WP-03
Institute of Health Policy & Management
HEFPA working paper
Marginal Benefit Incidence of Public
Health Spending: Evidence from
Indonesian sub-national data
Ioana Kruse
Menno Pradhan
Robert Sparrow
Marginal Benefit Incidence of Public Health Spending:
Evidence from Indonesian sub-national data
Ioana Kruse a, Menno Pradhan b and Robert Sparrow c
a
Johns Hopkins University Bloomberg School of Public Health
b
VU University Amsterdam & University of Amsterdam
c
International Institute of Social Studies of Erasmus University Rotterdam & IZA
August 2011
Abstract
We examine the marginal effects of decentralized public health spending by
incorporating estimates of behavioural responses to changes in health spending in benefit
incidence analysis. The analysis is based on a panel dataset of 207 Indonesian districts
over the period from 2001 to 2004. We show that district public health spending is
largely driven by central government transfers, with an elasticity of around 0.9. We find a
positive effect of public health spending on utilization of outpatient care in the public
sector for the poorest two quartiles. We find no evidence that public expenditures crowd
out utilization of private services or household health spending. Our analysis suggests
that increased public health spending improves targeting to the poor, as behavioural
changes in public health care utilization are pro-poor. Nonetheless, most of the benefits of
the additional spending accrued to existing users of services, as initial utilization shares
outweigh the behavioural responses.
JEL: H72, H75, I18
Key words: Decentralization, public spending, health care utilization, benefit incidence,
Indonesia.
Corresponding author: Robert Sparrow, International Institute of Social Studies,
Erasmus University Rotterdam, P.O. Box 29776, 2502 LT The Hague, The Netherlands.
Tel. +31-70-4260574. Fax +31-70-4260507. Email: [email protected].
Acknowledgements: We thank Wolfgang Fengler, Kai Kaiser, Blane Lewis, Bambang
Sjahrir Putra, Denis Raynaud and two anonymous referees for useful comments, as well
as seminar participants at the HEFPA workshops, Brunel University, Freiburg University,
IRDES, Maastricht University, University of Frankfurt, University of Göttingen,
University of Indonesia and VU University Amsterdam. Ellen Tan provided invaluable
assistance with the data. Robert Sparrow gratefully acknowledges financial support from
EU-FP7 grant 223166.
1. Introduction
Benefit incidence analysis of public spending analyzes how the recipients of public
spending are distributed across the population. It is widely used in policy circles (Shah
2005), including public health (McIntyre & Ataguba 2010). Noting that public health
spending is typically received in kind through obtaining services from public service
providers, a standard procedure is to assume these “benefits” are proportional to
utilization rates. The share an income group receives of total benefits is then directly
deduced from their observed use of public services relative to others.
Several authors have noted that this average benefit incidence may not be a good
predictor of the marginal benefit incidence, that is, how the benefit distribution changes
as a result of changes in public spending. The behaviour of administrators and (potential)
users could induce demand responses that differ by income group. Lanjouw and
Ravallion (1999) argue that the distributional effects of public spending reforms may
differ from static average patterns due to political processes that drive such reforms,
while Van de Walle (1998) notes the importance of considering behavioural responses in
benefit incidence analysis. Younger (2003) notes that a lot of progress can be made by
looking in more detail at which spending categories the changes in spending takes place.
Additional spending used to finance an expansion of services, for instance, will benefit a
new group of users compared to a general quality upgrade, which will largely benefit
existing users.
This paper proposes a method to analyze marginal benefit incidence of public
health spending that incorporates behavioural responses to changes in public health
spending. We show that our proposed measure of marginal benefit incidence fits within a
1
general framework of benefit incidence analysis that also encompasses existing measures
of average and marginal benefit incidence. We argue that these measures differ only in
the assumptions they impose regarding the causal effects of changes in public spending
on health care utilization, and in particular how spending elasticity’s of utilization vary
across the population. Our proposed measure relaxes these implicit assumptions, by
explicitly estimating these elasticity’s.
The focus on the causal effects of public health spending places this paper in a
broader literature on the effectiveness of public health spending (World Bank 2004). This
literature generally looks at the effects on health outcomes. Based on cross country
variation, Filmer and Pritchett (1999) note the lack of correlation between public health
spending and child mortality and conclude that governance, or the way in which
resources translate into actual programs, and crowding out of the private sector by the
public sector are the missing chains that explain the low correlation.
Analyzing the distributional effects of public health spending on the utilization of
public health services can thus be viewed as an analysis of the first chain, whether
spending translated into programs that people use. McGuire (2006) shows that in a cross
section of developing countries, access to maternal and infant health programs is
correlated with decreased under 5 mortality, while public health spending is not. This
indicates that it is the quality of the implemented programs that matter, and not the
spending per se. The effects of public health spending may also vary by population
segment. Using cross-country data, Gupta et al. (2003) show that higher public health
spending reduces child mortality among the poorest quintile, while no effects can be
detected among the rich.
2
This paper also contributes to the empirical evidence relating to the second chain,
whether changes in public spending crowds out private spending in health. This claim is
supported by several authors, who note that total health spending is closely correlated
with GDP, with an elasticity around one (Gerdtham & Jönsson 2000; van der Gaag &
Štimac 2008). But these studies also show that the share of public spending devoted to
health however varies a lot across governments, which could suggest that the size of the
private sector adjusts to the size of the public sector.
However, conclusions regarding the determinants and effectiveness of public
health spending based on cross-country analysis such as referenced above should be
interpreted with caution. Omitted variable bias, resulting from country specific
unobserved historic and institutional factors that influence both public spending decisions
and health outcomes make it difficult to interpret the estimated relationships as causal.
Moreover, cross-country studies are typically prone to measurement error, due to
inconsistencies between countries in data quality, data collection tools and underlying
sources of (micro) data. Perhaps the most serious problem is that public health spending
is endogenously determined by a fiscal and public health policy that could be influenced
by the outcome variables of interest.
Analyzing sub-national expenditures in a decentralized context overcomes many
of the problems associated with cross country analysis. As sub national governments
operate within the same institutional setting, and often share data collection tools, the
analyses are less plagued by omitted variable biases. Bhalotra (2007) uses this strategy to
analyze the effects of state health expenditures in India on child mortality using a 29 year
panel of 15 states, taking a state fixed effects approach.
3
This paper follows a similar strategy, taking Indonesia as case study and
analyzing sub-national public health care spending, its determinants and impacts, in 207
districts. The analysis covers the period from 2001 to 2004, immediately after Indonesia
decentralized, and district governments received far reaching authorities to set the size
and composition of their spending, including how much of their total budget to allocate to
the health sector. This period just after decentralization, when budgets were still in flux,
provides a unique opportunity to analyze the determinants and impacts of public health
spending, across local governments that inherited a similar institutional setting.
We first analyze the determinants of district public health spending, and how
changes in central transfers affect the size and composition of district health spending.
Next, by linking the fiscal data to district income specific utilization rates of public
services obtained from household surveys, we estimate the elasticity of demand for
public health services with respect to changes in public spending. Armed with these
results, we develop a method to analyze the marginal benefit incidence of changes in
public health spending that takes account of behavioural responses to changes in public
spending, avoiding arbitrary assumptions used in earlier studies. We show that marginal
benefit incidence can be calculated by multiplying average benefit incidence rates with a
simple correction factor based on these elasticity’s.
Further, we test for whether changes in public spending crowd out private
spending in the health sector. We investigate whether public services crowd out private
services by estimating the cross elasticity of utilization of services in the private sector
with respect to changes in public spending directed to the public sector. We also test
4
whether changes in public health spending crowd out out-of-pocket (OOP) health
expenditures by households.
We find that public health spending is elastic with respect to district revenues,
which is mainly driven by untied transfers from the central government and district
revenues. Public health spending has a positive effect on utilization of outpatient care in
the public sector for the poorest two quartiles. Increased public health spending improves
targeting to the poor, as behavioural changes in public health care utilization are pro-poor.
But these behavioural changes are relatively small compared to initial utilization shares.
Hence, most of the benefits of additional spending accrue to existing users of services.
We find no impact of changes in public health spending on utilization of health services
from the private sector or OOP health expenditures, thus casting doubt on the earlier
suggestions that the public sector crowds out the private sector in health.
The paper is organized as follows. Section 2 provides an overview of
decentralization in the health sector in Indonesia. We describe the sources of revenues of
districts, and the trends in central and district government expenditures over the period of
study. Section 3 analyzes the determinants of district public health spending. Section 4
presents a behavioural benefit incidence analysis of changes in district health spending,
evaluates the impact of public spending and provides a sensitivity analysis. Section 5
concludes.
2. Institutional Setting and Data
Districts inherited an organizational structure to deliver public health services which was
implemented nationwide during the centralist Suharto Regime, under which district
5
health offices implemented centrally determined policies, and a large network of public
health clinics (puskesmas) was set up to deliver primary outpatient care. In 2001,
Indonesia’s health sector decentralized, following far-reaching reforms that involved
fiscal, administrative and political decentralization. Bossert (1998) developed a
classification method to characterize the extent of centralization in the health sector.
Table 1 shows where Indonesia stands in this classification. Districts have the legal
responsibility to provide basic health care. They are free to set user fees for public health
services (to be used as a revenue stream for local government operations) and there are
no rules or guidelines for allocating resources and carrying out particular programs.
Districts are not required to justify local spending to the central government based on
outputs or pre-defined objectives. Instead, district governments are accountable to district
parliaments. Indonesia’s health care system also retains important centralist features. The
central government sets employment conditions for civil servants, including those
working in public health service providers financed by district governments. It also
finances and runs social safety net programs for the poor, such as targeted price subsidies
for public care. Total health spending is split almost evenly between the
central/provincial level on one hand and the district level on the other hand; in 2005, they
accounted for 48 percent and 52 percent of public health expenditures respectively
(World Bank 2008).
In spite of their high share of expenditures, districts remain highly dependent on
the central government for their revenues, 90 percent of which they receive as transfers
from the center (World Bank 2008). The largest transfer, 56 percent of total revenues, is
the general allocation grant (Dana Alokasi Umum – DAU), which is a formula based
6
untied grant (Hofman et al. 2006). The other main transfers are shared tax revenues - 11
percent of total revenues - and shared non-tax revenues - 12 percent of total revenues.
The former consists largely of property and income taxes that are administered by the
central government and transferred back to the districts. The shared non-tax revenue is
largely a natural-resource revenue that is distributed back to the districts (World Bank
2007). Finally, there is the specific allocation grant (Dana Alokasi Khusus – DAK), a tied
resource whose use is determined centrally but which only accounts for a modest share of
district revenues (3 percent in 2005). Districts own revenues are non-negligible and have
been increasing as a share of total district revenues from 10 to 16 percent from 2001 to
2004 (World Bank 2007), but they are unequally distributed.
The period just after decentralization saw several revisions of the rules that
determined district budgets (World Bank 2003). While the DAU was supposed to be
formula driven based on the fiscal gap of the district, a ‘hold harmless’ clause
predetermined 80 percent of the DAU in the first year, and 50 percent in the year
thereafter. The motivation for this clause was to warrant that public services would not
suffer disruptions due to sudden budgetary shortages with the onset of decentralization.
Also the fiscal gap formula underwent two revisions over the period of investigation
(Hofman et al. 2006). Coupled with a rising oil price affecting the shared non tax
revenues, the district resource transfers were in considerable flux during this period.
