Part I: Technical Report. Research Report. Asep Suryahadi Wenefrida Widyanti Daniel Perwira Sudarno Sumarto. Chris Elbers. Menno Pradhan.

Transcription

1 Research Report Asep Suryahadi Wenefrida Widyanti Daniel Perwira Sudarno Sumarto (SMERU ) Chris Elbers (Vrije Uniersity, Amsterdam) Menno Pradhan (World Bank) Part I: Technical Report May 2003 The findings, iews, and interpretations published in this report are those of the authors and should not be attributed to the SMERU Research Institute or any of the agencies proiding financial support to SMERU. For further information, please contact SMERU, Phone: ; Fax: ; smeru@smeru.or.id; Web:

2 Deeloping a Poerty Map for Indonesia: An Initiatory Work in Three Proinces Asep Suryahadi, Wenefrida Widyanti Daniel Perwira, Sudarno Sumarto The SMERU Research Institute Chris Elbers Vrije Uniersiteit, Amsterdam Menno Pradhan The World Bank Part I: Technical Report The SMERU Research Institute Jakarta May 2003

3 Table of Contents Abstract ii I. Introduction 1 II. The Method 4 A. The Consumption Model 4 B. The Estimators 4 III. Data Sources 6 IV. Model Application 7 A. Stage 1: Matching Variables in the Surey and the Census 7 B. Stage 2: Selecting Explanatory Variables for the Consumption Model 7 C. Stage 3: Estimating the Consumption Model 8 D. Stage 4: Simulations on Census Data 9 E. Stage 5: Calculation of Poerty and Inequality Indicators 11 V. Poerty and Inequality Maps 12 A. Poerty Estimates and Their Standard Errors 12 B. District, Subdistrict, and Village Poerty Maps 17 C. Examples for Further Applications 27 D. Conformity with Other Measures 30 VI. Concluding Remarks 32 Appendix 33 References 77 i

4 Deeloping a Poerty Map for Indonesia: An Initiatory Work in Three Proinces Asep Suryahadi, Wenefrida Widyanti Daniel Perwira, Sudarno Sumarto The SMERU Research Institute Chris Elbers Free Uniersity, Amsterdam Menno Pradhan The World Bank Abstract This report presents the results of a pilot study to apply a recently deeloped technique for obtaining high-resolution poerty maps, using data from three proinces in Indonesia: Jakarta, East Jaa, and East Kalimantan. The purpose of this pilot study is to try out the applicability of the poerty mapping method gien the aailable data in Indonesia and, furthermore, to test the feasibility of deeloping a poerty map for the whole country at arious administratie leels (proince, district, subdistrict, and illage). The report is consisted of two parts. Part I is a technical report describing the steps that hae been implemented in the exercise and discussions on the results. Part II presents the results of the exercise in the forms of tables of poerty and inequality point estimates and standard errors at the proincial, district, subdistrict, and illage leels for the three proinces. The results indicate that the currently aailable data in Indonesia are sufficient to deelop a poerty map with reasonable standard errors, at least for the proincial, district, and subdistrict leels. Meanwhile, the results for illage leel need to be used with caution as the standard errors of the estimates for a large fraction of the illages are relatiely large. Oerall, the results appear to support the extension of the method s application to the rest of the country. ii

5 I. Introduction Experience shows that locating the poor is one of the most crucial and difficult problems in the implementation of programs aimed at targeting the poor. 1 In Indonesia, a country which is ery large in size and where poerty statistics are reliable only up to the proincial-urban/rural leel, geographic targeting of the poor is een more difficult. Figure 1 shows the poerty map of Indonesia based on the aailable estimates of poerty rates at the proincial leel. 2 While this map is useful for identifying poerty differential across broad regions for example, it shows that the proinces at the eastern part of Indonesia are the poorest regions in the country it is less useful for the purposes of practical budget allocation or program targeting. As poerty reduction efforts will continue to be an important endeaor in Indonesia een long into the future, there is clearly a need to deelop tools for more effectie geographic targeting than those that hae been used in the past. Ideally, geographic targeting would be based on a description of poerty incidence and other indicators of economic welfare at small areas or low administratie leels. More generally, the analysis of poerty and welfare in a country could benefit tremendously from detailed and disaggregated data on the distribution of economic welfare. In the context of Indonesia, administratie leels go from the national leel all the way down to the illage leel (desa/kelurahan). 3 One could of course obtain illage leel information on the distribution of economic welfare by carrying out a household surey with a sample which is representatie for all illages in Indonesia. Howeer, with a total of almost 70,000 illages in Indonesia, such a household surey is prohibitiely huge and expensie. For comparison, the current poerty statistics in Indonesia are based on the consumption module of the National Socio-Economic Surey (SUSENAS), which has a sample size of around 65,000 households. 1 See Bigman and Fofack (2000), Raallion (2000), an de Walle (1998). 2 The estimates of poerty rates are taken from Pradhan et al. (2001). 3 The hierarchy of goernment administratie units in Indonesia below the central goernment is proinces (propinsi), districts (kabupaten) or cities (kota), subdistricts (kecamatan), and illages. A illage which is located in a rural area is called a desa, while a illage which is located in an urban area is called a kelurahan. 1

