## _wp08228

## Source details

**Canonical URL:** [_wp08228](https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2008/_wp08228.pdf)

## Other formats

- [Markdown version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2008/_wp08228.pdf.md)
- [Structured JSON version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2008/_wp08228.pdf.json)

---

### I. Introduction — purpose, dataset, and key contributions
- Purpose: assess government spending efficiency across countries and over time, focusing on health and education.
- Two main contributions:
  - Compile a large cross-country panel dataset of government efficiency: 114 countries, 1980 to 2006, with about 1,800 country-year observations for education and about 900 observations for health.
  - Examine policy and environmental determinants of efficiency using a much broader universe of regressors than previous studies.
- Context and motivation:
  - Efficiency of government spending is critical for advanced economies, transition countries, and low-income countries (e.g., to reach the Millennium Development Goals).
- High-level empirical findings (paper summary):
  - Public sector performance and efficiency have been trending upward over time.
  - "The most resounding conclusion is that higher expenditure relative to GDP tends to be associated with lower efficiency."
  - In education, "even the relationship between spending and performance is tenuous, in stark contrast to the health sector."
  - Institutions matter: greater government accountability improves education performance; worse corruption deteriorates performance and efficiency.
  - Demographic and geographic correlates: a relatively larger youthful population reduces education performance and efficiency; higher population density improves performance.

### II. Methodology — measures, computation, and econometric approach
- Key concepts and definitions:
  - Public sector performance (PSP): outcome-only scalar computed as PSP_i = sum_j ω_j PSP_ij with equal weights assumed.
  - Public sector efficiency (PSE): PSE = PSP / EXP where EXP is expenditure in percent of GDP.
  - Data Envelopment Analysis (DEA) efficiency scores: non-parametric technical efficiency relative to an empirical frontier; scores in [0,1[ with 1 indicating the frontier.
- Aggregation and normalization:
  - Outcome indicators divided by their standard deviation and then set to mean 1 using the mean and standard deviation of the pooled observations.
  - Equal weights imposed for aggregation; different weights generate similar rank correlations.
- DEA specifics and limitations:
  - Input approach: compares actual spending with the minimum spending sufficient to produce the same outcome.
  - Pooling observations to construct a piecewise linear frontier to avoid period-specific degrees-of-freedom effects.
  - Limitations: random errors interpreted as inefficiency (sensitivity to outliers); DEA maximizes each unit’s score via endogenous weights.
  - Both endogenous-weight DEA and exogenously-weighted PSE computed for comparison.
  - Simar and Wilson (2007) methods discussed; traditional DEA used given N much greater than 100 and Monte Carlo evidence of small differences for larger N.
- Formal and econometric approach:
  - Variable returns-to-scale DEA formulation presented (optimization in equation (3) and slack-adjusted form in equation (4)).
  - Correlates estimated via regressions y_i = f(z_i) + ε_i (equation (5)).
  - Estimation methods:
    - OLS for PSP and PSE baseline.
    - Truncated regressions for DEA scores (following Simar and Wilson (2007)).
    - Fixed and random effects; Hausman tests guide towards fixed effects but random effects retained for checks and for time-invariant regressors.
    - Regressions run with and without per capita income due to collinearity and potential endogeneity.
    - System GMM used where endogeneity concerns are justified after Hausman-augmented tests.
    - Heteroskedasticity-consistent standard errors and year fixed effects included in all regressions.

### III. Dataset and empirical scope; sample sizes and indicators
- Sectoral focus: health and education.
- Coverage and time span:
  - Up to 114 advanced and developing economies.
  - Panel: 1980–2006 overall; health observations concentrated in the period 1997– (most health observations fall into the period of 1997–).
- Sample sizes (dataset and methodology section):
  - Overall roughly 1,800 observations for education and 900 observations for health.
  - Education: 470 observations for advanced economies and 1,350 observations for developing economies.
  - Health: 470 observations for advanced economies and 450 observations for developing economies.
- Output indicators used:
  - Education: (i) primary school enrolment rate, (ii) secondary school enrolment rate, (iii) primary teacher to student ratio.
  - Health: (i) DPT immunization rate, (ii) physicians to population ratio, (iii) share of public spending in total health spending.
- Inputs: health and education expenditure of general government where available, central government otherwise.
- Data sources: WHO World Health Statistics; UNESCO Global Education Digest; IMF Government Finance Statistics; OECD Analytical/ Education/ Health Databases; European Commission AMECO and Government Finance Statistics; IMF WEO and DEME; World Bank World Development Indicators.
- Data construction: ratio splicing when series overlap; linear interpolation for short gaps.
- Six scores per country: PSP, PSE, and DEA for each sector. Note: DEA scores sensitive to degrees of freedom; focus is on determinants rather than levels.

