results will capture the combined effects of the full
package of treatment for both HIV and TB. My estimates also capture the effects of any other additional
investments in healthcare that are provided alongside ARV therapy, and of improvements to the rest of the
healthcare system that result from reducing congestion. As a result, I will compare the estimated benefits
of ARV therapy coverage expansion to conservative estimates of HIV spending that capture all types of
spending, not simply spending on delivering ARV therapy.
In conclusion, although there is suggestive evidence that some of these alternative mechanisms have some
61 The malaria ecology index is highly skewed and has many zeros. To reduce skewness, I transform the variable by taking
the arcsinh. This transformation increases the explanatory power of the malaria ecology variable as a control; without this
transformation, the malaria ecology index has little predictive power and less influence on the coefficients of interest.
28
explanatory power for changes in growth rates in per capita GDP over the time period of the study, there
is no compelling evidence to suggest that the measured effect of ARV therapy can be attributed to any of
these alternative mechanisms.
5.6
Evidence on potential long-run results: population growth
The main results provide evidence that expanding access to ARV therapy increased growth in per capita
GDP in the short run i.e. in the years following expansion of access. I cannot provide direct evidence on the
medium or long-run effects on economic growth, for the simple reason that they have yet to happen. However,
the framework outlined in Section 2 motivates an analysis of the effect of ARV therapy on population growth.
I can then compare the magnitude of these effects to magnitude of the short-run effects on growth in GDP
per capita, to provide a first order indication on whether these short-run increases in GDP per capita will
translate into persistent increases in the future.
Figure 9 replicates the graphical preliminary analysis I conducted for the main results, for both the
main sample and sub-Saharan Africa alone. I show results for both population growth and its constituent
components: the crude birth rate, the crude death rate, and the net migration rate. The non-parametric
evidence shown in Figure 9 shows that broadly, population growth is relatively declining in countries with
high HIV prevalence relative to other low and middle income countries, until the expansion of ARV therapy
coverage commences in 2001. The reversal in trends in population growth rates is primarily driven by changes
in the death rate and net migration rate, with birth rates experiencing relatively small changes in trends.
Several caveats on the quality of the data on birth rates, death rates and migration rates are worth noting.
Estimates of the values of these components are aggregated by the World Bank but once again, where registry
data is not available, the estimates may be adjusted to model the effect of HIV/AIDS. While HIV/AIDS
may suppress individual fertility, aggregate birth rates may be mechanically increased by HIV/AIDS because
deaths among middle-aged women increase the proportion of fertile women in the population. Net migration
is only available at 5 year intervals, and in some cases where more reliable data is not available, net migration
is calculated taking the difference between the population growth rate and the natural increase rate (birth
rate minus death rate).
The overall patterns observed in Figure 9 are confirmed by analysis shown in Table 6. Table 6 shows the
estimated effect of ARV therapy on population growth rates, and its components, for both the main sample
(Panel A) and sub-Saharan Africa alone (Panel B). The results shown in column 1) confirm that population
growth rises: the coefficients imply that a 1 percentage point increase in ARV therapy coverage leads to
a 0.39 percentage point increase in population growth rates. A leave-one-out analysis that replicates the
29
exercise shown in Figure 7 for these results suggests the effects on population growth, birth rate and death
rate are robust to the exclusion of individual countries.62
Column 2) of Table 6 shows that there is a small positive effect of ARV therapy on the crude birth rate,
equivalent to an increase in the birth rate of 0.06 percentage points for every additional 1% of the population
receiving ARV therapy. This effect is precisely estimated and robust to excluding individual countries within
sub-Saharan Africa, but accounts for only 15% of the overall effect on the population growth rate. The
effect of HIV/AIDS on fertility rates was previously the subject of considerable debate and prior empirical
evidence is mixed.63 My estimates suggest that HIV most likely does depress aggregate fertility rates, but
that the resulting effect on population growth is small compared to the effect on mortality.
Column 3) of Table 6 shows that the crude death rate falls robustly in response to increasing coverage
of ARV therapy coverage. Increasing access to ARV therapy coverage to an additional 1% of the population
reduces the crude death rate by 0.14 percentage points. This effect is robust and accounts for around a third
of the effect on population growth.
Finally, the remaining component of population growth is net migration. The results on net migration are
shown in Column 4 of Table 6. The estimated effect on net migration is positive and large, but imprecisely
estimated and insignificant. Further, leave-one-out analysis shows that Zimbabwe is a particularly extreme
outlier in driving this result.
The potential long-run effects of increasing access to ARV therapy may also be larger if there are increases
in human capital investment that take several years to generate a return. Primary school enrollment seems
a natural first place to look for these effects. I find no evidence of increase in primary school enrollment
associated with increasing levels of ARV therapy. In fact, within sub-Saharan Africa, trends in gross and net
school enrollment are largely stable in high HIV prevalence countries with respect to other low and middleincome countries.64 These results suggest that increases in investment in primary schooling are small, not
well measured, or offset by other changes in the composition of the population.
6
Conclusions and Discussion
In this section, I first carry out an accounting exercise which uses the estimates of the impact of ARV
therapy coverage expansion on growth rates in per capita GDP to provide a first order estimate of the value
of ARV therapy coverage expansion in sub-Saharan Africa. I then use the framework described in Section 2
to compare the magnitudes of the effects on GDP per capita and population growth to provide a preliminary
62 See
Appendix Figure A1.
e.g. Lewis, Ronsmans, Ezehc, and Gregson (2004), Oster (2005), Fortson (2009).
64 See Appendix Figure A2. Further details available on request.
63 See
30
indication of the likely sign of the long-run effect on GDP per capita. Finally, I summarize the main results,
and conclude.
6.1
Estimates of total value of antiretroviral therapy coverage expansion
Policy-makers continue to debate whether and how international organizations and national governments
should continue to fund ARV therapy coverage and expansion. For this reason, it seems essential to carry
out an overall valuation exercise of the benefit of expanding ARV therapy coverage to date. However, the
estimates I will discuss in this section should be taken as indicative of magnitude, rather than precisely
estimated, and should be interpreted carefully. I will discuss several specific limitations of this accounting
exercise below. Nonetheless, the accounting exercise that I will discuss suggests that expanded coverage of
ARV therapy led to an additional US$33 bn in GDP in sub-Saharan Africa in 2012.
I obtain this estimate as follows, focusing on sub-Saharan Africa alone, as the primary region of interest.65
The main estimates for sub-Saharan Africa (presented in column 3, panel C of Table 2) suggest that a 1
percentage point increase in coverage of a country’s population with antiretroviral drugs increases the growth
rate of per capita GDP by 1.3 percentage points.66 This compares to an observed annual mean real growth
rate in per-capita GDP across countries in sub-Saharan Africa of 1.3% between 1990 and 2012. This mean
growth rate obscures considerable heterogeneity in growth rates across time and space. Mean growth rates
averaged only 0.3% between 1990 and 2001, and rose to 2.2% between 2001 and 2012.67
I estimate the value of the contribution of ARV therapy coverage expansion, and show the results in Table
7, for each country in sub-Saharan Africa. The valuation exercise suggests that expanded ARV therapy
coverage added approximately 71 US$68 in growth in GDP per capita to countries in sub-Saharan Africa in
2012, equivalent to each person receiving ARV therapy in sub-Saharan Africa adding on average US$2820
to a country’s economy. In some high HIV prevalence countries, the estimated component of increase in
GDP per capita due to the expansion of ARV therapy coverage is larger than observed growth. For these
countries, the counterfactual implied by the model is that GDP per capita would have been decreasing by
2012, as a result of an unchecked HIV/AIDS epidemic. Recall, however, that a key limitation of the model
is that it projects linear trends in growth rates experienced in the 1990s after the expansion of ARV therapy
coverage.
65 Valuation
estimates outside of sub-Saharan Africa are sensitive to the inclusion of countries with high per capita GDP.
natural question to ask is whether the estimate from the main regression is the right estimate to apply to a valuation
exercise like this one. While the main estimate reflects the average effect on a country’s growth rate, across countries, we might
instead be interested in the average effect per $ of GDP exposed to treatment when we turn to overall valuation. In this case,
estimates from a regression weighted approximately by a measure of the fraction of GDP exposed to treatment — here, the
interaction between lagged GDP and ARV therapy coverage in 2012 — are very similar to the main results.
67 Recall that the analysis implies that growth rates in countries with high HIV prevalence appear to be depressed by the
presence of HIV/AIDS itself in the first half of the study period.
68 All valuations in 2012 US$.
66 A
31
In total, this additional growth sums to US$33 bn in GDP in sub-Saharan Africa in 2012. I compile
this overall estimate by multiplying the coefficient on growth in GDP per capita by the observed level of
ARV therapy coverage and lagged GDP per capita, measured in 2012 US$, to give an approximation of
the increase in GDP per capita due to ARV.69 I then multiply this growth rate by the contemporaneous
population70 to calculate the additional GDP added to a country’s economy as a result.
The total impact is largely driven by the estimated impact in South Africa, which accounts for approximately two-thirds of the total impact. South Africa’s importance is a combined product of its relatively
high per capita GDP and its large HIV positive population, which makes up around 1/4 of the total HIV
positive population in sub-Saharan Africa. Botswana, Kenya, Namibia, Nigeria and Zambia also contribute
significant amounts to the total valuation.
I estimate the mean effect of ARV therapy coverage expansion to an additional 1% of the population
across the years following the change in access. Note that the cumulative effect of increasing coverage to
an additional 1% of the population would be the sum of these effects over the years for which effects are
greater than zero. However, since the cost of providing ARV therapy is a yearly expense in the short- and
medium-run if ARV therapy coverage is to be maintained,71 it seems reasonable to consider the mean annual
effect as the appropriate policy variable.
For comparison, UNAIDS estimates suggest that around US$9.4 billion was spent on HIV programs in
sub-Saharan Africa, including ARV therapy provision to around 7.5 million people.72 This implies that a
conservative comparable figure for the costs of providing ARV therapy in sub-Saharan Africa is US$1250
per person per year. This figure comprises not only the costs of providing ARV therapy, but also all
spending associated with HIV/AIDS. However, the figure is appropriate for comparison with the estimated
benefits, since I cannot empirically separate the effects of other HIV-related programs provided alongside
ARV therapy.73 The total estimated benefits (US$33 billion) therefore exceed the maximum total costs
(US$9.4 billion) by a factor of around 3.5.
