Decentralization Without Representation (or Mobility): Implications for Rural Public Service Delivery Tewodaj Mogues International Food Policy Research Institute (IFPRI) (Paper with Katrina Kosec, IFPRI) WIDER Development Conference: Public Economics for Development 5-6 July 2017, Maputo, Mozambique
Introduction Decentralization and Public Service Provision Decentralization can create more efficient and effective public service delivery through several mechanisms: Inter-jurisdictional competition: Labor and capital sort themselves into localities that provide their preferred package of services, taxes, and regulations, inducing competition between governments (Tiebout, 1956; Brennan and Buchanan, 1980; Hatfield, 2012) Yardstick competition: Voters can better monitor local policymakers performance against those in similar jurisdictions, and accordingly elect into/out of office (Besley and Case, 1995; Seabright, 1996) Central government benchmarking: Central government can monitor and reward/ sanction local leaders (e.g. through transfers, promotion, etc) based on relative performance Asymmetric information: Local officials have better information about local conditions and thus can target public funds more efficiently (Hayek, 1945)
Introduction Study Context: Ethiopia A predominantly poor, agrarian, and autocratic country In this context, we expect: Limited inter-jurisdictional competition (due to limited mobility) Limited yardstick competition due to voters (given nondemocratic setting); central government benchmarking is still possible Information asymmetry is still fully operative We analyze two types of public services: Agricultural service delivery (high reported priority for central government) as a type of 'productive service provision' Drinking water service delivery (low priority for central government; high priority for citizens (IFPRI-WB, 2010)) as a type of 'social service provision'
Introduction Preview of the Results How does decentralization impact service delivery in the absence of yardstick and inter-jurisdictional competition (i.e. when there is limited mobility and limited political competition)? We employ a spatial regression discontinuity approach, with geographic and temporal placebo tests and other robustness checks Decentralization improves services of high priority to the central government: Access to agricultural extension services (farmer training) Access to local government meetings on farming practices Use of modern agricultural inputs No impact on services of lower priority to the central government: Drinking water access, distance, and quality Decentralization decreases the alignment of policy priorities between local policymakers and citizens, but increases alignment among local government policymakers
Context Agricultural and Water Service Provision in Ethiopia Agriculture 85% of the labor force and 84% of exports A primary concern for Ethiopian government, reflected in its Agricultural Development-Led Industrialization (ADLI) strategy; share of expenditure devoted to agriculture among 4 highest in Africa Drinking water Access improves labor productivity (health human capital) but also contributes to non-material well-being (intrinsic value of good health) A relatively low priority for government (share of spending below African median) (MOFED, 2006; van Ginneken et al., 2011) High priority sector for citizens: Individuals in our dataset ranked drinking water as No. 1 concern out of 9 areas of service provision
Context Decentralization Process in Ethiopia Decentralization to the district level in 2001, in 4 out of the 9 regions Reasons motivating selection of regions: Decentralized regions were less remote/ more accessible via road networks, and had higher agricultural potential (i.e. more favorable climatic and geographic conditions) Our identification strategy addresses these factors that were used to select regions into decentralization Administrative and fiscal features: Transfer of block-grants to districts Greater retention by districts of their internally generated revenues Districts appoint their own technical staff Districts make personnel decisions for district bureaucratic positions Greater service provision responsibilities in agriculture, rural infrastructure, and primary health and education