Overall public resources for health increased considerably between 2001 and
2004 (Table 2). Total health expenditures increased on average 23 percent on a year to
year basis. For comparison, the average inflation rate equalled 9 percent over this period
and the average increase in nominal total public expenditures of all levels of government
7
10 percent.1 Indonesia is no anomaly in this respect; other countries that decentralized
also increased spending in the public sector (Granado 2005). Both local and central
governments contributed to rising health expenditures.2 The elasticity’s reported in this
paper thus reflect mostly the impact of increases in public spending, which may differ
from the ones resulting from downward adjustments (Lago-Penas 2008).
The empirical analysis in this paper draws on two main data sources. The first is
Indonesia’s national household survey, Susenas, fielded every year and representative at
the district level. It contains information on household socio-economic characteristics,
health services utilization, and private expenditures, including on health. Regarding
health care utilization, the Susenas survey records for each household member whether
they have sought outpatient care in the last month, which health care providers have been
visited and the number of visits per provider. The second source, compiled by the
Ministry of Finance, contains detailed records of local government revenues and
expenditures for the post decentralization years; both routine and development
expenditures can be broken down by sector, including health. Routine expenditures
consist of salaries, subsidies and operational costs of providing health services at public
facilities. Development expenditures refer to financing of development projects, such as
constructing or upgrading of health facilities and training. This may include capital
1
Source: IMF 2009 World Economic Outlook and World Bank (2008), respectively.
Note that central government spending saw a decline in 2002, as the process of decentralizing
responsibilities and budgets to districts was not immediate, but took some time to materialize. Hence, some
budgets transferred from the central government to district governments after 2001, which is reflected by
the decline in central spending in 2002. By 2003 central spending was rising again due to central ministries
lobbying for regaining part of their budget lost under decentralization during a period of increased
government resources as a result of a rising oil price.
2
8
spending, but also operation, maintenance and personal associated with a project.3
However, the data do not allow a facility level stratification.
We combined the two data sources to construct a district level panel. Since the
household survey data are collected around February, while the fiscal data reflect
expenditures during the calendar year, the effects of changes in public spending are
observed in the Susenas of the subsequent year. Therefore we constructed a panel that
contains district revenues and spending data of 2001 to 2004, linked with the Susenas
data for 2002 to 2005.
During the 4 years analyzed, new districts emerged as a result of district splits. In
such cases, we aggregated the data from the split districts, and assigned those to the
original district definition.4 We applied the 2000, pre-decentralization, district definition
frame, which comprised 305 districts. Unfortunately some of these districts had to be
dropped from the sample, for several reasons. First, the capital Jakarta comprises 6
districts but is treated as one observation, since the budget data is consolidated for the
larger metropolis. Second, over the period under investigation, Indonesia faced a number
of local conflicts that made it unsafe for surveyors to collect information. Only those
districts for which we have complete Susenas data are included in the analysis. Third,
budget data is not available for all districts, and those with missing entries were therefore
dropped from the analysis. Finally, two provinces, Aceh and Papua, are excluded since
both have been granted a special autonomy status in 2001 and their budgets are not
included in the dataset compiled by the Ministry of Finance. The balanced panel therefore
3
See World Bank (2008) for a detailed description.
Districts splits followed almost entirely sub-district boundaries within the relevant district, and did not
affect borders with neighboring districts. See Fitrani, Hofman and Kaiser (2005) for a more complete
account of this process. Statistics Indonesia maintains a full list of district codes over time (see
http://www.bps.go.id/mstkab/mfkab_03_09.pdf).
4
9
contains data from 207 districts, which represent 70 percent of total public health
spending by districts at the time of decentralization and include 40 districts that split after
2001.5
Table 3 shows descriptive statistics for the balanced panel of districts6, including
(i) per capita district revenues by source, and health spending for budget years 2001 to
2004, and (ii) average district utilization rates and OOP health spending by households in
the month prior to the survey in 2002 to 20057.
Both district revenues and public health spending increased strongly during the
first four years of decentralization in Indonesia. Total per capita district revenues
increased from 415,987 Rupiah in 2001 to 563,934 Rupiah in 2005. The bulk of district
income comes from DAU allocations, but its share decreased from 75 percent in 2001 to
66 percent in 2004. This is mainly due to increases in shared non tax revenue and DAK
allocations. Public health spending by districts also increased per capita, from 26,057
Rupiah in 2001 to 41,959 Rupiah in 2004. This change is driven by development health
spending, its share increasing from 22 to 42 percent, respectively.
Average utilization of public outpatient care in districts increased from 0.073 out
patient visits per person per month (vppm) in 2002 to 0.094 vppm in 2005, with a slight
dip in 2004. This trend is in contrast to private health care utilization, which decreased
slightly from 0.079 vppm in 2002 to 0.072 vppm in 2005. We observe somewhat similar
patterns for the poorest quarter of the Indonesian population. District average for
5
Throughout the paper we use the balanced panel for the analysis, mainly because the descriptive trends in
public health spending and health care utilization reflect the same sample of districts as the estimation
results. The choice of sample does not affect the analysis itself, as the results are very similar. The results
for the unbalanced panel are reported in the supplemental appendix.
6
District revenue, public spending and OOP health payments are reflected in 2001 constant prices. Rupiah
– USD exchange rate for 2001 is 10,246 (IMF article IV consultation 2004).
7
Note that utilization and OOP spending do not reflect national averages, but the average of the district
averages (i.e. the observations for our balanced panel).
10
utilization of public health care by the poorest quartile increased from 0.074 vppm in
2002 to 0.083 vppm in 2005, most of which occurred after 2003. Utilization of private
care decreased from 0.053 vppm in 2002 to 0.046 vppm in 2005. This gradual move from
private to – largely subsidised – public utilization is reflected in average per capita OOP
health spending by the poorest, which decreased from 3,339 Rupiah per capita per month
to 2,386 Rupiah in 2005.
3. Determinants of decentralized public health spending
In recent years, many developing countries have introduced decentralization policies,
which to varying degrees delegate the provision of local services (including key health
services) to sub-national governments. Given the significance of this trend and of the new
responsibilities vested in local administrations, it is important to understand how subnational government revenues – and their composition – translate into health spending
and this in turn translates into benefits for their populations. In this context, the
Indonesian decentralization experience provides an interesting case study.
Public spending on health is closely correlated with overall levels of district
government revenues (Figure 1). Comparing the pattern of 2004 with that of 2001, the
slope is steeper and the fit of the curve improves, indicating that districts are converging
towards a common spending pattern. This is also reflected in simple cross section
regressions, as the gradient increases from 0.91 to 0.99 and the R-squared from 0.58 to
0.70. When we compare the changes in district ranking by revenue with changes in
ranking by health spending, the Spearman rank correlation coefficient between district
per capita revenue in 2001 and 2004 is 0.706, while for district health spending it is 0.674,
11
indicating that district revenue rank is more stable over time than spending rank.
Although this cannot be interpreted as a causal effect, it does suggest that spending is
adjusting to revenues, rather than the other way around.
To investigate the causal relationship between local revenues and public health
spending patterns, model (1) relates district per capita public health spending (Hit) to
district per capita revenues (Rit), time variant district observable variables, time invariant
district unobservable variables, a time trend and an error term
6
log H it  c   log Rit    r s rit   X it   i   t   it
(1)
r 2
where i denotes the district and t the year. Imposing this log-log specification, we
interpret β as the elasticity of health spending with respect to revenues at the district level.
We also investigate whether the source of revenue matters for health spending, by
including the share of each main revenue source, srt in the regression, excluding the share
of DAU funding. The case of perfect fungibility of district revenue sources would imply
the testable hypothesis that γr = 0, for all revenue sources r. The set of control variables
Xit includes demographic variables (average age, household size, and percentage of
female population), the fraction of the population living in a rural area, and region fixed
effects8. Time dummy variable, δt, picks up aggregate time shocks, while αi is modelled
as a district fixed or random effect. For the fixed effect models, the region fixed effects
are absorbed by the district fixed effects. Since the panel is based on the district definition
of 1998, we also include a variable that tracks the number and timing of district splits.
Our aim is to identify the causal effect of revenues on health spending. However,
district unobservable factors, such as the number of civil servants employed at the time of
8
We define 5 regions: (i) Java and Bali, (ii) Sumatra, (iii) Sulawesi, (iv) Kalimantan and (v) Other Islands.
12
decentralization, could influence both revenues and health spending and if ignored could
lead to false inferences. By treating αi as a district fixed effect, we correct for such time
invariant district specific omitted variables. Nevertheless, this fixed effects approach
leaves a potential source of bias through endogeneity in changes in revenues and
spending over time. But the scope for time variant confounding effects is small, as all
major sources of district revenue are determined exogenously with respect to public
health spending. The only revenue source that is potentially susceptible to endogeneity is
own revenues, for example if increased public spending would be used to reduce user
fees. We therefore estimated equation (1) excluding own revenues from total revenues. If
endogenous own revenues are indeed driving part of the correlation between spending
and revenues, then we would expect the coefficients to be smaller when own revenues are
excluded. We find the estimates to be robust to this sensitivity test, suggestion little scope
for any remaining bias.9
Table 4 summarises the estimates of the elasticity of per capita district health
expenditures with respect to per capita district revenue. Equation (1) is estimated first
without the interaction terms, and separately for total public health spending, routine
spending and development spending. A Hausman test rejects the random effects model in
favour of the fixed effects model, although the elasticity’s are fairly robust to the choice
of specification. The elasticity of total health spending with respect to revenue is slightly
below one, at 0.88. Development spending is more sensitive to district revenue than
routine spending: a one percent increase in revenue is associated with a 1.12 percent
increase in development health spending whereas routine expenditure increases by only
9
These results are not shown here, but are reported in a supplemental appendix.
13
0.83 percent. We conclude that the share of development health expenditures increases as
districts have access to more resources.
We examined the relationship between categories of expenditure and sources of
revenue in Table 5, by including the share of revenue source srt in the specification.
Economic theory predicts that, money being fungible, the revenue sources should not
affect expenditure allocation decisions. Even for earmarked grants, such as the DAK, an
increase should not necessarily translate in an equal increase in the associated sector
spending, as government can adjust the allocations to other budget lines (Dye & McGuire
1992).
However, the results suggest that the source of funding does matter for public
health spending. A higher share of own revenues increases routine spending on health.
Possibly local governments feel that the long term commitments (hiring staff) associated
with routine spending are more securely covered if there are own revenues. A one percent
increase in the share of own revenues increases health spending by 2 percent. For
development spending we find a positive effect of the share of DAK spending, which is
not unexpected. In the period under investigation, the sectoral coverage of the DAK was
limited to education, health, local infrastructure (roads and irrigation), and government
office buildings for newly created local governments (World Bank 2007, p 123). The
results suggest that, indeed, these tied grants resulted in a net increase in health spending.
Variation in public spending is driven mainly by DAU transfers and local tax
revenue. We find higher public health spending for districts with relatively large DAU
revenue vis-à-vis shared tax and not tax revenues. The negative effect for shared non tax
revenue is interesting, as it suggests that differences in natural resource endowment do
14
not lead to divergence in public health spending. This implies that central government
can have considerable indirect influence on local health budgets, through its decentralized
fiscal instruments.
Districts splits tend to reduce public health spending. A district that splits into two
districts over the period of investigation – and is treated as one observation throughout –
has on average 6 percent lower public per capita health expenditures after the split. It is
not possible to separately identify the effect of a split on routine and development
spending, as both elasticity’s are not statistically significant. Possibly, the cost of the
district splits, such as from establishing new offices and hiring new civil servants, put a
pressure on the regular budgets.10 The health sector may have been prone to cuts as
Governments of split districts will be less inclined to consider the benefits of public
health spending that benefit the other district it was formerly part of, such as from
hospitals and communicable disease interventions, which both can accrue benefits to
residents of the other district.