6 Figure 1. Poerty Map of Indonesia Based on Proincial Poerty Rates Fortunately, as a result of recent methodological adances in this area, a new methodology has been deeloped to estimate such description from statistical data collections that are normally aailable in a country. 4 The core of the method is to combine the information obtained from a household surey with the information collected through a population census. A household surey usually collects ery detailed information on household characteristics, including consumption leel, but the coerage is generally limited and only representatie at a relatiely large geographical unit. On the other hand, a population census has a complete coerage of all households, but usually collects ery limited information on household characteristics. Hence, the method tries to combine the adantage of detailed information on household characteristics obtained from a household surey with the complete coerage of a population census. Essentially, the method imputes estimates of per capita consumption for each household in the population. This is done by applying obsered correlation patterns between household characteristics and household per capita consumption to census 4 See Elbers et al. (2001), Hentschel et al. (2000). 2

7 data on household characteristics. The correlation patterns are estimated on the basis of household surey data. This study is a pilot and the first attempt to apply the method in Indonesia. The objectie is to obtain estimates of poerty incidence at geographical units smaller than a proince-urban/rural area, which is the lowest leel of aggregation for which reliable (but still ery imprecise) poerty statistics are currently aailable. The study is planned to be conducted in two stages. The first stage is the current pilot study to test the feasibility of the method in the context of Indonesia. It uses data from three proinces out of 32 proinces in Indonesia: East Kalimantan, Jakarta, and East Jaa. The results of this pilot study are summarized in this report. The pilot study has been carried out by the SMERU Research Institute. The next scale will entail a larger-scale application to Indonesia s remaining proinces and will be carried out by Statistics Indonesia (BPS), building on the experience gained during the pilot phase. The rest of the report is organized as follows. Chapter two discusses in brief the method used to obtain these estimates. Chapter three discusses the sources of data utilized in this exercise. Chapter four discusses the model application and the procedures for implementing it. Chapter fie presents the results of the exercise in the forms of poerty and inequality maps from the proince leel down to the illage leel. Finally, chapter six proides the concluding remarks. In addition, a separate Part II of this report proides the complete poerty mapping results for the three pilot proinces in table form. 3

8 II. The Method The method used in this study basically inoles a two-step procedure. First, a model of consumption determination is estimated using the data from household surey. In the second step, the parameters estimated in the first step are then transferred to the data from the population census to simulate the consumption leel of each and eery household enumerated in the population census. The simulated household consumption is then used as the basis for calculating poerty and other welfare indicators. A. The Consumption Model Following Elbers et al. (2001, 2002), the empirical model of household consumption is defined as: ln y = E( yν x ) + u (1) h h h h where ln yh is the logarithm of per capita consumption of household h in illage, x h is a ector of obsered characteristics of this household (including illage leel ariables), and u h is the error term. Note that u h is uncorrelated with x h. This model is simplified by using a linear approximation to the conditional expectation E( y ν h xh ) and decomposing u h into uncorrelated terms: uh = η + ε (2) h where η represents a illage leel error term common to all households within the illage, and ε h is a household specific error term. It is further assumed that the η are uncorrelated across illages and the ε h are uncorrelated across households. With these assumptions, equation (1) reduces to ln y = β + η + ε h. (3) h x h Estimation of the parameters underlying this equation, in particular the ector of parameters β and the distributional characteristics of the error terms, can be done by using standard tools from econometric analysis (see Elbers et al., 2002). B. The Estimators The consumption model specification in equation (3) allows for an intra-illage correlation in the error terms. Household income or consumption is certainly affected by the location where the household lies. Een though x h has some ariables representing illage leel characteristics, it is quite plausible that some of the location effects will remain unexplained. The consequence of failing to take into account this within-illage correlation of the error terms can result in biased welfare 4

9 5 estimates (in particular for inequality indicators) and will generally lead to underestimation of the standard errors of welfare estimates. Take illage aerages oer equation (2): + = u ε η (4) where a subscript indicates an aerage oer the index. Since the two error components are uncorrelated, then: [ ] ( ) ( ) ar ar E + = + = u τ σ ε η η (5) An unbiased estimator for 2 η σ can be defined as: ( ) ( ) ( ) = j j j j j j w w w w w w w u τ σ η (6) where: ( ) ( ) = h h n n ε ε τ (7) and w is a set of non-negatie weights summing to one. Elbers et al. (2001, 2002) gie the following formula for the sampling ariance of 2 ˆη σ : ( ) + b u a ar ar ar τ σ η, n b a τ τ σ τ σ η η (8) where = j j j w w w a ) (1 and = j j j w w w w b ) (1 ) (1.

