Network Reporting Methods

Abstract

This dissertation is about how the science of sampling and survey research can

be generalized beyond the individual: while traditional surveys ask respondents

to report about themselves, we study network reporting surveys in

which respondents are asked to report about others.

Despite many technical challenges, understanding how to design and analyze network

reporting studies is worth the effort: network reporting surveys can be used to

study many important rare and hidden populations for which traditional survey

methods are inadequate.

We begin by introducing the network reporting framework, a general toolkit that

can help researchers develop new estimators and new data collection methods.

The framework can also help researchers better understand many existing estimators.

We apply the framework to analyze network scale-up, an existing method for

studying epidemiologically important hidden populations like sex workers and

drug injectors.

We derive the precise conditions required for the basic scale-up

estimator to have desirable statistical properties;

we also introduce a new, generalized scale-up estimator, which we

expect to outperform the existing estimator in many settings.

Next, we turn to a key question for all network reporting studies: which

personal network should respondents be asked to report about? We conjecture

that there may be a trade-off between the quantity and the quality of

information obtained from different personal networks.

We test this conjecture by embedding an experiment in a large,

nationally-representative household survey that we conducted in Rwanda.

Our results show that there may indeed be a trade-off, but future work

is required to understand this possibility in more detail.

Finally, we apply the network reporting framework to a critical, unsolved

problem in demography: estimating adult death rates in countries that lack

complete vital registration systems.

We introduce a new estimator and data collection procedure called

network survival, and we test our new approach using the Rwanda survey.

Our results demonstrate that a network reporting study is

feasible in an environment where adult mortality estimates are sorely needed,

and the estimated network survival rates have plausible levels and age-patterns.

However, more work is required before we can fully assess the accuracy of this new

approach.