Health and Health Care Delivery in Developing Countries

1
Health and Health Care Delivery in Developing
Countries

• Michael Kremer
• Economics 1386
• Fall 2006

2
The Determinants of Mortality

• Cutler, Deaton, Llenas-Muney (2005)
• Life expectancy increased 30 last century
• Hunter-gatherers 25 years
• England 1700 37 years
• England 1820-1870 41 years
• UK today 77 years
• South Asia 2004 63 years (WB WDI, 2006)
• Vietnam today higher life expectancy (69 years)
  than US in 1900 (47 years) but US in 1900 had 10
  times higher per capita income than Vietnam today
  (Kremer 2002)
• Sub-Saharan Africa 2004 46 (WB WDI, 2006)
• Why? Possible explanations
• Improved nutrition
• Public health
• Urbanization
• Vaccination
• Medical treatments
• Early life factors

3
The Determinants of Mortality (II)

• Still - differences between rich and poor
  countries
• Progress in health until HIV/AIDS
• Correlation between per capita income and
  mortality
• Correlation within countries richer, better
  educated people live longer. Why?
• Unclear
• Access to medical care
• Resources for non-health care
• Differences in health-related behaviors
• Social structures, stress and health
• Authors
• Scientific advance and technical progress
• Less on direct causality of income

4
Health Care Delivery in Rural Rajasthan

• Banerjee, Deaton and Duflo (2004)
• Survey of 100 hamlets in Udaipur, collaboration
  with Seva Mandir (NGO) and Vidhya Bhavan (schools
  group)
• Village survey, facility/traditional healer
  survey, weekly visit to facilities, and
  household/individual survey
• Very poor region
• Adult literacy men 46, women 11
• 21 of households have electricity

5
Health Care Delivery in Rural Rajasthan (II)

• Poor health status
• gt50 of men and women anemic
• 93 men and 88 women below US BMI cut-off for
  low nutrition
• High self-reported disease symptoms
• Health facility use
• Adults visit 0.51 times per month
• More for wealthier households
• 0.12 to public facilities, 0.11 to bhopas, 0.28
  to private facilities
• 7 of household budget spent on health

6
Health Care Delivery in Rural Rajasthan (III)

• Public Health Care Facilities
• Official policy one subcentre with one nurse per
  3000 people, open 6 days per week, 6 hours per
  day, free care
• 3600 people per subcentre. One primary health
  centre (PHC) with 5.8 medical personnel and 1.5
  doctors per 48,000 people
• 45 absenteeism in subcentre, 35 in larger PHCs
• Subcentres closed 56 of time, unpredictable
• 32 given injection, 6 given drip, 3 given test
• 75 report that visit made them feel better

7
Health Care Delivery in Rural Rajasthan (IV)

• Private Care
• Poor training as doctors 41 have no medical
  degree, 18 have no medical training, 17 did not
  graduate from high school
• 68 given injection, 12 given drip, 3 test
• 81 report that visit made them feel better
• Almost comparable costs even though public
  supposed to be free
• Public Rs 71
• Private Rs 84
• Bhopa Rs 61
• Reported satisfaction compared with poor health
  outcomes suggests need for state involvement

8
Do Conditional Cash Transfers Improve Child
Health?

• Gertler (2004) on PROGRESA
• Requirements for cash transfer
• Clinic visits for infants, young children,
  pregnant/lactating women
• Vaccinations for infants
• Yearly clinic checkups for all family members
• Meetings on health, hygiene, nutrition for adults
• Illness rate of treatment children 39.5 lower
  than control group
• Treatment children almost 1 cm taller
• Treatment children 25.5 less likely to be anemic

9
Contracting for HealthEvidence from Cambodia
Elizabeth King Michael Kremer Benjamin
Loevinsohn Brad Schwartz

• Indu Bhushan
• Erik Bloom
• David Clingingsmith
• Rathavuth Hong

10
Contracting in Cambodia

• Many developing countries have mix of salaried
  government and private fee for service provision.
  Lots of problems
• Weak incentives of government providers (high
  absence rates)
• Glucose drip problem misalignment of private
  providers incentives with
• patients interest (asymmetric information)
• public health (inadequate attention to
  externalities from infectious disease, drug
  resistance)
• Lack of risk sharing
• In 1999, Cambodia tried experiment of tendering
  contracts for management of health services in
  certain districts
• Monitored performance against 8 targeted outcomes
• Increased public funding (offset by decreased
  private expenditure)
• District level contracting could potentially
• Tie incentives to public health objectives
• Allow benchmark competititon
• Pool risks with limited adverse selection (in
  rural areas)

11
Estimating program impact

• Call for bids in 8 districts, randomly chosen
  from 12 candidates
• Acceptable bids in only some districts
• IV with random assignment
• Small number of districts
• Clustering
• Randomization inference
• Average effects

12
Main results

• Delivery of targeted services improved
• Non-targeted services were unchanged
• Some evidence of health improvement
• Management of health centers improved.
• Patients shifted to public providers
• Effect on total spending (public and private) was
  neutral to negative.