An alternative measure of the importance of ARV therapy coverage expansion to growth in sub-Saharan
Africa would be to evaluate to what extent ARV therapy coverage expansion helps to explain sub-Saharan
69 This approach has the advantage of simplicity and transparency, but is somewhat ad hoc, as the lagged level of GDP per
capita is endogenous to prior provision of ARV therapy. However, I obtain comparable results with a more involved approach
estimating growth in GDP in each country and year, and aggregating the cumulative effect via a simulation of GDP trajectories.
70 A similar caveat applies to the endogeneity of population.
71 If ARV therapy coverage is maintained at sufficiently high levels, the HIV/AIDS epidemic will in the long run be reversed,
and possibly halted, through interrupted transmission rates.
72 UNAIDS (2013) estimated that in 2012 total spending on HIV was US$18.9 billion across all low and middle countries.
Applying spending shares from 2011, just below half of this was spent in sub-Saharan Africa (UNAIDS, 2012), or around US$9.4
billion. 9.7 million people were receiving ARV therapy in 2012, of which approximately 7.5 million were in sub-Saharan Africa.
73 The cost of providing ARV therapy alone is significantly lower, with facility-level costs averaging US$200 per patient-year
in 2012 in Ethiopia, Malawi, Rwanda and Zimbabwe; and US$682 in South Africa (Tagar et al., 2012). Prices of antiretroviral
drugs also continue to fall; to US$115 per person per year for first line ARV therapy and US$330 per person per year for second
line ARV therapy by 2013 (WHO, 2014). Delivery costs are likely falling over time as well, as the fixed costs of establishing
delivery systems are paid.
32
Africa’s ‘growth miracle’, observed since around the year 2000. Undoubtedly, other factors have also been at
play in sub-Saharan Africa’s recovery. Mean real per-capita income in sub-Saharan Africa as a whole reached
its nadir in the mid-1990s and Rodrick (2014) flags declining civil wars and trends towards democracy and
electoral competition in sub-Saharan Africa since the early 1990s. However, my estimates imply that the
expansion of ARV therapy coverage represents an important explanation for this reversal of prior trends.
Table 8 compares the mean estimated component of growth in GDP per capita attributable to the expansion
of coverage of ARV therapy to mean observed growth in the same year. Overall, the magnitude of the effect
on growth in GDP per capita is comparable to approximately one third of the value of observed growth
between 2004 and 2012, recalling that the counterfactual scenario in countries with high HIV prevalence is
one of decreasing GDP per capita as a result of an unchecked HIV epidemic. These approximate comparisons
are somewhat approximate support the qualitative conclusion that the expansion of ARV therapy coverage
helps to explain the sub-Saharan Africa growth miracle.
The above valuation exercise come with several further important caveats. I estimate the average value of
covering an additional 1% of the population with ARV therapy over time. If people receiving ARV therapy
are less sick as time goes on — for example, because the threshold for treatment with ARV therapy is
decreasing — or if there are otherwise diminishing marginal returns to covering larger numbers of people
with ARV therapy, then my estimates will tend to overestimate the benefits in later years.74 My estimates
also reflect the average of contemporaneous and lagged effects of increasing the fraction of the population
receiving ARV therapy on growth in GDP per capita. If lagged effects are decreasing with time, then this
will similarly cause me to overestimate the effects of ARV therapy in later years. Bearing these caveats in
mind, the evidence nonetheless largely suggests that the benefits of ARV therapy, as monetized in GDP per
capita, outweigh the costs; and that the expansion of ARV therapy in sub-Saharan Africa is an important
cause of the sub-Saharan Africa ‘growth miracle’.
6.2
Prospects for long-term growth
The theoretical framework outlined in Section 2 motivates a simple test of the prospects for sustained
growth in GDP per capita. In particular, Equation 6 provides a way to compare the magnitude of the
estimated effects on short-run growth in GDP per capita and population growth. Once again, I focus on
sub-Saharan Africa. These estimates suggest that, conditional on two further conservative assumptions, the
positive effects on long-run growth will at least outweigh the negative effects.
The motivating framework implies that the sign of the long-run effect of expansion in access to ARV
therapy should be given by (γ + η) −
74 But
1−α
α λ,
recalling that α is labour’s factor share, α (γ + η) is the short-
correspondingly, to underestimate them in earlier years.
33
run effect of increasing access to ARV therapy on per capita income, via the effects on human capital and
TFP, and λ is the semi-elasticity of access to ARV therapy on the population growth rate.
The estimated short-run effect on the population growth rate is 0.39 percentage points for each additional
1% of the population covered by ARV therapy. This is a linear effect, which converted into a semi-elasticity
suggests a value of λ of approximately 14.5.75 Making the conservative assumption that this effect on
population growth rates persists indefinitely, the motivating framework implies that the effect of this increase
in population growth (λ) on capital accumulation should translate into a long-run effect on per-capita income
with approximate magnitude
1−α
α λ.
Taking α, labour’s factor share, as approximately 0.7 and applying
this transformation suggests that long-run per capita income should be 6.2 percentage points lower due to
increased population growth as a result of a 1 percentage point increase in ARV therapy coverage.
The corresponding short-run effect on growth rates in the first few years after expansion of ARV therapy
coverage in GDP per capita is 1.3 percentage points, an average of contemporaneous and lagged effects.
The total cumulative short-run effect on GDP per capita should be the summed effects over the years for
which lagged effects are positive. Assuming conservatively that these effects are only positive over 4 years76 ,
a cumulative short-run effect of 5.2 percentage points seems plausible. The long-run effect on income per
capita through increases in human capital and TFP is expected to be larger than the corresponding short-run
effect, as while capital is fixed in the short run, capital adjusts and accumulates in the long run, yielding
further increases in per capita income. Again taking α, labour’s factor share, as approximately 0.7, this
translates into predicted long-term gains due to the positive effects on TFP, labour force participation and
human capital of 7.4 percentage points.77
In summary, making the conservative assumptions that the effect on the population growth rate is persistent, but that the short-run effects on growth rates in GDP per capita only persist for 4 years, the positive
effects on long-run growth at least outweigh the negative effects. These order-of-magnitude comparisons
tentatively suggest that the prospects for sustained long-run growth are reasonably strong.
6.3
Summary and discussion
This paper shows that antiretroviral (ARV) therapy coverage expansion to HIV positive patients in lowand middle-income countries has resulted in substantial increases in life expectancy and increased growth
75 In the model, I defined λ as the elasticity of the population growth rate with respect to the parameter
(1 − φHIV it + φARV it ), roughly equivalent to the semi-elasticity via the linear approximation of ln (1 − φHIV it + φARV it )
and therefore to the semi-elasticity of population growth to φARV it . To convert the linear estimated effect to a semi-elasticity,
note that 0.39 percentage points reflects a 0.145 log difference at mean population growth 2.5% (in countries reporting HIV
prevalence in 1990 in sub-Saharan Africa, between 1990 and 2012), suggesting a semi-elasticity of 14.5.
76 The estimated effect is an average of lagged and contemporaneous effects of increasing ARV therapy coverage over an
average of 6 years.
77 The motivating framework implies that the estimated effect measures α (γ + η), the short-run effect of ARV therapy on
growth through human capital and TFP, while the long-run effect should be given by γ + η.
34
rates in per capita GDP. ARV therapy coverage expansion of an additional 1% of a country’s population
leads to a 5.3% increase in life expectancy, and to a short-run 1.4 percentage point increase in the growth
rate of GDP per capita. A first order estimate of the value of providing antiretroviral therapy in low and
middle income countries suggests that ARV therapy added US$33 billion to sub-Saharan Africa’s economy
in 2012, in comparison to a conservative cost estimate of US$9.4 billion. The corresponding effect on the
population growth rate is relatively modest — 0.39 percentage points — and a preliminary test, based on the
neoclassical (Solow) growth model, suggests that any associated Malthusian effects are unlikely to outweigh
the positive effect on growth in GDP per capita in the long run.
While the estimated effects are substantial, mean per-capita GDP in the low and middle-income countries
between 2004 and 2012 was on average only $2410 (current US$) in the main sample, and $1,618 (constant
2005 US$) in sub-Saharan Africa. These figures imply that additional growth of per-capita GDP of 1.4
percentage points translates into an additional US$33 in annual per-capita income in the main sample, or
US$22 in sub-Saharan Africa (both in current US$). Given mean life expectancy of 63 years in the main
sample, and 55 years in sub-Saharan Africa, the increases in life expectancy translate to increases of 3.3 and
2.9 years, respectively. While these are substantial increases, the magnitudes of the effects are nonetheless
plausible.
The effect sizes are large, however, in comparison to the microeconomic literature which estimates the
effect of ARV therapy on individuals and households, and are too large to be explained simply by the return
to labour market participation of HIV patients and their carers. The pattern of small estimated relative
effects and large estimated aggregate effects suggests mechanisms that affect the HIV negative, as well as the
HIV positive. Plausible channels include: direct program impacts on local employment and procurement;
reductions in absenteeism due to funeral attendance; and changes in investment decisions resulting from
changes in beliefs about the risk and consequences of infection with HIV/AIDS.78 Reports from the field
also suggest that ARV therapy programs have reduced congestion in other public health services, leading to
possible indirect improvements in health and labour market participation for patients suffering from other
conditions (Médecins Sans Frontières, 2015). In some countries, the expansion of ARV therapy coverage,
partly funded with international support, may have permitted the reallocation of government spending
towards other productive investments. Unfortunately, data limitations largely prevent me from exploring
further the mechanisms via which improved health translates into increased economic growth.79
78 A recent study by Baranov et al. (2012) finds far-reaching impacts of expanded ARV therapy coverage on productivity,
even among the HIV negative, supporting the case for mechanisms beyond labour market participation.