4. Marginal benefit incidence of decentralized public health spending
4.1. Marginal benefit incidence analysis
To analyze the marginal benefit incidence of public health spending we propose a method
that takes account of changes in utilization rates that may result from changes in spending,
and apply this in the Indonesian setting. Following standard practice we write the benefit
incidence of public health spending on outpatient care for quartile q as the unit cost times
10
Combining all sectors, DSF (2007) finds that district splits induced recruitment of new civil servants, as
many existing civil servants were reluctant to move to the split off district. Whereas districts that did not
split over the period 2001-2005 reduced their civil servant force by 1 percent, it increased by 6.4 percent for
districts that split over this period (DSF 2007, table A5.4).
15
the utilization rate observed in the quartile.11 This setup is similar to traditional static
benefit incidence analysis, except that we let the utilization rate be a function of public
health spending
Bq H , X  
H o (H )
Dq H o H , X 
DH o H , X 
(2)
where H o is the per capita public health spending on outpatient care, which is a function
of total public health spending H; D is the per capita outpatient care utilization of public
services, which is a function of public health spending on outpatient care Ho and a set of
other determinants of health care demand, X. In the standard set up, Ho/D reflects the unit
cost (or average unit subsidy) of an out patient visit, and Bq is interpreted as the per capita
subsidy for utilization of outpatient care allocated to quartile q.
The marginal benefit incidence of public health spending for quartile q is then
defined as
Bq
H

H o Dq Dq H o H o D H o H o Dq


H D H o H D H o H D 2

Dq  H o Dq H o H o D H o H o 




D  H H o H Dq H o H D 

Dq  H o Dq H H o D H H o 




D  H
H Dq H H D H 
or:


B q  d q H o'  ho  q   
'
(3)
where  and q denote the elasticity of outpatient care utilization with respect to public
health spending for the full population and quartile q, respectively; and dq is the initial
11
See Demery (2003) for a general discussion of average benefit incidence analysis.
16
utilization share of quartile q, and ho is the share of the health budget devoted to
outpatient care.
Unfortunately, we do not have data on public spending on outpatient care
provided in the public sector so we cannot estimate H o' and ho. If public spending on
outpatient care is a linear function of total health care spending, H o  H , then (3)
reduces to
Bq'   d q 1   q   
(4)
If Ho would be observed directly then α drops out altogether. The marginal benefit
incidence of quartile q thus depends on the initial share of the quartile q in total
utilization, and the demand response of the quartile to public health spending, relative to
the demand response of the entire population.
This formula encompasses the two cases distinguished in Younger (2003) of a
marginal benefit incidence of spending which only affects existing users (the average
benefit incidence), and the marginal benefit incidence which is used entirely to finance an
expansion of services. In the former, the demand response is assumed to be equal across
the population, i.e. θq = θ, and the benefit incidence is entirely determined by existing
differences in utilization rates between quartiles dq. In the latter, the unit cost remains
constant, implying that θ = 1 and the benefits of additional public spending are absorbed
by increases in demand. In that case the marginal benefit incidence equals

D H  Dq  Dq H
Dq  Dq H

 1
1 

1 
D  H Dq H D  D  H Dq 
H
H Dq D H Dq Dq



D H
H D D
D
Bq

17
which is the change in demand of quartile q relative to that of the entire group. This is
what Lanjouw and Ravallion (1999) estimate when they relate the average enrolment in a
region to the income group specific enrolment using data from 62 regions in India’s
national sample survey. Hence, to interpret the Lanjouw and Ravallion marginal benefit
incidence results as marginal effects of public spending, one has to assume constant unit
costs.
4.2. Demand responses to district public health spending
To calculate the marginal benefit incidence of decentralized public health spending for
Indonesia we need estimates of the national and quartile specific elasticity’s of demand
with respect to decentralized health spending. We model the utilization rate of public
services as a log-linear function of one-year lagged per capita district health spending and
a set of control variables:
u it  c   log( H it 1 )   dk d it 1  X it   i   t   it
(5)
Where uit is the number of visit per person per month to a public provider. Recall that uit
is recorded each year round February and Hit-1 refers to the preceding calendar year. We
present the effect on utilization both in levels and logarithms, where the latter is used for
the marginal benefit incidence analysis. The control variables Xit are the same as in
equation (1). We investigate the differential effect of development and routine spending,
by including the lagged share of development spending in overall district health spending,
dit-1. The equation above is also used to investigate crowding out effects, by estimating
the impact of public health spending on the utilization of private services and on OOP
health spending.
18
Again we need to consider possible endogeneity bias that may result from
unobserved district specific effects, omitted variables related to local welfare and
economic activity that drive tax revenues and health care demand, or even direct reverse
causality if increased utilization of public care drives up district health budgets. We
address these sources of bias by means of the combination of district fixed effects and
lagged spending. Time invariant district effects that affect both health spending and
utilization are controlled for by including district fixed effects. Dynamic effects, such as a
sudden increase in utilization resulting in a sudden increase in health expenditures, are
corrected to a large extent by using the previous year’s budget as the explanatory variable
for this year’s utilization. Nevertheless, confounding time variant unobservables could
still frustrate identification through serial correlation in vit. We therefore test for
endogeneity using a Durbin-Wu-Hausman test, instrumenting Hit-1 with the shares of
different revenues from central government (srt-1): DAU transfers, shared non tax revenue
and DAK transfers. These seem suitable instruments as the previous section has shown
these to be determinants of public health spending, while there is no reason to expect
correlation of lagged revenue source shares with current health care utilization. The
instruments provide sufficient support for identification as they are jointly significant at
the 5 percent level and the validity of the exclusion restriction is supported by a Sargan
test.12 Finally the Durbin-Wu-Hausman test results show no evidence of endogeneity of
Hit-1 with respect to uit.13 We therefore choose the fixed effects specification for the
remainder of the analysis. We also estimated a random effects model, but Hausman tests
12
The Sargan Chi-squared test statistics vary between 0.061 to 2.406, with a critical value of 4.61 at 10
percent level and 2 degrees of freedom.
13
The results are not shown here, but are reported in a supplementary appendix.
19
rejected this in favour of the fixed effects specification for the effect of public spending
on public utilization and OOP health payments.14
The results of the fixed effect regressions are summarized in Table 6. Panel A
presenting the spending coefficients for a specification that excludes the share of
development spending, while panel B includes this share. The results suggest that public
spending indeed affects the overall level of health care utilization. A one percent increase
in district public health spending leads to an increase of 0.016 outpatient visits to a health
care provider per person per month. This result is mainly due to the positive effect on
public utilization, which increases by 0.011 vppm. The effect on private utilization is
positive but imprecise, but appears statistically significant when the share of development
spending is included. This contradicts the crowding out hypothesis, which would imply a
negative sign. The positive effect could be explained by the fact that many physicians in
public health centres operate a private practice on the side, often referring public care
patients to their private practice. Hence, increased public utilization driven by increased
routine budgets for public health clinics appears to have positive spillovers for the private
sector. An increase in the share of development spending, on the other hand, is associated
with a decrease in private care utilization, probably due to the specific nature of
development spending. There seems to be no differential effect of development spending
for public health care utilization.
The increase in public sector utilization associated with additional public
spending, and the absence of substitution effects, does not lead to increased household
health expenditures. Rather, we find negative but statistically insignificant effect in the
data. This would suggest that either increased local public health budgets have been
14
For detailed estimation results we refer again to the supplemental appendix.
20
partly used to reduce the direct costs of public care for patients through reduction of user
fees, that the mix of services has changed through an increase in routine low-cost services,
or that prices in the private sector have been cut in response to public sector investments.
We next investigate the distributional effects of public health spending on
utilization, by taking the utilization rate of different per capita expenditure quartiles as
outcome variable in equation (5)15. The fixed effect results are given in Table 7.
Additional district health spending increases health care utilization mainly for the poorest
half of the population. A one percent increase in public spending increases the utilization
rate by 0.014 vppm for the poorest quartile and 0.020 vppm for the second quartile. This
mainly occurs at public centers, with no differential effect between routine and
development spending. We find no effect of public spending on OOP health spending by
households. The coefficients are negative, but not statistically significant.
The marginal benefit incidence estimates can be directly calculated from the
quartile specific demand elasticity relative to the demand elasticity of the entire
population and the average utilization rates per quartile. The elasticity’s are given in
Table 9, corresponding to the patterns found in Table 7. Elasticity’s are estimated
following equation (5), with log(uit) as dependent variable.
Demand for public outpatient care utilization is relatively inelastic with respect to
public spending. A one percent increase in district health spending is associated with a
statistically significant increase of 0.09 percent. This implies that the unit cost increases
with increased public spending; that is, the average benefit of public outpatient care in
15
The quartiles reflect a household’s rank in the distribution of per capita expenditure across the full
population (and not the district population). Since we only use districts from the balanced panel for which
survey data contains at least 50 observation per quartile, for the quartile analysis we lose 8 districts from
the balanced panel, reducing it to 199 districts, with 796 observations from 2002-2005.
21
terms of public resource increases, for example through increased quality of public care
or reduced user fees. The demand elasticity is higher for the poorest quartiles, at 0.14,
while the hypothesis of perfectly inelastic demand is not rejected for two richest quartiles.
We find no statistically significant estimates for private outpatient care or OOP.
Table 10 presents the estimated marginal benefit incidence following (4). Because
we are primarily interested in relative benefits across quartiles, we omit α for
convenience, since this is a constant across quartiles. The distributional pattern in demand
elasticity’s yields a pro-poor marginal behavioural response (1 + q – ), while the initial
utilization share does not. The combined result suggests that the benefits from public
spending are spread fairly equally across the population. The positive demand response
of the poor to public spending results in a marginal benefit incidence that is more propoor than the average benefit incidence. But despite this pro-poor behavioural response,
the initial utilization shares dominate the distribution of a marginal increase in public
spending.
4.3. Sensitivity analysis
There seems to be little scope for endogeneity bias in the empirical results in the previous
section. However, a potential source of bias could be due to spurious correlation between
local public health budgets and utilization through omitted variables. Two variables raise
particular concerns. The first is economic welfare of the district population, as this is
likely to drive demand for health care as well as local revenue raising capacity. Our
analysis of health spending patterns finds that district own revenues affect routine health
spending, indicating scope for omitted variable bias through district variation in changes
22
in economic welfare. A second possible omitted variable is the social safety net health
card scheme, which was introduced during the economic crisis in 1998, and is managed
centrally by the ministry of health. This health card was targeted to the poorest
households, entitling recipient to price subsidies at public health care facilities. Public
health care providers were compensated through lump sum transfers. If the health card
program crowds out local spending, this may cause an upward bias in the observed public
spending effects, since the health cards are targeted towards poorer districts with
relatively smaller health budgets and lower utilization rates.
Information on both variables is available in Susenas. Per capita household
consumption serves as indicator of economic welfare, and questions on health card
receipt are included in the questionnaire for all years. However, both variables are
potentially endogenous in the utilization regressions, while obvious instrumental
variables are not at hand. Therefore, we opt for a sensitivity analysis, evaluating the
robustness of the results to including these variables. If indeed there was an omitted
variable bias, we would expect the health spending coefficients to be sensitive to
specification.
The results of this sensitivity analysis are summarized in Table 8, showing that
our results are robust to including health card coverage or per capita consumption. Panel
A repeats the estimates of the main specification (Table 6). Panel B shows the
specification that includes per capita consumption, and the specification for panel C
includes health card coverage. The health spending coefficients are very similar to the
specification of Table 6, and are well within one standard error.