Theoretical Model Model A government of type k {L, C} (local or central) chooses share of budget 0 δ k 1 to allocate to agricultural projects, allocating rest to water projects The probability that a selected project is of good quality is a k with 0 a C < a L 1. With the budget normalised to 1, government thus produces output O 1 k = δa k in sector 1 (agriculture) and O 2 k = (1 δ)a k in sector 2 (water). Citizen i s utility U i (Y, h) depends on income Y = f (Ok 1, O2 k ) taxed at the rate τ, and non-material well-being (e.g. mental health) h = g(ok 2) Government s utility is U k (Y ) = τα k Y, where α C = 1, α L (0, 1) Functional forms for: income, Y = O 1 k + (O2 k )π ; health, h = O 2 k ; and citizen s utility, U i = [(1 τ)y ] q h q ; where {π, q} (0, 1) Then, U i = {(1 τ)[δa k + ((1 δ)a k ] π } q [(1 δ)a k ] q and U k = τα k {δa k + [(1 δ)a k ] π }
Theoretical Model Model First order condition for the citizen: { q(1 τ) q (1 δ) q 1 a 2q k [δ + aπ 1 k (1 δ) π ] q 1} {1 2δ (1 + π)a π 1 k (1 δ) π } = 0 ] First order condition for a government of type k: T k [a k πaπ k (1 δ) = 0, 1 π which leads to the optimal allocation rule for both types of government: δk = 1 π 1 π 1 a k First prediction of the model: Government always allocates more to agriculture and less to water than citizen s optimal choice, i.e. δ k > δ i Second prediction: decentralization leads to greater [smaller] share of budget allocated to agriculture [water] than centralization: δ L> δ C Third prediction: decentralization leads to more agricultural services and unchanged water services, i.e. OL 1 and O2 > O1 C L = O2 C
Empirical Strategy Data Ethiopia Governance and Rural Services Survey (2008 09) Individual, household, and village level data (1,899 individuals in 1,117 households in 32 villages) Captures supply of, access to, satisfaction with, and prioritization of public services (focus is agriculture and drinking water) Includes 4 district pairs; each pair straddles a regional boundary For 3 pairs: one district in a decentralized region, one not 4th pair: both districts in decentralized regions (used in falsification test) Other data (for temporal placebo analyses): Agricultural Sample Survey 1999/2000 Demographic and Health Survey 2000
Empirical Strategy Study Area Full sample within 75km of regional boundary; 84% within 25km
Empirical Strategy Empirical Strategy and Econometric Specification RD approach: Factors that vary across regions and may influence service delivery (e.g. soil quality, climatic conditions) change smoothly over space Following Holmes (1998) (with adjustments based on Gelman and Imbens, 2014), we estimate: { 2 } q ivr = δd r + W i β + [γ 1,j A 1 + γ 2,j A 2 + γ 3,j A 3]Bv j + j=0 { 2 } [γ 4,j A 1 + γ 5,j A 2 + γ 6,j A 3]Bv j S v + ɛ i j=0 where qivr is a measure of rural service delivery quality for household or individual i in village v in region r D r is a dummy for local level decentralization S v = distance of village v to region border ( forcing variable); B v = (continuous) boundary marker for village v; A = fixed effects for the three district-pairs W i are control variables of household/individual characteristics
Results Balance on Key Climatic Variables at Policy-change Borders Mean monthly temperature (degrees celsius) Annual rainfall (mm) 75km 50 km 25 km 75 km 50 km 25 km All policy-change borders in the country Decentralized 0.0392-0.0428-0.0456 34.0091 18.1977 9.4391 (0.173) (0.162) (0.152) (34.529) (31.475) (29.171) Observations 411,030 310,954 178,321 410,023 310,378 178,246 Policy-change borders in the study area Decentralized -0.1488-0.1210-0.0260 78.1310 42.7282 6.8850 (0.121) (0.096) (0.025) (60.932) (39.102) (11.527) Observations 22,931 20,248 13,460 22,918 20,243 13,453 Notes: Standard errors are clustered at the zone level in the Ethiopia-wide analysis and at the district level in the analysis on the study area, and appear below the coefficient in parentheses. Coefficient estimates statistical significance is indicated at the 10, 5, and 1 percent levels with *, **, and ***, respectively. Source: CSA (2007).
Results Access to and Learning from Agricultural Extension Visited a demonstration plot, home, or research station Learned new farming practices from district agricultural extension agent/expert Decentralized 0.5486*** 0.5987*** 0.2064*** 0.1896*** (0.007) (0.035) (0.021) (0.027) A, B, S, interactions Yes Yes Yes Yes Male 0.0263** 0.0213** (0.011) (0.010) Literate 0.0451** 0.0084 (0.019) (0.010) Farming household -0.0305 0.0214*** (0.027) (0.007) Household size -0.0006 0.0012 (0.002) (0.001) Land holder 0.0348*** -0.0045 (0.011) (0.016) Age -0.0008 0.0005 (0.002) (0.002) Age squared 0.0000-0.0000 (0.000) (0.000) Obs. 1,436 1,433 1,440 1,437 R 2 0.065 0.093 0.072 0.079 Notes: Standard errors are clustered at the village level, and appear below the coefficient in parentheses. Coefficient estimates statistical significance is indicated at the 10, 5, and 1 percent levels with *, **, and ***, respectively.
