Are sibling models a suitable tool in analyses of how reproductive factors affect child mortality?

Øystein Kravdal

Volume 42 - Article 28 | Pages 777–798

Are sibling models a suitable tool in analyses of how reproductive factors affect child mortality?

By Øystein Kravdal

Date received:

24 Jan 2020

Date published:

5 May 2020

Word count:

4353

Keywords:

bias, child mortality, reproductive factors, sibling models

DOI:

10.4054/DemRes.2020.42.28

Additional Files:

42-28 additional material (pdf file, 28 kB)
demographic-research.42-28 (zip file, 5 kB)
readme.42-28 (text file, 1022 Byte)

Download PDF
Submit a Response Letter

Abstract

Background: Several studies of how reproductive factors affect child mortality or other child outcomes have been based on sibling comparisons. With such models one controls for unobserved determinants of the outcome that are shared by the siblings and linked to the reproductive process. However, it has been shown mathematically that estimates from sibling models are biased when the outcome for one sibling affects the exposure for another, and this is precisely the situation when the outcome is child mortality and the exposure is aspects of the mother’s reproductive behaviour. The goal of this analysis was to find out, by means of simulation, whether the bias really matters in practice.

Results: All simulation experiments showed that, when there was an effect of infant mortality on subsequent fertility, the estimated effect of higher maternal age was considerably more adverse than the true effect, while the effects of higher birth order and very short or very long birth interval were biased in the opposite direction.

Contribution: Although it is possible that the bias is unimportant in other situations than those examined here, a reasonable conclusion is that one should have serious doubts about sibling model estimates of effects of reproductive factors on infant or child mortality. Stated differently, we may know less about these effects than we tend to think and need other alternatives to a ‘naïve’ regression model than the sibling approach. Obviously, there may be problems also when analysing other child outcomes that affect subsequent fertility, through mortality or otherwise.