23
5. Conclusion
This paper proposes a measure of marginal benefit incidence that takes into account
behavioural responses to changes in public spending. This measure is presented with a
general framework of benefit incidence analysis that also encompasses existing measures
of average and marginal benefit incidence. We show that these measures differ only the
assumptions regarding the causal effects of changes in public spending on utilization, and
the associated behavioural responses.
The empirical application analyses spending patterns and utilization of health
services during the first years of decentralization in Indonesia. We looked in particular at
the relationship between local revenues and health spending categories (development and
routine), and their effect on health care utilization. Indonesia’s partial decentralization
brought a large proportion of the health budget under control of district authorities and
induced a massive redistribution of resources across districts. Local government health
spending increased sharply with decentralization, reflecting the transfer of responsibility
and authority from the centre to the districts. Health care utilization increased from 2001
to 2004, in particular in the public sector.
Public health spending appears close to unit elastic (around 0.9) with respect to
local revenues. District revenues do not seem completely fungible with respect to health
spending, as the source of financing matters. Spending is mostly driven by DAU transfers,
while inequalities in local revenues sources play an important role for routine health
spending, suggesting divergence in spending due to differences in local endowments.
Transfers from the central government (DAU and DAK) are important source of
financing for development spending, while we do not find that resource rich districts
24
allocate more funds to development health spending. These results suggest that the
central government retains a strong role in influencing local health budgeting and
addressing horizontal inequalities in district public health spending. Tied transfers (DAK)
seem particular effective for regulating development spending, as these do not appear to
crowd out local health spending.
Local public health spending seems to increase overall health care utilization, in
particular for the poorest half of the population, without affecting OOP health payments
or utilization of the private sector, once we control for confounding factors. This result
provides no support for the hypothesis that public spending crowds out private spending
in health. Increased routine spending seems to have positive effects on both public and
private health care utilization.
Translating our results to a marginal benefit incidence analysis, taking into
account behavioural responses to changes in publics spending, suggests that increased
public health spending improves targeting of public funds to the poor. At the margin,
increased local public health spending leads to net public resource transfers from the
richest to the poorest quarter of the population, as it increases both public health care
utilization by the poor and the average benefit of public funds through using these
services. However, the initial utilization shares still dominate the marginal benefits, such
that the bulk of the benefits accrue to the two middle quartiles. Hence, for effective
targeting of public resources to the poor, increased public health spending induced
though reallocation of central resources could be complemented with more directly
targeted demand side interventions, for example price subsidies for the poor or social
health insurance.
25
References
Bhalotra, S., 2007. Spending to save? State health expenditure and infant mortality in
India. Health Economics 16, 911-928.
Bossert, T., 1998. Analyzing the decentralization of health systems in developing
countries: decision space, innovation, and performance. Social Science and
Medicine 47, 513-27.
Demery, L., 2003. Analyzing the Incidence of Public Spending. In: Bourguignon F &
Pereira da Silva LA (eds.) Evaluating the Poverty and Distributional Impact of
Economic Policies (Techniques and Tools). World Bank and Oxford University
Press, Washington DC.
DSF, 2007. Costs and Benefits of New Region Creation in Indonesia. DSE, Jakarta.
Dye, R.F., McGuire, T.J., 1992. The Effect of Earmarked Revenues On the Level and
Composition of Expenditures. Public Finance Review 20, 543-556.
Filmer, D., Pritchett, L., 1999. The impact of public spending on health: does money
matter? Social Science & Medicine 49, 1309-1323.
Fitrani, F., Hofman, B., Kaiser, K., 2005. Unity in diversity? The creation of new local
governments in de decentralizing Indonesia. Bulletin of Indonesian Economic
Studies 41, 57-79.
Gerdtham, U.-G., Jönsson, B., 2000. International comparisons of health expenditure:
Theory, data and econometric analysis. In: Culyer AJ & Newhouse JP (eds.)
Handbook of Health Economics, 11-53.
Granado, F.J.A.d., Martinez-Vazquez, J., McNab, R., 2005. Fiscal Decentralization and
The Functional Composition of Public Expenditures. International Studies
26
Program Working 54, Andrew Young School of Policy Studies, Georgia State
University.
Gupta, S., Verhoeven, M., Tiongson, E.R., 2003. Public spending on health care and the
poor. Health Economics 12, 685-696.
Hofman, B., Kadjatmiko, Kaiser, K., Sjahrir, B.S., 2006. Evaluating Fiscal Equalization
in Indonesia. World Bank Policy Research Working Paper 3911, Washington DC.
Lago-Penas, S., 2008. Local Governments' Asymmetric Reactions to Grants: Causes and
Consequences. Public Finance Review 36, 219-242.
Lanjouw, P., Ravallion, M., 1999. Benefit Incidence, Public Spending Reforms, and the
Timing of Program Capture. World Bank Economic Review 13, 257-273.
McGuire, J.W., 2006. Basic health care provision and under-5 mortality: A cross-national
study of developing countries. World Development 34, 405-425.
McIntyre, D., Ataguba, J.E., 2010. How to do (or not to do)…a benefit incidence analysis.
Forthcoming in Health Policy and Planning.
Shah, A., 2005. Public Expenditure Analysis. World Bank, Washington DC.
van de Walle, D., 1998. Assessing the welfare impacts of public spending. World
Development 26, 365-379.
van der Gaag, J., Štimac, V., 2008. Towards a New Paradigm for Health Sector
Development.
Working
Paper,
Amsterdam
Institute
for
International
Development.
World Bank, 2003. Decentralizing Indonesia. World Bank, Jakarta.
World Bank, 2004. World Development Report: Making Services Work for Poor People.
World Bank, Washington DC.
27
World Bank, 2007. Spending for Development: Making the most of Indonesia's new
oppurtunities. In: Indonesia Public Expenditure Review 2007. World Bank,
Jakarta.
World Bank, 2008. Investing in Indonesia’s Health: Challenges and Opportunities for
Future Public Spending. World Bank, Jakarta.
Younger, S.D., 2003. Benefits on the Margin: Observations on Marginal Benefit
Incidence. World Bank Economic Review 17, 89-106.
28
Tables
Table 1 Decision space of Indonesia's decentralization in health
Function
Finance
Indicator
Narrow (centralized)
Intergovernmental transfers as % of total
local health spending
% of local spending that is explicitly
Allocation of expenditures earmarked by higher authorities
Range of prices local authorities are
allowed to choose
Fees
Contracts
Number of models allowed
Service organization
X
Source of revenue
Hospital autonomy
Insurance plans
Payment mechanism
Required programs
Human resources
Salaries
Contract
civil service
Access rules
Targeting
Governance rules
facility boards
district offices
Community participation
Wide (highly decentralized)
DAU2001/8614.3 billion
Many vertical programs remain.
Central/local health exp =72%
No rules
No rules
A number of vertical hospitals
remain, Many hospital doctors
financed out of central budget
Choice of range of autonomy for
hospitals
Health card for the poor,
social insurance for civil
Choice of how to design insurance plans servants remain central
choice of how providers will be paid
(incentives and non-salaried)
specificity of normal for local programs
X
choice of salary range
Centrally decided
contracting non-permanent staff
Hiring and firing permanent staff
X
defining priority populations
X
Size and composition of boards
Size and composition of local offices
Size, number, composition and role of
community participation
Range of choice
Moderate
Freedom
Functions are specified
Difficult under civil servant rules
Central guidelines local
implementation
Freedom
Old system still in place
Wide variation
29
Table 2 Public spending on health by level of Government (Nominal, in billion Rupiah)
Year
Central Government
Provincial spending
District Government
Total public health spending
2001
3,119
1,745
4,387
9,251
2002
2,907
2,372
5,725
11,004
2003
5,752
2,821
7,473
16,046
2004
5,595
3,000
8,108
16,703
Source: World Bank (2008).
Table 3 Descriptive statistics for balanced panel (district averages)
2001
2002
2003
2004
415,987
316,289
31,218
20,258
5,296
27,429
15,497
26,057
5,611
20,446
508,375
369,773
39,180
36,496
2,728
38,724
21,472
32,329
7,735
24,594
557,883
377,826
44,927
37,821
17,559
41,112
38,637
39,033
15,830
23,203
563,934
371,202
50,786
40,204
19,180
41,767
40,795
41,959
17,514
24,445
2002
2003
2004
2005
Household utilization and spending on health
Full population
Outpatient utilization rate – public
Outpatient utilization rate – private
Per capita OOP health spending
0.0727
0.0793
8,368
0.0874
0.0786
6,526
0.0827
0.0753
6,664
0.0944
0.0724
7,242
Poorest quartile
Outpatient utilization rate – public
Outpatient utilization rate – private
Per capita OOP health spending
N
0.0736
0.0530
3,339
207
0.0716
0.0464
2,668
207
0.0892
0.0476
2,498
207
0.0825
0.0459
2,386
207
Per capita district revenues and spending
Total revenues
DAU revenues
Shared tax revenue
Shared non tax revenue
DAK revenue
Own revenues
Revenues from other sources
Total public spending on health
Development spending on health
Routine spending on health
Note: District revenue, public spending and OOP health payments in 2001 constant prices Rupiah.
Outpatient utilization rates reflect the average number of outpatient visits in the last month. OOP health
spending reflects per capita household OOP health payments in the last month.
Source: Revenue from fiscal data from ministry of finance, utilization and OOP spending from Susenas
household survey.
30
Table 4 Elasticity of per capita district public health spending w.r.t per capita district
revenue
Routine spending
Development health spending
Total health spending
Random effects
0.8431**
[0.0465]
1.2371**
[0.0723]
0.9468**
[0.0344]
Fixed effects
0.8284**
[0.0657]
1.1192**
[0.1376]
0.8789**
[0.0449]
Hausman test
0.0022
0. 5717
0. 0133
Statistical significance: + at 10 percent, * at 5 percent, and ** at one percent level.
Note: balanced panel of 207 districts, 828 observations, 2001-2004. All models control for demographic
characteristics (average age, household size and percentage female population), percentage rural population,
district splits, region fixed effects and aggregate time shocks. The Hausman test reports p-values for
rejecting the hypothesis that the difference in coefficients is not systematic. See the supplemental appendix
for detailed estimates. Standard errors in brackets.
Table 5 Elasticity of per capita district public health spending w.r.t per capita district
revenue source (district fixed effects)
Routine
Development
Total
Total district revenue
0.8653**
1.0534**
0.8810**
[0.0663]
[0.1386]
[0.0453]
Revenue shares
Own revenue
2.0303**
1.2494
1.4428**
[0.7373]
[1.5422]
[0.5037]
Shared tax revenue
0.3574
-3.3658**
-0.9871**
[0.5445]
[1.1390]
[0.3720]
Shared non tax revenue
-0.8711
-0.1989
-0.7042+
[0.6053]
[1.2662]
[0.4135]
DAK revenue
-1.1145
3.0818*
0.1264
[0.7247]
[1.5158]
[0.4951]
Revenue from other sources
-0.5009
0.4164
-0.2940
[0.3551]
[0.7427]
[0.2426]
District splits
-0.0218
-0.0027
-0.0612*
[0.0416]
[0.0871]
[0.0284]
Statistical significance: + at 10 percent, * at 5 percent, and ** at one percent level.
Note: balanced panel of 207 districts, 828 observations, 2001-2004. All models control for demographic
characteristics (average age, household size and percentage female population), percentage rural population,
district splits, aggregate time shocks and district fixed effects. See the supplemental appendix for detailed
estimates. Standard errors in brackets.