79 For example, the only globally available data on labour participation rates is modelled by the International Labor Organisation (ILO), based where possible on nationally representative surveys and using census data or national estimates where this
data is unavailable. Appendix Figure A3 shows non-parametric trends in labour force participation for high HIV prevalence
countries relative to other low- and middle-income countries from the main sample or in sub-Saharan Africa alone. These
estimates show decreasing labour force participation that is correlated with higher levels of ARV therapy coverage. However, if
35
A few limitations are worth re-stating. First, effects are measured relative to previous linear trends in
growth rates, and the counterfactual — no ARV therapy — scenario effectively extrapolates these linear
trends by just over ten years. If negative trends would have weakened, my estimates will tend to overstate
the effects of ARV therapy coverage expansion.80 The rate of increase in HIV prevalence rates in subSaharan Africa had in fact started to decline by the end of the 1990s, although this was largely a product
of deaths due among the HIV positive, which continued to increase at a constant rate through the end of
the 1990s. However, there is no particular evidence that negative trends in health or growth associated with
HIV prevalence weakened throughout the 1990s.81
Second, the results results on life expectancy come with the important caveat that life expectancy estimates rely on modelled data as well as statistical evidence. As a result, it is possible that the results are
at least partially recovering a modelled impact of the HIV/AIDS epidemic. However, the findings on life
expectancy are, at least, consistent with other evidence. In particular, the results are similar in magnitude
to those of Bor et al. (2013), who studied the impact of antiretroviral therapy on life expectancy in a high
HIV prevalence region of South Africa. Their findings correspond to a 3.0% increase in life expectancy associated with an additional 1% of the population being treated with ARV therapies.82 This estimate is within
the confidence interval of my estimate on life expectancy. The authors note, however, that their estimate
is biased downwards by not considering the reversal of the clear prior downward trend in life expectancy,
whereas my results take this into consideration; an approximate adjustment to their results to reflect this
would imply a figure of 4.2%, still closer to my estimated treatment effect.83
Third, I am unable to separate out the effect of ARV therapy alone from the overall effect of the full
package of interventions delivered alongside ARV therapy, including HIV testing programs and TB treatment
for patients with TB-HIV coinfection. However, the variation in time that I exploit stems directly from the
fall in prices and resultant expansion of ARV therapy in the years following 2001, suggesting that the results
would not have been produced without the component of ARV therapy.84
The impact of treating an additional 1% of the population with ARV therapy can be interpreted as a
it is difficult to separate out a mechanical effect of reducing the death rate and increasing the surviving population. The labour
force participation rate could also be reduced by increasing participation in secondary and tertiary schooling.
80 Although the reverse is also true: if they would have worsened, my estimates will tend to understate the effects of ARV
therapy coverage expansion.
81 In regressions of the main outcome variables on linear and quadratic trends interacted with HIV prevalence in 2001, country
and region-year fixed effects, for the years prior to 2001, the quadratic term is negative and significant when life expectancy is
the outcome variable and is positive, but very small in magnitude and insignificant when growth rates in per capita GDP is the
outcome variable.
82 The authors find a 21% increase in life expectancy associated with coverage of 7% of the population by antiretroviral
therapy
83 But note that if these estimates in turn were exploited in simulating the modelled data, I would naturally recover similar
estimates.
84 The cost of TB treatment is probably substantially less than the cost of ARV therapy, so interpreting the results as the
combined impact of TB treatment and ARV therapy do not substantially alter the main conclusions regarding the benefit to
cost ratio, especially since my cost estimates are derived from total spending on HIV rather than spending on ARV therapy
alone.
36
lower-bound estimate of the impact of an additional 1% of the population suffering from full-blown AIDS.85
My results are therefore inconsistent with those of Ahuja et al. (2009), who found no impact of HIV/AIDS
on economic growth, using male circumcision rates as an instrument for HIV infection rates. Their results
may differ from mine if the exclusion restriction in their study is not valid i.e. if male circumcision rates are
correlated with economic growth through alternative channels such as institutions. My results are consistent
in sign with those of Cahu and Fall (2011) and comparable in magnitude; their results suggest that a 1
percentage point increase in HIV prevalence induces a 12% long-run decrease in GDP per working age adult.
My results are also largely comparable to a projection of the benefit/cost ratio of scaling up ARV therapy
carried out in 2011, which estimated a benefit/cost ratio of approximately 2.4:1 (Resch et al., 2011), despite
the authors only considering three specific channels.86
My estimates necessarily focus on the short-run impacts of increasing ARV therapy coverage, and yield
the average of both contemporary and lagged changes in the level of ARV therapy coverage. A simple test
compares adjusted magnitudes of the short-run effects on growth rates in per capita GDP and on population
growth rates and indicates that the long-run effects on per-capita income are likely to also be positive. While
this conclusion differs from that of Acemoglu and Johnson (2007) regarding the international epidemiological
transition, the most likely explanation is the nature of the disease examined. HIV/AIDS strongly affects
morbidity and mortality among working adults, and indeed was the single largest cause of death among
adults age 15 to 69 in Africa at the turn of the Millennium (WHO, 2004). In contrast, the diseases covered
by the international epidemiological transition primarily, if not exclusively, affected infant mortality. As
a result, basic intuition suggests that the channels of labour force participation, TFP and human capital
accumulation are likely to be relatively more important, and Malthusian effects following from population
growth relatively less important. My conclusions also contrasts with those of Young (2005), who modelled
the effect of the HIV/AIDS epidemic on fertility rates and human capital investment, and concluded that the
long-run effect of the epidemic on per capita income would be positive, driven by a large predicted negative
fertility response to HIV/AIDS. The small positive impact of ARV therapy on fertility seen here is more
consistent with conclusions drawn elsewhere by Kalemli-Ozcan (2011) and Juhn, Kalemli-Ozcan, and Turan
(2013).
A basic accounting exercise implies that the expansion of coverage of ARV therapy, and its apparent
reversal of the negative impacts of HIV/AIDS, makes a substantial contribution to explaining the subSaharan ‘growth miracle’ of the early 2000s. Clearly, I do not assert that ARV therapy coverage expansion is
85 It is an estimate of the impact of AIDS, rather than HIV, because current policy is to treat HIV patients only after functioning of their immune system falls below a specified threshold, or if they fall within other distinct groups. As of 2013: children
below 5 years, pregnant women, people coinfected with TB and HIV, people with chronic severe liver disease, serodiscordant
couples.
86 Labour force participation, avoided costs of orphan care, and avoided end of life health expenses.
37
the only mechanism at play. Nonetheless, the overall magnitude of the effects, combined with the distinctive
pattern of changes in trends between countries with high and low HIV prevalence, together suggest that
the expansion of ARV therapy coverage is an important factor behind the reversal in trends in sub-Saharan
Africa. In other words, that the Lazarus drug helps to explain the ‘growth miracle’.
In summary, this paper provides the first robust empirical evidence of positive short-run effects of health
improvements on aggregate growth in per-capita income. The paper also provides the first robust evaluation
of the global effects on per capita income resulting from the expansion of coverage of ARV therapy to low
and middle-income countries. The results point to the reversal of the negative effects of the HIV/AIDS
epidemic as a previously neglected factor in sub-Saharan Africa’s ‘growth miracle’ in the first decade of
the new millennium. My findings also confirm that over and above the direct benefits of improving health,
saving lives and avoiding orphaning children, ARV therapy has substantial and important effects on economic
growth, that outweigh the costs by a factor of approximately 3.5 to 1. I conclude that overall the results
provide compelling evidence for the continued and extended support of ARV therapy programs in low and
middle-income countries.
38
References
Acemoglu, D. & Johnson, S. (2007). Disease and development: the effect of life expectancy on economic
growth. Journal of Political Economy, 115 (6), 925–985.
Acemoglu, D. & Johnson, S. (2014, April). Disease and development: A reply to Bloom, Canning, and Fink.
NBER Working Paper No. 20064.
Ades, V., Mwesigwa, J., Natureeba, P., Clark, T. D., Plenty, A., Charlebois, E., . . . Ruel, T. D. (2013).
Neonatal mortality in HIV-exposed infants born to women receiving combination antiretroviral therapy
in rural Uganda. Journal of Tropical Pediatrics, 59 (6).
Ahuja, A., Wendell, B., & Werker, E. (2009). Male circumcision and AIDS: The macroeconomic impact of
a health crisis. Harvard Business School Working Paper 07-025.
Baranov, V., Bennett, D., & Kohler, H.-P. (2012). The indirect impact of antiretroviral therapy. PSC Working
Paper Series, PSC 12-08.
Barré-Sinoussi, F., Chermann, J., Rey, F., Nugeyre, M., Chamaret, S., Gruest, J., . . . Montagnier, L. (1983).
Isolation of a T-lymphotropic retrovirus from a patient at risk for acquired immune deficiency syndrome
(AIDS). Science, 220 (4599).
Bertrand, M., Duflo, E., & Mullainathan, S. (2004). How much should we trust differences-in-differences
estimates? Quarterly Journal of Economics, 119 (1), 249–275.
Bhargava, A., Jamison, D. T., Lau, L. J., & Murray, C. J. (2001). Modeling the effects of health on economic
growth. Journal of Health Economics, 20.
Bleakley, H. (2007). Disease and development: Evidence from hookworm eradication in the American South.
Quarterly Journal of Economics, 122 (1).
Bleakley, H. (2010a). Health, human capital, and development. Annual Review of Economics, 2.
Bleakley, H. (2010b). Malaria eradication in the Americas: A retrospective analysis of childhood exposure.
American Economic Journal: Applied Economics, 2.
Bloom, D. E., Canning, D., & Fink, G. (2009, July). Disease and development revisited. NBER Working
Paper No. 15137.
Bloom, D. E., Canning, D., & Sevilla, J. (2004). The effect of health on economic growth: A production
function approach. World Development, 32 (1).
Bor, J., Herbst, A., Newell, M.-L., & Bärnighausen, T. (2013). Increases in adult life expectancy in rural
South Africa: valuing the scale-up of HIV treatment. Science, 339, 961–965.
Bor, J., Tanser, F., Newell, M.-L., & Bärnighausen, T. (2012, July). In a study of a population cohort
in South Africa, HIV patients on antiretrovirals had nearly full recovery of employment. Health Aff
(Millwood), 31 (7), 1459–69. doi:10.1377/hlthaff.2012.0407
Cahu, P. & Fall, F. (2011, December). Accounting for the effects of AIDS on growth in Sub-Saharan Africa.
Documents de Travail du Centre d’Economie de la Sorbonne.
Chinese Ministry of Commerce. (2009). 2009 Statistical bulletin of China’s outward foreign direct investment.
Corno, L. & de Walque, D. (2012, February). Mines, migration and HIV/AIDS in Southern Africa. World
Bank Policy Research Working Paper 5966.
Cutler, D., Fung, W., Kremer, M., Singhal, M., & Vogl, T. (2010). Early-life malaria exposure and adult outcomes: Evidence from malaria eradication in India. American Economic Journal: Applied Economics,
2 (2).
Evans, D. K. & Miguel, E. (2007). Orphans and schooling in Africa: A longitudinal analysis. Demography,
44 (1).
Forson, J. G. (2011). Mortality risk and human capital investment: The impact of HIV/AIDS in Sub-Saharan
Africa. Review of Economics and Statistics, 93 (1).
Fortson, J. G. (2009). HIV/AIDS and fertility. American Economic Journal: Applied Economics, 1 (3).