### IV. Indicators, summary statistics, and correlations (selected exact figures)
- Table 1 — Education:
  - PSPE mean 1.01; Coefficient of variation 0.34; Minimum 0.25; 1st quartile 0.77; Median 1.03; 3rd quartile 1.25; Maximum 2.38.
  - PSEE mean 1.21; Coefficient of variation 0.62; Minimum 0.26; 1st quartile 0.79; Median 1.08; 3rd quartile 1.44; Maximum 15.19.
  - DEAE mean 0.32; Coefficient of variation 0.65; Minimum 0.03; 1st quartile 0.16; Median 0.27; 3rd quartile 0.43; Maximum 1.00.
- Table 1 — Health:
  - PSPH mean 1.15; Coefficient of variation 0.28; Minimum 0.23; 1st quartile 0.93; Median 1.23; 3rd quartile 1.41; Maximum 1.81.
  - PSEH mean 1.38; Coefficient of variation 0.64; Minimum 0.53; 1st quartile 0.89; Median 1.10; 3rd quartile 1.54; Maximum 8.04.
  - DEAH mean 0.34; Coefficient of variation 0.63; Minimum 0.08; 1st quartile 0.18; Median 0.25; 3rd quartile 0.45; Maximum 1.00.
- Selected Spearman rank correlations (Table 2):
  - Education PSP vs Health PSP correlation coefficient 0.81.
  - Education PSE cross-sector correlation 0.37.
  - DEA correlation between education and health reported as -0.03 in parts of the table (negative/low DEA cross-sector correlation).
  - Within education: PSPE and PSEE correlation 0.70; PSPE and DEAE correlation 0.81; PSEE and DEAE correlation 0.83.
- Observed patterns:
  - Education and health outputs are more strongly correlated than their efficiency measures.
  - Coefficient of variation for efficiency measures about twice as large as for performance measures.
  - Performance is positively correlated with efficiency in education but negatively correlated in health.

### V. Trends and graphical findings
- Performance highest in advanced economies in both education and health.
- Efficiency highest in advanced economies for education but lowest in advanced economies for health.
- Both performance and efficiency have trended upwards on average across income groups.
- Education spending shows larger fluctuation relative to GDP over a longer horizon, contributing to volatility in PSE/DEA scores while outcomes remain more stable.

### VI. Determinants of performance and efficiency — empirical strategy and estimation practice
- Three-stage empirical approach:
  - Stage 1: Univariate regressions for all potential determinants.
  - Stage 2: Multivariate regressions combining determinants significant in univariate regressions.
  - Stage 3: Tested-down regressions keeping regressors significant at least at the 10 percent level.
- Regression frameworks: random effects and fixed effects; endogeneity tested using augmented regressions and two-step system GMM (Arellano and Bover (1995); Blundell and Bond (1998)). Windmeijer (2005) correction applied where appropriate.

### VII. Main empirical findings — economic determinants (exact reported conclusions)
- Spending and efficiency:
  - Larger education and health spending reduces public sector efficiency in each sector.
  - In education, spending does not bear a significant relationship with public sector performance.
  - Augmented regression tests:
    - Do not suggest endogeneity of spending in education efficiency regressions.
    - Do suggest endogeneity for health efficiency (PSE and DEA), implying lower efficiency may lead to higher spending.
  - Two-step system GMM:
    - For PSE: significant negative effect of health spending confirming OLS result.
    - For DEA: coefficient on health expenditure becomes insignificant.
  - Conclusion: nexus between higher expenditure and lower efficiency is robust, though causality in health remains somewhat unclear.
- Income per capita:
  - Several tested-down regressions show higher income associated with better performance; some negative signs appear in univariate/multivariate settings and disappear when tested down.
  - Augmented regression tests suggest income per capita effect is exogenous.
- Commodity exporter dummy:
  - Significant in univariate regressions but becomes insignificant in tested-down multivariate specifications.
- Inflation:
  - Expected negative effect reversed in health sector (positive effect in most cases), possibly due to budget process mechanics (unexpected inflation lowering expenditure/GDP denominator and improving measured efficiency temporarily).
- Openness/trade liberalization:
  - Not significant in final tested-down specifications.