13
Previous work

• Keller and Schwartz (2001)
• Bhushan, Keller, and Schwartz (2002)
• Schwartz and Bhushan (2004)
• Based on 2001 survey of districts where program
  actually implemented

14
Overview

• Background on project and context
• Model of health care provision
• Empirical methods
• Health services results
• Targeted outcomes
• Non-targeted outcomes
• Health outcomes
• Choice of provider
• Health facility management
• Consumer perception of care quality
• Public and private health spending
• Conclusion

15
Cambodian health context

• Political background
• 1975-1979 Khmer Rouge
• 1979-1993 Vietnamese-backed regime
• 1993 Elections adoption of market economy
• 1998 End of fighting
• Health care system
• Government health worker salary 85 GDP/cap
  politicization of civil service, governance
  issues
• Boom in private medical practice after 93
• government staff moonlighting
• drug sellers get about 33 of curative visits in
  1997
• Traditional understanding of health, disease
• Spending high health service coverage poor
• Project runs 1999-2003
• Huge improvements in health over study period.
  Health center construction.

16
Contracting Project I

• Covered 11 of population
• Responsible for Minimum Package of Activities
• Targets on improvement of child and maternal
  health service coverage.
• Prevention oriented

17
Two program variants

• Contracting in (CI)
• Management authority, but cant hire/fire,
  procure outside government
• Operating budget through government
• Contracting out (CO)
• Full control of staffing--hire and fire
• Full control of procurement
• Operating budget through ADB/WB loans

18
Bidding

• Fixed price per capita bids increased public
  spending
• Technical criteria and price
• 3 districts got no technically acceptable bids

19
HR practices under contracting I

• Example Peareng district, contracting in (CI)
• Facilities signed annual contracts with NGO,
  workers 3-mo subcontracts. Private practice
  banned.
• Staff motivation viewed as key problem
• Additional payment on top of government salary
• Fixed supplement, attendance, facility
  performance
• Staff incentives based on targeted outcomes,
  patient satisfaction, quality of care, and no
  fraud

20
HR practices under contracting II

• All contractors chose to use some expatriates.
• Between 0.5 and 3.0 expatriate staff per
  contracted district.
• Local staff between 90 and 150 per district.

21
HR practices under contracting III

• Additional compensation for workers in all
  treated districts
• Two officially banned private practice, three
  allowed it
• Contractor compensation choices
• Contracting in (CI) Base salary plus performance
  bonus, no provision for firing
• Contracting out (CO) high fixed salaries, with
  possibility of firing non-performers
• CO attracted some staff from outside district,
  outside government service

22
Randomization procedure

• Three provinces with 3, 4, and 5
  treatment-eligible districts respectively
• Randomly assign CI, CO, and comparison district
  within each province.
• Remaining 3 districts randomized in capital
• Each district had equal probability of being CI,
  CO, comparison

23
Data

• Baseline household survey in 1997, follow-up in
  2003 facility survey in 2003
• 30 randomly selected villages in each of 12
  districts 7-14 households per village randomly
  chosen in each survey year
• Household census, recent illnesses and treatment,
  program outcomes
• Follow-up included health service quality module

24
Targeted outcomes at baseline
Baseline Percentage by Random Assignment Baseline Percentage by Random Assignment Baseline Percentage by Random Assignment
Comp. Contracting In Contracting Out Goal
Fully immunized child 34 28 31 70
Children get Vitamin A 42 46 41 70
Antenatal care 9 11 13 50
Delivery by trained personnel 24 27 32 50
Delivery in a health facility 3 6 5 10
Use modern birth control method 13 12 18 30
Knowledge of birth control 22 27 20 70
Use of public health care facilities 4 4 3 Increase
25
Targeted outcomes at baseline
Baseline Percentage by Random Assignment Baseline Percentage by Random Assignment Baseline Percentage by Random Assignment
Comp. Contracting In Contracting Out Goal
Fully immunized child 34 28 31 70
Children get Vitamin A 42 46 41 70
Antenatal care 9 11 13 50
Delivery by trained personnel 24 27 32 50
Delivery in a health facility 3 6 5 10
Use modern birth control method 13 12 18 30
Knowledge of birth control 22 27 20 70
Use of public health care facilities 4 4 3 Increase
26
Randomization quality

• Will see baselines as go through each outcome
• Of baseline levels for 22 outcomes
• three significant for each of CI and CO under
  clustering
• one significant for each of CI and CO under
  randomization inference