10 III. Data Sources Four sources of data are used in this study: (i) Consumption Module SUSENAS 1999, (ii) Core SUSENAS 1999, (iii) Population Census 2000, and (i) PODES (Village Potential) In the consumption model estimation, the data on household consumption is obtained from the Consumption Module SUSENAS, the data on household characteristics is obtained from the Core SUSENAS, and the data on illage-leel characteristics is obtained from the PODES and illage means of the population census. SUSENAS, the National Socio-Economic Surey, is a nationally representatie household surey, coering all areas of the country. A part of the SUSENAS is conducted eery year in the month of February, collecting information on the characteristics of oer 200,000 households and oer 800,000 indiiduals. This part of the SUSENAS is known as the Core SUSENAS. Another part of the SUSENAS is conducted eery three years, specifically collecting information on ery detailed consumption expenditure from around 65,000 households. These households are a randomly selected subset of the 200,000 households in the Core SUSENAS sample of the same year. This consumption module part of the SUSENAS is commonly known as the Module SUSENAS. Population census 2000 is the fifth population census conducted in Indonesia after independence. The preious censuses were conducted in 1961, 1971, 1980, and The 2000 population census was conducted in the month of June, coering all people liing in the territory of Indonesia, including foreigners. Data on 15 demographic, social, and economic ariables at both indiidual and household leels were collected in the census. PODES, meanwhile, is a complete enumeration of illages throughout Indonesia. The information collected through this surey only includes illage characteristics such as size of area, population, infrastructure, and local industries characteristics. The questionnaires are filled out by the local subdistrict officials who are responsible for collecting statistical data (mantri statistik). The information is obtained from official illage documents as well as interiews with illage officials. The PODES surey is usually conducted three times eery ten years, usually prior to and as a preparation for an agricultural census, an economic census, or a population census. A PODES surey was conducted in the months of September and October 1999 as a preparation for the population census in In total, the 1999 PODES enumerates 68,783 illages. 5 5 Officially it is called PODES

11 IV. Model Application This chapter outlines the stages and procedures implemented in applying the model to obtain poerty maps for three proinces: East Kalimantan, Jakarta, and East Jaa. For each proince, the estimations for urban and rural areas are implemented separately, except for Jakarta which is a wholly urban area. The poerty line for each region is taken from Pradhan et al. (2001). A. Stage 1: Matching Variables in the Surey and the Census In order to obtain rigorous estimates of consumption leels of the households in the census, the explanatory ariables selected in the consumption determination model hae to exist and are measured in the same way in both the household surey and in the census. If the sample of the household surey was randomly selected and nationally representatie, the distribution of each explanatory ariable in the household surey can be expected to be the same as its distribution in the census. The means and standard deiations of the matched ariables in SUSENAS and Population Census data are shown in the Appendix: Table A1 for urban East Kalimantan, Table A2 for rural East Kalimantan, Table A3 for Jakarta, Table A4 for urban East Jaa, and Table A5 for rural East Jaa. B. Stage 2: Selecting Explanatory Variables for the Consumption Model The procedure in selecting the explanatory ariables of equation (3) starts by running a regression of log consumption on the matched ariables identified in Stage 1, plus some ariables that can be created from those ariables such as the square and cube of household size or the square and cube of the age of the household head. 6 In order to obtain a robust specification, ariables are only selected for inclusion in equation (3) if they contribute significantly to the explanation of (log) per capita consumption. Hence ariables with low t-alues are dropped. After a promising set of ariables has been selected in this way, the regression is run again and the residuals of this regression are saed. These residuals need to be scrutinized to check if there are some outliers in the obseration. If indeed there are some residual alues which are far out of the range of most residual alues, then these obserations must be checked for coding or other errors. Ultimately it may be necessary to delete them from the data. Fortunately, this is extremely rare. 6 Experience with poerty mapping in other countries suggests that these regressions should be weighted using cluster expansion factors. In the case of SUSENAS, cluster expansion factors within urban or rural areas in a proince are all equal. Since the estimations are implemented at this leel, the issue of weighting does not arise. 7

12 The next step is to select illage-leel independent ariables to complete the consumption model specification. The illage leel ariables are obtained from either the census data aggregated at the illage leel (for example the total number of indiiduals in the population or means of the ages of household heads in each illage) or from the PODES data. These ariables are then grouped into seeral sets such as demographic ariables, illage infrastructure ariables, and illage economic ariables. The residuals of the last regression are then aggregated at the illage leel to calculate the mean of these residuals for each illage. The ariable selection is then done by running separate regressions of the illage-leel mean of residuals on each set of the illage-leel ariables. The ariables with significant t-alues are selected as the candidates for inclusion in the consumption model. The feasibility of including these candidates for illage-leel ariables in the consumption model is tested by running regressions of illage dummy ariables on these ariables. One regression is run for each independent ariable candidate. If the coefficient of a certain ariable in a regression is one, it shows that there is a perfect multicollinearity between this ariable and the illage dummy ariable. This will happen if, for example, a illage has a certain infrastructure which no other illages hae, or on the other hand, all illages except one hae a certain infrastructure. Such ariables are necessarily excluded from the model. This test may explain why, for example, electricity is included in the model for rural areas but excluded from the model for urban areas. C. Stage 3: Estimating the Consumption Model The result of stage 2 is a complete specification of the consumption model, incorporating both household-leel and illage-leel independent ariables of the model. The next step is to test whether there is heteroscedascity in the data. This will determine the method to be employed to estimate the model. The first step to do this is to estimate the model of equation (3) using Ordinary Least Squares (OLS) and sae the residuals as a ariable uνh. Based on equation (2) the residuals components as uνh are then decomposed into uncorrelated u ˆ νh = u + u h u = η + eh (9) To inestigate the presence of heteroscedasticity in the data, a set of potential ariables that best explain the ariations in e 2 ν h are used to estimate the following logistic model: 2 e h T ln = zh α+ rh (10) 2 A eh 8

