TL;DR
This paper introduces the "Conditional Random Slope" (CRS) model, a novel statistical approach for estimating individual child growth velocity. The CRS effectively addresses "regression to the mean" (RTM) in longitudinal data, a common issue in epidemiological studies of child development, leading to more accurate associations between early growth and later health outcomes.
What Problem Does This Paper Address?
Understanding how children grow is crucial for identifying developmental origins of health and disease. Estimating a child's "growth velocity" – the rate of change in physical size over time – is a key metric. However, a significant challenge in this field is "regression to the mean" (RTM). RTM is the statistical tendency for extreme measurements to move closer to the average on subsequent measurements. In child growth, this means a child who is exceptionally tall or short at one point might appear to have a more typical growth rate later, simply due to this statistical artifact, masking true underlying growth patterns.
Existing methods for estimating growth velocity often fall into two camps: "unconditional" approaches that don't explicitly account for RTM, and "classical conditional" approaches that do but are limited to using only two data points per child (e.g., birth and 12 months). This limitation is problematic because modern longitudinal studies often collect multiple measurements per child within an interval, providing richer data. The paper aims to bridge this gap by developing a method that leverages multiple data points while robustly addressing RTM.
Methods
The core innovation is the Conditional Random Slope (CRS) model. This approach is built upon a linear mixed-effects (LME) model, a standard tool for analyzing longitudinal data with repeated measurements. The CRS model introduces a specific interaction term: a baseline size-time interaction. This term allows the model to disentangle the effect of a child's initial size from their subsequent growth rate, thereby directly accounting for RTM.
Key aspects of the CRS model:
- Leverages LME flexibility: It can handle multiple, irregularly timed measurements per child, unlike classical methods.
- Addresses RTM: The baseline-time interaction explicitly models and corrects for the influence of initial size on observed growth.
- Generalization: It extends the classical conditional standard deviation score (SDS) model.
The authors compared the CRS model against four alternative metrics:
- Unconditional change in LAZ: Simple difference in length-for-age z-score (LAZ) between two time points.
- Classical conditional change in LAZ: The difference between observed and expected LAZ at the end of an interval, based on baseline LAZ.
- Fixed slope estimate (unconditional): A linear regression of LAZ on time for each child, without accounting for RTM.
- Random slope estimate (unconditional): An LME model that estimates individual-specific slopes but doesn't explicitly correct for RTM.
Datasets
The study utilized data from two birth cohorts:
- Cohort A: Vellore, India
- Cohort B: Karachi, Pakistan
These cohorts provided longitudinal length measurements from birth up to 12 months of age, with varying frequencies of data collection. The authors ensured comparability by selecting data points to achieve consistent "data density" (number of measurements per child) for analyses.
Key Insights and Ideas
- Increasing data density alone doesn't fix RTM: Simply having more measurements within an interval (e.g., at 3, 6, 9 months) did not significantly attenuate RTM in unconditional models. The influence of baseline and end-point measurements remained strong.
- Unconditional models misclassify growth: Unconditional velocity metrics showed only moderate correlation and poor agreement with conditional metrics (like CRS) in classifying children with "abnormal" growth (defined as being in the bottom 10% of velocity). This suggests that unconditional methods can lead to different conclusions about which children are growing slowly.
- Conditional approaches are crucial for health associations: The association between infant growth velocity and later stunting (at 2 years) was significantly attenuated or even non-existent with unconditional metrics, especially in Cohort B. Conditional metrics, including CRS, consistently showed stronger and more reliable associations.
- CRS and classical conditional SDS are similar with two points: When only two data points were used, the CRS model produced nearly identical results to the classical conditional SDS, validating the CRS as a robust extension.
Key Results
- RTM persistence: Even with all available data points, unconditional metrics showed moderate negative correlations between baseline size and subsequent growth rate, indicating persistent RTM.
- Classification discordance: Kappa coefficients (measuring agreement) between CRS and unconditional metrics ranged from 0.38 to 0.78, indicating poor to moderate agreement in classifying children with slow growth. Discordance rates could be as high as 10.5%.
- Velocity-stunting association: For Cohort B, two-point unconditional velocity estimates were not associated with stunting at 2 years, while conditional metrics yielded robust associations. For example, the odds ratio (OR) for stunting per standard deviation decrease in velocity was 1.30 (95% CI: 0.96, 1.76) for unconditional delta vs. 8.76 (95% CI: 4.90, 15.69) for conditional delta. The CRS model also showed strong associations (OR: 9.03, 95% CI: 5.01, 16.3).
- Predictive accuracy: Receiver Operating Characteristic (ROC) curves showed that conditional models (including CRS) had substantially higher areas under the curve (AUC) for predicting stunting at 2 years compared to unconditional models. For Cohort A, AUC was 0.71 for conditional vs. 0.61 for unconditional. For Cohort B, AUC was 0.89 for conditional vs. 0.57 for unconditional.
Limitations and Future Work
- Gestational age: The study did not account for gestational age at birth, which could bias growth trajectory classifications, especially for preterm infants.
- Limited outcome: The study focused on length-for-age z-score (LAZ) and stunting. Future work should explore the applicability of CRS to other anthropometric variables and different health outcomes (e.g., obesity, neurodevelopment).
- Simplicity of data structure: The analyses assumed a relatively clean data structure with consistent baseline and follow-up measurements. Future research should investigate how to manage missing baseline measurements (e.g., using multiple imputation) and explore the impact of variable timing of baseline measurements more deeply.
- The "why" of associations: While the study demonstrated that conditional metrics are better for predicting stunting, further research is needed to fully understand why these associations differ for other health outcomes.
Summary updated