### VIII. Main empirical findings — institutional determinants
- Accountability:
  - Higher government accountability consistently increases education performance.
  - Association with health sector efficiency less robust and sometimes negative.
- Corruption control:
  - Robust evidence that better control of corruption increases both efficiency and performance in education and health.
  - Corruption control effects appear in tested-down regressions when controlling for country effects and per capita income.
- Other institutions:
  - Democracy, regime durability, social infrastructure, and schooling largely not significant in tested-down specifications (with some exceptions in certain specifications).

### IX. Main empirical findings — demographic and geographic determinants
- Age structure:
  - Education sector:
    - Younger population (higher share <14) reduces efficiency; older population (higher share >65) increases efficiency.
    - Performance also appears to suffer with younger population (possible crowding effects).
  - Health sector:
    - Share of younger population does not matter much.
    - Some evidence that a larger elderly population positively affects health performance and efficiency.
- Population density:
  - Positive effects on performance and efficiency consistent with economies of scale in service provision.
- Fractionalization and climate:
  - Linguistic fractionalization not significant in tested-down specifications.
  - Malaria index and distance-from-equator do not show significant effects in final tested-down regressions.

### X. Policy implications and conclusions (preserve phrasing)
- Two key policy implications emphasized:
  - Efficiency declines with the level of spending in education and health; in education, spending is not significantly associated with better performance. Policy takeaway: increasing funding alone often fails to yield expected improvements unless accompanied by efficiency-enhancing reforms.
  - Institutional improvements matter: government accountability and controlling corruption can improve government efficiency and performance in education and health.
- Governments face demographic and geographic constraints that matter for performance and efficiency but are largely exogenous and must be managed rather than changed.

### XI. Limitations and caveats
- Data shortcomings: presented "only as a first step to a better understanding of broad global trends in expenditure efficiency and its determinants" due to important but currently irresolvable data limitations.
- Lag structure assumption: impact of expenditure on outcomes is assumed contemporaneous; authors note most inputs and outcomes exhibit high persistence, limiting the impact of this assumption.
- DEA caveats: sensitivity to outliers and random errors interpreted as inefficiency; DEA yields an upper bound of "true" efficiency since best performers in the sample may themselves be inefficient.

* _wp08228 - References (IMF PDF content provided)._

### References..............................................................................................................

### References

### I. Introduction — purpose, dataset, and key contributions
- Purpose: assess government spending efficiency across countries and over time, focusing on health and education.
- Two main contributions:
  - Compile a large cross-country panel dataset of government efficiency: 114 countries, 1980 to 2006, with about 1,800 country-year observations for education and about 900 observations for health.
  - Examine policy and environmental determinants of efficiency using a much broader universe of regressors than previous studies.
- Context and motivation:
  - Efficiency of government spending is critical for advanced economies, transition countries, and low-income countries (e.g., to reach the Millennium Development Goals).
  - Prior literature measures public sector efficiency by relating government expenditure to socio-economic indicators; recent work has begun examining determinants of cross-country efficiency differences.
- High-level empirical findings (summary from the paper):
  - Public sector performance and efficiency have been trending upward over time.
  - Higher expenditure relative to GDP tends to be associated with lower efficiency.
  - In education, the relationship between spending and performance is tenuous; in health, spending-performance links are stronger.
  - Institutions matter: greater government accountability improves education performance; worse corruption deteriorates performance and efficiency.
  - Demographic and geographic correlates: a relatively larger youthful population reduces education performance and efficiency; higher population density (allowing for returns to scale) improves performance.

