Strategies for assessing the impact of loss to follow-up on estimates of neurodevelopmental impairment in a very preterm cohort at 2 years of age

Very preterm birth (VPT, < 32 weeks of gestation) is associated with increased infant mortality and morbidity. Survivors have higher risks of poor physical health, neurodevelopmental impairment and psychological disorders than children born at term [1‐3]. Many countries have constituted longitudinal very preterm birth cohorts using population-based designs, to evaluate the longer-term health burden and to investigate the determinants of adverse long-term outcomes [4‐6]. One problem facing these cohorts is loss to follow-up which varies from 25 to 50% in most cohorts, but can be up to 70% [4‐7].

There are many reasons for loss to follow-up, including difficulties tracing the location of families who move, lack of time due to other family obligations or work, financial barriers and not wanting to be reminded of the circumstances of the child’s birth. Studies have shown that patients who are lost to follow-up differ from those who remain in the study, with most finding that they have lower socioeconomic status [8‐11]. Some studies have also found that children who are not included are more likely to be impaired [11, 12], although having data on the full sample to investigate this bias is uncommon. Despite limited empirical evidence, there are reasons to be concerned about attrition related to the child’s health condition since the time and psychosocial consequences of raising an impaired child may make families less mobile or willing to participate in research [13]. Loss to follow-up can undermine the representativeness of estimates and introduce selection biases when the factors affecting follow-up are associated with health and developmental outcomes [14, 15].

Most studies of very preterm cohorts include a description of the characteristics of children lost to follow-up, but several analytical strategies are available for going beyond a qualitative assessment and adjusting for attrition, notably inverse probability weighting (IPW) and multiple imputation (MI) [16‐19]. IPW creates a pseudo-population where individuals are weighted to represent the inverse of the probability of follow-up conditional on baseline covariates; individuals who are under-represented in the follow-up sample will be assigned a larger weight and those over-represented will have a lower weight. MI replaces each missing value with a set of plausible values based on the distribution conditional on the observed data.

These methods assume that all possible variables associated with the missingness are included in the model so that data are missing at random (MAR). Otherwise, data are missing not at random (MNAR). Checking for the MAR assumption is difficult in general because the data needed to perform such checks are missing. Sensitivity analysis can be performed in these cases. For example, we can impute under MAR, shifts to the imputations by an amount delta to account for the imperfect MAR assumption, and re-analyze the data. If the result does not change under the values for delta, then we may conclude that the estimate is robust to violations of the MAR assumption. This method is called the delta method [20, 21].

In this study, we assess how the estimated prevalence of neurodevelopmental impairment at 2 years of corrected age (CA) among very preterm infants varies under two statistical methods (IPW and MI) and study the robustness to the MAR condition.

At two-years, parental questionnaires were not returned for 45.8% (n = 796) children, varying from 32.5% in Portugal and 52.9% in the UK (Table 1). Among responders, neurodevelopmental impairment, our principal outcome, was missing for 11.1% of children meaning the prevalence of the outcome was computed on only 48.2% of children surviving at 2 years old. These missing values were mainly due to cases in UK regions (85 of the 104 missing observations). Table 1 also illustrates the impact of having missing data on perinatal variables. The complete case dataset represented 85.5% of the total sample, leading to a 14.5% loss of data with a distribution of 17.7% for non-responders and 11.7% for responders to the 2 year follow-up. The deprivation index had few missing observations (4.2%); it was well correlated with individual measures of socioeconomic disadvantage that were collected at 2 years and therefore available for the follow-up sample only (Table S1).

Table 2 shows the perinatal characteristics associated with follow-up and impairment. Neonatal morbidities and postnatal care were not associated with non-response except for children having received surfactant and those breastfeeding at discharge, who were more likely to be followed-up. In contrast, maternal factors such as younger age, having more than one child, foreign-birth/ethnicity, having had a previous cesarean section and PPROM during pregnancy were associated with non-response. Unlike the association with follow-up, neonatal morbidities and care were most strongly associated with the presence of neurodevelopmental impairment. Males were more likely to have an impairment as were children with more than three siblings. Living in the most deprived neighborhoods was strongly associated with loss to follow-up, but the association with neurodevelopmental impairment was not statistically significant. Having a missing outcome among responders was associated with being foreign-born, but not other baseline characteristics or small area deprivation (Table S2).

Table 3 compares the prevalence of neurodevelopmental impairment before and after correction for loss to follow-up. All three approaches, IPW using the complete-case dataset, IPW using the imputed dataset, and multiple imputation, provided higher prevalence estimates, with relative increases of about 10% over the crude prevalence (range from 5.4 to 10.9%). This table also provides results separately for responders and non-responders, derived from the MI models: 18.6% (95CI (16.0%; 21.2%)) for responders versus vs 23.0% (95CI (17.4; 28.7) for non-responders.

Table 4 presents the sensitivity analyses simulating MNAR mechanism with values ranging from 0.8 to 1.5 which produced prevalence estimates from 19.0% (95% CI: 16.2 21.8) to 23.6% (95% CI: 20.1 27.1) versus 20.4% (95% CI: 17.3 23.4) under MAR. The UK regions varied from 20.5% (95% CI: 16.6 24.3) to 26.1% (95% CI: 21.3 30.8) compared to 21.9% (95% CI: 17.9 26.0), Portugal regions from 16.3% (95% CI: 12.5 20.0) to 19.0% (95% CI: 15.0 23.0) compared to 17.4% (95% CI: 13.4 21.4).

