1 Using the Results from Poverty Mapping Similar problem as to - - PDF document

SLIDE 1

1

Poverty Mapping in the World Bank: Our Work and Lessons Learned

Qinghua Zhao Development Research Group The World Bank (202) 473-1273 Qzhao@worldbank.org

ch c ch ch

y    ~ ~ ~ ' ~ ln   x

Believing in PovMap ? Why ?

? : People with same characteristic have same income : earning differs by location : even people with same characters may earn differently

 '

ch

x

c



ch

 ~

Is the Solution Robust?

• No necessary. Depend on the the implementation.
• Good model indeed give stable result
• Example: Chinese agriculture census and national

census Possible Factors:

• it is the match variables matter
• variations in variable help
• basic living conditions always important

How to Verify the Result ?

• Not an easy job.
• Few existing statistical data can be used to verify
• not detail enough
• no standard error given
• data collected differently
• Oversampled survey could provide a solution but very

costly.

• Kenya: great story on sub-district dimension

What Makes a Good Model ?

• Matched variable between survey and census
• Variables on living condition (wall material, toilet, roof,

kitchen,…)

• Interactive term (may not be meaningful)
• Variables with significant deviation
• Correct weighting

Rethinking the income model with sub-population

• The well being of handicapped population
• Whose income model? Everybody or handicapped
• The result is still over-estimate the well being, but what

can we do ?

SLIDE 2

2

Using the Results from Poverty Mapping

• Similar problem as to 2SLS estimate. Adjustment

needed.

• Spatial analysis
• Deforestation vs. poverty
• Poverty and crime

Part II: Software Tools for Poverty Mapping Typical Simulation Model

• The fully specified simulation model is defined as follows:
• where
• is a random variable (normally distributed or T-distributed)
• is a random variable (normally distributed or T-distributed)

ch c ch ch

y    ~ ~ ~ ' ~ ln    x

) ˆ , ˆ ( ~ ~



   N

c

 ~

ch

 ~

) ~ ~ exp( 

T ch

Z ) ˆ , ˆ ( ~ ~



   N

                

3 2 ,

) 1 ( ) 1 ( ) ( 2 1 1 ˆ B B AB r ar V B AB

ch







Where B= and is function of

T ch

Z ~

 ~ y ~ and

ch ch T ch

X Z 

Challenges

• Storage
• large Dataset (about 80M per 1 mil

household with 20 variables)

• Speed
• Computing poverty measurements

such as Gini index and quintile

• Random number generating

Design Goals

• Highest speed
• Least memory usage
• Database connectivity
• Parallel processing
• Flexibility
• Zero Installation

Basic Framework of Poverty Mapping

Evaluator

B A C H

Estimator

census

Estimated result

estimated y

nObs*nVars nObs*nSim

Survey

Measurement Computation

result

SLIDE 3

3

Approach 1: Complicated SAS Macro

Evaluator

B A C H

Estimator

census

Estimated result

estimated y

nObs*nVars nObs*nSim

Survey

Measurement Computation

result

SAS Program

Performance: 3 hours per 1M

Approach 2: Distributed Computing Model

census

survey result

Estimator Estimated result dispatcher

Measurement Computation

Evaluator

B A C H

Computer 1 Computer 2 Computer 3 Computer n

Server Side Client Side

Approach 3: Basic Structure of PovMap

Measurement Computation

Evaluator

B A C H

census

nObs*nVars nObs*nlRecL

survey result Packed census

Estimator Estimated result packer

Step 1 Data Prep. Step 2. PovMap.exe Model Specification Simulation Configuration

Why Pack the Intermediate Data ?

Three memory modes

• 1. N*8*(m+2)
• 2. N*(lrecl+16)
• 3. N*16

8 bits

Approach 2a: Packer Written in SAS Macro

Measurement Computation

Evaluator

B A C H

Estimator

census

Estimated result

survey result

packer

Packed census

SAS macro DataPrep FileAux PovMap.exe SAS format SAS format Model Specification Simulation Configuration SAS

Approach 2b: Packer Not Dependent on SAS

Measurement Computation

Evaluator

B A C H

Estimator

census

Estimated result

survey result

packer

Packed census

PovMapPacker PovMap.exe Stata dBase ASCII Simulation Configuration Model Specification

SLIDE 4

4

srvdata=HLANDS_SURV.dta lhs=ltexpae rhs=lsize youth mf1550 adultread food animal i218 sch_inc_201 sch_inc_218 n_rooms hh_size_m married_m hdschyr_m hdschyr_m2 food_m rvalue rain_metre slopedum arhs= _Yhat_ sch_inc_218 tree_m food_m rain_m_sq hdschyr_m Cluster=CLUSTER sWeight=PERSWEIG cendata=hlands_census.dta cWeight=n_ad_eq cKeyVar=cluster dataout=hlands.pda LOCERR=YES

Modeling Specification

nSim=100 CDist=n HDist=n PovLine=195 memorysize=128 MinImpute=auto maximpute=auto abound=none bbound=.99 cbound=auto hbound=auto seed=12345678 INDICES=FGT0 FGT1 FGT2 GE00 GE05 GE10 GE15 GE20 ATK2 GINI Dist:20 ydump=1 Simulation=0 3 5 end

Simulation Specification Performance: What Determines the Speed

• memory size
• indicator
• random number generating

nObs Size(M) Random Memory Mode Indicator Time(sec) 166625 5.12 T(8),N 128 1 HC+GINI 41 166625 5.12 T(8),N 20 2 HC+GINI 65 166625 5.12 T(8),N 5 3 HC+GINI 73 166625 5.12 N,N 5 3 HC+GINI 58 1.8m 74 N,N 128 2 HC only 647 1.8m 74 N,N 128 2 HC+GINI 961

Unsolved Problems

• parameterize the error term
• outlier and trimming
• sensitivity analysis
• awkward structure
• inflexible

Complications…

• small survey into big survey
• census collected with a fixed ratio
• survey into a subset of census
• sensitivity analysis
• variance decomposition
• income modeled by simultaneous equations
• income estimated by maximum likelihood model
• better estimation of the cluster effect
• better estimation of idiosyncratic effect

Solutions: New Structure for More Flexibility

• Modular design for easy expansion.
• A single simulator can’t satisfy all different needs--

user can build their own simulator with C++ compiler.

• Multithread between disk I/O and numeric

computation.

• Random number generating trigged by data event.
• Flexible function form and looping.
• Adding free C++ compiler.