Fowler, M. G., Qin, M., Shapiro, D., Fiscus, S., Currier, J., Makanani, B., . . . Mofenson, L. (2015, February). PROMISE: efficacy and safety of 2 strategies to prevent perinatal HIV transmission (abstract).
Abstract, Conference on Retroviruses and Opportunistic Infections, Seattle, February 2015.
Fuchs-Schündeln, N. & Hassan, T. A. (2015, April). Natural experiments in macroeconomics. Chapter prepared for the Handbook of Macroeconomics.
Gallup, J. L. & Sachs, J. D. (2001). The economic burden of malaria. American Journal of Tropical Medicine
and Hygiene, 64 (1,2).
39
Gulick, R. M., Mellors, J. W., Havlir, D., Eron, J. J., Gonzalez, C., McMahon, D., . . . Condra, J. H. (1997).
Treatment with Indinavir, Zidovudine, and Lamivudine in adults with Human Immunodeficiency Virus
Infection and prior antiretroviral therapy. New England Journal of Medicine, 337.
Habyarimana, J., Mbakile, B., & Pop-Eleches, C. (2010). The impact of HIV/AIDS and ARV treatment on
worker absenteeism: Implications for African firms. Journal of Human Resources, 45 (4).
Hammer, S., Squires, K., Hughes, M., Grimes, J., Demeter, L., Currier, J., . . . Fischl, M. (1997). A controlled trial of two nucleoside analogues plus indinavir in persons with Human Immunodeficiency Virus
infection and CD4 cell counts of 200 per cubic millimeter or less. AIDS Clinical Trials Group 320 Study
Team. New England Journal of MEdicine, 337 (11).
Hansen, B. E. (2001). The new econometrics of structural change: Dating breaks in U.S. labor productivity.
Journal of Economic Perspectives, 15 (4).
Hansen, G. D. & Prescott, E. C. (2002). Malthus to solow. American Economic Review, 92 (4).
Heritage Foundation. (2004). China global investment tracker. Retrieved from http://www.heritage.org/
research/projects/china-global-investment-tracker-interactive-map
Information Office of the State Council. (2013, August). China-Africa economic and trade cooperation. White
Paper.
Juhn, C., Kalemli-Ozcan, S., & Turan, B. (2013). HIV and fertility in Africa: first evidence from populationbased surveys. Journal of Population Economics, 26 (3).
Kalemli-Ozcan. (2011). AIDS, reversal of the demographic transition and economic development: Evidence
from Africa. Journal of Population Economics, 25 (3).
Keele, B. F., Van Heuverswyn, F., Li, Y., Bailes, E., Takehisa, J., Santiago, M. L., . . . Hahn, B. H. (2006,
July). Chimpanzee reservoirs of pandemic and nonpandemic HIV-1. Science, 313 (5786), 523–6.
Kher, U. (2003, March). A name for the plague. Time.
Kiszewski, A., Mellinger, A., Spielman, A., Malaney, P., Sachs, S. E., & Sachs, J. (2004). A global index
representing the stability of malaria transmission. American Journal of Tropical Medicine and Hygiene,
70 (5).
Korber, B., Muldoon, M., Theiler, J., Gao, F., Gupta, R., Lapedes, A., . . . Bhattacharya, T. (2000). Timing
the ancestor of the HIV-1 pandemic strains. Science, 288 (1789).
Lewis, J. J., Ronsmans, C., Ezehc, A., & Gregson, S. (2004). The population impact of HIV on fertility in
sub-Saharan Africa. AIDS, 18 (suppl 2).
McCord, G. (n.d.). Malaria ecology index dataset.
Médecins Sans Frontières. (2001, November). A matter of life and death: the role of patents in access to
essential medicines.
Médecins Sans Frontières. (2015, February). Hivbehandling i låginkomstländer inkl. prevention av
mödrabarnsmitta ( hiv treatment and prevention of maternal-child hiv transmission in developing countries). Lecture given by Kristina Bolme at Noah’s Ark, Stockholm, February 18th 2015.
Mertens, T. & Low-Beer, D. (1996). HIV and AIDS: where is the epidemic going? Bulletin of the World
Health Organization, 74 (2).
Moore, R. & Chaisson, R. (1999). Natural history of HIV infection in the era of combination antiretroviral
therapy. AIDS, 13 (14).
Nahmias, A., Weiss, J., Yao, X., Lee, F., Kodsi, R., Schanfield, M., . . . Essex, M. (1986). Evidence for human
infection with an HTLV III/LAV-like virus in Central Africa, 1959. The Lancet, 327 (8492).
Oster, E. (2005). Sexually transmitted infections, sexual behavior, and the HIV/AIDS epidemic. Quarterly
Journal of Economics, 120 (2).
Oster, E. (2012). Routes of infection: exports and HIV prevalence in sub-Saharan Africa. Journal of the
European Economic Association, 10 (5).
Palmer, S., Maldarelli, F., Wiegand, A., Bernstein, B., Hanna, G. J., Brun, S. C., . . . King, M. S. (2008).
Low-level viremia persists for at least 7 years in patients on suppressive antiretroviral therapy. PNAS,
105 (10).
Pinkovskiy, M. & Sala-i-Martin, X. (2014). Africa is on time. Journal of Economic Growth, 19.
Popovic, M., Sarngadharan, M., Read, E., & Gallo, R. (1984). Detection, isolation, and continuous production
of cytopathic retroviruses (HTLV-III) from patients with AIDS and pre-AIDS. Science, 224 (4648).
40
Reich, M. R. & Bery, P. (2005). The AIDS pandemic: impact on science and society. In K. H. Mayer & H. F.
Pizer (Eds.), (Chap. 13 Expanding Global Access to ARVs: The Challenges of Prices and Patents).
New York: Academic Press.
Resch, S., Korenromp, E., Stover, J., Blakley, M., Krubiner, C., Thorien, K., . . . Atun, R. (2011). Economic
returns to investment in AIDS treatment in low and middle income countries. PLOS One, 6 (10).
Rodrick, D. (2014, June). An African growth miracle? NBER Working Paper No. 20188.
Strauss, J. & Thomas, D. (1998). Health, nutrition, and economic development. Journal of Economic Literature, 36 (2).
Tagar, E., Sundaram, M., Condliffe, K., Over, M., Matatityo, B., Chimbwandira, F., . . . Gwinnell, D. (2012).
Multi-country analysis of treatment costs for HIV/AIDS (MATCH): Unit costing at 161 representative
facilities in Ethiopia, Malawi, Rwanda, South Africa and Zambia. Clinton Health Access Initiative.
Thirumurthy, H., Zivin, J. G., & Goldstein, M. (2008). The economic impact of AIDS treatment: Labor
supply in Western Kenya. Journal of Human Resources, 43 (3).
Thomas, D. & Frankenberg, E. (2002). Health, nutrition and prosperity: a microeconomic perspective. Bulletin of the World Health Organization, 80 (2).
UNAIDS. (2012). Global report: UNAIDS report on the global AIDS epidemic 2012.
UNAIDS. (2013). Global report: UNAIDS report on the global AIDS epidemic 2013.
UNAIDS. (2014, September). The gap report.
UNDESA. (2014). World population prospects: The 2012 revision: Methodology of the united nations population estimates and projections world population prospects the 2012 revision methodology of the United
Nations population estimates and projections. United Nations Department of Economic and Social
Affairs: Population Division.
United Nations Department of Economic and Social Affairs. (2004). The impact of AIDS. United Nations.
WHO. (2004). Global burden of disease: 2004 update.
WHO. (2011). Global HIV/AIDS response: epidemic update and health sector progress towards universal
access: progress report 2011.
WHO. (2013, June). Global update on hiv treatment 2013: results, impact and opportunities.
WHO. (2014, July). Access to antiretroviral drugs in low- and middle-income countries.
Wooldridge, J. M. (2001). Econometric analysis of cross section and panel data. MIT Press.
Young, A. (2012). The African growth miracle. Journal of Political Economy, 120.
Young, A. (2005). The gift of the dying: The tragedy of AIDS and the welfare of future African generations.
Quarterly Journal of Economics, 120 (2).
41
Figures and Tables
Figure 1: Differences between high and low HIV-prevalence countries before and after ARV scale-up
2010
0
.2
.4
.6
.8
1
% HIV-Positive patients on ARV therapy (L/MIC)
2010
0
.2
.4
.6
.8
1
% HIV-Positive patients on ARV therapy (L/MIC)
Differences between high HIV-prevalence countries
and other low- and middle-income countries
Difference between high HIV-prevalence
countries and other L/MIC
-.3
-.2
-.1
0
a) Life expectancy
1990
1995
2000
Year
2005
Difference between high HIV-prevalence
countries and other L/MIC
-.04
-.02
0
.02
Annual growth rate in GDP per capita
1990
1995
2000
Year
2005
Graphs show results from a local linear regression of the difference in the outcome variable
between 15 high HIV prevalence countries (HIV prevalence > 5% in 2001) and 71 other low and
middle income countries (HIV prevalence < 5% in 2001). The vertical grey line indicates the
year in which the price of antiretroviral therapy fell; the dashed line shows the proportion of
HIV positive individuals receiving antiretroviral therapy in all low and middle income countries.
Country-clustered bootstrapped 90% confidence intervals are shown in grey.
42
Figure 2: HIV Prevalence in 2001
Legend
HIV Prevalence, adult population, 2001
< 0.5%
0.5% - 1.0%
1.0% - 2.0%
2.0% - 4.0%
4.0% - 8.0%
8.0% - 16.0%
16.0% - 32.0%
Missing Data
Notes World borders shape file from thematicmapping.org. Estimated HIV prevalence data among population aged 15-49
from World Bank.
43
Figure 3: Fall in ARV prices in 2001
Notes Figure reproduced from Médecins Sans Frontières (2001).
Original caption: Figure 1: The effects of generic competition. Sample AIDS triple-combination: lowest world prices per
patient per year (stavudine (d4T) + lamivudine (3TC) + nevirapine).
44
Figure 4: Antiretroviral coverage among people infected by HIV in low and middle-income countries (%),
2012
Legend
ARV coverage, % of HIV positive population
< 10%
10% - 20%
20% - 30%
30% - 40%
40% - 50%
> 50%
Missing Data
Notes World borders shape file from thematicmapping.org. Data is calculated using numbers of people receiving antiretroviral
therapy (UNAIDS AIDSinfo, 2013), and number of people living with HIV estimated using World Bank data on HIV
prevalence among adults and total population.
