Measuring Income and Poverty Using Proxy Means Tests
Measuring Income and Poverty Using Proxy Means Tests
Measuring Income and Poverty Using Proxy Means Tests
Measuring
income and poverty using
Proxy Means Tests
OVERVIEW
This document will serve as an introduction to proxy means testing (PMT).
The concept of proxy means testing is usually understood as using
observable characteristics of the household or its members to estimate their
incomes or consumption, when other income data (salary slips, tax returns)
are unavailable or unreliable.
Household
owns
the
house
100
Number
of
household
members
in
50
employment
One
child
in
the
household
40
Number
of
persons
per
room
-‐20
Two
children
in
the
household
30
Three
or
more
children
in
the
household
20
In this case, we simply have to multiply the
Household
owns
cattle
200
coefficient listed in the table by the value applicable
Household
owns
a
bicycle
300
to the household. For example, in a household
Household
owns
a
car
800
where two persons are employed and, on average,
three persons share a room, we modify the baseline
Dwelling
walls
made
of
brick
100
by:
Dwelling
walls
made
of
tin
0
Dwelling
walls
made
of
clay
-‐100
2 ∗ 50 + 3 ∗ −20 =
100 − 60 = 40
Constant
1000
In the case of this household, we increase their
estimated consumption by 40 takas.
Those results allow us to be more precise: let us
assume that our household: Choosing data source
• Is composed of two adults and three children;;
As we mentioned, the essence of the proxy means
• Owns its house;; test method is using household characteristics to
• Owns a bicycle, but not a car;; predict household welfare. Therefore, we need a
• Doesn’t have any cattle;; data source that would contain information about
• And that the walls of the house are made of clay. both household characteristics (household
composition, dwelling qualities, assets) and
In order to calculate the estimated consumption household welfare (normally measured by
expenditure of the household, we should: consumption volume), and that would allow us to
draw correspondence between the two.
Start with the value listed in the “constant” row.
This will be our baseline. In our case, this amounts This means that we normally need results of a
to 1000 takas (Bangladeshi currency) household survey. Household Income and
Expenditure Surveys (sometimes called household
• Find all rows that apply to the household
economic surveys) or Living Standards
situation and add or subtract the corresponding
Measurement Study surveys are examples of
coefficients to or from the “baseline” value.
surveys that contain data that could be used in
• Consequently, the estimated value of
developing a PMT model.
consumption for the example household will be
Regardless of the exact type of the survey that we
equal to 1000 (baseline) + 100 (house
want to use, there are two important criteria that our
ownership) + 20 (three or more children) + 300
data source has to meet:
(ownership of bicycle) – 300 (clay walls) = 1320
takas. • The data should be as recent as possible—as the
living conditions of the population change, the
relationship between particular household
Not all of the variables will refer to simple yes/no
characteristics and its welfare is also prone to
questions. It could be the case that an entry in the
change. Ideally, the survey should have been
table will read:
conducted in the last five years.
• The survey results should be representative for
the area. If we want to use the PMT scores for
targeting social safety nets beneficiaries model. The simplest way of validating the model
nationwide, we should use a survey that is involves using the data from the same survey that
nationally representative. has been described above. Next to the survey, we
will also need to know the poverty line. In its
Choosing the variables absence, we can select a percentage and assume
Results of a typical LSMS survey will contain that the poor comprise the bottom 30 or 40% of
hundreds of variables. Among them, we need to households with lowest consumption.
choose the ones that will be used in the PMT
model. While doing that, we should take into Following that, we conduct the validation by:
account two basic criteria: • using the survey data, calculate PMT scores for
every household in the survey dataset;;
• “Verifiability” of the variable • listing the households below the poverty line,
• and correlation between the variable and according to their consumption (not the PMT
household welfare (consumption levels). score);;
• listing the households below the poverty line,
Good PMT variables will be easy to verify by according to their PMT scores;;
enumerators or social workers. For example, the • comparing the two lists.
type of walls in the dwelling can be easily
discerned by the enumerator visiting the Since the actual consumption data is more reliable
household—brick is visibly different from wood or than the PMT data, if a household was included in
tin. This is generally the case with other variables the consumption-based list, but not in the PMT-
describing housing quality. Although household based list, we speak of exclusion error. The
characteristics (number of members, their ages, household was poor, but the PMT model has not
education or occupation) could be misrepresented, identified it as such.
this does not happen often, in particular when the
enumerators or social workers belong to the same
On the other hand, if a household was not identified
local community. Caution should be exercised with
as poor according to its consumption measured by
including ownership of small assets in the model—
the household survey, but the PMT model has
these could be removed or concealed.
identified is as poor, we speak of inclusion error.
The household was not poor, but the PMT model
The relationship between the selected variables considered it poor.
can be determined through regression analysis;;
while comparing results of regressions with
Ideally, we would like to minimize both inclusion and
different variables included or omitted, the authors
2
exclusion errors. In practice, however, there is often
of the PMT model can compare R and statistical
trade-off between the two, with typical PMT models
significance of the coefficients for different models.
featuring inherent 30-40% inclusion and exclusion
Once one model has been chosen, the regression errors.
coefficients will become the coefficients for the
PMT model—they will determine the extent to Different models
which each variable affects the household
It is not required that a single model is used for
consumption.
entire country. Sometimes it might improve accuracy
of prediction if separate models for different regions
Validating the PMT model
are created. Often a separate model for urban and
After choosing the model, it is important to see how rural poor is developed, as welfare indicators differ
it performs in identifying the poorest households. between cities and the countryside (for instance,
This is often referred to as ‘validating’ the PMT
livestock ownership, which often accurately predict good data quality and the fact that quite details
welfare in the countryside, is irrelevant for the city household characteristics variables were included. In
dwellers). addition to that, the survey was used to compute
More advanced PMT models can use multi-staged official poverty figures and poverty maps for the
approach, where the results of one model’s country;; consequently, its use meant that a
prediction are ‘plugged in’ to another model, in consistent definition of poverty could be applied
order to minimize inclusion and exclusion errors. across the board.
Updating the model The country uses a single model, although the final
As societies change, PMT models may become score is adjusted based on the region. That
outdated. Intuitively, the models can and should be approach simplifies the model, which is important
updated when more recent data become available. when it comes to promoting its use across different
The household surveys used to produce the government agencies. Most of the variables selected
models (HEIS, LSMS) are often released in 5- to for the models can be directly observed by the
10-year cycles, which suggests a similar pattern for enumerator visiting the household, minimizing the
PMT model updates. risk of fraud or misrepresentation.
Bangladesh experience The variables used by the Bangladesh model and
To finish this note, it may be useful to present the the corresponding scores are included in the
experience of building the PMT model for following table. Undercoverage, for the cut-off point
Bangladesh to see how the principles described of 40%, is equal to 36%. While the value could still
above can be applied in practice. be considered a sizeable exclusion error, it has been
proven through pilots that targeting benefits based
In terms of the data source, the mode has been on the model is still more accurate compared to the
based on the results of the most recent 2010 methods previously used for social safety net
Household Expenditure and Income Survey. programs in Bangladesh.
The main rationale behind the choice has been