劑量與效應反應的統合分析之改良與應用：以血糖值與牙周病的關係為例

Abstract

背景 系統性文獻回顧以及統合分析在證據整合上應用的非常廣泛，研究者利用傳統統合分析對從系統性文獻回顧中找到的證據進行兩組之間的統合比較，但許多的觀察性研究會探討多個劑量類別與結果（效應）之間的關係。劑量與效應反應的統合分析（dose-response meta-analysis）中，每一劑量組均會與同一基準組進行比較，此時劑量類別與劑量類別的勝算比間具有相關性，現今常使用的統計軟體中需要每一劑量類別下的人數及事件發生人數來計算此相關性。如此，研究者則可藉由此方法了解藥物在何等劑量之下能達到最大的治療效益；或者疾病風險在何種暴露量下會達到最大的風險。 目的 本論文的研究目的主要發展一方法以補齊無提供每一劑量類別下之人數及事件發生人數的文獻其結果與結果間之相關係數，具體的目標如下： 1. 發展generalized least square (GLS)模型於固定效應的劑量與效應反應統合分析上，以算出無提供人數之文獻其結果與結果間之相關係數。 2. 發展R程式以應用GLS模型，基於每一劑量類別下之人數及事件發生人數已知以及未知的情況。 3. 發展R程式以應用GLS模型，以研究血糖值與牙周病風險之間的關聯。 資料來源 本論文實際進行系統性文獻回顧，收集血糖值與牙周病間的關係之文獻以及使用酒精與心血管疾病間的關係之資料以應用於本論文的統計方法。

結果 1. 本研究以GLS方法以及restricted cubic splines模型分析資料「血糖值與牙周病間的關係」：發現隨著血糖值的上升；牙周病的勝算亦是增加。 2. 本研究中發展R程式基於'pool-first'或是'pool-last'的統合架構下應用GLS，所統合得到的結果與Orsini等人利用Stata軟體以及Liu與Rota等人利用SAS軟體統合分析之結果相同。 3. 本研究假設暴露類別與暴露類別其勝算比間之相關係數為0.5來統合得到之結果，與透過實際人數所求得之相關係數而所統合的結果一致，此舉對於未來進行系統性文獻回顧時，能留下符合主題卻無報告每一劑量類別下的人數以及事件發生人數之文獻進入統合分析。 結論 Greenland與Longnecker所提出之GLS方法能有效的於統合估計時將劑量組與劑量組其勝算比間的相關性考慮進去，但需要每一劑量類別下之人數及事件發生人數來計算此相關性。而Orsini等人提出之restricted cubic splines模型提供研究者有效地於劑量與效應之間估計非線性的關係。此外，本研究所提出之相關係數的假設則供研究者能增加於統合分析時之文獻樣本。

Background Systematic reviews and meta-analysis are widely used for evidence synthesis. Traditional meta-analysis compare difference in outcomes between two interventions/treatments, but in observational studies, comparisons are sometimes made for groups with different levels of exposure to risk factors. Dose-response meta-analysis transforms the discrete levels of an exposure back to a continuous variable and then estimates a linear or nonlinear relation between the outcome (the response) and continuous exposure (the doses). However, as multiple levels of exposure are usually reported in a study, the outcomes are therefore not independent, and current statistical model for dose-response meta-analysis requires the input of the number of subjects with or without the outcome event at different levels of exposure to calculate the correlations between the outcomes. Objectives The main objective of this dissertation is to develop an alternative approach by imputing missing information for correlations between outcomes. The specific objectives are: 1. Developing generalized least squares (GLS) models for fixed effects dose-response analysis, when the correlations between outcomes of different exposure levels are unknown. 2. Implementing GLS models in R software package for known and unknown correlation structures between the outcomes. 3. Testing the relation between glucose levels and the risk of periodontal diseases by the means of proposed GLS models Source of Data Systematic reviews and data extraction are undertaken for the relation between glucose and periodontal disease. Another dataset from literature on the relation between alcohol and vascular disease data are also used to illustrate the application of the statistical method. Results 1. Generalized least square method and restricted cubic splines model are applied to analyze the relation between glucose and periodontal disease: the increase in glucose level could lead to the increasing of the OR of the periodontal disease. 2. Fixed effects and random effects generalized least square method implemented in R yields the same results obtained by using the software package Stata or SAS. 3. The result obtained by fixing the correlations between log(OR) of the exposure level at 0.5, is very similar to those obtained by the actual number of the cases and controls. Conclusions The generalized least squares method proposed by Greenland and Longnecker requires the input of the correlations of the log(OR) between the exposure level rank when pooling the results, and the restricted cubic splines model proposed by Orsini can efficiently estimate the nonlinear relationship between the dose and response. Researchers can gain a greater sample size using our approach.