45
Figure 5: Antiretroviral coverage among total population (%), 2012
Legend
ARV coverage, % of total population
< 0.5%
0.5% - 1.0%
1.0% - 2.0%
2.0% - 4.0%
4.0% - 8.0%
> 8.0%
Missing Data
Notes World borders shape file from thematicmapping.org. Data is calculated using numbers of people receiving antiretroviral
therapy (UNAIDS AIDSinfo, 2013), and total population from World Bank data.
46
Figure 6: Structural tests for timing of trend-break
Tests for timing of trend-break associated with HIV Prevalence
0
2.4
2.6
F-test, trend only
2
4
2.8
3
F-test, trend and level
3.2
6
3.4
a) Chow Tests
1996
1998
2000
2002
2004
2006
Break year
F-test, trend only
F-test, trend and level
6.268
6.266
6.262
6.264
Residual sum of squares
6.27
b) Residual sum of squares test
1996
1998
2000
2002
2004
2006
Break year
Graphs show results from structural tests of timing of trend break associated with HIV prevalence. See text for discussion.
47
Figure 7: Leave-one-out estimates: Life expectancy and growth in GDP per capita
.05
P-value
.1
.15
a) Log life expectancy
0
4
SWZ
LSO
NAM
ZMB
ZAF
MWI
RWA
MDG
COM
NER
GMB
MUS
MRT
AGO
CPV
GNB
BEN
MLI
LBR
KEN
DJI
ETH
BFA
GHA
SLE
THA
BLZ
UKR
SUR
LAO
NIC
BDI
GIN
SEN
MOZ
TZA
IRN
EGY
ARM
GEO
MAR
AZE
PER
PAN
BTN
ECU
IND
UGA
YEM
IDN
VNM
BOL
MDA
MEX
TJK
BGD
COL
GUY
MYS
PRY
SLV
TUN
UZB
DOM
FJI
HND
LKA
PAK
PNG
VEN
GTM
KGZ
NPL
CRI
PHL
MNG
BLR
GAB
TCD
NGA
CMR
TGO
CIV
COG
ZWE
5
BWA
6
7
Coefficient
.15
b) Growth in GDP per capita
LSO
LBR
AGO
.1
NGA
MDG
TCD
NAM
MWI
MLI
ZMB
CPV
SEN
RWA
CMR
PNG
THA
AZE
NIC
PAN
MNG
KEN
GEO
GIN
PHL
YEM
ETH
ARM
FJI
LAO
UZB
CRI
NPL
UKR
VNM
TJK
PAK
MDA
BTN
BGD
PER
VEN
MEX
MYS
SLV
ECU
COL
GTM
MAR
UGA
GUY
TUN
LKA
DOM
BOL
IND
EGY
IRN
IDN
KGZ
BLR
PRY
CIV
TZA
HND
DJI
BLZ
SUR
COG
MOZ
MRT
GAB
BDI
GMB
SLE
TGO
NER
COM
BFA
ZAF
BEN
MUS
GNB
GHA
BWA
SWZ
0
.05
P-value
ZWE
.5
1
1.5
2
Coefficient
Notes Graph shows coefficients on antiretroviral therapy coverage and associated p-values from repeated estimates, in each
case excluding the labelled country. Estimates are from a 2SLS regression where the outcome variable is regressed on
antiretroviral therapy coverage, instrumented with the interaction between HIV prevalence in 2001 and global ARV coverage,
region-year and country fixed effects, and HIV-prevalence-specific linear time trends. Standard errors clustered at the country
level. The main estimate is shown in red.
48
Figure 8: Collapsed reduced form regression
Changes in long-run growth trends post ARV scale-up vs HIV prevalence (2001)
.6
Life expectancy, SSA only
.6
Life expectancy, all LMICS
KEN
CIV
TZA
COG
UGA
SLE
GAB
TGO
NGA
CMR
SEN
GIN
GHA
TCD
DJI
BDI
BLR
BFA
ETH
UKR
BLZ
MLI
BEN
THA
SUR
LBR
MDA
TJK
GUY
KGZ
MNG
AZE
UZB
VEN
MRT
CPV
COM
GNB
PHL
GEO
LKA
PAN
PNG
FJI
DOM
AGO
MUS
GMB
CRI
TUN
MYS
PRY
COL
IND
PAK
YEM
SLV
MEX
BOL
IDN
MAR
HND
VNM
NIC
ARM
BGD
PER
EGY
IRN
ECU
NER
BTN
MOZ
GTM
NPL
LAO
MDG
ZMB
NAM
ZWE
LSO
SWZ
MWI
ZAF
BWA
0
.1
KEN
CIV
TZA
COG
UGA
SLE
GAB
TGO
NGA
CMR
SEN
GIN
GHA
TCD
BDI
BFA
ETH
MLI
BEN
LBR
MRT
CPV
COM
GNB
AGO
MUS
GMB
NER
MOZ
MDG
.2
.3
ZMB
NAM
LSO
SWZ
MWI
ZAF
BWA
RWA
-.3
-.3
RWA
Difference in 10 yr growth rate
0
.3
Difference in 10 yr growth rate
0
.3
ZWE
0
.1
.2
.3
GDP per capita, all LMICS
GDP per capita, SSA only
1
HIV Prevalence in 2001
1
HIV Prevalence in 2001
GNB
ZWE
LSO
SWZBWA
AGO
TCD
KGZ
BLR
MDA
ARM
UKR
GEO
TJK
0
.1
.2
.3
LBR
0
HIV Prevalence in 2001
LSO
SWZBWA
AGO
TCD
-2
-2
AZE
LBR
ZWE
ETH
GAB
MWI
GHA UGA
MUS
CPV
CIV
TGO
BENCOGKEN ZMB
COM
GMB
ZAF
NER
MRT
BFA TZA
MDG
NAM
GIN
SEN
SLE
MLI
BDI
RWA
CMRMOZ
NGA
D-in-D in 5 yr growth rate
-1
0
D-in-D in 5 yr growth rate
-1
0
PNG
GNB
IDN
GUY
THA
PRY
MYS
ETH
BTN
GAB
PER
MWI
COL
PAN
GHA UGA
BOL
NPL
VNM
MUS
LKA
DOM
IND
CPV
CIV
TGO
LAO
EGY
PHL
GTM
SLV
BGD
ECU
KEN ZMB
FJI
BEN
VEN
PAK
CRI
COM
COG
GMB
ZAF
NER
HND
TUN
MAR
NIC
IRN
DJI TZA
MEX
MRT
BFA
MDG
NAM
MNG
SUR
GIN
SEN
SLE
MLI
BDI
BLZ
YEM
RWA
CMRMOZ
UZBNGA
.1
.2
.3
HIV Prevalence in 2001
Notes Graphs show scatterplots of data collapsed to cross-sectional observations against HIV prevalence in 2001. Life
expectancy graphs show the difference in difference in log life expectancies, comparing the decade following 2001 to the decade
preceding 2001. GDP per capita graphs show the difference in difference in long-run growth rates, comparing the difference
between the first and second halves of the decade following 2001, to the difference between the first and second halves of the
decade preceding 2001. Linear regression lines are shown.
49
Figure 9: Differences between high and low HIV-prevalence countries
before and after ARV therapy coverage expansion
Population growth and its components
Population growth rate and its components
before and after expansion of ARV therapy coverage
Difference between high HIV-prevalence
countries and other L/MIC
0
.002
.004
.006
All L/MICs reporting HIV prevalence in 2001
1990
1995
2000
Year
2005
2010
2005
2010
Difference between high HIV-prevalence
countries and rest of SSA
-.004
-.002
0
.002
Sub-Saharan Africa only
1990
1995
2000
Year
Population growth rate, direct
Birth rate
Population growth rate, indirect
Death rate
Net migration rate
Graphs show results from a local linear regression of the difference in the outcome variable
between 15 high HIV prevalence countries (HIV prevalence > 5% in 2001) and 71 or 24 other
low and middle income countries (HIV prevalence < 5% in 2001).
50
Table 1: HIV Prevalence in 2001 and post-2001 trends in growth rates
Log life expectancy
Growth in per-capita
GDP
(3)
(4)
(1)
(2)
HIV Prevalence2001 x Post2001 x t
Coefficient
s.e.
0.155***
(0.028)
0.149***
(0.033)
0.042**
(0.018)
0.041**
(0.019)
HIV Prevalence2001 x Post2001
Coefficient
s.e.
-0.260***
(0.056)
-0.304***
(0.064)
-0.079
(0.121)
-0.002
(0.165)
HIV Prevalence2001 x t
Coefficient
s.e.
-0.108***
(0.011)
-0.123***
(0.013)
-0.025**
(0.010)
-0.028**
(0.014)
N
1978
897
1892
858
Sample
Main
SSA
Main
SSA
Note: Coefficients from regressions of outcome variable on listed variables, year and country fixed effects, and other time controls
where indicated. Main sample consists of a balanced panel of 86 countries for which HIV prevalence data is available, for 1990 to 2012.
Standard errors clustered at the country level. *** p<0.01, ** p<0.05, * p<0.1.
Table 2: Impact of increasing ARV Coverage on life expectancy and growth in per capita GDP
OLS
(1)
OLS
(2)
OLS
(3)
IV
(4)
IV
(5)
5.26***
(1.60)
5.28***
(1.61)
Panel A: % of country population treated with ARVs
HIV Prevalence2001 x global treatment
coveraget
Coefficient
s.e.
1.01***
(0.11)
HIV Prevalence2001 x t
Coefficient
s.e.
0.000
(0.001)
N
1966
Panel B: Log life expectancy
% of population treated with ARVs
Coefficient
s.e.
HIV Prevalence2001 x global treatment
coveraget
Coefficient
s.e.
HIV Prevalence2001 x t
Coefficient
s.e.
N
-4.04***
(0.70)
2.06*
(1.22)
5.33***
(1.08)
-0.090***
(0.017)
-0.123***
(0.013)
-0.124***
(0.012)
1966
1966
1978
1966
1966
0.24*
(0.13)
0.56
(0.41)
1.36*
(0.81)
1.38*
(0.81)
Panel C: Change in log GDP per capita
% of population treated with ARVs
Coefficient
s.e.
HIV Prevalence2001 x global treatment
coveraget
Coefficient
s.e.
HIV Prevalence2001 x t
Coefficient
s.e.
N
1.34*
(0.70)
-0.013**
(0.007)
-0.021**
(0.009)
-0.022**
(0.009)
1880
1880
1892
1880
1880
No
No
Yes
No
Yes
No
Yes
No
Yes
Yes
Controls
Country linear trends
Note: Coefficients from regressions of outcome variable on listed variables and controls where indicated. Controls include region-year
and country fixed effects. Main sample consists of a balanced panel of 86 low and middle-income countries for which HIV prevalence
data is available, for 1990 to 2012. 10 countries are missing ARV coverage data for 1 year and 1 country is missing ARV coverage for
2 years. In columns 3) and 5), % of population treated by ARVs is instrumented with global coverage of ARVs interacted with HIV
prevalence in 2001. Standard errors clustered by country. *** p<0.01, ** p<0.05, * p<0.1.
