In early life, a child's physical growth and stature (e.g., weight, height, and head circumference) are shaped by genetic, nutritional, socioeconomic and other environmental factors. They are often examined as an outcome or exposure in population health research. In the framework of life course epidemiology or studies of the developmental origins of health and disease, various child growth patterns—such as 'impaired', 'excessive' or 'catch-up' growth—and the identification of 'critical' or 'sensitive' periods of growth related to health outcomes in later life have also been of particular interest [1]. In the current issue of Paediatric and Perinatal Epidemiology, Giacomini and colleagues [2] examine sex-specific associations between the growth of the head circumference in the first 5 years of life and potential behavioural problems at age 5 in a longitudinal birth cohort of 303 girls and 318 boys in Brazil. The study provides an opportunity to consider crucial but frequently overlooked methodological issues in studies of postnatal growth and its associations with later childhood outcomes. 'Growth' inherently means changes in a body size measured over a specified period. There are multiple ways to conceptualise and quantify growth in epidemiological research [3, 4], and importantly, the choice of growth metric and analytic approach to defining growth affects inferences and interpretations [5]. Gioacomini et al. [2] defined a child's head circumference growth in the first 5 years as the change in sex- and age-specific z-scores (HCZs) at birth and age 5 based on WHO standards, reflecting the absolute change in a child's relative size compared with an external reference measured at the two-time points, which may be referred to as a 'change score'. Gioacomini et al. [2] then categorised the change score into quintiles, with the lowest quintile of the change score as the reference group, which they referred to as 'impaired growth.' The change between two time points (t0 and t1) is an intuitive measure of growth in the absence of variability in actual measurement timing, but its use in regression modelling is not as straightforward as it might seem. First, the change score is mathematically coupled to the baseline head circumference (at t0). Due to regression to the mean, larger magnitudes of change in head circumference between birth and age 5 are expected for children whose head sizes at birth are closer to the tails of the HCZ distribution. This is reflected in their eTable 3 where the largest HCZ growth between birth and age 5 occurred among children whose HCZ at birth was categorised as small for gestational age, and the least 'growth' (in fact, the decline in the change score) among those whose HCZ at birth was large for gestational age. Gioacomini et al. [2] appropriately attempted to address regression to the mean by adjusting for head size at birth in the multivariable models so that the inferences about head growth are unaffected by baseline variation. Second, the coefficients for change scores as a measure of 'growth' do not provide clear interpretations, particularly when multiple age time points or intervals are simultaneously modelled. When there are only two measures of body sizes to quantify growth, the regression coefficient for the change score between time t1 and t0, adjusting for the size at time t0, is identical to those estimated using two commonly used alternative approaches: (i) regression of the outcome on size at time t1, adjusting for the size at time t0 (often referred to as a life-course model); and (ii) a two-stage approach whereby individual child-level model residuals are first generated by regressing size at time t1 on the size at time t0, then the residuals from the first stage are used as the exposure measure of interest in a second-stage regression model (referred to as conditional growth modelling). However, interpretations and meanings of results from these three approaches for each parameter are not identical. Interpretations of the coefficients for change scores across multiple time intervals are not intuitive, as each interval-specific change score estimate represents the cumulative 'effect' of growth up to the interval and later intervals [4]. Thus, change scores should be considered and interpreted cautiously, particularly with the increasing number of interval-specific growth measures. Another issue for consideration concerns generating z-scores as a body size or 'growth' measure and identifying 'impaired' growth. Expression of an anthropometric measurement as a z-score represents the relative standing of each child in comparison with a reference distribution. Therefore, changes in z-scores quantify growth differently from changes in absolute body sizes expressed using raw measures. The use of z-scores has an advantage in statistical modelling in that it accounts for different degrees of variances of body size (or change in size) distributions measured across ages due to differential rates of growth with age in childhood. However, the choice of reference or standards used to generate the z-scores can influence the interpretations of empirical analyses. The WHO growth standards used in Gioacomini et al. [2] are prescriptive—they define 'how children should grow under optimal conditions', including free of disease and healthy practices such as breastfeeding and nonsmoking [6]. These standards are widely used and recommended, yet numerous studies have demonstrated discrepancies in body size distributions, including head circumference, between the WHO standards and samples of comparable children in many countries [7, 8]. These results warrant a clear rationale for using WHO standards (or other references) to assess head growth within a study population. The choice of reference becomes more critical when investigators aim to categorise some children as having 'impaired' or 'excessive' growth, which implies that there are known clinical implications of memberships in such categories. Defining children in the bottom 20% of the distribution of change z-scores as 'impaired' growth is not only arbitrary but also unintuitive, given the defined 'impaired' growth reflects regression to the mean as discussed above. Contextual factors affect the distribution of anthropometric measures in a study population and, therefore, may also need to be considered when choosing a reference and identifying a particular change as 'impaired' growth. For instance, the Zika virus epidemic, first reported in May 2015 and spread across Brazil since then, has been linked to neonatal microencephaly [9]. The HCZ change score based on the WHO standards and the quintile-based category of 'impaired' growth among children born in 2015–2016 in Gioacomini et al. [2] may not reliably represent 'impaired' HC growth in other settings. Finally, Gioacomini et al. [2] also examined whether a 'sensitive' period of head circumference growth during the first 5 years of life may be more important for later problematic behaviours. They observed that the association between HCZ change from birth to age 5 is mainly driven by HCZ growth in the first 2 years. They separately assessed the associations of HCZ change from birth to age 2 and HCZ change from ages 2 to 5 with the behaviour scores. While this is an additional insight, a more comprehensive assessment could have been made by using all head circumference measures available in the study—at birth, 1, 2 and 5 years. Given the most rapid growth occurs shortly after birth followed by deceleration, growth in infancy (the first year) may be a more critical time window than the second year for behavioural problems in later childhood. The added complexity for the 1-year measurement can be managed using alternative growth metrics and analytical methods. For instance, the conditional growth modelling approach mentioned above could use a child's head circumference at each age after birth, adjusted for all prior measurements, as a growth metric to estimate interval-specific associations simultaneously, avoiding collinearity. This represents deviations from the expected size based on the child's earlier growth. Monitoring child physical growth and examining its long-term sequelae is essential in clinical practice and our efforts to improve population health. However, growth metrics that seem intuitive for describing a change in an individual child (e.g., change score) may not be the optimal approach in epidemiological analyses. Different definitions of growth and analytical approaches to estimating the effects of growth can yield different results, and they are complementary. Therefore, a comprehensive and robust understanding of child growth and its effects on later outcomes is most likely achieved by multiple approaches to quantifying and analysing growth in any given study. Both authors contributed to the conception and the content of the commentary. S.Y. drafted the manuscript and D.E.R. provided critical revisions. Both authors approved the final version. The authors declare no conflicts of interest. Data sharing not applicable as no datasets were generated or analysed for this commentary.