13 2 where we take A equal to.05* max{ } 1 e h as in Elbers et al., (2002). This 2 specification puts bounds on the predicted ariance of ε νh. The results of the OLS and heteroscedasticity regressions are shown in the Appendix: Table B1 for urban East Kalimantan, Table B2 for rural East Kalimantan, Table B3 for Jakarta, Table B4 for urban East Jaa, and Table B5 for rural East Jaa. In the case where homoscedasticity is rejected, a household specific ariance estimator for ε h is calculated as: 2 AB 1 AB(1 B) σ ε, h = + Var() r 3 (11) 1+ B 2 ( 1+ B) T where B = exp z h α. The consumption model is then re-estimated using the Generalized Least Squares (GLS) method, utilizing the estimated ariancecoariance matrix, Σ, resulting from equation (11) and weighted by the population weight, l h. The estimated parameters, β GLS, and their ariance, Var β GLS, are saed for use in the simulation. The results of these GLS regressions are shown in the Appendix: Table C1 for urban East Kalimantan, Table C2 for rural East Kalimantan, Table C3 for Jakarta, Table C4 for urban East Jaa, and Table C5 for rural East Jaa. D. Stage 4: Simulations on Census Data The purpose of this procedure is to apply the parameters estimated in the preious procedure to the census data. Howeer, since the alues of these parameters are obtained through estimations, they are not precise alues of these parameters and are subject to sampling error. This needs to be taken into account in applying the parameters to the census data, i.e. by incorporating the standard errors of the coefficient estimates in the application process. To start, recall that the purpose is to calculate the simulated ersion of equation (3): ln y = x ε (12) s h s s hβ + η + s h where the superscript s refers to simulated ersion of each parameter or ariable and now x h refers to characteristics of the households in the population census data. Simulation of β The simulated alue of β is obtained through a random draw, assuming β ~ N β GLS, Var β GLS. Note that the draw has to take into account the s coariance across β s. The randomly drawn parameter is defined as β. The next 9

14 step is then to apply this simulated parameter to each household in the census data s to calculate the alue of x β. Simulation of η h The process of obtaining the simulated alue of η requires two steps of simulations. This is because the ariance of η itself is estimated with error. Hence, the first step is to obtain the simulated ariance of η, σ 2s η. Elbers et al. (2002) propose to draw σ from a gamma distribution: σ ( ) η ~ G σ η, Var σ η. Accordingly, a random draw of 2s the ariance for the whole sample is exercised and its mean is defined as σ η. Then s the second step is to randomly draw η for each illage in the census data, assuming 2s η ~ N 0, σ. ( ) Simulation of ε h The process of obtaining the simulated alue of ε h requires the use of the results of estimation of equation (10). Assuming α ~ N α, Var α, a random draw of α is s made and defined as α. Like in the case of β, the draw has to take into account the coariance across α s. The simulated parameter is then used to simulate the household specific ariance estimator for ε h as defined in equation (11) for each household in the census data. Finally, the simulated alue of household specific idiosyncratic shock, ε, for eery household in the census data is obtained by taking s h 2s a random draw, assuming ~ N ( 0, σ ) Collecting ε. 7 h h Now all the three components of equation (12) hae been simulated, the alue of s ln y h for all households in the census data can be calculated by summing up the s s s alues of xhβ, η, and ε h that hae been obtained. The whole set of simulations is then repeated a number (100) of times, so that in the end a database of 100 simulated alues of (log) per capita household expenditure of all the households in the census data is created. 2s η 7 Elbers et al. (2002) mention alternaties for the assumption that the error component terms follow normal distributions. In separate sets of simulations we hae experimented with these alternatie assumptions. In no case did this lead to significantly different results. 10