31
Table 6 Effect of (lagged) log per capita district public health spending on outpatient
health care utilization rates and household out-of-pocket health spending (district fixed
effects)
Public
Private
Total
OOP
A. Without source of health spending
District health spending
0.0114** 0.0042
0.0156** -94.42
[0.0039]
[0.0036]
[0.0060]
[459.56]
B. With source of health spending
District health spending
0.0111** 0.0059+
0.0170** 1.40
[0.0039]
[0.0036]
[0.0060]
[465.43]
Share of development spending
0.0037
-0.0234** -0.0197
-1,269.52
[0.0084]
[0.0076]
[0.0129]
[995.61]
Statistical significance: + at 10 percent, * at 5 percent, and ** at one percent level.
Note: balanced panel of 207 districts, 828 observations, 2002-2005. All models control for demographic
characteristics (average age, household size and percentage female population), percentage rural population,
district splits, aggregate time shocks and district fixed effects. A Hausman test rejects the random effects in
favour of the fixed effects specification in all cases except for public utilization. See the supplemental
appendix for detailed estimates. Standard errors in brackets.
Table 7 Effect of (lagged) log per capita district public health spending on outpatient
health care utilization rates and household out-of-pocket health spending, by per capita
expenditure quartile (district fixed effects)
Public
Private
Total
OOP
Quartile 1 (poorest)
0.0175** -0.0032
0.0143+
-65.80
[0.0065]
[0.0039]
[0.0083]
[215.74]
Quartile 2
0.0164** 0.0032
0.0197**
64.38
[0.0055]
[0.0042]
[0.0075]
[270.08]
Quartile 3
0.0063
0.0005
0.0068
-216.31
[0.0060]
[0.0050]
[0.0087]
[543.65]
Quartile 4 (richest)
-0.0055
-0.0048
-0.0104
-1,685.68
[0.0085]
[0.0090]
[0.0149]
[1,878.55]
Statistical significance: + at 10 percent, * at 5 percent, and ** at one percent level.
Note: balanced panel of 199 districts, 796 observations, 2002-2005. The number of districts differs from
Table 6 as we only use districts from the balanced panel for which the survey data contains at least 50
observations for each quartile. Model specification similar to Table 6. See the supplemental appendix for
detailed estimates. Standard errors in brackets.
32
Table 8 Sensitivity analysis of the effect of (lagged) log per capita district public health
spending on outpatient health care utilization rates and household out-of-pocket health
spending (district fixed effects)
Specification
A. Original (as in Table 6)
B. Including log per capita
household consumption
C. Including district health
card coverage
Public
0.0114**
[0.0039]
0.0116**
[0.0039]
0.0110**
[0.0039]
Private
0.0042
[0.0036]
0.0022
[0.0035]
0.0040
[0.0035]
Total
0.0156**
[0.0060]
0.0138*
[0.0060]
0.0150*
[0.0059]
OOP
-94.42
[459.56]
-478.65
[441.50]
-117.07
[458.97]
Statistical significance: + at 10 percent, * at 5 percent, and ** at 1 percent level.
Note: balanced panel of 207 districts, 828 observations, 2002-2005. All models control for demographic
characteristics (average age, household size and percentage female population), percentage rural population,
district splits, aggregate time shocks and district fixed effects. See the supplemental appendix for detailed
estimates Standard errors in brackets.
Table 9 Elasticity of outpatient health care utilization and household out-of-pocket health
spending with respect to district public health spending, by per capita expenditure quartile
(district fixed effects estimates)
Quartile 1 (poorest)
Quartile 2
Quartile 3
Quartile 4 (richest)
Total
Public
0.1441+
0.1418*
0.0821
-0.0339
0.0897+
Private
-0.0922
0.0006
-0.0028
-0.0516
0.0071
Total
0.0446
0.0838
0.0173
-0.0326
0.0583
OOP
-0.0738
0.0224
-0.0088
-0.0723
-0.0173
Statistical significance: + at 10 percent, * at 5 percent, and ** at one percent level.
Table 10 Relative marginal benefit incidence of spending on outpatient services
Quartile 1 (poorest)
Quartile 2
Quartile 3
Quartile 4 (richest)
1 + q – 
1.054
1.052
0.992
0.876
dq,2002
0.232
0.257
0.272
0.243
dq (1 + q – )
0.245
0.270
0.270
0.213
Note: results based on statistically significant estimates are reported in bold.
33
Figures
Figure 1 Correlation between district log per capita health spending and log per capita
revenues in 2001 and 2004
8
Log district health expenditure (per capita)
9
10
11
12
2001
12
12.5
13
13.5
Log total district revenue (per capita)
14
14.5
bandwidth = .8
8
Log district health expenditure (per capita)
9
10
11
12
13
2004
12
13
14
15
Log total district revenue (per capita)
16
bandwidth = .8
34
December 2010
Marginal Benefit Incidence of Public Health Spending:
Evidence from Indonesian sub-national data
Supplemental appendix
[Not intended for publication, but will be made available online]
Table I Elasticity of per capita district routine health spending w.r.t. per capita district
revenue, district fixed effects (FE) and random effects (RE) specification
Log total district revenue
(1)
RE
0.8431**
[0.0465]
(2)
FE
0.8284**
[0.0657]
(3)
RE
0.9256**
[0.0472]
(4)
FE
0.8653**
[0.0663]
-0.0158
[0.0221]
2.8323**
[0.6945]
-1.6168
[1.3792]
-0.0178
[0.0934]
-0.0335
[0.0410]
1.1069*
[0.5291]
-0.8551*
[0.3878]
-1.7780**
[0.2862]
-1.5380*
[0.6984]
-1.0021**
[0.3280]
0.0073
[0.0129]
0.0940
[0.0904]
-1.0861
[1.2229]
0.0100
[0.0598]
0.0056
[0.0381]
2.0303**
[0.7373]
0.3574
[0.5445]
-0.8711
[0.6053]
-1.1145
[0.7247]
-0.5009
[0.3551]
-0.0088
[0.0223]
2.6654**
[0.6956]
-1.8723
[1.3736]
0.0200
[0.0939]
-0.0218
[0.0416]
Revenue shares
Own revenue
Shared tax revenue
Shared non tax revenue
DAK revenue
Revenue from other sources
Average age
Rural share of population
Female share of population
Average household size
District splits
Region
Sumatra
Sulawesi
Kalimantan
Other islands
Year
2002
2003
2004
Constant
R-squared (within)
Hausman test (p-value)
0.0120
[0.0132]
0.0617
[0.0855]
-0.2583
[1.2421]
0.0313
[0.0622]
-0.0352
[0.0376]
-0.0243
[0.0862]
0.0378
[0.0975]
-0.0564
[0.1070]
-0.0414
[0.1198]
0.0936
[0.0836]
0.0151
[0.0928]
0.0651
[0.1021]
-0.0253
[0.1134]
0.0622*
[0.0274]
-0.0910**
[0.0294]
-0.0384
[0.0292]
-1.4422
[0.9205]
0.3227
0.0995**
[0.0294]
-0.0891**
[0.0326]
-0.0071
[0.0329]
-1.2044
[1.4138]
0.3428
0.0022
0.0435
[0.0286]
-0.0596+
[0.0345]
0.0134
[0.0367]
-1.8088*
[0.9110]
0.3393
0.0613+
[0.0320]
-0.0885*
[0.0389]
-0.0079
[0.0423]
-1.9298
[1.4190]
0.3609
0.0024
Statistical significance: + at 10 percent, * at 5 percent, and ** at 1 percent level.
Note: balanced panel of 207 districts, 828 observations, 2001-2004. Standard errors in brackets.
i
Table II Elasticity of per capita district development health spending w.r.t. per capita
district revenue, district fixed effects (FE) and random effects (RE) specification
Log total district revenue
(1)
RE
1.2371**
[0.0723]
(2)
FE
1.1192**
[0.1376]
(3)
RE
1.1856**
[0.0790]
(4)
FE
1.0534**
[0.1386]
-0.0632
[0.0464]
0.8789
[1.4555]
3.2317
[2.8904]
-0.1123
[0.1958]
0.0412
[0.0860]
1.4054
[0.8954]
-2.0020**
[0.6897]
0.1423
[0.4460]
2.8004*
[1.3945]
0.0044
[0.6371]
-0.1081**
[0.0206]
0.3566**
[0.1350]
6.4367**
[2.2670]
-0.0031
[0.0955]
0.0612
[0.0753]
1.2494
[1.5422]
-3.3658**
[1.1390]
-0.1989
[1.2662]
3.0818*
[1.5158]
0.4164
[0.7427]
-0.0433
[0.0466]
0.9997
[1.4551]
2.9781
[2.8733]
-0.0582
[0.1965]
-0.0027
[0.0871]
Revenue shares
Own revenue
Shared tax revenue
Shared non tax revenue
DAK revenue
Revenue from other sources
Average age
Rural share of population
Female share of population
Average household size
District splits
Region
Sumatra
Sulawesi
Kalimantan
Other islands
-0.1046**
[0.0199]
0.3832**
[0.1138]
6.4425**
[2.2533]
-0.0046
[0.0940]
0.1000
[0.0730]
-0.6737**
[0.1183]
-0.7590**
[0.1343]
-0.6413**
[0.1465]
-0.3368*
[0.1630]
-0.6337**
[0.1245]
-0.7493**
[0.1386]
-0.6098**
[0.1513]
-0.3702*
[0.1678]
Year
2002
0.0800
0.0895
0.0887
0.1043
[0.0560]
[0.0616]
[0.0578]
[0.0669]
2003
0.7605**
0.7486**
0.7236**
0.7001**
[0.0585]
[0.0684]
[0.0680]
[0.0813]
2004
0.7494**
0.7503**
0.7173**
0.7112**
[0.0580]
[0.0691]
[0.0720]
[0.0884]
Constant
-7.7589**
-5.8993*
-6.9718**
-5.6780+
[1.4660]
[2.9628]
[1.5232]
[2.9682]
R-squared (within)
0.4981
0.5009
0.5109
0.5164
Hausman test (p-value)
0.5717
0.2761
Statistical significance: + at 10 percent, * at 5 percent, and ** at 1 percent level.
Note: balanced panel of 207 districts, 828 observations, 2001-2004. Standard errors in brackets.
ii
Table III Elasticity of per capita district total health spending w.r.t. per capita district
revenue, district fixed effects (FE) and random effects (RE) specification
Log total district revenue
(1)
RE
0.9468**
[0.0344]
(2)
FE
0.8789**
[0.0449]
(3)
RE
0.9766**
[0.0349]
(4)
FE
0.8810**
[0.0453]
-0.0303*
[0.0151]
1.5398**
[0.4749]
-0.5748
[0.9432]
-0.0480
[0.0639]
-0.0569*
[0.0281]
1.0364**
[0.3894]
-1.3814**
[0.2805]
-1.0272**
[0.2214]
-0.1527
[0.4817]
-0.5319*
[0.2283]
-0.0202*
[0.0098]
0.0945
[0.0721]
-0.2950
[0.8641]
-0.0187
[0.0452]
-0.0225
[0.0265]
1.4428**
[0.5037]
-0.9871**
[0.3720]
-0.7042+
[0.4135]
0.1264
[0.4951]
-0.2940
[0.2426]
-0.0199
[0.0152]
1.4415**
[0.4752]
-0.8391
[0.9384]
-0.0092
[0.0642]
-0.0612*
[0.0284]
Revenue shares
Own revenue
Shared tax revenue
Shared non tax revenue
DAK revenue
Revenue from other sources
Average age
Rural share of population
Female share of population
Average household size
District splits
Region
Sumatra
Sulawesi
Kalimantan
Other islands
-0.0174+
[0.0099]
0.1003
[0.0687]
0.2059
[0.8735]
-0.0152
[0.0465]
-0.0350
[0.0261]
-0.1979**
[0.0682]
-0.1367+
[0.0772]
-0.2352**
[0.0848]
-0.1153
[0.0953]
-0.1108+
[0.0665]
-0.1458*
[0.0738]
-0.1406+
[0.0816]
-0.1195
[0.0908]
Year
2002
0.0551**
0.0824**
0.0479*
0.0671**
[0.0188]
[0.0201]
[0.0198]
[0.0219]
2003
0.1292**
0.1424**
0.1399**
0.1376**
[0.0204]
[0.0223]
[0.0241]
[0.0265]
2004
0.1655**
0.1948**
0.1911**
0.1947**
[0.0203]
[0.0225]
[0.0257]
[0.0289]
Constant
-1.6311*
-0.7956
-1.6033*
-1.0925
[0.6803]
[0.9668]
[0.6779]
[0.9694]
R-squared (within)
0.6542
0.6609
0.6639
0.6709
Hausman test (p-value)
0.0133
0.0083
Statistical significance: + at 10 percent, * at 5 percent, and ** at 1 percent level.
