The DHS Program User Forum      
Discussions regarding The DHS Program data and results
Home » Topics » Service Provision Assessment (SPA) » Weighted multilevel models using Tanzania SPA
Weighted multilevel models using Tanzania SPA [message #14167] Wed, 28 February 2018 14:51 Go to next message
cmoucheraud is currently offline  cmoucheraud
Messages: 2
Registered: April 2015

I am trying to run a multilevel model using Tanzania SPA data (including facility, provider and client levels) -- but I cannot get any of these models to converge when the multiple levels' sampling weights are included. We have isolated the problem as due to the extremely skewed distribution of the facility-level sampling weight (it has a very long right tail with interruptions).

Has anyone found a way to work with multilevel models & multiple levels of survey weights, with the Tanzania SPA dataset or any of the other SPA datasets that have a similarly skewed distribution of facility weights?

Thanks in advance!,
Re: Weighted multilevel models using Tanzania SPA [message #14439 is a reply to message #14167] Fri, 06 April 2018 09:18 Go to previous message
Messages: 1236
Registered: February 2013
Senior Member
Dear User,
A response from sampling experts, Dr. Ruilin Ren:

Firstly, about the multilevel modeling with SPA data, we do not know if other people did the same research or not, this question was never asked before. When talking about the multilevel weights, it needs additional attention because the weights we developed in the SPA data may not be "multilevel" since the provider weights and client weights all use the facility weights as base weights. This means the provider weights and client weights already take the multi-stage selection into account, they do not just reflect the selection of providers and clients in the second stage.
Secondly, about the skewed distribution of the sampling weights, this should be the case for SPA surveys by facility type because the important health facilities such as hospitals and health centers are included with probability one in the sample, so their sampling weight should be small compared with facilities of other types which are sampled. If the skewed weight distribution is a problem for your modeling work, you may need to take the facility type as a co-variable. This may make more sense for your analysis because of the big difference on the role played by different types of facility in the health system.

Previous Topic: Exit survey question development
Goto Forum:

Current Time: Thu Jun 21 06:26:16 Eastern Daylight Time 2018