Results Attendance of Local Gov't Agriculture-Related Meetings Attended a community meeting Attended an important Attended an important to discuss agricultural issues district government-organized village government organized meeting on agricultural issues meeting on agricultural issues Decentralized 0.4918*** 0.3973*** 0.2681*** 0.1972*** 0.2701*** 0.2200*** (0.038) (0.061) (0.028) (0.059) (0.017) (0.035) A, B, S, interactions Yes Yes Yes Yes Yes Yes Male 0.1795*** 0.1174*** 0.0875*** (0.031) (0.023) (0.020) Literate 0.1155*** 0.1040*** 0.0183 (0.022) (0.025) (0.018) Farming household 0.1179*** 0.1123*** 0.0150 (0.035) (0.029) (0.021) Household size -0.0026-0.0017 0.0014 (0.006) (0.006) (0.003) Land holder -0.0108-0.0173 0.0308** (0.042) (0.040) (0.014) Age 0.0154*** 0.0097*** 0.0050* (0.003) (0.003) (0.002) Age squared -0.0002*** -0.0001*** -0.0001** (0.000) (0.000) (0.000) Obs. 1,435 1,432 1,435 1,432 1,435 1,432 R 2 0.260 0.348 0.224 0.288 0.106 0.149 Notes: Standard errors are clustered at the village level, and appear below the coefficient in parentheses. Coefficient estimates statistical significance is indicated at the 10, 5, and 1 percent levels with *, **, and ***, respectively.
Results Use of Modern Agricultural Inputs Use of urea fertilizer Use of improved seed Use of pesticide Decentralized 0.2342*** 0.1429* -0.0782 (0.045) (0.071) (0.095) A, B, S, interactions Yes Yes Yes Male head -0.0198 0.0191 0.0694** (0.019) (0.015) (0.033) Literate head 0.0209* 0.0128 0.0013 (0.011) (0.020) (0.018) Farming household -0.0091 0.0128 0.1127** (0.018) (0.025) (0.045) Household size 0.0031 0.0021 0.0024 (0.003) (0.003) (0.002) Land holder -0.0210 0.1174*** -0.0184 (0.014) (0.039) (0.039) Age of head -0.0021-0.0018 0.0009 (0.002) (0.003) (0.002) Age of head squared 0.0000 0.0000-0.0000 (0.000) (0.000) (0.000) Obs. 837 837 837 R 2 0.098 0.302 0.687 Notes: Standard errors are clustered at the village level, and appear below the coefficient in parentheses. Coefficient estimates statistical significance is indicated at the 10, 5, and 1 percent levels with *, **, and ***, respectively.
Results Drinking Water Source and Quality Dep. variable: Use of improved drinking Time to get to water Dummy satisfied with water source during the: (min.) during the: water quality during the: wet season dry season wet season dry season wet season dry season (1) (2) (3) (4) (5) (6) Decentralized 0.3009*** 0.0423 6.2199 1.3031-0.6580** -0.2864 (0.071) (0.069) (6.440) (14.957) (0.237) (0.201) Obs. 837 837 738 742 824 824 R 2 0.153 0.103 0.568 0.641 0.597 0.575 Notes: Columns (1) - (4) are household-level outcomes and include our full set of household-level regression controls, while columns (5) - (6) are individual-level outcomes and include our full set of individual-level regression controls. Standard errors are clustered at the village level, and appear below the coefficient in parentheses. Coefficient estimates statistical significance is indicated at the 10, 5, and 1 percent levels with *, **, and ***, respectively.
Results Matching of Public Service Priorities Between Citizens & Local Leaders, & Among Leaders Priority-matching correlation based on: Kendall s τ-b Spearman s ρ Pearson s coefficient (1) (2) (3) (4) (5) (6) Panel A: Citizen - leader priority matching Citizen Village -0.8380*** -0.8500*** -0.8597*** -0.8712*** -1.2514*** -1.2984*** (HH chair (0.114) (0.116) (0.107) (0.110) (0.183) (0.176) head) Village council -0.4618-0.5470*** -0.4626-0.5500*** -0.4673*** -0.5468*** member (0.000) (0.070) (0.000) (0.071) (0.069) (0.097) Panel B: Inter-leader priority matching Village Village council 0.7676*** 0.8177*** 0.8295*** 0.8883*** 0.8066*** 0.8009*** chair member (0.182) (0.196) (0.206) (0.221) (0.031) (0.040) Notes: All presented estimates are those for decentralization. The specification of each underlying regression is identical to that of the analysis of public services, presented in the earlier tables. Regressions use doubly-censored tobit estimation given the censored nature of the matching correlations. Standard errors are clustered at the village level, and appear below the coefficient in parentheses. Coefficient estimates statistical significance is indicated at the 10, 5, and 1 percent levels with *, **, and ***, respectively.