15 E. Stage 5: Calculation of Poerty and Inequality Indicators The final output of stage 4, a database of 100 simulated alues of household expenditure of all households in the census data, is used as the basis for calculating arious poerty and inequality measures at the proincial, district, subdistrict, and illage leels. The point estimate of each measure is the mean of the calculated measure oer the 100 simulation alues. Meanwhile, the standard error of this estimate is equal to the standard deiation of the calculated measure oer the 100 simulation alues. 8 A word of warning should be issued here on interpreting the results obtained from this exercise. Suppose a headcount poerty indicator of 0.10 is listed for a location, along with a standard error of This should be taken to mean that if there were to be found other locations, with similar patterns of household characteristics, and if one had direct measurements of poerty headcounts in these locations, then we would predict that the poerty headcount in these locations are likely to fall between 0.07 and 0.13 (with a 70% confidence interal). In particular, we do not claim that all these similar locations share the same headcount, nor is there a good reason to attach too much significance to the point estimate of The pair of point estimate and standard error express that, conditionally on the information about the location that we hae, it is just as likely that its headcount is between 0.07 and 0.13 as it would be centered in the slightly narrower interal between and This uncertainty in the poerty estimates reflects the fact that the parameters of the consumption model (3) cannot be estimated with infinite precision, and that there is no way to deduce the error terms u ν from the aailable data. Similarly, to conclude that the headcount in one location (A) is bigger than in another (B), it is not sufficient to note that the point estimate for the headcount in A is higher than the one for B. Again, one has to take into account the error margins on the point estimates. For example, suppose that the headcount in A is ha with a standard error of s A and similarly for location B with h B and S B, where A s point estimate is higher: h A > h B. Then one can only conclude with reasonable confidence (more than 70%) that A s true headcount is higher than B s if h A s A > hb + sb. In other words one should account for the possibility that the estimated headcount for A is an oerestimate, while B s estimate is an underestimate. h 8 The application of this poerty mapping exercise from stage 3 to 5 is implemented using a software package called PoMap (Version 1.0 BETA), deeloped by Qinghua Zhao at the World Bank. 11

16 V. Poerty and Inequality Maps Poerty analysis is often based on national leel indicators that are compared oer time or across regions. The broad trends that can be identified using aggregate information are useful for ealuating and monitoring the oerall performance of a country. For many policy and research applications, howeer, the information that can be extracted from aggregate indicators is not sufficient, since they hide significant local ariations in liing conditions within countries. The detailed poerty maps at small administratie areas that are the ultimate output of this exercise proide benefits to help address this shortcoming of aggregate poerty analysis. This chapter proides the poerty and inequality maps at arious administratie leels as a result from this exercise. A. Poerty Estimates and Their Standard Errors Part II of this report proides the complete results of this pilot study in the forms of tables of arious poerty and inequality measures. The poerty measures calculated are the poerty headcount index (P0), poerty gap index (P1), and poerty seerity index (P2), commonly known as the FGT family of poerty measures. 9 Meanwhile, the inequality measure calculated is the Gini ratio. In addition to the estimates of poerty and inequality indicators as usually presented, the results of this poerty mapping exercise also proide the standard errors of these estimates as a measure of their precision. Table 1 compares the estimated headcount poerty rate for East Kalimantan, Jakarta, and East Jaa as calculated directly from the SUSENAS data and those estimated from the Population Census data through the poerty mapping method. Note the increase in precision of the census-based estimates compared to the SUSENAS-based estimates. This is a well-known phenomenon, employed extensiely in the statistical technique of small area estimation Foster et al. (1984). 10 Howeer, when the sample size in the SUSENAS is sufficiently large, such as in the case of East Jaa, the increase in the precision of the estimates is not large. 12

17 Table 1. Estimates of Headcount Poerty Rates in Jakarta, East Jaa, and East Kalimantan Based on SUSENAS and Poerty Mapping Method Area Jakarta: SUSENAS 1999 Poerty Mapping: Poerty Standard Error (%) Sample Size Rate (%) Points Proportion Household Indiidual ,959 12, ,204,219 8,246,736 East Jaa: SUSENAS 1999: - Urban ,250 12,535 - Rural ,285 19,593 - Total ,535 32,128 Poerty Mapping: - Urban ,703,652 13,761,133 - Rural ,655,930 20,730,848 - Total ,359,582 34,131,981 East Kalimantan: SUSENAS 1999: - Urban ,882 - Rural ,409 - Total ,003 4,291 Poerty Mapping: - Urban ,323 1,399,814 - Rural ,593 1,062,777 - Total ,916 2,462,591 Source: Authors computations. The standard errors on the SUSENAS-based headcounts are calculated by bootstrapping. 13

18 Table 1 shows the adantage of using the poerty mapping method to increase the precision of poerty estimates. Howeer, the real adantage of the method is its ability to produce poerty estimates and other welfare indicators at much smaller areas than the one presented in Table 1. A separate olume as a part of this report proides point estimates and standard errors of poerty headcount (P 0 ), poerty gap (P 1 ), poerty seerity (P 2 ), and Gini ratio at the proincial, district, subdistrict, and illage leels in the three pilot proinces. 11 Table 2 proides the summary of the precision of poerty headcount estimates at arious leels of areas. The numbers in this table show summary statistics of the standard errors as a proportion of the point estimates. For example, the table shows that the mean of standard errors across districts within a proince ranges from 11 percents of the point estimates for East Jaa to 27 percent for Jakarta. For the subdistrict leel, the mean of standard errors ranges from 23 percent for East Jaa to 63 percent for Jakarta. Meanwhile, for the illage leel, the mean of standard errors ranges from 53 percent for East Kalimantan to 128 percent for Jakarta. Table 2. Summary Statistics of Standard Error as a Proportion of Point Estimate for the Poerty Headcount Measure Region Mean Std. De. Minimum Maximum N Jakarta: - Proince District Subdistrict Village East Jaa: - Proince District Subdistrict Village ,412 East Kalimantan: - Proince District Subdistrict Village , See Part II: Tables of Poerty and Inequality Estimates. 14

