利用流行病學資料建構因果圓派

Abstract

Rothman提出的因果圓派模式(causal-pie model)是流行病學家非常熟悉的流行病學方法。每一個因果圓派(causal pie)以一系列的組成因子(component cause)組成而成，只要集合所有因果圓派內的組成因子，就意味著疾病的必然產生。因此，在疾病的成因探討上，利用此因果圓派模式能夠同時考慮多因子作用且釐清所有可能危險因子對於疾病錯綜複雜的共同作用。然而，此模式一直以來只是一個概念示意圖，並未真正被應用在實證資料的分析。直到Hoffmann等人在2006年提出相關因果圓派分析方法，但該方法卻有重大缺失。有鑑於此，本研究即著手發展統計方法進行因果圓派的建構。 本研究建立一個因果圓派量化指標，稱其為因果圓派權重(causal-pie weight; CPW)。由於Rothman所提出的因果圓派模式中指出，各組成因子應皆為危險因子且各組成因子間的交互作用假定為加成性及協同性(additive and synergistic)，為了符合該假設，我們採用非負係數加成性迴歸模式來進行危險性的估計，並計算族群可歸因危險性(population attributable risk; PAF)。因果圓派權重可成功的經由可歸因危險性矩陣方程式求得，另採用拔靴法(bootstrap)即可估計其信賴區間。 為使此因果圓派模式分析的概念能夠有較為廣泛的應用，本研究探討相關之方法學議題，包括：(1) 二等級或多等級危險因子之因果圓派分析方法；(2) 探討適用於不同研究設計及資料型態之加成性迴歸模式； (3) 干擾因子的調整以獲得危險指標的不偏估計。 此外，本研究亦與Hoffmann等人所提出的定量因果圓派方法進行比較。我們同時利用兩方法進行一筆虛擬資料的分析，結果發現Hoffmann等人提出之方法不但無法反映資料的真實情況，甚至因果圓派值產生不合常理的負值；而本方法卻沒有此缺失。 我們利用電腦模擬以評估本研究方法的正確性，不論因果圓派權重指標值、信賴區間及危險性估計值(特別是危險勝算比)，模擬結果皆與情境真值呈現令人滿意的一致性。 最後，我們真正將此套方法應用在流行病學資料分析上，透過三種不同研究設計的資料形式，包括病例對照研究及世代研究，探討多因子對疾病的作用，建構疾病的因果圓派。同時，實證資料分析亦突顯了因果圓派權重估計值的穩定性。 透過一系列的研究，我們不難發現，因果圓派模式確實能夠發揮其優勢，讓研究者釐清各危險因子對於疾病錯綜複雜的共同作用。

Epidemiologists are familiar with Rothman’s model of causal pies. A causal pie contains a combination of component causes and people will inevitably develop disease with the completion of it. Rothman’s causal-pie model helps to clarify the multi-factorial and complex interactive nature in disease causation. However, the model has been more of a theoretical concept than a practical tool of data analysis. Until 2006, Hoffmann et al. proposed a method for calculating the ‘proportion of diseased subjects who develop the disease due to specific classes of causal pies’ (PDCs), based on actual case-control data. However, their method is flawed in various ways. The aim of this study is to develop the statistical methodology for causal-pie modeling. We construct a new index, the ‘causal-pie weights’ (CPWs). The CPWs serve to quantify the relative importance of each and every class of causal pies. To be conforming to Rothman’s model, a nonnegative additive model is used to constrain all the risk ratios to be equal to or greater than one, and the interactions between them to be additive or supra-additive. Based on these constrained risk estimations, we then calculate the population attributable fractions (PAFs). The CPWs can then be calculated from the PAFs using a matrix equation. To calculate the confidence intervals of CPWs, the bootstrap technique is applied. A number of issues related to causal-pie modeling is introduced here to give the methodology a more extensive application, such as (1) developing methods specific for dichotomous or for polytomous risk factors; (2) using various additive regression model for pure-count or person-time data; (3) adjusting confounding effects to acquire the unbiased risk estimates. Besides, the methods proposed by us and by Hoffmann et al. are compared in the study using a hypothetical data. CPWs and PDCs are then calculated. And we find that the association between exposure and disease will be mistakenly claimed and the PDC value will be negative, which is clearly inadmissible in Rothman’s causal-pie model. We conduct computer simulation to evaluate the validity of the procedure for constructing causal pies, in which simulated risk estimates, especially for OR, CPW, and confidence interval are calculated. It is satisfied that the comparable results between simulation and the scenario are acquired. In the last part of the study, we try to apply our causal-pie modeling technique to analyze epidemiological data, including case-control and cohort data. Through a series of studies, we conclude that the methodology for causal-pie modeling can indeed help researchers to clarify the multi-factorial and complex interactive nature in disease causation.