以階層結構方程模型探討社會流行病學相互關聯因子對於肥胖相關指標之影響

Abstract

背景 雖然許多理論透過基於隨機分派試驗或者觀察性研究的介入計畫發展改善肥胖的指引（例如身體自我效能理論、計畫行為理論、跨理論模型，理性行為理論等）；然而，這些介入計畫是否能夠成功推廣運用到其他場域、族群、國家或地區，事實上相當依賴當地導致肥胖的潛在流行病學因素。 許多研究探討導致肥胖的相關因子之間的關係，卻鮮少探究流行病學特徵與人口群體特徵影響肥胖相關因子的路徑關係。多數先前的研究透過傳統兩階段的迴歸模型檢視導致肥胖的相關因子之間的關聯性，但該方法並無法考量變數之間的中介過程。而從統計的觀點出發，在有限樣本之下，傳統的統計方法會產生多重比較和過度參數化的問題；更重要的是，傳統方法在分析上僅使用觀察變數而非潛在的因子。 目標 本研究的目標是應用以概似值為基礎以及貝氏取向的結構方程模式，試圖釐清導致肥胖的相關因子與錯綜複雜的社會流行病學特徵之間的關係，並以社區整合式篩檢資料進行分析；該資料提供多種癌症和慢性病的篩檢服務。進一步採用貝氏階層結構方程模型(Bayesian hierarchical SEM)同時考量個人、區域層級以及不同層級之間的交互作用對於肥胖的影響。 材料與方法 本研究利用2002至2010基隆社區整合式篩檢資料進行分析。基隆市社區整合式篩檢每年邀請20歲以上民眾參加，篩檢服務內容包括五項癌症及三項慢性病篩檢。人體生理相關檢測由通過訓練之公共衛生護士進行量測，量測項目包括身高、體重、腰圍、臀圍等。人口學特性、飲食行為、頻率及食物種類及攝取量、健康或不健康生活型態等則利用結構式問卷以面對面方式進行訪視並記錄。生化相關指標則於篩檢現場同時採8小時禁食血液進行生化檢驗。 本研究利用傳統多變量迴歸分析模式調整相關因素後，探討影響肥胖相關因子之變項。本研究利用探索性因素分析進行組成構面指標之探討，並進一步利用路徑分析模式進行構面與構面間之關聯性分析，以建立影響肥胖相關因子之結構方程模型。利用AIC及BIC模式檢定指標以挑選出以目前資料所表現最適之結構方程模型。進一步利用多階層貝氏結構方程模型同時考慮個人及地區層級之資料，探討相關構面對於肥胖相關因子的影響及不同層級之間的交互作用。 結果 本研究共有75077位成年人，包含30042位男性（49.54±16.06歲） 和45035位女性（46.58±14.4歲）。本研究透過個人特質、肥胖特徵、身體代謝指標、飲食型態和習慣，透過標準化的方法得到七個顯著的潛在因子：社經地位（SES）、健康動機（HM）、飲食型態（EP）、規律飲食（RD）、多樣性攝取（DI）、不健康習慣（UH）和肥胖相關生物標記（OB）。在考慮社經地位、健康動機、飲食型態和規律飲食間的相關後，得到研究中最佳的結構方程模型。模型中的路徑關係包含兩個層次、13條顯著的路徑關係。在第一層中，社經地位（SES）所直接影響規律飲食（RD）、多樣性攝取（DI）、和肥胖相關因子（OB）的路徑係數分別為-0.055、0.219和-0.078；由不健康習慣（UH）所直接影響的路徑係數分別為0.295、0.350和0.253；由飲食型態（EP）所直接影響的路徑係數分別為-0.080、0.149和0.074。而由健康動機（HM）所直接影響多樣性攝取和肥胖相關因子的路徑係數分別為-0.656和-0.311。在第二層的多樣性攝取（DI）以及規律飲食（RD）對肥胖相關因子（OB）的影響中，路徑係數分別為-0.359以及-0.024。 透過貝氏方法亦可得到相似的結果，不過更重要的是，貝氏階層結構方程模型指出區與區之間在肥胖相關因子的變異上有很強的隨機截距效果（基礎值的影響），且區域層級的社經地位與個人的層級的因子之間（例如健康動機等等）亦有顯著的交互作用關係。 結論 本研究透過七個潛在構面釐清導致肥胖的相關因子與錯綜複雜的社會流行病學特徵之間的關係，包含社會經濟地位、健康動機、不健康習慣、規律飲食、飲食型態、多樣性攝取和肥胖相關因子，且這七個潛在構面對於肥胖相關因子的效果亦受區域層級社經地位的脈絡因子所影響。貝氏階層結構方程模型及其應用在社會流行病學的資料分析上，提供在探討多階層潛在變數之間相互關係上的新典範，且對於肥胖相關因子的預防上有重要的意涵。