19 Table 2 indicates that the standard errors at the proincial, district, and subdistrict leels are reasonably acceptable. At the illage leel, howeer, there are great ariations in the precision of poerty headcount estimates across illages within a proince. In East Jaa, the standard errors at the illage leel range from 9 percent of the point estimates to 169 percent. In East Kalimatan, they range from 11 percent to 441 percent, while in Jakarta the range is from 75 percent to 223 percent. This implies that the poerty mapping results for the illage leel need to be used with caution. For illages with high standard errors, other information is required to erify the estimates. In interpreting the statistics in Table 2, a word of caution is warranted. The proportion of standard error from point estimate can be high due to two different reasons: large magnitude of the standard error or small magnitude of the point estimate. A good example of the latter is the statistics for Jakarta. It appears that the estimates for Jakarta at arious leels always hae higher standard errors compared to the other two proinces. This, howeer, is due to the fact that Jakarta has much smaller poerty headcount point estimates than other proinces in Indonesia. In such cases, it is better to examine the absolute magnitudes of the standard errors rather than their proportions from the point estimates. The absolute magnitudes of the standard errors are clearly related to the population size. Figure 2 shows the plots between the magnitude of standard error with the population size at the illage leel in the fie estimation areas. The figure clearly indicates that the two ariables hae a negatie relationship, implying that the larger the population size of a illage the smaller the standard error of the estimate. This also suggests that where the standard error of poerty estimate for a illage is considered too large, the standard error can be made smaller by lumping that illage with its adjacent neighbors in one estimation. 15

20 Figure 2. Standard Error and Population on Size Village Leel East Jaa Urban Villages East Jaa Rural Villages Standard Error Standarrd Error ,000 10,000 15,000 20,000 25,000 30,000 35,000 40,000 Number of Population ,000 4,000 6,000 8,000 10,000 12,000 14,000 16,000 18,000 Number of Population East Kalimantan Urban Villages East Kalimantan Rural Villages Standard Error Standard Error ,500 5,000 7,500 10,000 12,500 15,000 17,500 20,000 22,500 25, ,000 2,000 3,000 4,000 5,000 6,000 7,000 8,000 9,000 Number of Population Number of Population Jakarta Urban Villages Standard Error ,000 20,000 30,000 40,000 50,000 60,000 70,000 Number of Population 16

21 B. District, Subdistrict, and Village Poerty Maps The first time aailability of accurate welfare indicators at district, subdistrict, and illage leels is already an achieement, but the real power of mapping is in presenting the outcomes in a geographical map, making it possible to oerlay the poerty data with all kinds of spatial characteristics. Figure 3a shows the distribution of poerty in the proince of East Kalimantan by district. Figure 3b proides the same information but calculated at subdistrict leel. Comparing the two figures clearly indicates that the heterogeneity of poerty within districts is quite large, making the information on the distribution of poerty in this proince coneyed by the two figures differ markedly. Figure 3c proides the information at an een finer illage leel, which differs een more markedly from Figure 3a. Figure 4a 4c show the same maps for the proince of Jakarta, while Figure 5a 5c for the proince of East Jaa. 17

22 Figure 3a. Poerty P Map of East Kalimantan District Leel 18

23 Figure 3b. Poerty Map of East Kalimantan Subdistrict Leel 19

24 Figure 3c. Poerty Map of East Kalimantan Village Leel 20

25 Figure 4a. Poerty Map of Jakarta District Leel 21

26 Figure 4b. Poerty Map of Jakarta Subdistrict Leel 22

27 Figure 4c. Poerty Map of Jakarta Village Leel 23

28 Figure 5a. Poerty Map of East Jaa District Leel 24

29 Figure 5b. Poerty Map of East Jaa Subdistrict Leel 25

30 Figure 5c. Poerty Map of East Jaa Village Leel 26

31 When inspecting these maps it should be kept in mind that they hae been created using the expected headcount. The true headcount for a location will differ from the expected headcount because of sampling and modeling error. The maps do not take errors into account. To show an example of what precision can be achieed at the subdistrict leel, Figure 6 shows the district leel predicted poerty headcount in urban East Kalimantan along with brackets giing a 70 percent confidence interal from one standard error below to one standard error aboe the point estimate. For reference, the proincial (urban) headcount has been included as a ertical line in the graph. Clearly, on the basis of this graph there is a large group of subdistricts for which one cannot tell with reasonable confidence that they hae below- or aboeaerage headcounts. Figure 6. The Precision of District Leel Poerty Estimates in Urban East Kalimantan Urban East Kalimantan Proincial headcount (headcounts with 2se error bounds) Predicted headcount estimates C. Examples of Further Applications Poerty mapping can be of great alue in policies targeted at the poor, but targeting is not the only possible use. For instance, the following Figure 7 could be used to illustrate the olatility of headcounts oer time. The figure depicts the (estimated) distribution of per capita expenditure of a particular subdistrict, with an estimated headcount of 0.3. The graph shows that the distribution function is ery steep in the neighborhood of the poerty line, implying that coariant consumption shocks (for 27 27