### II. Methodology — measures, computation, and econometric approach
- Performance and efficiency concepts used:
  - Public sector performance (PSP): measures outcomes only; defined (following Afonso et al. (2005) with slight changes) as PSP_i = sum_j ω_j PSP_ij where ω_j are weights; equal weights assumed.
  - Public sector efficiency (PSE): ratio of PSP to the respective expenditure EXP (in percent of GDP); PSE = PSP / EXP.
  - Data Envelopment Analysis (DEA) efficiency scores: non-parametric measurement of technical efficiency relative to an empirical frontier.
- Aggregation and normalization:
  - Outcome indicators divided by their standard deviation and then set to mean 1, using the mean and standard deviation of the pooled observations.
  - Equal weights imposed for aggregation due to unobserved societal welfare function; different weights generate similar rank correlations.
- DEA specifics:
  - Measures technical efficiency: ability to produce a given set of outcomes with minimal inputs; efficiency scores in [0,1[, with 1 indicating the frontier.
  - Uses an input approach: compares actual spending with the minimum spending theoretically sufficient to produce the same actual outcome.
  - Non-parametric DEA chosen to construct a piecewise linear frontier; frontier calculated for the pool of observations to avoid varying degrees-of-freedom effects across periods.
  - Limitations of DEA noted: interprets random errors as inefficiency (sensitivity to outliers) and chooses weights to maximize each unit’s score; therefore both endogenous-weight DEA and exogenously-weighted PSE are computed for comparison.
  - Simar and Wilson (2007) methods discussed; traditional DEA used given N much greater than 100 and Monte Carlo results indicating small differences for larger N.
- Formal DEA formulation (as provided):
  - Variable returns-to-scale efficiency problem sketched as the optimization in equation (3) and slack-adjusted form in equation (4).
- Econometric strategy for correlates:
  - Regress PSP, PSE, and DEA scores on a set of potential correlates using forms y_i = f(z_i) + ε_i (equation (5)).
  - Estimation methods:
    - OLS for PSP and PSE baseline.
    - Truncated regressions for DEA scores (following Simar and Wilson (2007) rather than Tobit).
    - Fixed and random effects used; Hausman tests guide preference toward fixed effects but random effects retained for checks and to estimate time-invariant dummies.
    - Regressions run with and without per capita income due to collinearity and potential endogeneity concerns.
    - System GMM used where endogeneity concerns are justified after Hausman-augmented tests.
    - Heteroskedasticity-consistent standard errors and year fixed effects included in all regressions.

### III. Dataset and empirical scope
- Sectoral focus: health and education (chosen because outcome data for other sectors are not widely available).
- Coverage and time span:
  - Up to 114 advanced and developing economies.
  - 1980–2006 panel for the overall dataset; health observations concentrated in the period 1997– (text indicates most health observations fall into the period of 1997–).
- Computed scores per country and sector:
  - PSP: outcome-based scalar per country-sector.
  - PSE: PSP divided by expenditure (percent of GDP).
  - DEA scores: technical efficiency relative to pooled empirical frontier.

### IV. Main empirical findings and policy implications
- Key empirical findings (preserve original phrasing where given):
  - "The most resounding conclusion is that higher expenditure relative to GDP tends to be associated with lower efficiency."
  - In education: "even the relationship between spending and performance is tenuous, in stark contrast to the health sector."
  - Implication: citizens face declining marginal "bang for the buck" as governments spend more on education, and possibly no additional "bang" at all.
- Institutional and structural correlates:
  - Government accountability strengthens public sector performance in education.
  - Corruption worsens public sector performance and efficiency.
  - Higher per capita income, and variables linked to governance and decentralization, are discussed elsewhere in the paper as correlated with efficiency (methodology notes discuss inclusion/exclusion of per capita income due to its strong correlation with institutional variables).
- Demographic and geographic correlates:
  - A relatively larger youthful population reduces performance and efficiency in the education sector.
  - Higher population density, allowing for returns to scale, improves performance.
- Policy relevance:
  - Findings caution against a reflexive policy of "throwing money at issues" (especially in education) without attention to efficiency, institutions, and demographic/geographic constraints.
  - Improving governance, accountability, and reducing corruption are highlighted as levers to improve performance and efficiency.