Background In spite of numerous theories, including physical self-efficacy theory, planned behavior theory, transtheoretical model, theory of reasoned action and so on, developed for the guidance of ameliorate obesity through intervention program either based on randomized controlled design or observational studies, whether the results of these intervention program can be generalized or applied to other settings, ethnics, countries, and areas is highly dependent on the underlying epidemiological causes leading to obesity in each of place. Although a body of evidence on the relationship of each correlate to obesity has been documented, elucidating the pathway of how these epidemiological and population-based characteristics are connected and affect obesity-related markers has been barely addressed. Most previous studies examined the relationship using the traditional two-state regression model, which cannot take intermediate process between variables into consideration. From the statistical viewpoint, traditional approach might have troubles in multiple comparisons and also over-parametrization given limited sample size. More importantly, only observed variables instead of latent contextual variables were implicated. Aim Our study aim was to apply likelihood-based and Bayesian-oriented structural equation model (SEM) to clarify the relationship interwoven with these socio-epidemiological characteristics leading to obesity-related phenotypes based on the community-based integrated screening data that offered various screening modalities for manifold cancers and chronic diseases. Bayesian hierarchical SEM was applied to assessing the influences among individual, district-level data (multilevel/ hierarchical), and interactions between latent constructs at different levels simultaneously. Materials and Methods The participants of Keelung Community-based Integrated Screening (KCIS) program during 2002 to 2010 were recruited for this study. Residents living in Keelung and aged 20 years or older have been invited to participate this program for mainly screening five neoplastic diseases and three non-neoplastic chronic diseases. Anthropometric measurements, including body weight, height, waist circumference and hip circumference, were measured by trained staff in each visitation. Information on demographic characteristics, dietary behaviors and intake diversity in frequency and quantity, health and unhealthy behaviors, and life styles were collected through face-to-face interview using a structured questionnaire conducted by trained interviewers. The biomarkers were also simultaneously collected and examined by central laboratory using the 8-hour fasting blood serum. The traditional multivariate regression method was generally employed in the studies to examine the relationship between observed variables and obesity-related factors. The exploratory factor analysis was conducted to cluster the indicators for each constructs and the path analysis was employed to constitute the candidate structural equation model (SEM). Both criteria of Akaike’s information criterion (AIC) and Bayesian information criterion (BIC) were used for parsimonious model selection. We also conducted hierarchical Bayesian SEM to examine the main effect and interaction between district-level and individual information. Results The overall 75077 subjects aged 20 years and older were recruited as study population, including 30042(age 49.54±16.06 and 45035(age 46.58±14.4) for male and female respectively. Based on the personal characteristics, obesity traits, metabolic biomarkers, life style of eating patterns, and habits, 7 clusters with significant factor loadings were generated by standardized approach, including socioeconomic status (SES), health motivation (HM), eating patterns (EP), regular diet (RD), diverse intakes (DI), unhealthy habits (UH), and obesity-related biomarkers (OB), which was nominated as best SEM after taking the significant associations among SES, unhealthy habits, health motivation, and eating patterns into account. There were two main pathways identified from constituted the SEM with 13 significant pathways. For the first layer, the path coefficients from SES were -0.055, 0.219, and -0.078, from unhealthy habits were 0.295, 0.350, and 0.253, from eating patterns (EP) were -0.080, 0.149, and 0.074, to regular diet (RD), diverse intake (DI), and obesity-related factors, respectively. The path coefficients from health motivation were -0.656 and -0.311 on diverse intake (DI) and obesity-related biomarkers. For the second layer, including diverse intakes and regular diet, both path coefficients on obesity-related biomarkers were -0.359 and -0.024, respectively. The similar findings were noted while Bayesian hierarchical SEM was used but , most importantly, there was a strong random intercept (baseline influence) effect of the variation of obesity-related phenotypes among districts and also significant interaction between district SES and other latent constructs (such as health motivation and so on). Conclusion Our study developed the seven latent constructs to clarify the relationship interwoven with these socio-epidemiological characteristics leading to obesity-related phenotypes, including socio-economic status (SES), health motivation (HM), unhealthy habit (UH), regular diet (RD), eating pattern (EP), diverse intakes (DI), and obesity-related phenotype (OB) and effect of these seven latent constructs on obesity-related phenotype was modified by the contextual factor of district SES. The development of Bayesian hierarchical SEM and its application to socio-epidemiological data here provide a new insight into a new paradigm on the pathway interplay with the underlying latent variables in multilevel, which has a significant implication for prevention of obesity-related phenotypes.