Targeted adaptive sampling strategies for prevalence surveys

Federico Andreis & Micheal Chipeta
Lancaster - 05/06/2019

Some background - prevalence surveys

Prevalence surveys are a type of health survey, typically carried out in low-resources settings.

They help to

  • answer questions on the epidemiology of diseases
  • monitor the health status of a population
  • produce maps to inform health policies.

Some background - associated costs

As in all surveys, costs are a critical aspect that needs to be carefully considered. Proper prevalence surveys require an enormous amount of resources to be organised and carried out.

Main sources of costs are

  • planning
  • local communities engagement
  • logistics
  • measurement costs
  • time!!

The team

Dr Federico Andreis, University of Stirling (tw: @Chicco_Stat)

  • adaptive designs for tuberculosis prevalence [Andreis et al (2017)]
  • targeted adaptive designs for polluted sites [Andreis et al (2018)]

Dr Micheal Chipeta, University of Oxford

  • adaptive geostatistical designs for malaria surveys in Malawi [Chipeta et al (2016)]
  • efficient identification of malaria hotspots [Kabaghe et al (2017)]

Adaptive designs

As opposed to non-adaptive ones, adaptive designs are characterised by strategies allowed to change 'as we go', incorporating sample information as it is collected.

Typically, adaptive strategies are implemented in steps, and the cumulated information can be used to direct the sampling effort according to some criterion.

A few examples:

  • inducing some degree of 'spatial balance'
  • oversampling cases
  • minimising predictive variance.

Adaptive geostatistical designs

Micheal and colleagues have developed a framework for model-based spatial sampling that aims at producing accurate predictive maps.

  1. initial non-adaptive sample
  2. geostatistical model fitted
  3. new locations priority: those with high prediction variance
  4. do not rinse, and repeat 2-3.

predictive variance

Targeted, design-based approach

Federico and colleagues have worked on a design-based approach that targets units that respond to prescribed characteristics.

  1. initial non-adaptive sample
  2. prediction for nonsampled locations
  3. new locations priority: those with predictions responding to target
  4. do not rinse, and repeat 2-3.

predictive variance

Our work in progress

We are developing a framework to allow the design and implementation of surveys that offer the 'best' trade-off across

  • targeting objectives (if any)
  • accuracy of predictive maps
  • resources spent on obtaining it.

How