Note: balanced panel of 207 districts, 828 observations, 2001-2004. Standard errors in brackets.
iii
Table IV Effect of (lagged) per capita district public health spending on outpatient health care utilization rates and household out-ofpocket health spending (district fixed effects)
L. Log district health spending
(1)
Public
0.0114**
[0.0039]
L. Development health spending share
Average age
Rural share of population
Female share of population
Average household size
L. District splits
Year
2003
-0.0009
[0.0020]
0.0757
[0.0991]
0.0431
[0.1167]
0.0137+
[0.0078]
-0.0018
[0.0032]
(2)
Public
0.0111**
[0.0039]
0.0037
[0.0084]
-0.0009
[0.0020]
0.0783
[0.0994]
0.0407
[0.1169]
0.0135+
[0.0078]
-0.0018
[0.0032]
(3)
Private
0.0042
[0.0036]
-0.0005
[0.0018]
0.0050
[0.0906]
0.2362*
[0.1067]
0.0147*
[0.0071]
-0.0006
[0.0030]
(4)
Private
0.0059+
[0.0036]
-0.0234**
[0.0076]
-0.0007
[0.0018]
-0.0115
[0.0902]
0.2512*
[0.1060]
0.0158*
[0.0071]
-0.0005
[0.0029]
(5)
Total
0.0156**
[0.0060]
-0.0014
[0.0031]
0.0807
[0.1523]
0.2793
[0.1793]
0.0284*
[0.0120]
-0.0024
[0.0050]
(6)
Total
0.0170**
[0.0060]
-0.0197
[0.0129]
-0.0016
[0.0031]
0.0668
[0.1524]
0.2919
[0.1793]
0.0293*
[0.0120]
-0.0023
[0.0050]
(7)
OOP
-94.42
[459.56]
-184.91
[239.16]
15,364.09
[11,725.69]
21,349.14
[13,798.70]
14.45
[921.16]
156.76
[382.18]
-0.0057*
-0.0057*
-0.0096**
-0.0095**
-0.0153**
-0.0152** -1,910.32**
[0.0028]
[0.0028]
[0.0025]
[0.0025]
[0.0043]
[0.0043]
[327.74]
2004
0.0073**
0.0068*
-0.0119**
-0.0089**
-0.0046
-0.0021
-1,647.51**
[0.0027]
[0.0029]
[0.0025]
[0.0026]
[0.0041]
[0.0045]
[318.59]
2005
0.0027
0.0023
-0.0152**
-0.0125**
-0.0125**
-0.0103*
-1,002.53**
[0.0030]
[0.0031]
[0.0027]
[0.0028]
[0.0045]
[0.0048]
[350.11]
Constant
-0.1426
-0.1409
-0.1289
-0.1394
-0.2716
-0.2804
-5,187.83
[0.1185]
[0.1187]
[0.1084]
[0.1077]
[0.1821]
[0.1820]
[14,021.00]
R-squared (within)
0.1024
0.1026
0.1042
0.1177
0.0668
0.0704
0.1000
Hausman test (p-value)
0.6995
0.7139
0.0001
0.0001
0.0182
0.0234
0.1208
Statistical significance: + at 10 percent, * at 5 percent, and ** at 1 percent level.
Note: balanced panel of 207 districts, 828 observations, 2002-2005. All models control for district fixed effects. Standard errors in brackets.
(8)
OOP
1.40
[465.43]
-1,269.52
[995.61]
-194.77
[239.16]
14,466.00
[11,740.84]
22,159.88
[13,806.29]
73.18
[921.84]
167.43
[382.08]
-1,904.50**
[327.60]
-1,481.59**
[343.99]
-856.33*
[368.24]
-5,757.98
[14,020.97]
0.1024
0.1936
iv
Table V Exogeneity test of per capita district public health spending w.r.t. outpatient
utilisation and OOP health spending (district fixed effects)
(1)
First stage
L. Log district health spending
Average age
Rural share of population
Female share of population
Average household size
L. District splits
Year
2003
2004
2005
L. Share of DAU revenue
L. Share of Shared non tax revenue
L. Share of DAK revenue
-0.0267
[0.0209]
-2.7688**
[1.0239]
0.4452
[1.2064]
-0.1006
[0.0806]
-0.3381**
[0.0312]
0.2370**
[0.0278]
0.3459**
[0.0302]
0.3900**
[0.0349]
0.2894
[0.2688]
0.5974
[0.4490]
1.9467**
[0.6382]
First stage residual
Constant
12.3855**
[1.1584]
0.4563
3.4015*
(2)
Public
0.0434
[0.0303]
-0.0001
[0.0022]
0.1707
[0.1333]
0.0274
[0.1176]
0.0165*
[0.0082]
0.0088
[0.0104]
(3)
Private
-0.0340
[0.0277]
-0.0015
[0.0020]
-0.1082
[0.1218]
0.2549*
[0.1074]
0.0113
[0.0075]
-0.0132
[0.0095]
(4)
Total
0.0094
[0.0466]
-0.0016
[0.0033]
0.0626
[0.2051]
0.2823
[0.1808]
0.0278*
[0.0127]
-0.0045
[0.0161]
(5)
OOP
0.0094
[0.0466]
-0.0016
[0.0033]
0.0626
[0.2051]
0.2823
[0.1808]
0.0278*
[0.0127]
-0.0045
[0.0161]
-0.0130+
[0.0073]
-0.0043
[0.0112]
-0.0103
[0.0125]
-0.0010
[0.0067]
0.0018
[0.0102]
0.0002
[0.0114]
-0.0139
[0.0113]
-0.0024
[0.0172]
-0.0101
[0.0192]
-0.0139
[0.0113]
-0.0024
[0.0172]
-0.0101
[0.0192]
-0.0326
[0.0305]
-0.5489
[0.3990]
0.1040
0.0388
[0.0279]
0.3548
[0.3646]
0.1070
0.0062
[0.0470]
-0.1941
[0.6137]
0.0669
0.0062
[0.0470]
-0.1941
[0.6137]
0.0669
R-squared (within)
F test joint sig. IVs
Sargan Chi squared statistic
0.0611
1.2377
0.6749
2.4064
Statistical significance: + at 10 percent, * at 5 percent, and ** at 1 percent level.
Note: balanced panel of 207 districts, 828 observations, 2002-2005. All models control for district fixed
effects. Standard errors in brackets. Critical value for Sargan Chi squared test statistic is 4.61 at 10 percent
level and 2 degrees of freedom.
v
Table VI Effect of (lagged) per capita district public health spending on public outpatient
health care utilization rates, by per capita expenditure quartile (district fixed effects)
L. Log district health spending
Average age
Rural share of population
Female share of population
Average household size
L. District splits
(1)
Q1
0.0175**
[0.0065]
0.0021
[0.0033]
-0.0471
[0.1633]
0.0159
[0.1937]
0.0125
[0.0130]
0.0025
[0.0053]
(2)
Q2
0.0164**
[0.0055]
-0.0034
[0.0028]
0.1069
[0.1395]
0.1272
[0.1655]
0.0222*
[0.0111]
-0.0027
[0.0045]
(3)
Q3
0.0063
[0.0060]
-0.0036
[0.0031]
0.0865
[0.1516]
-0.0605
[0.1799]
-0.0011
[0.0121]
-0.0034
[0.0049]
(4)
Q4
-0.0055
[0.0085]
-0.0030
[0.0044]
-0.0949
[0.2144]
0.5347*
[0.2544]
0.0138
[0.0171]
-0.0062
[0.0069]
Year
2003
-0.0004
-0.0082*
-0.0087*
0.0016
[0.0046]
[0.0039]
[0.0043]
[0.0060]
2004
0.0119**
0.0038
0.0072+
0.0132*
[0.0045]
[0.0038]
[0.0041]
[0.0059]
2005
0.0038
0.0020
0.0016
0.0118+
[0.0049]
[0.0042]
[0.0045]
[0.0064]
Constant
-0.2047
-0.2225
0.1068
-0.0535
[0.1970]
[0.1683]
[0.1829]
[0.2587]
R-squared (within)
0.0651
0.0737
0.0427
0.0270
Statistical significance: + at 10 percent, * at 5 percent, and ** at 1 percent level.
Note: balanced panel 199 districts, 796 observations, 2002-2005. The number of districts differs from the
district level regressions as we only use districts from the balanced panel for which the survey data contains
at least 50 observations for each quartile. All models control for district fixed effects. Standard errors in
brackets.
vi
Table VII Effect of (lagged) per capita district public health spending on private
outpatient health care utilization rates, by per capita expenditure quartile (district fixed
effects)
L. Log district health spending
Average age
Rural share of population
Female share of population
Average household size
L. District splits
(1)
Q1
-0.0032
[0.0039]
0.0009
[0.0020]
-0.0371
[0.0992]
0.0183
[0.1177]
0.0241**
[0.0079]
-0.0010
[0.0032]
(2)
Q2
0.0032
[0.0042]
0.0021
[0.0022]
-0.0537
[0.1058]
0.2125+
[0.1255]
0.0179*
[0.0084]
-0.0037
[0.0034]
(3)
Q3
0.0005
[0.0050]
0.0019
[0.0026]
0.0518
[0.1273]
0.1719
[0.1511]
0.0236*
[0.0101]
-0.0028
[0.0041]
(4)
Q4
-0.0048
[0.0090]
-0.0019
[0.0047]
0.1618
[0.2279]
0.5153+
[0.2705]
0.0263
[0.0182]
-0.0052
[0.0074]
Year
2003
-0.0042
-0.0085**
-0.0100**
-0.0210**
[0.0028]
[0.0030]
[0.0036]
[0.0064]
2004
-0.0047+
-0.0110**
-0.0161**
-0.0183**
[0.0027]
[0.0029]
[0.0035]
[0.0062]
2005
-0.0057+
-0.0154**
-0.0161**
-0.0215**
[0.0030]
[0.0032]
[0.0038]
[0.0068]
Constant
-0.0403
-0.1744
-0.1873
-0.2274
[0.1197]
[0.1276]
[0.1536]
[0.2750]
R-squared (within)
0.0401
0.0756
0.0751
0.0630
Statistical significance: + at 10 percent, * at 5 percent, and ** at 1 percent level.