Robustness Robustness and Placebo Tests - Placebo Test 1: Shows No Impact of Decentralization in year prior to implementation (Ethiopia AgSS, 1999/2000; DHS 2000) - Placebo Test 2: Shows No Impact of Decentralization in 2009 at Placebo Boundary (Both Sides Decentralized) - Robustness to Quadratic Polynomial in Longitude and Latitude (Dell 2010) - Robustness to Only Comparing Households in the Same Province During All of 1942 95
Conclusions Conclusions How does decentralization impact service delivery in the absence of yardstick and inter-jurisdictional competition (i.e. when there is limited mobility and limited political competition)? We contribute to a very limited theoretical and empirical literature on how decentralisation in authoritarian systems affects resource allocation Decentralization improves services of high priority to the central government: Access to agricultural extension services (farmer training) Access to meetings and information on new farming practices Use of modern agricultural inputs No impact on services of lower priority to the central government: Drinking water access, distance, and quality Decentralization decreases the alignment of policy priorities between local government policymakers and citizens, but increases alignment among local policymakers
Robustness Placebo Test 1: Shows No Impact of Decentralization In Year Prior to Implementation (Ethiopia AgSS, 1999 2000) Panel A: 50 km bandwidth (1) (2) (3) (4) (5) (6) Uses Amount of Uses improved Amount of Uses Amount of fertilizer fertilizer seed improved seed pesticides pesticides Decentralized 0.1715* 4,559.2702 0.0214 457.5324 0.0030-1,045.2489 (0.091) (6,793.475) (0.031) (705.116) (0.018) (2,324.785) Observations 10,160 10,160 10,160 10,160 10,160 10,160 R 2 0.430 0.298 0.198 0.0810 0.194 0.0984 Panel B: 75 km bandwidth Decentralized 0.2242** 7,484.8175 0.0313 594.4499 0.0181 754.8277 (0.088) (6,313.329) (0.032) (614.297) (0.019) (2,234.212) Observations 11,774 11,774 11,774 11,775 11,774 11,774 R 2 0.437 0.289 0.194 0.0759 0.190 0.0867
Robustness Placebo Test 1: Shows No Impact of Decentralization In Year Prior to Implementation (Ethiopia DHS, 2000) Panel A: 50 km bandwidth (1) (2) (3) Dummy Piped Dummy Improved Time to get to main water source water source water source (min.) Decentralized -0.0576-0.0631-39.0352 (0.035) (0.082) (26.091) Observations 4,027 4,027 3,993 R 2 0.261 0.407 0.529 Panel B: 75 km bandwidth Decentralized -0.0647-0.0489-41.5457* (0.039) (0.092) (20.850) Observations 4,556 4,556 4,539 R 2 0.239 0.407 0.534 Notes: All coefficients reflect the coefficient on a dummy for decentralization in a regression of the listed outcome variable on the decentralization dummy and the full set of control variables from Eq. (1). Standard errors are clustered at the zone level, and appear below the coefficient in parentheses. Coefficient estimates statistical significance is indicated at the 10, 5, and 1 percent levels with *, **, and ***, respectively.
Robustness Placebo Test 2: Exploit the one District Pair in Which Both Districts are Decentralized (Use 2009, Post-D Data)
Robustness Placebo Test 2: Shows No Impact of Decentralization in 2009 at Placebo Boundary (Both Sides Decentralized) Dependent Variable Coeff. on S.E. on Obs. R 2 placebo placebo decent. decent. dummy dummy Visited a demonstration- -0.1499 (0.101) 457 0.117 plot, home, or research station Learned new farming practices -0.3404 (0.331) 459 0.325 from district agricultural extension agent/expert Attended a community -0.3201 (0.213) 458 0.598 meeting to discuss agricultural issues Attended meeting district government 0.0914 (0.174) 458 0.471 on agricultural issues organized by: village government -0.2549 (0.255) 458 0.354 Use of fertilizer -0.1500 (0.386) 280 0.590 Use of improved seed -0.3238 (0.211) 280 0.323 Use of pesticides -0.1380 (0.303) 280 0.287 Notes: All coefficients reflect the coefficient on a dummy for decentralization when estimating Eq. (1). Standard errors are clustered at the village level, and appear below the coefficient in parentheses. Coefficient estimates statistical significance is indicated at the 10, 5, and 1 percent levels with *, **, and ***, respectively.
Robustness Placebo Test 2: Shows No Impact of Decentralization in 2009 at Placebo Boundary (Both Sides Decentralized) Dependent Variable Coeff. on S.E. on Obs. R 2 placebo placebo decent. decent. dummy dummy Use of improved Wet season -0.2177 (0.204) 280 0.301 drinking water during the: Dry season -0.2358 (0.202) 280 0.297 Time to get Wet season -23.5148 (25.116) 221 0.621 to water (min.) during the: Dry season -17.0204 (26.577) 221 0.657 Dummy satisfied Wet season -0.2109 (0.405) 220 0.848 with water quality during the: Dry season -0.5892 (0.514) 221 0.841 Notes: All coefficients reflect the coefficient on a dummy for decentralization when estimating Eq. (1). Standard errors are clustered at the village level, and appear below the coefficient in parentheses. Coefficient estimates statistical significance is indicated at the 10, 5, and 1 percent levels with *, **, and ***, respectively.