51
52
HIV prevalence
Sample
Time trends
N
Coefficient
s.e.
Coefficient
s.e.
Observed 2001
Main
Region-year
F.E.
1880
-0.02**
(0.01)
1.36*
(0.81)
1966
Observed 2001
Main
Year F.E.
1880
-0.02***
(0.01)
1.57*
(0.81)
1966
-0.10***
(0.01)
5.72***
(1.53)
(2)
Observed 2001
SSA
Year F.E.
852
-0.02**
(0.01)
1.34*
(0.81)
891
-0.12***
(0.01)
5.22***
(1.60)
(3)
Observed 2001
Trimmed
Region-year
F.E.
1814
-0.01
(0.01)
1.08*
(0.58)
1897
-0.11***
(0.02)
6.54***
(0.59)
(4)
Observed 2001
SSA
Linear
XY-year F.E.
852
-0.03*
(0.02)
2.36
(1.48)
891
-0.13***
(0.02)
6.41***
(2.48)
(5)
Observed 2001
SSA
Quadratic
XY-year F.E.
852
-0.06***
(0.02)
3.57*
(1.85)
891
-0.13***
(0.02)
6.83**
(3.23)
(6)
Observed 1990
Main
Region-year
F.E.
1880
-0.04*
(0.02)
3.58*
(2.14)
1966
-0.02
(0.07)
9.06***
(3.38)
(7)
Simulated
2001
SSA
Year F.E.
888
-0.02*
(0.01)
1.30
(0.87)
929
-0.10***
(0.02)
4.30***
(1.50)
(8)
Note: Coefficients from IV regressions of outcome variable on listed variables, year fixed effects (unless region-year fixed effects are specified) and country fixed effects. % of population
treated with ARVs is instrumented with the interaction between global ARV coverage and the listed measure of HIV prevalence. Main sample consists of a balanced panel of 86 low and
middle-income countries for which HIV prevalence data is available, for 1990 to 2012. 10 countries are missing ARV coverage data for 1 year and 1 country is missing ARV coverage for 2
years. Extended sample consists of all low- and middle-income countries, assuming that HIV prevalence and hence ARV coverage is zero where HIV prevalence data is missing. Trimmed
sample is constructed by dropping the three countries with highest HIV prevalence in 2001. Simulated data for HIV prevalence in 2001 generously provided by Cahu and Fall (2011).
Standard errors clustered at the country level. *** p<0.01, ** p<0.05, * p<0.1.
HIV Prevalence x t
% of population treated with ARVs
N
-0.12***
(0.01)
Coefficient
s.e.
HIV Prevalence x t
Panel B: Change in log GDP per capita
5.26***
(1.60)
Coefficient
s.e.
(1)
% of population treated with ARVs
Panel A: Log life expectancy
Table 3: Robustness of main results to changes in the sample and specification
Table 4: HIV Prevalence in 2001 and changes in long-run trends post ARV scale-up
Log life expectancy
(1)
(2)
HIV Prevalence2001
Coefficient
s.e.
N
Sample
Differencing
Change in log GDP per capita
(3)
(4)
1.35***
(0.24)
1.28***
(0.28)
1.18**
(0.50)
1.16*
(0.58)
86
39
86
39
All
D in D
SSA
D in D
All
Triple D
SSA
Triple D
Note: In columns 1) and 2), dependent variable is the difference in differences in log life expectancy, comparing the decade before 2001
(1991 – 2001) to the decade following 2001 (2001-2011). In columns 3) and 4), dependent variable is the difference in differences in 5
year growth rates, comparing the difference between 2001 and 1996, and 1996 and 1991, to the difference between 2011 and 2006, and
2006 and 2001. Main sample consists of all low and middle-income countries reporting HIV prevalence in 2001. Robust standard errors
in parentheses. *** p<0.01, ** p<0.05, * p<0.1.
53
Table 5: Impact of ARVs on growth in per-capita GDP: Alternative explanations
(1)
% of population treated with
ARVs
Coefficient
s.e.
Fraction GDP mineral rents2001
xt
Coefficient
s.e.
-0.048**
(0.023)
-0.044*
(0.024)
Fraction GDP mineral rents2001
x Post 2001
Coefficient
s.e.
0.237*
(0.133)
0.189
(0.210)
Fraction GDP mineral rents2001
x Post 2001 x t
Coefficient
s.e.
0.006**
(0.003)
0.003
(0.002)
N
1.36*
(0.81)
Outcome variable: Growth in per-capita GDP
(2)
(3)
(4)
(5)
1.46*
(0.82)
1.46*
(0.82)
1.34*
(0.81)
1.39*
(0.83)
1880
1880
852
852
852
1.36*
(0.81)
1.27
(0.79)
1.27
(0.79)
1.34*
(0.81)
1.27
(0.79)
1.26
(0.79)
Coefficient
s.e.
Fraction GDP petroleum
rents2001 x t
Coefficient
s.e.
-0.001
(0.007)
0.001
(0.008)
Fraction GDP petroleum
rents2001 x Post 2001
Coefficient
s.e.
0.085**
(0.034)
0.068*
(0.040)
Fraction GDP petroleum
rents2001 x Post 2001 x t
Coefficient
s.e.
-0.001
(0.001)
-0.001
(0.001)
1880
1880
1880
852
852
852
1.36*
(0.81)
1.27
(0.89)
1.26
(0.89)
1.34*
(0.81)
1.52
(0.93)
1.53
(0.94)
% of population treated with
ARVs
Coefficient
s.e.
Fraction GDP exports2001 x t
Coefficient
s.e.
-0.0052
(0.0035)
0.0005
(0.0053)
Fraction GDP exports2001 x
Post 2001
Coefficient
s.e.
0.0284
(0.0191)
0.0090
(0.0306)
Fraction GDP exports2001 x Post
2001 x t
Coefficient
s.e.
0.0003
(0.0004)
-0.0005
(0.0007)
N
1.38*
(0.83)
1880
% of population treated with
ARVs
N
(6)
1880
1880
1880
852
852
852
1.36*
(0.81)
1.61*
(0.89)
1.62*
(0.89)
1.34*
(0.81)
1.53*
(0.93)
1.52
(0.93)
% of population treated with
ARVs
Coefficient
s.e.
Malaria ecology x t
Coefficient
s.e.
-0.0023
(0.0039)
0.0001
(0.0054)
Malaria ecology x Post 2001
Coefficient
s.e.
-0.0060
(0.0293)
-0.0352
(0.0443)
Malaria ecology x Post 2001 x t
Coefficient
s.e.
0.0004
(0.0003)
0.0002
(0.0005)
N
1880
1880
1880
852
852
852
Additional heterogeneous trends
Sample
None
Main
Linear
Main
Flexible
Main
None
SSA
Linear
SSA
Flexible
SSA
Note: Coefficients from IV regressions of outcome variable on listed variables, region-year and country fixed effects, and HIV prevalence
in 2001 interacted with a linear time trend. % of population treated with ARVs is instrumented with the interaction between global
ARV coverage and HIV prevalence in 2001. Main sample consists of a balanced panel of 86 low and middle-income countries for which
HIV prevalence data is available, for 1990 to 2012. 10 countries are missing ARV coverage data for 1 year and 1 country is missing
ARV coverage for 2 years. Where flexible trends are included, these comprise interactions between the relevant cross-sectional variable
and year dummies. Standard errors clustered at the country level. *** p<0.01, ** p<0.05, * p<0.1.
54
Table 6: Effects of ARV therapy coverage expansion on population growth and its components
Population
growth rate
Crude birth
rate
Crude death
rate
Net
migration
rate
(1)
(2)
(3)
(4)
Panel A: Main sample
% of population treated
with ARVs
Coefficient
s.e.
0.39**
(0.16)
0.06***
(0.02)
-0.14***
(0.04)
0.15
(0.11)
HIV Prevalence2001 x t
Coefficient
s.e.
-0.0072***
(0.0022)
-0.0009*
(0.0005)
0.0038***
(0.0005)
-0.0015
(0.0017)
1880
1966
1966
430
N
Panel B: Sub Saharan Africa
% of population treated
with ARVs
Coefficient
s.e.
0.39**
(0.16)
0.06***
(0.02)
-0.14***
(0.04)
0.15
(0.11)
HIV Prevalence2001 x t
Coefficient
s.e.
-0.0073***
(0.0022)
-0.0009*
(0.0005)
0.0038***
(0.0005)
-0.0015
(0.0017)
852
891
891
195
N
Note: Coefficients from IV regressions of outcome variable on listed variables, region-year and country fixed effects. % of population
treated with ARVs is instrumented with the interaction between global ARV coverage and HIV prevalence in 2001. Main sample consists
of a balanced panel of 86 low and middle-income countries for which HIV prevalence data is available, for 1990 to 2012. 10 countries
are missing ARV coverage data for 1 year and 1 country is missing ARV coverage for 2 years.
55
Table 7: Approximate valuation of expanded ARV therapy coverage in sub-Saharan Africa (2012)
Country
Angola
Benin
Botswana
Burkina Faso
Burundi
Cabo Verde
Cameroon
Chad
Comoros
Congo, Rep.