Note: balanced panel 199 districts, 796 observations, 2002-2005. The number of districts differs from the
district level regressions as we only use districts from the balanced panel for which the survey data contains
at least 50 observations for each quartile. All models control for district fixed effects. Standard errors in
brackets.
vii
Table VIII Effect of (lagged) per capita district public health spending on total outpatient
health care utilization rates, by per capita expenditure quartile (district fixed effects)
L. Log district health spending
Average age
Rural share of population
Female share of population
Average household size
L. District splits
(1)
Q1
0.0143+
[0.0083]
0.0030
[0.0043]
-0.0842
[0.2093]
0.0342
[0.2484]
0.0365*
[0.0167]
0.0015
[0.0068]
(2)
Q2
0.0197**
[0.0075]
-0.0013
[0.0039]
0.0532
[0.1894]
0.3397
[0.2247]
0.0402**
[0.0151]
-0.0064
[0.0061]
(3)
Q3
0.0068
[0.0087]
-0.0017
[0.0045]
0.1383
[0.2203]
0.1113
[0.2614]
0.0225
[0.0175]
-0.0061
[0.0071]
(4)
Q4
-0.0104
[0.0149]
-0.0049
[0.0077]
0.0668
[0.3751]
1.0500*
[0.4451]
0.0401
[0.0299]
-0.0114
[0.0121]
Year
2003
-0.0046
-0.0167**
-0.0187**
-0.0194+
[0.0059]
[0.0053]
[0.0062]
[0.0106]
2004
0.0072
-0.0072
-0.0089
-0.0051
[0.0057]
[0.0052]
[0.0060]
[0.0103]
2005
-0.0019
-0.0134*
-0.0145*
-0.0097
[0.0063]
[0.0057]
[0.0066]
[0.0112]
Constant
-0.2450
-0.3969+
-0.0805
-0.2809
[0.2525]
[0.2285]
[0.2658]
[0.4526]
R-squared (within)
0.0362
0.0647
0.0318
0.0342
Statistical significance: + at 10 percent, * at 5 percent, and ** at 1 percent level.
Note: balanced panel 199 districts, 796 observations, 2002-2005. The number of districts differs from the
district level regressions as we only use districts from the balanced panel for which the survey data contains
at least 50 observations for each quartile. All models control for district fixed effects. Standard errors in
brackets.
viii
Table IX Effect of (lagged) per capita district public health spending on household outof-pocket health spending, by per capita expenditure quartile (district fixed effects)
L. Log district health spending
Average age
Rural share of population
Female share of population
Average household size
L. District splits
(1)
Q1
-65.8046
[215.7378]
82.9553
[111.1076]
352.9615
[5,441.54]
-3,085.2477
[6,456.696]
582.5544
[433.3209]
-90.9058
[175.9780]
(2)
Q2
64.3785
[270.0778]
140.7639
[139.0933]
13,018.08+
[6,812.15]
-2,770.4594
[8,083.006]
1,198.430*
[542.4656]
-120.8166
[220.3032]
(3)
Q3
-216.3061
[543.6525]
-359.3405
[279.9876]
4,029.6630
[13,712.51]
20,120.65
[16,270.67]
2,613.764*
[1,091.955]
-394.5591
[443.4589]
(4)
Q4
-1,685.6826
[1,878.551]
606.2411
[967.4766]
2,280.7629
[47,382.56]
27,236.88
[56,222.10]
3,040.2572
[3,773.170]
333.8933
[1,532.34]
Year
2003
-658.08**
-1,613.42**
-1,231.67**
-3,681.72**
[153.2796]
[191.8876]
[386.2597]
[1,334.69]
2004
-893.64**
-1,673.75**
-1,669.18**
-1,314.8332
[148.8109]
[186.2933]
[374.9987]
[1,295.78]
2005
-995.73**
-1,496.43**
-996.95*
-1,044.14
[162.9629]
[204.0099]
[410.6614]
[1,419.01]
Constant
287.2167
-11,231.05
-4,635.78
-10,964.67
[6,565.17]
[8,218.81]
[16,544.03]
[57,166.68]
R-squared (within)
0.1233
0.2042
0.0881
0.0301
Statistical significance: + at 10 percent, * at 5 percent, and ** at 1 percent level.
Note: balanced panel 199 districts, 796 observations, 2002-2005. The number of districts differs from the
district level regressions as we only use districts from the balanced panel for which the survey data contains
at least 50 observations for each quartile. All models control for district fixed effects. Standard errors in
brackets.
ix
Table X Specification and robustness checks: Elasticity of per capita district total health
spending w.r.t. per capita district revenue excluding own revenues (district fixed effects)
Log total district revenue
(excluding own revenues)
Average age
Rural share of population
Female share of population
Average household size
District splits
(1)
Routine
0.7974**
[0.0654]
-0.0193
[0.0223]
2.8583**
[0.6993]
-1.5521
[1.3887]
-0.0345
[0.0941]
-0.0413
[0.0413]
(2)
Development
1.0872**
[0.1365]
-0.0680
[0.0465]
0.9123
[1.4583]
3.3109
[2.8958]
-0.1345
[0.1961]
0.0340
[0.0861]
(3)
Total
0.8511**
[0.0450]
-0.0340*
[0.0153]
1.5664**
[0.4812]
-0.5103
[0.9555]
-0.0656
[0.0647]
-0.0634*
[0.0284]
Year
2002
0.1171**
0.1120+
0.1005**
[0.0292]
[0.0609]
[0.0201]
2003
-0.0719*
0.7697**
0.1595**
[0.0324]
[0.0676]
[0.0223]
2004
0.0130
0.7752**
0.2150**
[0.0326]
[0.0680]
[0.0224]
Constant
-0.6191
-5.2345+
-0.2392
[1.4094]
[2.9389]
[0.9698]
R-squared (within)
0.3336
0.4989
0.6519
Statistical significance: + at 10 percent, * at 5 percent, and ** at 1 percent level.
Note: balanced panel of 207 districts, 828 observations, 2001-2004. All models control for district fixed
effects. Standard errors in brackets.
x
Table XI Specification and robustness checks: Effect of (lagged) per capita district public health spending on outpatient health care
utilization
Log district health spending (per capita)
Log per capita household consumption
Health card coverage
Average age
Rural share of population
Female share of population
Average household size
District splits
Public
(1)
(2)
0.0116**
0.0110**
[0.0039]
[0.0039]
-0.0065
[0.0131]
0.0726**
[0.0207]
-0.0008
-0.0011
[0.0020]
[0.0020]
0.0738
0.0805
[0.0993]
[0.0982]
0.0453
0.0426
[0.1168]
[0.1156]
0.0132+
0.0139+
[0.0078]
[0.0077]
-0.0017
-0.0016
[0.0032]
[0.0032]
Private
(3)
(4)
0.0022
0.0040
[0.0035]
[0.0035]
0.0576**
[0.0117]
0.0306
[0.0191]
-0.0012
-0.0006
[0.0018]
[0.0018]
0.0216
0.0071
[0.0890]
[0.0905]
0.2167*
0.2360*
[0.1048]
[0.1065]
0.0187**
0.0148*
[0.0070]
[0.0071]
-0.0014
-0.0006
[0.0029]
[0.0030]
Total
(5)
(6)
0.0138*
0.0150*
[0.0060]
[0.0059]
0.0511*
[0.0200]
0.1032**
[0.0319]
-0.0020
-0.0017
[0.0031]
[0.0031]
0.0954
0.0875
[0.1518]
[0.1512]
0.2620
0.2786
[0.1786]
[0.1779]
0.0319**
0.0287*
[0.0120]
[0.0119]
-0.0031
-0.0022
[0.0049]
[0.0049]
OOP
(7)
(8)
-478.65
-117.07
[441.50]
[458.97]
11,473.37**
[1,475.03]
4,328.51+
[2,469.73]
-308.39
-195.17
[228.87]
[238.83]
18,673.09+
15,651.23
[11,202.20]
[11,707.04]
17,466.69
21,321.75
[13,182.61]
[13,775.41]
798.35
26.44
[885.16]
[919.63]
14.36
167.83
[365.31]
[381.59]
Year
2003
-0.0056*
-0.0054*
-0.0106**
-0.0095**
-0.0162**
-0.0149**
-2,104.98** -1,894.65**
[0.0028]
[0.0027]
[0.0025]
[0.0025]
[0.0043]
[0.0042]
[313.88]
[327.30]
2004
0.0071*
0.0069*
-0.0097**
-0.0121**
-0.0026
-0.0052
-1,201.05** -1,672.49**
[0.0027]
[0.0027]
[0.0025]
[0.0025]
[0.0042]
[0.0041]
[309.51]
[318.37]
2005
0.0032
0.0034
-0.0196**
-0.0149**
-0.0165**
-0.0116*
-1,881.65**
-961.59**
[0.0031]
[0.0029]
[0.0028]
[0.0027]
[0.0048]
[0.0045]
[352.83]
[350.30]
Constant
-0.0664
-0.1484
-0.8076**
-0.1314
-0.874**
-0.2797
-140470.6**
-5,531.11
[0.1945]
[0.1175]
[0.1744]
[0.1082]
[0.2973]
[0.1808]
[21,946.65]
[13,998.70]
R-squared (within)
0.1027
0.1200
0.1382
0.1079
0.0767
0.0826
0.1811
0.1045
Statistical significance: + at 10 percent, * at 5 percent, and ** at 1 percent level.
Note: balanced panel of 207 districts, 828 observations, 2002-2005. All models control for district and year fixed effects. Standard errors in brackets.
xi
Table XII Elasticity of public outpatient health care utilization with respect to (lagged)
per capita district public health spending (district fixed effects)
L. Log district health spending
Average age
Rural share of population
Female share of population
Average household size
L. District splits
(1)
Q1
0.1441+
[0.0866]
-0.0057
[0.0447]
-2.7683
[2.1909]
3.1957
[2.5981]
0.1058
[0.1738]
-0.0592
[0.0706]
(2)
Q2
0.1418*
[0.0664]
-0.0271
[0.0342]
1.9177
[1.6759]
2.2560
[1.9885]
0.2120
[0.1335]
-0.0799
[0.0542]
(3)
Q3
0.0821
[0.0667]
-0.0206
[0.0343]
1.8697
[1.6815]
-1.1113
[1.9952]
0.0975
[0.1339]
-0.0471
[0.0544]
(4)
Q4
-0.0339
[0.0809]
-0.0254
[0.0418]
-1.1192
[2.0458]
3.3324
[2.4276]
0.2749+
[0.1627]
-0.0253
[0.0660]
(5)
Total
0.0897+
[0.0461]
-0.0167
[0.0240]
0.4977
[1.1768]
0.7563
[1.3848]
0.1481
[0.0924]
-0.0607
[0.0384]
Year
2003
0.0260
-0.1168*
-0.1494**
0.0158
-0.0645+
[0.0618]
[0.0472]
[0.0474]
[0.0576]
[0.0329]
2004
0.1959**
0.0545
0.0662
0.1381*
0.0997**
[0.0597]
[0.0458]
[0.0460]
[0.0558]
[0.0320]
2005
0.1235+
0.0404
0.0185
0.1266*
0.0625+
[0.0654]
[0.0502]
[0.0504]
[0.0612]
[0.0351]
Constant
-4.6146+
-6.6362**
-3.8837+
-3.9053
-4.4986**
[2.6401]
[2.0219]
[2.0287]
[2.4653]
[1.4071]
Number of observations
793
796
796
794
828
R-squared (within)
0.0692
0.0751
0.0643
0.0333
0.1069
Statistical significance: + at 10 percent, * at 5 percent, and ** at 1 percent level.