### V. Limitations and caveats
- Data shortcomings acknowledged: the paper is presented "only as a first step to a better understanding of broad global trends in expenditure efficiency and its determinants" due to important but currently irresolvable data limitations.
- Lag structure assumption: impact of expenditure on outcomes is assumed contemporaneous due to practical difficulties in choosing lag/averaging windows; authors note most inputs and outcomes exhibit high persistence, limiting the impact of this assumption.
- DEA caveats: sensitivity to outliers and random errors interpreted as inefficiency; DEA yields an upper bound of "true" efficiency since best performers in the sample may themselves be inefficient.

*Italic source attribution: _wp08228 - References (IMF PDF content provided)._

### 2003. Overall, we have roughly 1,800 observations for education and 900 for health. The sample

### _wp08228 - 2003

### Dataset and methodology
- Sample sizes:
  - Overall roughly 1,800 observations for education and 900 observations for health.
  - Education: 470 observations for advanced economies and 1,350 observations for developing economies.
  - Health: 470 observations for advanced economies and 450 observations for developing economies.
- Key methodological points:
  - Cross-country comparisons assume homogeneity in production functions and ignore data heterogeneity; Baumol (1967) effect and heterogeneity in input quality are emphasized as potential violations.
  - Outcome indicator choice driven by data availability: three indicators per sector chosen with broadest coverage; outputs used as proxies for outcomes because outcome indicators (e.g., test scores, healthy life expectancy) have much more limited coverage.
  - Output indicators used:
    - Education: (i) primary school enrolment rate, (ii) secondary school enrolment rate, (iii) primary teacher to student ratio.
    - Health: (i) DPT immunization rate, (ii) physicians to population ratio, (iii) share of public spending in total health spending.
  - Inputs: health and education expenditure of general government where available, central government otherwise.
  - Data sources: WHO World Health Statistics; UNESCO Global Education Digest; IMF Government Finance Statistics; OECD Analytical/ Education/ Health Databases; European Commission AMECO and Government Finance Statistics; IMF WEO and DEME; World Bank World Development Indicators.
  - Data construction procedures: ratio splicing when series from two sources overlap; linear interpolation for short gaps.
  - Six scores calculated per country (PSP, PSE, DEA for each sector). DEA scores noted to be sensitive to degrees of freedom; focus is on determinants of scores rather than levels.

### Indicators, summary statistics, and correlations
- Table 1 (summary of scores) — Education:
  - PSPE mean 1.01; Coefficient of variation 0.34; Minimum 0.25; 1st quartile 0.77; Median 1.03; 3rd quartile 1.25; Maximum 2.38.
  - PSEE mean 1.21; Coefficient of variation 0.62; Minimum 0.26; 1st quartile 0.79; Median 1.08; 3rd quartile 1.44; Maximum 15.19.
  - DEAE mean 0.32; Coefficient of variation 0.65; Minimum 0.03; 1st quartile 0.16; Median 0.27; 3rd quartile 0.43; Maximum 1.00.
- Table 1 — Health:
  - PSPH mean 1.15; Coefficient of variation 0.28; Minimum 0.23; 1st quartile 0.93; Median 1.23; 3rd quartile 1.41; Maximum 1.81.
  - PSEH mean 1.38; Coefficient of variation 0.64; Minimum 0.53; 1st quartile 0.89; Median 1.10; 3rd quartile 1.54; Maximum 8.04.
  - DEAH mean 0.34; Coefficient of variation 0.63; Minimum 0.08; 1st quartile 0.18; Median 0.25; 3rd quartile 0.45; Maximum 1.00.
- Spearman rank correlations (Table 2) — selected coefficients:
  - Education PSP vs Health PSP correlation coefficient 0.81 (PSP cross-sector correlation).
  - Education PSE cross-sector correlation 0.37.
  - DEA correlation between education and health is negative (reported as -0.03 in parts of table and negative/low for DEA cross-sector).
  - Within education, PSPE and PSEE correlation 0.70; PSPE and DEAE correlation 0.81; PSEE and DEAE correlation 0.83 (Table 2 entries).
- Observed patterns:
  - Education and health outputs are more strongly correlated than their efficiency measures.
  - Coefficient of variation for efficiency measures is about twice as large as for performance measures.
  - Performance is positively correlated with efficiency in education but negatively correlated in health (noted in text).