Robustness Robustness to Quadratic Polynomial in Longitude and Latitude (Dell 2010): Agriculture Outcomes Dependent Variable Coeff. S.E. Obs. R 2 Visited a demonstration- 0.0751* (0.041) 1,433 0.077 plot, home, or research station Learned new farming practices 0.1298* (0.073) 1,437 0.079 from district agricultural extension agent/expert Attended a community 0.2526* (0.139) 1,432 0.339 meeting to discuss agricultural issues Attended meeting district government 0.0181 (0.153) 1,432 0.264 on agricultural issues organized by: village government 0.2535*** (0.070) 1,432 0.143 Use of fertilizer 0.0722* (0.039) 837 0.084 Use of improved seed 0.4762*** (0.154) 837 0.266 Use of pesticides 0.2791*** (0.083) 837 0.672 Notes: Standard errors are clustered at the village level, and appear below the coefficient in parentheses. Fertilizer uses refers to commercial (urea) fertilizer. Coefficient estimates statistical significance is indicated at the 10, 5, and 1 percent levels with *, **, and ***, respectively.
Robustness Robustness to Quadratic Polynomial in Longitude and Latitude (Dell 2010): Water Outcomes Dependent Variable Coeff. S.E. Obs. R 2 Use of improved Wet season -0.2626** (0.104) 837 0.138 drinking water during the: Dry season -0.1834* (0.091) 837 0.090 Time to get Wet season -30.3484** (13.015) 738 0.525 to water (min.) during the: Dry season -26.3105 (20.160) 742 0.570 Dummy satisfied Wet season 0.0681 (0.243) 824 0.575 with water quality during the: Dry season 0.1977 (0.250) 824 0.549 Notes: Standard errors are clustered at the village level, and appear below the coefficient in parentheses. Statistical significance is indicated at the 10, 5, and 1 percent levels with *, **, and ***, respectively. Source: Gender and Rural Services Survey (2008 09).
Robustness Robustness to Only Comparing Households in the Same Province During All of 1942 95 All households in two of our three district pairs (i.e. 2/3 of sample observations) were in the same province for all of 1942 1995, only being separated into different regions (one of them to be decentralized and one not) in 1995 (6 years prior to decentralization) To ensure that our results are attributable to decentralization, we show that they still hold when restricting attention to only these two district pairs
Robustness Historic and Current Top-Tier Administrative Divisions
Robustness Robustness to Only Comparing Households in the Same Province During All of 1942 95: Agriculture Outcomes Dependent Variable Coeff. S.E. Obs. R 2 Visited a demonstration- 0.6210*** (0.047) 963 0.110 plot, home, or research station Learned new farming practices 0.1610*** (0.024) 966 0.056 from district agricultural extension agent/expert Attended a community 0.4222*** (0.065) 961 0.331 meeting to discuss agricultural issues Attended an district government 0.2302*** (0.064) 961 0.268 important meeting on agricultural issues organized by: village government 0.2490*** (0.026) 961 0.157 Use of fertilizer 0.2573*** (0.055) 557 0.103 Use of improved seed -0.0851 (0.119) 557 0.695 Use of pesticides 0.0674 (0.073) 557 0.359 Notes: Standard errors are clustered at the village level, and appear below the coefficient in parentheses. Statistical significance is indicated at the 10, 5, and 1 percent levels with *, **, and ***, respectively. Source: Gender and Rural Services Survey (2008 09).
Robustness Robustness to Only Comparing Households in the Same Province During All of 1942 95: Water Outcomes Dependent Variable Coeff. S.E. Obs. R 2 Use of improved Wet season 0.2464*** (0.028) 557 0.174 drinking water during the: Dry season -0.0209 (0.034) 557 0.106 Time to get Wet season 9.9121 (7.339) 475 0.598 to water (min.) during the: Dry season 10.5406 (16.138) 479 0.681 Dummy satisfied Wet season -0.7846*** (0.214) 527 0.629 with water quality during the: Dry season -0.3509* (0.179) 527 0.603 Notes: Standard errors are clustered at the village level, and appear below the coefficient in parentheses. Statistical significance is indicated at the 10, 5, and 1 percent levels with *, **, and ***, respectively. Source: Gender and Rural Services Survey (2008 09).