32 example, price shocks), which will shift the distribution to the left (negatie shock) or to the right (positie shock) will lead to a strong response of the headcount. Figure 7. Cumulatie Distribution Function of Consumption Illustrating ulnerability to shocks Urban East Kalimantan, subd= Poerty line 0.8 Simulated expenditure distribution fraction negatie shocks positie shocks consumption An obious application of the newly created data on economic welfare at disaggregated scale, is to correlate the data to other disaggregated statistics. For instance, a long-standing debate in deelopment concerns the relatie importance of a pro-growth policy and a policy aimed at reducing inequality. The following Figures 8a and 8b show that in urban East Kalimantan there is a strong negatie relationship between aerage per capita consumption expenditure and the poerty headcount, while the relationship between poerty and inequality is irtually nonexistent. The Gini coefficients are generally fairly low, suggesting that the scope for poerty reduction by redistributing income is limited. Note howeer that such graphics, suggestie as they are, cannot substitute for careful economic research into such important issues. 28

33 Figure 8a. Relationship between Poerty and Aerage Consumption 0.8 Headcount against aerage consumption Urban East Kalimantan by subdistrict 0.6 Estimated hc Aerage p/c consumption Figure 8b. Relationship between Poerty and Inequality Headcount and Gini 0.8 Urban East Kalimantan (subdistricts) 0.6 Headcount Gini 29

34 D. Conformity with Other Measures The indicator most widely used in Indonesia to rank regions for targeting purposes is based on the classification of family socio-economic status created by the Family Planning Coordinating Agency (BKKBN). Under this classification system, families are grouped into four socio-economic leels: pre-prosperous families ( keluarga prasejahtera or KPS), prosperous families leel I ( keluarga sejahtera I or KS I), leel II (KS II), and leel III (KS III). A family is defined as pre-prosperous if it fails to satisfy one of the following 5 conditions: (i) All household members are able to practice their religion; (ii) All household members are able to eat at least twice a day; (iii) All household members hae different sets of clothing for home, work, school, and isits; (i) A large part of the floor in the house is not made of earth; () The household is able to seek modern medical assistance for sick children and family planning serices for birth control. This BKKBN indicator was used extensiely in the targeting for arious Social Safety Net (SSN) Programs during the recent economic crisis. 12 Another large scale program, the Kecamatan Deelopment Program (KDP), also uses this indicator along with a composite of arious education, health, infrastructure, and economic indicators to rank subdistricts all oer the country. The subdistrict rank correlations of the poerty mapping results with these measures are shown in Table 3. Since the data are only for subdistricts within districts (kabupaten), and do not include subdistrict within cities (kota), the correlation tests can be implemented only for the proinces of East Jaa and East Kalimantan. 13 Table 3. Rank Correlations of Subdistricts Based on Poerty Mapping Results with BKKBN and KDP Indicators Proince Indicator BKKBN KDP East Jaa East Kalimantan See Pritchett, Sumarto, and Suryahadi (2002). 13 This is because KDP is implemented only in kabupatens. The data were used for the implementation of the second stage of KDP (KDP 2) starting in 2002 and proided by the World Bank Jakarta office. 30

35 Table 3 shows that all the correlation coefficients are not high. The correlation with BKKBN indicator in East Kalimantan is particularly low at around 21 percent. The other three correlation coefficients stand between 40 and 50 percent. This indicates that there is a wide scope to improe targeting of regions by incorporating the results of poerty mapping into the targeting decision. 31