Cote d’Ivoire
Ethiopia
Gabon
Gambia, The
Ghana
Guinea
Guinea-Bissau
Kenya
Lesotho
Liberia
Madagascar
Malawi
Mali
Mauritania
Mauritius
Mozambique
Namibia
Niger
Nigeria
Rwanda
Senegal
Sierra Leone
South Africa
Swaziland
Tanzania
Togo
Uganda
Zambia
Zimbabwe
ARV Coverage
0.002
0.003
0.106
0.003
0.003
0.002
0.006
0.003
0.000
0.004
0.006
0.003
0.009
0.002
0.003
0.002
0.004
0.014
0.045
0.001
0.000
0.025
0.002
0.000
0.001
0.012
0.052
0.001
0.003
0.010
0.001
0.001
0.041
0.071
0.009
0.005
0.012
0.034
0.041
Increase in GDP
per capita due to
ARV
Value per person
receiving ARV
Population
Valuation
(2012 US$)
(2012 US$)
(Thousands)
(2012 US$bn)
14.14
2.59
1096.45
2.43
0.98
8.90
9.53
4.42
0.04
18.17
9.75
1.50
145.18
1.39
5.88
1.42
3.34
18.71
74.22
0.66
0.01
12.42
1.91
0.74
13.86
8.87
387.22
0.36
9.77
7.70
1.55
0.92
431.59
325.78
6.43
3.55
7.12
79.67
45.28
6911
998
10360
871
331
5092
1683
1348
1083
4573
1753
479
15795
698
2135
608
910
1337
1642
505
611
488
989
1496
11720
722
7498
520
3359
770
1451
668
10489
4581
710
777
590
2332
1099
20821
10051
2004
16460
9850
494
21700
12448
718
4337
19840
91729
1633
1791
25366
11451
1664
43178
2052
4190
22294
15906
14854
3796
1291
25203
2259
17157
168834
11458
13726
5979
52275
1231
47783
6643
36346
14075
13724
294.46
26.00
2197.18
39.97
9.63
4.40
206.69
55.03
0.03
78.82
193.53
137.96
237.02
2.49
149.19
16.22
5.55
807.71
152.26
2.77
0.23
197.59
28.44
2.81
17.90
223.64
874.87
6.14
1649.40
88.27
21.32
5.52
22561.12
401.03
307.14
23.56
258.95
1121.37
621.47
Mean
2820
Total
33027.68
Table 8: Approximate contribution of ARV therapy coverage to growth in sub-Saharan Africa between
2004 and 2012
Year
2004
2005
2006
2007
2008
2009
2010
2011
2012
Total/Mean
ARV
Coverage
0.001
0.002
0.004
0.005
0.007
0.008
0.011
0.012
0.014
Component of
GDP/capita
growth due to
ARV
Observed
GDP/capita
growth
%
%
0.2
0.3
0.5
0.6
0.9
1.1
1.4
1.6
1.9
3.1
2.6
3.0
3.2
2.3
0.6
2.8
1.9
2.9
8.5
22.4
56
Fraction of
observed
increase due
to ARV
Increase in
total GDP due
to ARV
Observed
increase in
GDP
(2012 US$bn)
(2012 US$bn)
0.05
0.13
0.16
0.21
0.37
1.98
0.52
0.83
0.64
0.8
2.3
3.6
5.3
8.2
11.1
14.6
18.1
23.1
50.5
31.0
40.4
42.8
35.3
14.2
38.1
34.1
38.9
0.02
0.07
0.09
0.12
0.23
0.78
0.38
0.53
0.59
0.38
87.1
325.3
0.27
Fraction of
observed due
to ARV
Appendices
Figure A1: Leave-one-out estimates: Population growth and its components
b) Birth rate
.015
.04
a) Population growth
SWZ
BWA
P-value
.005
.01
LSO
SLELBRNAM
RWA
ZWE
MRT
MLI
TCD
MUS
ZMB
COM
AGO
MDG
BFA
NER
BEN
BDI
GHA
TGO
DJI
GMB
TZA
KEN
BTN
CPV
UKR
EGY
IRN
ARM
HND
IND
NPL
CRI
COG
NGA
KGZ
PHL
CMR
IDN
SLV
UZB
GNB
BOL
TJK
COL
ECU
GUY
DOM
UGA
LKA
VEN
MYS
VNM
GTM
GAB
PRY
AZE
BLR
PER
TUN
GEO
MAR
BGD
MDA
ETH
MOZ
MNG
PAK
PAN
PNG
MEX
SUR
FJI
YEM
LAO
NIC
CIV
ZAF
BLZ
THA
SEN
MWI
GIN
ZWE
COM
SLE
LSO
ZAF
NAM
AGO
LBR
BFA
ETH
TCD
MWI
BEN
MLI
MUS
COG
MRT
EGY
PHL
YEM
NER
IDN
FJI
BLZ
CIV
GEO
NGA
CMR
HND
MEX
BLR
ARM
SLV
GAB
UZB
BOL
ECU
GUY
LKA
COL
MYS
VEN
DOM
BGD
PER
IND
GTM
BTN
PAK
GNB
TZA
PRY
MAR
TJK
VNM
DJI
TUN
CRI
AZE
KGZ
MDA
GMB
PAN
UGA
NPL
GIN
TGO
MDG
SUR
IRN
UKR
LAO
PNG
NIC
GHA
KEN
MNG
RWA
THA
SEN
MOZCPV
BDI
ZMB
0
BWA
0
.01
P-value
.02
.03
SWZ
.3
.35
.4
Coefficient
.45
.5
.045
.065
d) Net migration rate
LSO
P-value
.3
.4
ZWE
ZWE
BWA
-.18
-.16
-.14
Coefficient
-.12
SLE
LBR
RWA
NAM
SWZ
BDI
LSO
MRT
ZMB
MLI
MOZ
TCD
GHA
TZA
TGO
KEN
MDA
BTN
COG
IRN
NPL
MNG
ARM
CRI
IND
EGY
HND
TJK
KGZ
AZE
SLV
UKR
DJI
UGA
BLR
GEO
UZB
GAB
COL
GUY
IDN
GTM
MYS
PRY
VEN
ECU
BOL
PNG
DOM
VNM
MAR
LKA
TUN
PAN
PHL
PAK
PER
BGD
SUR
LAO
CMR
BFA
MEX
FJI
BEN
NGA
CIV
NIC
YEM
AGO
MDG
GMB
NER
THA
BLZ
GNB
ETH
COM
SEN
CPV
MUS
ZAF
.2
NAM
ZMB
ZAF
MWI
RWA
MDG
NER
CPV
MRT
AGO
COM
GMB
MUS
BDI
BEN
GNB
LBR
ETH
MLI
KEN
GHA
DJI
THA
UKR
GIN
BFA
BLZ
NIC
LAO
MOZ
SUR
IRN
TJK
MNG
DOM
PRY
BOL
PNG
GEO
MEX
ARM
EGY
MAR
AZE
CRI
ECU
IND
PER
BTN
COL
MDA
VNM
YEM
PAN
FJI
KGZ
PAK
VEN
UZB
BGD
MYS
SLV
PHL
IDN
LKA
NPL
GUY
GTM
SEN
HND
TUN
BLR
UGA
SLE
TZA
GAB
TCD
TGO
CMR
NGA
CIV
COG
-.1
BWA
MWI
GIN
.1
0
P-value
.001 .002 .003 .004 .005
SWZ
.055
.06
Coefficient
.5
c) Death rate
.05
.05
.1
.15
Coefficient
.2
Notes Graph shows coefficients on antiretroviral therapy coverage and associated p-values from repeated estimates, in each
case excluding the labelled country. Estimates are from a 2SLS regression where the outcome variable is regressed on
antiretroviral therapy coverage, instrumented with the interaction between HIV prevalence in 2001 and global ARV coverage,
region-year and country fixed effects, and HIV-prevalence-specific linear time trends. Standard errors clustered at the country
level. The main estimate is shown in red.
57
Figure A2: Differences in primary school enrollment between high and low HIV-prevalence countries before
and after ARV scale-up
-.01
Difference between high HIV-prevalence
countries and other L/MIC
-.005
0
.005
Primary school enrollment rates
1990
1995
2000
Year
2005
Net, all LMICs
Gross, all LMICs
Net, SSA only
Gross, SSA only
2010
Graphs show results from a local linear regression of the difference in the outcome variable
between 15 high HIV prevalence countries (HIV prevalence > 5% in 2001) and either 71 or 24
other low and middle income countries (HIV prevalence < 5% in 2001).
58
Figure A3: Differences in labour force participation between high and low HIV-prevalence countries before
and after ARV scale-up
Difference between high HIV-prevalence
countries and other L/MIC
-.001
-.0005
0
.0005
.001
Labour force participation rates (ILO estimates)
1990
1995
2000
Year
2005
15 plus, all LMICs
15-64, all LMICs
15 plus, SSA only
15-64, SSA only
2010
Graphs show results from a local linear regression of the difference in the outcome variable
between 15 high HIV prevalence countries (HIV prevalence > 5% in 2001) and either 71 or 24
other low and middle income countries (HIV prevalence < 5% in 2001).
59
Table A1: Testing for spatial and serial correlation
Outcome variable: ûi,t
(2)
(3)
(1)
(4)
Panel A: Test for serial correlation
ûi,t−1
Coefficient
s.e.
0.301***
(0.024)
0.183***
(0.024)
0.178***
(0.024)
0.108***
(0.036)
ûi,t−2
Coefficient
s.e.
0.071***
(0.025)
0.037***
(0.024)
0.004**
(0.024)
0.020*
(0.035)
ûi,t−3
Coefficient
s.e.
0.015**
(0.025)
-0.018
(0.024)
-0.012
(0.024)
0.004
(0.035)
ûi,t−4
Coefficient
s.e.
-0.053
(0.025)
-0.078
(0.025)
-0.071
(0.024)
-0.083
(0.036)
ûi,t−5
Coefficient
s.e.
-0.055
(0.024)
-0.097**
(0.024)
-0.072
(0.024)
-0.082
(0.036)
N
1880
1880
1880
852
Coefficient
s.e.
0.692***
(0.201)
0.448***
(0.144)
0.107
(0.077)
-0.002
(0.041)
N
1880
1880
1880
852
No
No
Main
Yes
No
Main
No
Yes
Main
No
No
SSA
Panel B: Test for spatial correlation
ûj∈N,j6=i,t
Country linear trends
Region year F.E.
Sample
Note: Coefficients from regressions of estimated residuals on lagged estimated residuals (panel A) or mean estimated residuals in
neighbouring countries (panel B). Residuals estimated from 2SLS regressions of change in log GDP per capita on % of population
treated with ARV, HIV prevalence specific linear trends, year and country fixed effects, and other controls where indicated. % of
population treated with ARVs is instrumented with the interaction between global ARV coverage and HIV prevalence in 2001. Main
sample consists of a balanced panel of 86 low and middle-income countries for which HIV prevalence data is available, for 1990 to 2012.
10 countries are missing ARV coverage data for 1 year and 1 country is missing ARV coverage for 2 years. Standard errors clustered
1
(Panel A) or − Nc1−1 (Panel B) to reflect
at the country level. Stars correspond to tests to reject equality of coefficient with − T −1
mechanical negative correlation due to the inclusion of panel and year fixed effects respectively (Wooldridge, 2001). *** p<0.01, **
p<0.05, * p<0.1.
60
61
Coefficient
s.e.
