A Sustainable Poverty Monitoring System for Policy Decisions

1
A Sustainable Poverty Monitoring System for
Policy Decisions

• Bjørn K. G. Wold, Astrid Mathiassen and Geir
  Øvensen
• Division for Development Cooperation, Statistics
  Norway
• IAOS October, 2008

2
Two parts

• Describing the Household Survey System Model,
  including the Poverty Monitoring System Case
  Malawi
• Testing the Poverty Model Case Uganda

3
Part 1The Household Survey System Model
4
Growing need for statistics for policy decisions

• Recent initiatives since Millenium Development
  Goals
• Paris21 Scaling up
• World Bank Better statistics for better
  results
• Three major challenges remain
• Design short questionnaire for fast monitoring of
  MDG and PRSP indicators
• Method for easy and accurate measurement of
  money-metric poverty
• Household survey system with annual core for MDG
  and PRSP, and a rotating program of specialized
  sector surveys

5
Suggested solution

• Identify indicators to measure progress on MDG
  and PRSP
• Household Budget Survey for initial poverty line
• A poverty model for monitoring in non-HBS years
• Annual, rotating sector surveys with common core
• Household Survey Program to cover all topics in
  5-10 years
• Statistical tools such as seasonal adjustment for
  consistent trends, and small areas estimation
• Fast and easily accessible results
• Active dialogue between donors and to ensure that
  all agencies accept integration of their
  surveys in program

6
Poverty monitoring based on light household
surveys

• Exploit statistical correlation only, not total
  consumption
• Find 10-15 poverty indicators in HBS (other than
  expenditure)
• Include exactly the same indicators in the light
  survey
• Collect light survey data and apply the
  correlation found in HBS
• Estimate annual regional/ district poverty
  headcount from the light survey, including the
  inaccuracy
• Standard error for estimate at similar levels as
  in traditional consumption aggregate approach

7
The Malawi Poverty Prediction Sequence
IHS2
8
Malawi Large data gaps if use HBS only!

• 1998 IHS1 Budget Survey, (with data problems)
• 2004 IHS2
• 2009(?) IHS 3

9
Malawi II Complement with Poverty Estimates from
light Surveys!

• 1998 IHS1 Budget Survey, with data problems
• 2004 IHS2
• Estimates from (light) WMS 2005, 2006, 2007
• 2009(?) IHS 3

10
Part 2Testing the Poverty model on Ugandan
Surveys
11
Testing the poverty models predictive ability

• Test the predictive ability of the poverty model
• ?Compare models poverty estimates relative to
  poverty estimated directly from consumption
  aggregates
• Use 7 comparable household expenditure surveys
  from Uganda from 1993 to 2006
• Comparable consumption aggregates and
  sufficiently number of (exactly) identical
  indicators
• Calculate urban/rural poverty models from each
  survey
• Cross-testing models from each survey onto the
  other surveys

12
Example Pairwise testing from 1995 survey onto
itself, and the 6 other surveys
Expenditure Survey 1995
13
Uganda baseline trend Using traditional
consumption aggregate approach only

• 85 rural ? national rural
• Falling headcount ratio, especially in late
  90-ies
• Urban rural poverty gap closing

14
Uganda Comparing actual poverty level
predictions from RURAL model

• Models capture most, but not all of reduction
• All models have similar patterns of changes
• Less capture of variability within trend

? Biases related to factors specific for specific
years? (e.g, omitted variables)
15
Uganda Comparing actual poverty level
predictions from RURAL model

• Models capture most, but not all of reduction
• All models have similar patterns of changes
• Less capture of variability within trend

? Biases related to factors specific for specific
years? (e.g, omitted variables)
16
Uganda Comparing actual poverty level
predictions from URBAN model

• Better capture of variability than rural
• Low poverty in base year ? low urban predictions
• 1999 survey bad base for urban models

• Combination of long time elapsed and large fall
  in poverty seriously shake the model? (ref. 2005
  predictions)

17
Uganda Take out two most problematic surveys in
urban and rural models

• Predictions now much more in line with trend
• Also good predictions at sub-regional level

• Lower ability to capture sudden changes ? Add new
  types of variables?

18
Conclusion

• Predictive ability on general trend proven, but
• Modelscarry on their base year poverty level
• Difficult to capture sudden changes
• Challenge with two individual surveys (survey
  issues)
• ?If two HBS available, and both lt10 years old,
  use average of both in predictions
• Possible improvements Number and level of assets
  and locality-level explanatory variables
• Statistically, a second best solution ? never
  perfect