36 VI. Concluding Remarks Poerty reduction and social deelopment efforts will continue to be an important endeaor in Indonesia, een long into the future. Learning from past experiences in targeting difficulties, there is clearly a need to deelop tools for more effectie geographic targeting than those that hae been used in the past. Ideally, geographic targeting would be based on a description of poerty incidence and other indicators of economic welfare at small areas or low administratie leels. This study is a pilot and the first attempt to apply the recently deeloped poerty mapping method in Indonesia. The objectie is to obtain estimates of poerty incidence at geographical units smaller than a proince-urban/rural area, which is the lowest leel of aggregation for which reliable (but still ery imprecise) poerty statistics are currently aailable. This pilot study uses data from three proinces: East Kalimantan, Jakarta, and East Jaa. The results of this pilot study hae strongly shown that the poerty mapping method deeloped to estimate poerty measures and other welfare indicators for small areas using data that are already aailable can be successfully applied in Indonesia. Using data from the three pilot proinces, this study has successfully calculated arious poerty and inequality indicators at the proincial, district, subdistrict, and illage leels with reasonable and better than SUSENAS based calculations of standard errors. In particular, the standard errors at the proincial, district, and subdistrict leels are reasonably acceptable. At the illage leel, howeer, there are great ariations in the precision of poerty headcount estimates across illages within a proince. The implication of this is that the poerty mapping results for the illage leel need to used with caution. For illages with high standard errors, other information is required to erify the estimates. Finally, the proen applicability and the usefulness of the poerty mapping results appear to support the extension of the application of the method to the remaining proinces. It is desirable that Indonesia has a complete poerty map for the whole country. 32

37 Appendix Table A1. Mean and Standard Deiation of Matched Variables, East Kalimantan - Urban Variables SUSENAS Census Mean S.D. Mean S.D. Household size Household liing in permanent house Household liing in owned house Household liing in rented house Housing facilities: - Clean water Toilet Electricity Household head characteristics: - Age (years) Female Married Education leel of household head: > Incomplete primary education or lower > Completed primary education > Lower secondary education > Upper secondary education > Tertiary education Years of education of household head Working status of household head: > Unemployed > Self employed/employer > Employee/salaried workers > Family workers/non salaried workers Occupation sector of household head: > Agriculture > Industry > Trade > Serices Spouse of household head characteristics: - Age (years)

38 Table A1. Continued nued Variables SUSENAS Census Mean S.D. Mean S.D. Education leel of spouse of household head: > Incomplete primary education or lower > Completed primary education > Lower secondary education > Upper secondary education > Tertiary education Years of education of spouse of household head Working status of spouse of household head: > Unemployed > Self employed/employer > Employee/salaried workers > Family workers/non salaried workers Occupation sector of spouse of household head: > Agriculture > Industry > Trade > Serices Aerage years of study for adult Proportion of adults who are employed Proportion of 6-24 year olds who are enrolled in schools Proportion of children 5 years old or younger Proportion of male Proportion of less than 15 year olds and 65 year olds or older (Dependency ratio)

39 Table A2. Mean and Standard Deiation of Matched Variables, East Kalimantan - Rural Variables SUSENAS Census Mean S.D. Mean S.D. Household size Household liing in permanent house Household liing in owned house Household liing in rented house Housing facilities: - Clean water Toilet Electricity Household head characteristics: - Age (years) Female Married Education leel of household head: > Incomplete primary education or lower > Completed primary education > Lower secondary education > Upper secondary education > Tertiary education Years of education of household head Working status of household head: > Unemployed > Self employed/employer > Employee/salaried workers > Family workers/non salaried workers Occupation sector of household head: > Agriculture > Industry > Trade > Serices Spouse of household head characteristics: - Age (years)

40 Table A2. Continued Variables SUSENAS Census Mean S.D. Mean S.D. Education leel of spouse of household head: > Incomplete primary education or lower > Completed primary education > Lower secondary education > Upper secondary education > Tertiary education Years of education of spouse of household head Working status of spouse of household head: > Unemployed > Self employed/employer > Employee/salaried workers > Family workers/non salaried workers Occupation sector of spouse of household head: > Agriculture > Industry > Trade > Serices Aerage years of study for adult Proportion of adults who are employed Proportion of 6-24 year olds who are enrolled in schools Proportion of children 5 years old or younger Proportion of male Proportion of less than 15 year olds and 65 year olds or older (Dependency ratio)

41 Table A3. Mean and Standard Deiation of Matched Variables, Jakarta Variables SUSENAS Census Mean S.D. Mean S.D. Household size Household liing in permanent house Household liing in owned house Household liing in rented house Housing facilities: - Clean water Toilet Electricity Household head characteristics: - Age (years) Female Married Education leel of household head: > Incomplete primary education or lower > Completed primary education > Lower secondary education > Upper secondary education \.49 > Tertiary education Years of education of household head Working status of household head: > Unemployed > Self employed/employer > Employee/salaried workers > Family workers/non salaried workers Occupation sector of household head: > Agriculture > Industry > Trade > Serices Spouse of household head characteristics: - Age (years)

42 Table A3. Continued Variables SUSENAS Census Mean S.D. Mean S.D. Education leel of spouse of household head: > Incomplete primary education or lower > Completed primary education > Lower secondary education > Upper secondary education > Tertiary education Years of education of spouse of household head Working status of spouse of household head: > Unemployed > Self employed/employer > Employee/salaried workers > Family workers/non salaried workers Occupation sector of spouse of household head: > Agriculture > Industry > Trade > Serices Aerage years of study for adult Proportion of adults who are employed Proportion of 6-24 year olds who are enrolled in schools Proportion of children 5 years old or younger Proportion of male Proportion of less than 15 year olds and 65 year olds or older (Dependency ratio)