HIV Prevalence2001 x
t
891
-0.124***
(0.0121)
5.22***
(1.60)
1966
-0.124***
(0.0120)
5.26***
(1.60)
891
0.004***
(0.0005)
-0.14***
(0.04)
1966
0.004***
(0.0005)
-0.14***
(0.04)
(2)
(1)
891
0.052**
(0.0226)
2.66
(1.77)
1966
0.054**
(0.0224)
2.62
(1.77)
(3)
Log neonatal
mortality
891
0.153***
(0.0325)
-5.69***
(1.45)
1966
0.156***
(0.0322)
-5.73***
(1.45)
(4)
Log infant
mortality
891
0.228***
(0.0388)
-8.51***
(1.54)
1966
0.231***
(0.0384)
-8.56***
(1.54)
(5)
Log under 5
mortality
891
0.253***
(0.0223)
-10.87***
(2.93)
1961
0.256***
(0.0222)
-11.09***
(2.97)
(6)
Log adult male
mortality
891
0.322***
(0.0282)
-12.87***
(3.66)
1961
0.326***
(0.0282)
-13.07***
(3.70)
(7)
Log adult female
mortality
Note: Coefficients from IV regressions of outcome variable on listed variables, year and country fixed effects. % of population treated with ARVs is instrumented with the interaction
between global ARV coverage and HIV prevalence in 2001. Main sample consists of a balanced panel of 86 low and middle-income countries for which HIV prevalence data is available, for
1990 to 2012. 10 countries are missing ARV coverage data for 1 year and 1 country is missing ARV coverage for 2 years.
N
Coefficient
s.e.
% of population
treated with ARVs
N
Coefficient
s.e.
HIV Prevalence2001 x
t
Panel B: Sub Saharan Africa
Coefficient
s.e.
% of population
treated with ARVs
Panel A: Main sample
Crude death rate
Log life
expectancy
Table A2: Alternative measures of improvements in health
62
–
–
Extended
2254
-0.02**
(0.01)
1.27
(0.78)
2349
-0.12***
(0.01)
5.16***
(1.57)
(2)
–
Trimmed
1815
-0.02**
(0.01)
1.12
(0.76)
1897
-0.12***
(0.01)
5.07***
(1.64)
(3)
–
Winsor
1880
-0.02**
(0.01)
1.06
(0.72)
1966
-0.12***
(0.01)
5.02***
(1.54)
(4)
Raw 90s
growth
Matched
1619
-0.02**
(0.01)
1.32
(0.84)
1693
-0.13***
(0.01)
5.23***
(1.63)
(5)
Residual 90s
growth
Matched
1620
-0.03***
(0.01)
1.59*
(0.87)
1694
-0.13***
(0.01)
5.25***
(1.65)
(6)
90s trend
Matched
1618
-0.02**
(0.01)
1.39*
(0.79)
1692
-0.12***
(0.01)
5.40***
(1.60)
(7)
–
HIV
813
-0.03**
(0.01)
1.63*
(0.96)
850
-0.13***
(0.02)
4.94***
(1.67)
(8)
Note: Coefficients from IV regressions of outcome variable on listed variables, region-year and country fixed effects. % of population treated with ARVs is instrumented with the interaction
between global ARV coverage and HIV prevalence in 2001. Main sample consists of a balanced panel of 86 low and middle-income countries for which HIV prevalence data is available,
for 1990 to 2012. 10 countries are missing ARV coverage data for 1 year and 1 country is missing ARV coverage for 2 years. Column 1) replicates the main results for comparison.
Extended sample consists of all low- and middle-income countries, assuming that HIV prevalence and hence ARV coverage is zero where HIV prevalence data is missing. Trimmed sample
is constructed by dropping the three countries with most extreme values for the outcome variable. In column 4), the top and bottom 1.5% of values for the outcome variable are winsorized.
In columns 5), 6) and 7), a matched control group is constructed by finding the nearest neighbour with replacement to each country with HIV prevalence > 1%. In column 8), only
countries with HIV prevalence > 1% are included in the sample. Standard errors clustered at the country level. *** p<0.01, ** p<0.05, * p<0.1.
Matching variable
1880
Main
N
-0.02**
(0.01)
Coefficient
s.e.
HIV Prevalence x t
Sample
1.36*
(0.81)
Coefficient
s.e.
1966
% of population treated
with ARVs
N
-0.12***
(0.01)
Coefficient
s.e.
HIV Prevalence x t
Panel B: Change in log GDP per capita
5.26***
(1.60)
Coefficient
s.e.
(1)
% of population treated
with ARVs
Panel A: Log life expectancy
Table A3: Robustness of main results to changes in the sample and specification
Additional checks
Table A4: List of countries with HIV Prevalence > 1% in 2001
Country
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
20
22
22
24
25
26
26
26
29
30
30
30
33
33
35
35
35
38
38
38
HIV Prevalence (2001)
Botswana
Swaziland
Zimbabwe
Lesotho
Malawi
South Africa
Zambia
Namibia
Mozambique
Kenya
Tanzania
Uganda
Cote d’Ivoire
Gabon
Cameroon
Congo, Rep.
Togo
Rwanda
Chad
Ethiopia
Equatorial Guinea
Bahamas, The
Nigeria
Burundi
Guinea-Bissau
Djibouti
Ghana
Liberia
Burkina Faso
Angola
Belize
Thailand
Benin
Mali
Dominican Republic
Guinea
Trinidad and Tobago
Honduras
Panama
Suriname
63
.281
.248
.243
.234
.155
.153
.151
.15
.09
.085
.075
.068
.064
.061
.052
.047
.045
.044
.038
.036
.036
.035
.035
.029
.028
.023
.023
.023
.022
.018
.018
.018
.016
.016
.013
.013
.013
.012
.012
.012
Table A5: Impact of ARVs on life expectancy: Alternative explanations
Outcome variable: Log life expectancy
(2)
(3)
(4)
(5)
(1)
(6)
Panel A: Chinese investment in mineral rich countries
% of population treated with
ARVs
Coefficient
s.e.
Fraction GDP mineral rents2001
xt
Coefficient
s.e.
-0.027
(0.020)
-0.064**
(0.025)
Fraction GDP mineral rents2001
x Post 2001
Coefficient
s.e.
-0.013
(0.047)
-0.139***
(0.050)
Fraction GDP mineral rents2001
x Post 2001 x t
Coefficient
s.e.
0.0004
(0.0010)
0.0004
(0.0023)
N
5.26***
(1.60)
5.26***
(1.61)
5.27***
(1.61)
5.22***
(1.60)
5.22***
(1.61)
5.22***
(1.61)
1966
1966
1966
891
891
891
5.26***
(1.60)
5.30***
(1.61)
5.30***
(1.61)
5.22***
(1.60)
5.29***
(1.61)
5.29***
(1.61)
Panel B: Chinese investment in petroleum rich countries
% of population treated with
ARVs
Coefficient
s.e.
Fraction GDP petroleum
rents2001 x t
Coefficient
s.e.
-0.004
(0.005)
-0.006
(0.006)
Fraction GDP petroleum
rents2001 x Post 2001
Coefficient
s.e.
0.014
(0.013)
0.019
(0.016)
Fraction GDP petroleum
rents2001 x Post 2001 x t
Coefficient
s.e.
0.000
(0.0)
0.001
(0.001)
N
1966
1966
1966
891
891
891
5.26***
(1.60)
5.19***
(1.65)
5.19***
(1.65)
5.22***
(1.60)
5.22***
(1.69)
5.22***
(1.69)
Panel C: Export boom
% of population treated with
ARVs
Coefficient
s.e.
Fraction GDP exports2001 x t
Coefficient
s.e.
-0.010**
(0.005)
-0.015
(0.010)
Fraction GDP exports2001 x
Post 2001
Coefficient
s.e.
-0.002
(0.010)
-0.020
(0.020)
Fraction GDP exports2001 x Post
2001 x t
Coefficient
s.e.
0.00017
(0.00038)
-0.00004
(0.00080)
N
1966
1966
1966
891
891
891
5.26***
(1.60)
5.68***
(1.86)
5.69***
(1.87)
5.22***
(1.60)
5.92***
(2.11)
5.93***
(2.12)
Panel D: Global Fund investment in fighting malaria
% of population treated with
ARVs
Coefficient
s.e.
Malaria ecology x t
Coefficient
s.e.
-0.0107
(0.0085)
-0.0192
(0.0141)
Malaria ecology x Post 2001
Coefficient
s.e.
0.0174*
(0.0105)
0.0288*
(0.0150)
Malaria ecology x Post 2001 x t
Coefficient
s.e.
0.0007
(0.0006)
0.0012
(0.0010)
N
Heterogeneous trends
Sample
1966
1966
1966
891
891
891
Linear
Main
Fixed
Main
Flexible
Main
Linear
SSA
Fixed
SSA
Flexible
SSA
Note: Coefficients from IV regressions of outcome variable on listed variables, region-year and country fixed effects, and HIV prevalence
in 2001 interacted with a linear time trend. % of population treated with ARVs is instrumented with the interaction between global
ARV coverage and HIV prevalence in 2001. Main sample consists of a balanced panel of 86 low and middle-income countries for which
HIV prevalence data is available, for 1990 to 2012. 10 countries are missing ARV coverage data for 1 year and 1 country is missing
ARV coverage for 2 years. Where flexible trends are included, these comprise interactions between the relevant cross-sectional variable
and year dummies. Standard errors clustered at the country level. *** p<0.01, ** p<0.05, * p<0.1.
64
Table A6: List of countries with highest coverage of ARV therapy (by total population) in 2012
Country
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
ARV Coverage as % of
population (2012)
Botswana
Swaziland
Namibia
Lesotho
Zimbabwe
South Africa
Zambia
Malawi
Kenya
Mozambique
Uganda
Rwanda
Gabon
Tanzania
Cameroon
Cote d’Ivoire
Guyana
Togo
Congo, Rep.
Belize
Guinea-Bissau
Thailand
Chad
Ethiopia
Burundi
Nigeria
Burkina Faso
Ghana
Benin
Suriname
Guinea
Dominican Republic
Angola
Gambia, The
Mali
Cabo Verde
Djibouti
Papua New Guinea
Brazil
Panama
Venezuela, RB
65
.1058346
.0711089
.0516453
.0452084
.041217
.0411455
.0341685
.0254696
.0139892
.012294
.0120658
.0100035
.0091916
.009047
.0056583
.0055631
.0046733
.0045629
.0039737
.0038913
.0036674
.00358
.0032797
.0031412
.0029566
.0029083
.0027892
.0027544
.0025904
.0025667
.0023286
.0021623
.0020464
.0019936
.0019356
.0017476
.001724
.0016414
.0015765
.001558
.0014366