Wills Andrew K, Strand Bjørn Heine, Glavin Kari, Silverwood Richard J, Hovengen Ragnhild
School of Clinical Sciences & School of Oral & Dental Sciences, University of Bristol, Bristol, UK.
Norwegian Institute of Public Health, Oslo, Norway.
BMC Med Res Methodol. 2016 Apr 8;16:41. doi: 10.1186/s12874-016-0143-1.
Regression models are widely used to link serial measures of anthropometric size or changes in size to a later outcome. Different parameterisations of these models enable one to target different questions about the effect of growth, however, their interpretation can be challenging. Our objective was to formulate and classify several sets of parameterisations by their underlying growth pattern contrast, and to discuss their utility using an expository example.
We describe and classify five sets of model parameterisations in accordance with their underlying growth pattern contrast (conditional growth; being bigger v being smaller; becoming bigger and staying bigger; growing faster v being bigger; becoming and staying bigger versus being bigger). The contrasts are estimated by including different sets of repeated measures of size and changes in size in a regression model. We illustrate these models in the setting of linking infant growth (measured on 6 occasions: birth, 6 weeks, 3, 6, 12 and 24 months) in weight-for-height-for-age z-scores to later childhood overweight at 8y using complete cases from the Norwegian Childhood Growth study (n = 900).
In our expository example, conditional growth during all periods, becoming bigger in any interval and staying bigger through infancy, and being bigger from birth were all associated with higher odds of later overweight. The highest odds of later overweight occurred for individuals who experienced high conditional growth or became bigger in the 3 to 6 month period and stayed bigger, and those who were bigger from birth to 24 months. Comparisons between periods and between growth patterns require large sample sizes and need to consider how to scale associations to make comparisons fair; with respect to the latter, we show one approach.
Studies interested in detrimental growth patterns may gain extra insight from reporting several sets of growth pattern contrasts, and hence an approach that incorporates several sets of model parameterisations. Co-efficients from these models require careful interpretation, taking account of the other variables that are conditioned on.
回归模型被广泛用于将人体测量大小的系列测量值或大小变化与后期结果联系起来。这些模型的不同参数化方式使人们能够针对有关生长影响的不同问题，但它们的解释可能具有挑战性。我们的目标是根据其潜在的生长模式对比来制定和分类几组参数化方式，并通过一个说明性示例来讨论它们的效用。
我们根据其潜在的生长模式对比（条件生长；更大与更小；变得更大并保持更大；生长更快与更大；变得并保持更大与更大）来描述和分类五组模型参数化方式。通过在回归模型中纳入不同的大小重复测量值集和大小变化集来估计这些对比。我们使用挪威儿童生长研究中的完整病例（n = 900），在将年龄别身高体重z评分的婴儿生长（在6个时间点测量：出生、6周、3、6、12和24个月）与8岁时儿童期超重联系起来的背景下说明这些模型。
在我们的说明性示例中，所有时期的条件生长、在任何时间段内变大并在婴儿期保持变大以及从出生就更大，都与后期超重的较高几率相关。后期超重几率最高的是那些经历高条件生长或在3至6个月期间变大并保持变大的个体，以及那些从出生到24个月都更大的个体。不同时期和不同生长模式之间的比较需要大样本量，并且需要考虑如何对关联进行缩放以使比较公平；关于后者，我们展示了一种方法。
对有害生长模式感兴趣的研究可能会从报告几组生长模式对比中获得额外的见解，从而从纳入几组模型参数化方式的方法中获得额外见解。这些模型的系数需要仔细解释，同时要考虑到作为条件的其他变量。