27
Holmstrom and Milgrom (1991)Model of health
service provision

• Suppose targeted and non-targeted outcomes
  produced by exerting various kinds of costly
  effort
• Suppose only targeted outcome T is contractable
  linear compensation contract in T
• Increasing incentives for T may lead to more or
  less NT, depending on whether effort types
  relevant for T and NT are complements or
  substitutes
• Either plausible

28
Econometric Issues

• Selection into treatment
• CO 4 districts tendered, 3 awarded
• CI 4 districts tendered, 2 awarded
• Previous analysis based on actual treatment
  status, not initial assignment
• Perhaps NGOs focused bids on districts where
  gains would be easiest
• Cluster-level intervention
• Clustering, randomization inference
• Family-level effects

29
Econometric Issues

• Selection into treatment
• CO 4 districts tendered, 3 awarded
• CI 4 districts tendered, 2 awarded
• Previous analysis based on actual treatment
  status, not initial assignment
• Perhaps NGOs focused bids on districts where
  gains would be easiest
• Cluster-level intervention
• Clustering, randomization inference
• Family-level effects

30
Empirical method I

• District-level intervention with individual
  outcomes
• Randomly-assigned eligibility an instrument for
  actual treatment.
• TOT for outcome k
• Instruments

31
Empirical method I

• District-level intervention with individual
  outcomes
• Randomly-assigned eligibility an instrument for
  actual treatment.
• TOT for outcome k
• Instruments

32
Empirical method II

• District-level intervention with individual
  outcomes
• Need to account for district level shocks
• Clustering may over-reject null with small number
  of clusters
• Randomization inference
• Create full set placebo random assignments using
  actual randomization process. (Rosenbaum 2002)
• Generate placebo treatment effect for each member
  of the set.
• Use distribution of placebo treatment effects as
  test distribution.
• Low power imposes no structure on error

33
Randomization Inference
34
Empirical method III

• Average effect size (AES) for family of K
  outcomes
• Kling, Katz, Leibman, and Sonbanmatsu (2003),
  OBrien (1984)
• Joint estimation of TOT for K outcomes
• Aggregate to get common unit of observation v
• VCM estimates cross-equation correlation of
  effects
• AES is the average treatment effect measured in
  standard deviation units.
• We use the standard deviation of the change in
  outcome for comparison group.

35
Empirical method III

• Average effect size (AES) for family of K
  outcomes
• Kling, Katz, Leibman, and Sonbanmatsu (2003),
  OBrien (1984)
• Joint estimation of TOT for K outcomes
• Aggregate to get common unit of observation v
• VCM estimates cross-equation correlation of
  effects
• AES is the average treatment effect measured in
  standard deviation units.
• We use the standard deviation of the change in
  outcome for comparison group.

36
Change in District Averages I
37
Change in District Averages II
38
TOT for changes in targeted outcomes
Notes IV regressions including province X year
fixed effects. Average effects are differential
increases caused by treatment in units of
baseline comparison-group standard deviations.
Standard errors presented in parentheses are
corrected for clustering at the district level.
Stars indicate significance under clustering
at 10 at 5 at 1. P-values for
treatment effects computed by randomization
inference.
39
TOT for changes in targeted outcomes
Notes IV regressions including province X year
fixed effects. Average effects are differential
increases caused by treatment in units of
baseline comparison-group standard deviations.
Standard errors presented in parentheses are
corrected for clustering at the district level.
Stars indicate significance under clustering
at 10 at 5 at 1. P-values for
treatment effects computed by randomization
inference.
40
Robustness check wealth controls
Notes All IV regressions in Panel B include
province X year fixed effects and wealth
controls. Standard errors presented in
parentheses are corrected for clustering at the
district level. Stars indicate significance under
clustering at 10 at 5 at 1.
P-values for treatment effects computed by
randomization inference.
41
TOT for non-contracted outcomes
Notes All regressions include province X year
fixed effects. Standard errors presented in
parentheses are corrected for clustering at the
district level. Stars indicate significance under
clustering at 10 at 5 at 1.
P-values for treatment effects computed by
randomization inference. Treatment effects are in
bold.
42
TOT for non-contracted outcomes
Notes All regressions include province X year
fixed effects. Standard errors presented in
parentheses are corrected for clustering at the
district level. Stars indicate significance under
clustering at 10 at 5 at 1.
P-values for treatment effects computed by
randomization inference. Treatment effects are in
bold.
43
TOT for non-contracted outcomes
Notes All regressions include province X year
fixed effects. Standard errors presented in
parentheses are corrected for clustering at the
district level. Stars indicate significance under
clustering at 10 at 5 at 1.
P-values for treatment effects computed by
randomization inference. Treatment effects are in
bold.
44
TOT for final health outcomes
Notes All regressions include province X year
fixed effects. Standard errors presented in
parentheses are corrected for clustering at the
district level. Stars indicate significance under
clustering at 10 at 5 at 1.
P-values for treatment effects computed by
randomization inference. Treatment effects are in
bold.
45
TOT for final health outcomes
Notes All regressions include province X year
fixed effects. Standard errors presented in
parentheses are corrected for clustering at the
district level. Stars indicate significance under
clustering at 10 at 5 at 1.
P-values for treatment effects computed by
randomization inference. Treatment effects are in
bold.
46
TOT for changes in care-seeking outcomes
Notes IV regressions with provinceXyear effects.
Standard errors presented in parentheses are
corrected for clustering at the district level.
Stars indicate significance under clustering
at 10 at 5 at 1. P-values for
treatment effects computed by randomization
inference. Average effect codes drug seller and
traditional healer visits as negative and
qualified private and public provider visits as
positive.
47
TOT for health center management I
Notes All columns are IV regressions in levels
with province fixed effects. Standard errors
presented in parentheses are corrected for
clustering at the district level. Stars indicate
significance under clustering at 10 at
5 at 1. P-values for treatment effects
computed by randomization inference.
48
TOT for health center management II
Notes All columns are IV regressions in levels
with province fixed effects. Standard errors
presented in parentheses are corrected for
clustering at the district level. Stars indicate
significance under clustering at 10 at
5 at 1. P-values for treatment effects
computed by randomization inference.
49
AES for 11 health center management outcomes
Notes Standard errors presented in parentheses
are corrected for clustering at the district
level. Stars indicate significance under
clustering at 10 at 5 at 1.
P-values for treatment effects computed by
randomization inference.