Note: balanced panel, 2002-2005. Some observations are lost in the quartile specific regressions as we only
use districts from the balanced panel for which the survey data contains at least 50 observations for each
quartile. All regressions control for district fixed effects. Standard errors in brackets.
xii
Table XIII Elasticity of private outpatient health care utilization with respect to (lagged)
per capita district public health spending (district fixed effects)
L. Log district health spending
Average age
Rural share of population
Female share of population
Average household size
L. District splits
(1)
Q1
-0.0922
[0.0974]
0.0143
[0.0500]
2.4404
[2.4563]
-2.1171
[2.8887]
0.6010**
[0.1938]
-0.0889
[0.0783]
(2)
Q2
0.0006
[0.0747]
0.0359
[0.0384]
2.2367
[1.8814]
2.9453
[2.2358]
0.2373
[0.1500]
-0.1556*
[0.0609]
(3)
Q3
-0.0028
[0.0682]
0.0400
[0.0351]
1.9717
[1.7197]
2.9123
[2.0405]
0.4524**
[0.1369]
-0.0680
[0.0556]
(4)
Q4
-0.0516
[0.0726]
-0.0217
[0.0374]
1.9716
[1.8315]
3.0954
[2.1732]
0.3468*
[0.1458]
-0.0273
[0.0592]
(5)
Total
0.0071
[0.0502]
-0.0098
[0.0261]
1.2885
[1.2821]
2.7500+
[1.5087]
0.2597*
[0.1007]
-0.0367
[0.0418]
Year
2003
-0.0492
-0.1331*
-0.1327**
-0.1319*
-0.0960**
[0.0693]
[0.0530]
[0.0484]
[0.0516]
[0.0358]
2004
-0.1346*
-0.1381**
-0.1632**
-0.0991*
-0.1182**
[0.0672]
[0.0514]
[0.0470]
[0.0501]
[0.0348]
2005
-0.0689
-0.1530**
-0.1249*
-0.1084*
-0.1219**
[0.0739]
[0.0564]
[0.0515]
[0.0549]
[0.0383]
Constant
-5.9369*
-7.7353**
-8.3957**
-5.3507*
-5.7627**
[2.9495]
[2.2713]
[2.0748]
[2.2097]
[1.5330]
Number of observations
781
794
796
796
828
R-squared (within)
0.0375
0.0431
0.0575
0.0491
0.0624
Statistical significance: + at 10 percent, * at 5 percent, and ** at 1 percent level.
Note: balanced panel, 2002-2005. Some observations are lost in the quartile specific regressions as we only
use districts from the balanced panel for which the survey data contains at least 50 observations for each
quartile. All regressions control for district fixed effects. Standard errors in brackets.
xiii
Table XIV Elasticity of total outpatient health care utilization with respect to (lagged) per
capita district public health spending (district fixed effects)
L. Log district health spending
Average age
Rural share of population
Female share of population
Average household size
L. District splits
(1)
Q1
0.0446
[0.0730]
-0.0038
[0.0377]
-1.3788
[1.8491]
0.6021
[2.1882]
0.2352
[0.1468]
-0.0425
[0.0596]
(2)
Q2
0.0838
[0.0512]
-0.0094
[0.0264]
1.7379
[1.2924]
2.8741+
[1.5335]
0.2196*
[0.1029]
-0.0941*
[0.0418]
(3)
Q3
0.0173
[0.0490]
0.0018
[0.0253]
1.8894
[1.2369]
0.8932
[1.4676]
0.2049*
[0.0985]
-0.0564
[0.0400]
(4)
Q4
-0.0326
[0.0598]
-0.0337
[0.0308]
1.1053
[1.5082]
3.5162*
[1.7896]
0.2596*
[0.1201]
-0.0331
[0.0488]
(5)
Total
0.0583
[0.0378]
-0.0148
[0.0197]
0.9866
[0.9655]
1.9378+
[1.1361]
0.1724*
[0.0758]
-0.0410
[0.0315]
Year
2003
-0.0169
-0.1273**
-0.1267**
-0.0863*
-0.0889**
[0.0522]
[0.0364]
[0.0348]
[0.0425]
[0.0270]
2004
0.0931+
-0.0267
-0.0291
-0.0030
-0.0043
[0.0504]
[0.0353]
[0.0338]
[0.0412]
[0.0262]
2005
0.0683
-0.0424
-0.0422
-0.0187
-0.0348
[0.0552]
[0.0387]
[0.0370]
[0.0452]
[0.0288]
Constant
-3.2004
-6.0738**
-4.5794**
-3.9641*
-4.4351**
[2.2277]
[1.5593]
[1.4923]
[1.8197]
[1.1545]
Number of observations
794
796
796
796
828
R-squared (within)
0.0334
0.0680
0.0470
0.0417
0.0628
Statistical significance: + at 10 percent, * at 5 percent, and ** at 1 percent level.
Note: balanced panel, 2002-2005. Some observations are lost in the quartile specific regressions as we only
use districts from the balanced panel for which the survey data contains at least 50 observations for each
quartile. All regressions control for district fixed effects. Standard errors in brackets.
xiv
Table XV Elasticity of household out-of-pocket health spending with respect to (lagged)
per capita district public health spending (district fixed effects)
L. Log district health spending
Average age
Rural share of population
Female share of population
Average household size
L. District splits
(1)
Q1
-0.0738
[0.0733]
0.0408
[0.0378]
2.1924
[1.8500]
-0.5554
[2.1951]
0.1808
[0.1473]
-0.0666
[0.0598]
(2)
Q2
0.0224
[0.0607]
0.0476
[0.0313]
4.7379**
[1.5319]
-0.4787
[1.8176]
0.2777*
[0.1220]
-0.0487
[0.0495]
(3)
Q3
-0.0088
[0.0675]
-0.0207
[0.0348]
3.1629+
[1.7022]
2.4604
[2.0198]
0.4137**
[0.1356]
-0.0717
[0.0550]
(4)
Q4
-0.0723
[0.0948]
0.0267
[0.0488]
3.0598
[2.3914]
1.4889
[2.8375]
0.3030
[0.1904]
0.0400
[0.0773]
(5)
Total
-0.0173
[0.0530]
0.0036
[0.0276]
3.2415*
[1.3513]
1.4261
[1.5902]
0.1358
[0.1062]
0.0008
[0.0440]
Year
2003
-0.2572**
-0.3997**
-0.2530**
-0.2916**
-0.2783**
[0.0521]
[0.0432]
[0.0479]
[0.0674]
[0.0378]
2004
-0.3300**
-0.4145**
-0.3106**
-0.1888**
-0.2809**
[0.0506]
[0.0419]
[0.0466]
[0.0654]
[0.0367]
2005
-0.3574**
-0.3652**
-0.1784**
-0.0998
-0.1865**
[0.0554]
[0.0459]
[0.0510]
[0.0716]
[0.0403]
Constant
5.7410*
3.0588+
4.4689*
5.6321+
5.7561**
[2.2320]
[1.8482]
[2.0537]
[2.8852]
[1.6159]
Number of observations
796
796
796
796
207
R-squared (within)
0.1447
0.2257
0.1394
0.0612
0.1577
Statistical significance: + at 10 percent, * at 5 percent, and ** at 1 percent level.
Note: balanced panel, 2002-2005. Some observations are lost in the quartile specific regressions as we only
use districts from the balanced panel for which the survey data contains at least 50 observations for each
quartile. All regressions control for district fixed effects. Standard errors in brackets.
xv
Unbalanced panel results
Estimates for equation (1)
Table U.1 Elasticity of per capita district public health spending w.r.t per capita district
revenue
Random
Fixed
Hausman
N
effects
effects
test
Routine spending
0.7656**
0.7531**
0.2981
1,046
[0.0379]
[0.0536]
Development health spending
1.2278**
1.1828**
0.6039
1,042
[0.0612]
[0.1165]
Total health spending
0.9144**
0.8705**
0.4704
1,031
[0.0289]
[0.0386]
Statistical significance: + at 10 percent, * at 5 percent, and ** at one percent level.
Note: replicating results in Table 4 for an unbalanced panel of 278 districts, 2001-2004. All models control
for demographic characteristics (average age, household size and percentage female population),
percentage rural population, district splits, region fixed effects and aggregate time shocks. The Hausman
test reports p-values for rejecting the hypothesis that the difference in coefficients is not systematic.
Standard errors in brackets.
Table U.2 Elasticity of per capita district public health spending w.r.t per capita district
revenue source (district fixed effects)
Routine
Development
Total
Total district revenue
0.7754**
1.1248**
0.8739**
[0.0544]
[0.1170]
[0.0387]
Revenue shares
Own revenue
1.2766*
2.8296*
1.6177**
[0.6294]
[1.3322]
[0.4371]
Shared tax revenue
-0.0270
-3.4055**
-1.1506**
[0.4338]
[0.9209]
[0.3060]
Shared non tax revenue
-0.7523
0.2555
-0.5881
[0.5102]
[1.0749]
[0.3600]
DAK revenue
-0.6485+
1.6883*
0.0059
[0.3580]
[0.7247]
[0.2477]
Revenue from other sources
-0.4114
-0.8105
-0.3848+
[0.3062]
[0.6420]
[0.2140]
District splits
-0.0582
0.0321
-0.0631*
[0.0354]
[0.0711]
[0.0249]
1,046
1,042
1,031
Statistical significance: + at 10 percent, * at 5 percent, and ** at one percent level.
Note: replicating results in Table 5 for an unbalanced panel of 278 districts, 2001-2004. All models control
for demographic characteristics (average age, household size and percentage female population),
percentage rural population, district splits, aggregate time shocks and district fixed effects. Standard errors
in brackets.
xvi
Estimates for equation (2)
Table U.3 Effect of (lagged) log per capita district public health spending on outpatient
health care utilization rates and household out-of-pocket health spending (district fixed
effects)
Public
Private
Total
OOP
A. Unbalanced panel
District health spending
0.0093**
0.0021
0.0114*
-151.34
[0.0035]
[0.0031]
[0.0052]
[400.97]
B. With source of health spending
District health spending
0.0093**
0.0044
0.0138*
-25.86
[0.0036]
[0.0031]
[0.0054]
[411.29]
Share of development spending
-0.0003 -0.0220** -0.0223+ -1,193.00
[0.0077]
[0.0067]
[0.0115]
[879.84]
Statistical significance: + at 10 percent, * at 5 percent, and ** at one percent level.
Note: replicating results in Table 6 for an unbalanced panel of 278 districts, 1,032 observations, 2002-2005.
All models control for demographic characteristics (average age, household size and percentage female
population), percentage rural population, district splits, aggregate time shocks and district fixed effects.
Standard errors in brackets.
Table U.4 Effect of (lagged) log per capita district public health spending on outpatient
health care utilization rates and household out-of-pocket health spending, by per capita
expenditure quartile (district fixed effects)
Public
Private
Total
OOP
Quartile 1 (poorest)
Quartile 2
Quartile 3
Quartile 4 (richest)
0.0177**
[0.0059]
0.0168**
[0.0050]
0.0061
[0.0054]
-0.0019
[0.0075]
0.0033
[0.0038]
0.0049
[0.0037]
0.0014
[0.0045]
-0.0057
[0.0079]
0.0210**
[0.0076]
0.0216**
[0.0067]
0.0075
[0.0078]
-0.0075
[0.0129]
-78.34
[191.28]
-7.86
[241.36]
-264.84
[474.17]
-854.72
[1,805.81]
Statistical significance: + at 10 percent, * at 5 percent, and ** at one percent level.
Note: replicating results in Table 7 for an unbalanced panel of 264 districts, 981 observations, 2002-2005.
The number of districts differs from Table U.3 as we only use districts for which the survey data contains at
least 50 observations for each quartile. Standard errors in brackets.
xvii