Loss to follow-up in this sample of children born very preterm was associated with most maternal sociodemographic factors (younger age, foreign born, multiparous) as well as area-based deprivation scores, but with few clinical and neonatal variables. Using statistical methods to account for loss to follow-up led to a modest relative increase, between 5.4 and 10.9%, in the estimated prevalence of moderate to severe neurodevelopment, with marginal differences in individual country estimates. Results were consistent using MI and IPW techniques. The estimated prevalence of moderate and severe neurodevelopmental impairment was 20.4% (95% CI: 17.3–23.4) and 20.0% (95% CI: 16.9–23.1) for MI and IPW models, respectively, versus the crude prevalence of 18.4% (95% CI 15.9–21.2). Sensitivity analyses adopting the extreme assumption that the prevalence of neurodevelopment delay was 1.5 times greater for those lost to follow-up or with a missing outcome gave an estimated prevalence of 23.6% (95% CI: 20.1–27.1). The relatively small differences between the more tempered scenarios (e.g., in the range 0.8 to 1.2) suggest that the prevalence estimate is robust to small to medium violations of the MAR assumption.

In our sample, the response rate was 54.2% which is at the lower end reported in other very preterm cohorts in early childhood which range from about 50 to 90% [6, 7, 12, 38, 39, 40]. Follow-up rates tend to be higher for studies from neonatal networks when compared to population-based studies and those that have regular contact with families after discharge. Our results on the factors associated with loss to follow-up are consistent with studies on very preterm birth [5, 6, 7, 39, 40] and other cohorts [8, 9, 41, 42] which find that non–responders have lower socioeconomic status, as measured by maternal education, social deprivation scores, migrant status and young maternal age. These studies also report differences in exposures associated with socioeconomic status, such as breastfeeding and tobacco use [7, 12, 40]. These factors, in particular, maternal age, migration, maternal education and breastfeeding have all been associated with neurodevelopmental outcomes after VPT birth [43, 44].

As in our study, fewer and less consistent associations have been observed between follow-up and perinatal variables which are the strongest predictors of poor outcome for VPT children [44], as we saw in this study; many of the studies cited above report similar neonatal characteristics and morbidities for responders and non-responders. Higher birthweight has been related to lower follow-up rates in some neonatal network studies which may reflect a closer connection between families and the NICU for higher risk low-birthweight infants [38, 39]. In contrast, in some studies, families with more impaired children were less likely to attend clinical assessments or responded by postal questionnaires only versus full participation. This attrition mechanism, i.e. loss to follow-up resulting from the child’s health status, is of particular concern for VPT research.

Both IPW and MI are appropriate statistical approaches to correct estimators with baseline information and to produce standard errors that consider the uncertainty caused by missing data [16, 17, 19] and have been recommended by research advisory boards [18]. In the presence of loss to follow-up, complete-case analysis will be unbiased when data are missing completely at random or if the outcome is not included in the missingness mechanism once accounting for all remaining variables. In our analysis, these two approaches yielded similar prevalence estimates after correction, however, multiple imputation has several advantages over weighting. First, it provides an estimate of the level of impairment among non-responders conditional on fact that the imputation model was well specified. MI also allows for sensitivity analyses of the MAR assumption based on the delta method. Finally, another feature, which we could not use in this study, is the time invariance of MI, meaning that if some children are included in future follow-up waves of a cohort, their data from these time points can be used to improve precision at earlier time points.

It is difficult to compare our results to the literature because this bias is rarely quantified. In the French Epipage 2 study of very preterm births < 32 weeks of gestation, MI was used to adjust for non-response at two-years; these adjustments were generally concordant with ours: estimates of cerebral palsy prevalence rose from 4.6 to 4.8% (4% increase) and, among children without severe motor impairments, the percentage with Ages and Stages Questionnaire (ASQ) scores below the screening threshold increased from 42.0 to 47.8% (a 14% increase) [40]. The larger adjustment for the neurodevelopmental measure, compared to cerebral palsy, may reflect the stronger impact of social factors on neurodevelopment than on severe gross motor impairment [45, 46]. Corrections for attrition may also have more impact on other developmental outcomes that are more socially patterned, such as language capabilities [47] or cognition in later childhood. Having continuous as opposed to dichotomous measurements of neurodevelopment, such as parent reported scores or standardized clinical measures of cognition may also lead to different results.

Despite the importance of socioeconomic factors in determining participation in follow-up, measures of social status are often not included in baseline data because of the absence of such information in medical records. We therefore carried out our analysis in two regions which had a measure of socioeconomic status in addition to other sociodemographic or behavioral variables associated with social status which are more often available in medical records (maternal age, country of birth, parity, breastfeeding). However, although neighborhood deprivation was strongly related to follow-up it was not related to neurodevelopment in this sample.

Despite high attrition, adjustment for loss to follow-up in our sample led to a modest approximately 10% relative increase in estimates of neurodevelopmental impairment. Simulation exercises showed these to be relatively insensitive to MAR violations. These results are likely applicable to other cohorts, given the similarity between the factors affecting follow-up in our data and those reported in the literature. Given the wide availability of IPW and MI techniques in most software packages, systematically addressing loss to follow-up in VPT cohorts using these methods should become standard practice to provide more accurate estimates, draw attention to this important issue and build up the evidence-base on the impact of attrition in varying contexts for multiple key developmental and health outcomes.