### Trends and graphical findings
- Figure 2 highlights:
  - Performance is highest in advanced economies in both education and health.
  - Efficiency is highest in advanced economies for education but lowest in advanced economies for health.
  - Both performance and efficiency have trended upwards on average across income groups.
  - Education spending shows larger fluctuation relative to GDP over a longer horizon, contributing to volatility in PSE/DEA scores while outcomes remain more stable.

### Determinants of performance and efficiency — empirical strategy
- Three-stage empirical approach:
  - Stage 1: Univariate regressions for all potential determinants.
  - Stage 2: Multivariate regressions combining determinants significant in univariate regressions.
  - Stage 3: Tested-down regressions keeping regressors significant at least at the 10 percent level.
- Regressions reported: tested-down regressions summarized in Table 3; signs of significant regressors across specifications summarized in Table 4.
- Regression frameworks: random effects and fixed effects specifications; endogeneity tested using augmented regressions and two-step system GMM (Arellano and Bover (1995); Blundell and Bond (1998)). Windmeijer (2005) correction applied where appropriate.

### Main empirical findings — economic determinants
- Spending and efficiency:
  - Larger education and health spending reduces public sector efficiency in each sector.
  - In education, spending does not bear a significant relationship with public sector performance.
  - Augmented regression tests:
    - Do not suggest endogeneity of spending in education efficiency regressions.
    - Do suggest endogeneity for health efficiency (PSE and DEA), implying lower efficiency may lead to higher spending.
  - Two-step system GMM:
    - For PSE: significant negative effect of health spending confirming OLS result.
    - For DEA: coefficient on health expenditure becomes insignificant.
  - Conclusion: nexus between higher expenditure and lower efficiency is robust, though causality in health remains somewhat unclear.
- Income per capita:
  - Results provide more support for the hypothesis that higher income is associated with better performance (several significant positive signs in tested-down regressions), though some negative signs appear in univariate/multivariate settings and disappear when tested down.
  - Augmented regression tests suggest income per capita effect is exogenous.
- Commodity exporter dummy:
  - Significant in univariate regressions but becomes insignificant in tested-down multivariate specifications.
- Inflation:
  - Expected negative effect reversed in health sector (positive effect in most cases), possibly due to budget process mechanics (unexpected inflation lowering expenditure/GDP denominator and improving measured efficiency temporarily).
- Openness/trade liberalization:
  - Not significant in final tested-down specifications.

### Main empirical findings — institutional determinants
- Accountability:
  - Higher government accountability consistently increases education performance.
  - Less robust and sometimes negative association with health sector efficiency.
- Corruption control:
  - Robust evidence that better control of corruption increases both efficiency and performance in education and health.
  - Results for corruption control appear in tested-down regressions when controlling for country effects and per capita income.
- Other institutions:
  - Democracy, regime durability, social infrastructure, and schooling largely not significant in tested-down specifications for the sectors examined (with some exceptions in certain specifications).

### Main empirical findings — demographic and geographic determinants
- Age structure:
  - Education sector:
    - Younger population (higher share <14) reduces efficiency; older population (higher share >65) increases efficiency.
    - Performance also appears to suffer with younger population (possible crowding effects).
  - Health sector:
    - Share of younger population does not matter much.
    - Some evidence that a larger elderly population positively affects health performance and efficiency.
- Population density:
  - Positive effects on performance and efficiency consistent with economies of scale in service provision.
- Fractionalization and climate:
  - Linguistic fractionalization not significant in tested-down specifications.
  - Malaria index and distance-from-equator (climate proxy) do not show significant effects in final tested-down regressions, despite potential theoretical channels.

### Policy implications and conclusions
- Two key policy implications emphasized:
  - Efficiency declines with the level of spending in education and health; in education, spending is not significantly associated with better performance. Policy takeaway: increasing funding alone often fails to yield expected improvements unless accompanied by efficiency-enhancing reforms.
  - Institutional improvements matter: government accountability and controlling corruption can improve government efficiency and performance in education and health.
- Governments face demographic and geographic constraints that matter for performance and efficiency but are largely exogenous and must be managed rather than changed.

*Source: Authors’ calculations.*

---


_Source: https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2008/_wp08228.pdf_