• Health center open with patients
• All scheduled staff present
• Child delivery service available
• User fees clearly posted
• Number of supervisor visits

• Number of outreach trips
• Index of equipment installed and functional
• Index of drugs and other supplies available
• All childhood immunizations available

50
TOT for consumer perception of quality
Notes All regressions include province X year
fixed effects. Standard errors presented in
parentheses are corrected for clustering at the
district level. Stars indicate significance under
clustering at 10 at 5 at 1.
P-values for treatment effects computed by
randomization inference. Treatment effects are in
bold.
51
TOT for health care spending I
Notes All outcomes in 2003 USD per capita. All
regressions include province X year fixed
effects. Standard errors presented in parentheses
are corrected for clustering at the district
level. Stars indicate significance under
clustering at 10 at 5 at 1.
P-values for treatment effects computed by
randomization inference. Treatment effects are in
bold.
52
TOT for health care spending I
Notes All outcomes in 2003 USD per capita. All
regressions include province X year fixed
effects. Standard errors presented in parentheses
are corrected for clustering at the district
level. Stars indicate significance under
clustering at 10 at 5 at 1.
P-values for treatment effects computed by
randomization inference. Treatment effects are in
bold.
53
TOT for health care spending II
Notes All outcomes in 2003 USD per capita. All
regressions include province fixed effects.
Standard errors presented in parentheses are
corrected for clustering at the district level.
Stars indicate significance under clustering
at 10 at 5 at 1. P-values for
treatment effects computed by randomization
inference. Treatment effects are in bold.
54
Effects of spending on health outcomes I
55
Effects of spending on health outcomes II
56
Conclusion I

• Increase in targeted outcomes of about one-half
  standard deviation.
• Improved management, particularly in contracting
  out.
• No evidence non-targeted outcomes decreased.
• Total spending neutral or decreased.

57
Conclusion II

• Institutional change plus increased public
  spending
• Only one of set of possible interventions
• Interesting for policy because feasible
• External validity difficult to assess
• Additional trials in post-conflict environments
  and others with serious health care delivery
  problems.

58
Intradistrict correlation and wealth controls in
1997
59
ITT effects on quantiles of individual spending
Notes Mean and quantile regressions with
provinceXyear effects. Standard errors presented
in parentheses are corrected for clustering at
the district level using bootstrap. Stars
indicate significance under clustering at 10
at 5 at 1. P-values for treatment
effects computed by randomization inference.
Treatment effects are in bold.
60
Results in a nutshell

• Contracting improved facility management
  availability of 24h care, staff presence,
  supervision
• Large, positive and significant effects on
  targeted outcomes, no effects were significantly
  negative
• Non-targeted outcomes showed gains or no effect.
  No average effect.
• Increased use of public facilities, mostly at
  expense of drug sellers
• Contracted quality of care perceived as worse
  than comparison
• Effect on total spending neutral to negative

61
Public spending breakdown