產前多重暴露與七歲兒童過動行為之相關性 – 各種統計分析方法比較

Abstract

背景 過去探討產前暴露對兒童注意力過動或缺陷症狀的關係主要針對單一暴露，雖近日多增加產前多重暴露對嬰幼兒出生體重或兒童肥胖的研究，然尚未有針對產前多重暴露與七歲兒童注意力過動或缺陷症狀的研究。 研究目的 本研究比較不同統計方法於產前多重暴露與七歲兒童缺乏注意力的相關性。 研究方法 從 2004 年到2005 年臺灣出生世代研究共收集486 名懷孕女性，測量臍帶血中18 種重金屬、9 種塑化劑代謝物、雙酚A 與全氟碳化物等。追蹤至七歲共有192名完成兒童注意力過動或缺陷症狀評估問卷，大於80 百分位則被視為有嚴重注意力過動或缺陷症狀，使用不同統計方法因素分析(Factor Analysis, FA)、主成分分析(Principal Component Analysis, PCA)、偏最小平方法(Partial Least Squares , PLS) 與彈性淨回歸法(Elastic Net regression, EN)，線性回歸模式得出估計值。 結果 利用偏最小平方法 (Partial Least Squares, PLS) 與彈性淨回歸法(Elastic Net regression, EN) 兩者方法所挑選與注意力行為有關相同的變數為：MEHP、Zn、PFUA；挑選不同的變數為：Cotinine, PFNA, NP。利用因素分析(Factor Analysis, FA) 與主成分分析(Principal Component Analysis, PCA) 兩者方法產生類似結果，主要在塑化劑與全氟碳化物與注意力行為顯著負相關。 結論 偏最小平方法 (Partial Least Squares , PLS) 能同時考慮變數間與結果的共變異；而彈性淨回歸法(Elastic Net regression, EN) 則能在線性模式同時解決變數間的高度相關性，因此依研究設計與目的，選擇合適適配模型。

Background: Previously, many studies focus on postnatal exposure to single chemical has association with attention deficit/hyperactivity symptoms (ADHD) in children. However, few of them studies prenatal environmental exposure may induce ADHD in 7 year-old children. Objective: To evaluate the association between prenatal multiple exposure and child attention deficit/hyperactivity symptoms at 7 years of age, we compared different statistical methods. Methods: 486 pregnant women were enrolled in a Taiwan Birth Panel Study (TBPS) cohort study between 2004 and 2005. We measured the concentrations of nine phthalate metabolites, bisphenol A, eighteen heavy metals including manganese, arsenic, and lead, cotinine in cord blood. Among 192 offspring at 7 years, attention deficit/hyperactivity symptoms were accessed based on parental reported, Chinese version of Swanson, Nolan, and Pelham (SNAP-IV). We used different statistical methods, including factor analysis, principal component analysis (PCA), partial least squares (PLS), and Elastic Net regression (EN) to fit models. Furthermore, inattention symptom was defined as score over 80 percentile. Results: There same variables (MEHP, Zn, PFUA) were selected by EN and PLS, but there were also three different variables (Cotinine, PFNA, NP). Both phthalate and PFASs were showed negative association with inattention symptom by FA and PCA. Conclusions: PLS takes maximum covariance between original variables (X’s) and outcome into consideration, on the other hand, EN selects few important variable to represent original variables. In our study, comparison dimension reduction methods between PLS and EN showed that EN had a better prediction base on char. Nevertheless, there is no guarantee that EN generated a better result. Still, it relies on study design and purpose to select a fit model.