Transcript Slide 1

Small Area Estimation of Child
Malnutrition
Assessing the Omission of Maternal Anthropometrics
Ericka G. Rascon-Ramirez
ISER, University of Essex
What is SAE in the Public Policy
context?
Set of statistical methods for obtaining small areas
indicators (at locality, town or LSOA level) not
represented by household surveys.
Objective of the study
Assessment of the SAE method of Elbers,
Lanjouw and Lanjouw (2003) to generate
child malnutrition indicators at the local level
when maternal anthropometrics is omitted in
the modelling stage.
Applications of SAE: ELL Method for
Poverty and Nutrition Mapping
Around 45 Poverty Maps and 5 Nutrition Maps
in the world.
Figure: Extreme Poverty and Stunting at the
municipality level in Mexico (2005).
Brief Description of the ELL Approach
ELL: Using a household survey and census
or administrative data, they derive statistical
properties of estimators of welfare indicators
to be imputed at small area levels not
representative in surveys.
Stages of ELL Methodology
• Stage Zero: Comparability between census
and survey variables.
• Stage One: Modelling of welfare indicator in
the survey according to the representativity.
• Stage Two: Computation of welfare indicator
in census records.
Stage One: Child’s Anthropometric Model
In the survey, we run a GLS model using ONLY
comparable variables between census and
survey:
Where
denotes child’s height,
is a matrix
of household and individual characteristics,
and xxx is the error component.
Stage Two: Computation of welfare indicator
In the census, using the best prediction model,
we obtain the welfare indicator as follows:
Drawn parameters:
and
Having R replications at the individual level, we
use their average for constructing the welfare
indicator at the local level.
Drawbacks of ELL methodology
• Area Homogeneity (Conditional Independence).
The conditional distribution of the welfare
indicator y given X covariates in the small
area A is the same as in the larger
geographical region G. (Deaton and Tarozzi,
RES 2009)
• Omitted Variable Bias. The use of ONLY
comparable variables between census and
survey restricts the inclusion of relevant
variables. (Focus of this study)
Methodological Exercise
Assessing the Omission of Maternal
Height using Monte Carlo
Simulations
Monte Carlo Exercises: Child’s DGP
Let’s assume the true DGP of the variable of
interest follows this structure:
Where
is maternal height and is not available
in census records. The bias of the final estimate
of
will depend on the influence of
on its
variance and/or the correlation with other
covariates.
Assessing the Omission of Maternal Height
Mean of Bias of Child Malnutrition at EA
Contribution of Maternal Height: 25 % of Child’s
Contribution of Maternal Height: 75 % of Child’s
Note: Mean of MSE when the prediction of the model is 45%.
Study’s Contribution: Two-Step Small
Area Estimation
When a “relevant” variable has been omitted as
a consequence of census-survey
comparability:
 Obtain SAE of the “relevant” omitted variable
at the individual level using ELL.
 Having the relevant variable in both sources,
use it as a covariate for the final model.
 Following the ELL approach, obtain the final
SAE of child height.
Empirical Evidence
Two-Step SAE for obtaining
malnutrition indicators at the
municipality level in Mexico
(Chiapas and Hidalgo)
Empirical Exercise: Two Step SAE
Slight differences between both approaches:
Figure: Height (z-scores) for Mexican Children under 5.
Empirical Exercise: Two Step SAE
Relevant differences between both approaches:
Figure: : Height (z-scores) for Mexican Children under 5.
Conclusions
Methodological Exercise: Higher contribution of
the omitted variable (with low correlation with
other covariates) may bias the final malnutrition
estimate.
Empirical Application: Empirical evidence support
a two-step SAE for obtaining less biased
estimates for highly heterogeneous
communities.