2. Outline
I. Geographically linking population and facility surveys:
methodological considerations
Authors: Skiles MP, Burgert CR, Curtis SL, Spencer J
Methods – data sets, linking techniques, analysis
Findings – take home messages
II. Relating Depo-Provera Access with Use in Malawi
Partners: Deliver/JSI – Inglis A, Cunningham M.
MEASURE Evaluation – Skiles MP, Wilkes B, Bardon-O’Fallon
J, Spencer J
Application of methods
Findings – take home messages
III. Discussion: where do we go from here?
5. What is Kernel Density Estimation
(KDE)?
Health Center: Depo Delivery Site
Health Center
KDE is a technique
employed to distribute
a value associated
with a discrete point
across a plane or
continuous surface.
6. Operationalizing the links
Health service variables associated with a facility
merged with DHS cluster data if “linked”
Absolute measures – type of facility, FP method avail.
Relative measures – FP Readiness Score, VCT Readiness
Score
Analysis
Descriptive: percent disagreement between distribution of key
variables in master data and the sample/displaced data
Associations: basic logistic regression to assess association between
relative health service measures and use of modern contraception
7. Effects of…
Facility Sampling:
Substantial misclassification for clusters (linked/not-linked)
Substantial misclassification when considering the relative service environment;
Differential error introduced – direction and magnitude of bias is unpredictable.
Cluster Displacement:
Non-trivial misclassification introduced at geographic areas smaller than admin
units, particularly if linking to the closest facility;
Descriptive analyses with DHS data at cluster-level is possible, BUT the measurement
error introduced from displacement will bias relationships between service
environment and individual level health outcomes
Linking Methods:
Admin link was least affected by sampling and displacement, but may be a less
relevant link;
Linking to the closest facility performed poorly across all methods;
KDE did not appear to perform better than less sophisticated methods.
8. Take Home Messages
Linking independently sampled DHS clusters and
a sample of facilities is NOT recommended at
low levels of geographic disaggregation.
Linking to the closest facility is inappropriate if
your analysis conceptually depends on linking an
individual with the facility presumably used;
better to link to a service environment.
Definition of service environment will influence
the amount of error introduced in linking.
9. Application with PRH & DELIVER/JSI:
Relating Depo-Provera Access with Use in
Malawi
Research Question
Is availability of Depo-Provera associated with use of or
demand for injectable contraceptives among married
women of reproductive age?
Data Available
Facility data – Master public health facility list
Contraceptive supply data - DELIVER LMIS 2010
Population data - MEASURE DHS 2010 Malawi
10. Defining Supply Reliability
Used KDE to create variables measuring access
Access 1 = distance from DHS cluster to a Depo-Provera service
delivery point
Access 2 = distance from DHS cluster to a Depo-Provera service
delivery point + reliability of Depo-Provera supply
Operationalizing Access
Algorithm for “proportion of months with Depo stock”:
Monthly closing balance
Dispensing in prior months
Volume dispensed was NOT considered
11. Kernel Density Estimation: Access
KDE Surface: 10Km radius around all Depo Service
Delivery Sites
Health Center: Depo Delivery Site
Health Center
Health Center: Depo Delivery Site
Health Center
KDE Surface: 10Km radius around Depo sites using
weighted variable representing Depo-Provera supply
13. Multivariate Model: Injectable Use and Access
KDE: Distance KDE: Distance + Supply
-0.04
-0.02
0
0.02
0.04
0.06
0.08
0.1
Less
Access
Most
Access
-0.04
-0.02
0
0.02
0.04
0.06
0.08
0.1
Less
Access
Most
Access
14. Conclusions
Access to Depo-Provera (Reliability + Distance) is:
+ associated with a married woman’s use of an injectable
not associated with an increase in family planning demand
In the Central Region, Depo-Provera supply is either
unreliable or suffers from incomplete reporting
Next Steps
Clarify and include role of CHW in depo provision
Consider comparative analysis of DELIVER vs country
LMIS system
16. MEASURE Evaluation is funded by the U.S. Agency for
International Development (USAID) through Cooperative
Agreement GPO-A-00-03-00003-00 and is implemented by
the Carolina Population Center at the University of North
Carolina in partnership with Futures Group, John
Snow, Inc., ORC Macro, and Tulane University.
Visit us online at http://www.cpc.unc.edu/measure.
Editor's Notes
Sampling:Substantial misclassification for clusters (linked/not-linked) which led to a substantial underestimation of the adequacy of health service environment – absolute link;Substantial misclassification error when considering the relative service environment;Differential error introduced – biases estimates and direction of bias is unpredictable.Cluster Displacement:1.No additional error introduced when linking to all facilities within an administrative boundary BUT non-trivial misclassification introduced at smaller geographic areas particularly if linking to the closest facility;
Facilities:930 = public and privateDROPPED: 234 facilities missing unique ID (mostly privates); 125 facilities missing GPSDepo:483 = depo delivery points per supply data infoDROPPED: 60 with missing GPSMerged 571 mapped facilities with 423 Depo service delivery sites
Use Master Facility lists:model the systematic misclassification error due to facility sampling and include in regressions to improve estimatesCalibrate facility sample data and use for small area estimationLook at other services, outcomes, diseases – Malaria (bednets, disease)