Do non-participation and drop-out from cohort studies bias estimates of socioeconomic inequalities in health?

Individuals of lower socioeconomic position (SEP) may be less likely to consent to participation at the start of a cohort study and more likely to drop out over time.1,2 It has also been shown in several cohorts that participants tend to be healthier than non-participants.1,2

Scandinavian studies exploiting record linkage have shown that despite cohorts being a relatively wealthy and healthy sample of the population, exposure-outcome associations are not biased by this selection, at least for certain well established relationships.2 Within ALSPAC, there is evidence that the prediction equations for measures of disruptive behaviour do not differ between i) individuals who were eligible to participate but did not consent to the study, ii) those who consented but dropped-out before the age of 8 years, and iii) those who continued to participate at 8 years3, again suggesting that selection bias may not affect estimates of exposure-outcome associations. It is not known, however, whether this is true when estimating socioeconomic inequalities in health. Since continued participation in the cohort is likely to be associated with both high SEP and better health, this may lead to estimates of health inequalities at later ages being underestimates.

Aim 1: To investigate whether nonparticipation and drop-out from cohort studies leads to underestimates of socioeconomic inequalities in health

We will make use of a set of outcomes for which data is available for (almost) the full cohort:

1. Perinatal factors: birth length, birth weight, gestational age at delivery, breastfeeding

2. Maternal factors: maternal obesity, maternal smoking during pregnancy

3. Child factors from routine data: Educational attainment at key stages 2 and 3

First, we will examine the association between these outcomes and participationat various stages of the cohort (e.g. attendance at the Focus@9 clinic and the Teen Focus 3 clinic) to explore whether those who participate at each stage have favourable levels of each outcome compared with those who do not participate. For educational attainment, we will also be able to explore whether those who consented to participate at the start of the study have different educational attainment compared to eligible children identified from the National Pupil Database who have never participated.

Next, we will calculate the socioeconomic inequality in each of these outcomes in the full ALSPAC cohort, using maternal education as the measure of SEP. We will then restrict our analysis firstly to those who attended the Focus@9 clinic, and second to those who attended the Teen Focus 3 clinic. For each of these restricted samples, we will again calculate an estimate of the association between maternal education and each outcome. We will be testing the hypothesis that our estimates of socioeconomic inequalities will attenuate towards the null as the sample gets more and more selected.

We will formally test whether the association between maternal education and each outcome differs between those who do and do not participate at each stage, using tests for interaction. We will then carry out sensitivity analysis to assess how different the associations would have to be between participants and non-participants for our analyses to lead to very different conclusions.

Aim 2: To identify whether problems caused by selection bias for the estimation of health inequalities are aggravated by using a categorical SEP indicator

The indicator of SEP most frequently used in epidemiological studies is maternal education. Within ALSPAC, we tend to use a four category variable - less than O-Level, O-Level, A-Level, and Degree or higher. This four category variable is unlikely to be capturing all of the very broad concept of SEP; it is possible that within each category of maternal education, those who drop out are actually of lower SEP. If so, this would result in the observed socioeconomic inequalities being an underestimate.

Our first step towards exploring this issue will be to examine whether, within each of the four categories of maternal education, there is an association between drop-out from the cohort and other measures of SEP (income, household social class, etc).

Subsequently, we will construct a multi-dimension measure of SEP using several variables (income, household social class, car ownership, index of multiple deprivation, reported financial difficulties, maternal age and parity, etc) using factor analysis. This SEP construct should be able to more finely differentiate between participants than maternal education alone can. We will then repeat the analysis outlined above for Aim 1, using this latent construct of SEP as the exposure instead of maternal education. That is, we will examine whether we still see an attenuation of estimates of inequality in the (almost) fully observed outcomes as the sample gets more and more restricted when using this latent construct of SEP.