臺灣博碩士論文加值系統

中文摘要
在醫學及公共衛生領域中，並非所有研究皆可使用隨機分派的試驗，所以在觀察性研究中控制干擾因子的影響成了重要的議題。處理可觀測干擾因子常見的方法大致分為兩種，一為在資料分析的階段，將干擾因子放入統計模型中進行校正，二為在研究設計的階段降低部分干擾因子的影響，其中配對(Matching)為常見的方法，配對後仍經常需要搭配統計模式校正其他干擾因子。當有無法觀測到的潛在重要干擾因子(unmeasured confounder)存在時，若有外在輔助的資料庫將可協助校正此未觀測干擾因子的影響。但當無法得到這個額外且適當的資料庫時，在流行病學上可以使用敏感度分析去量化干擾因子對結果影響的程度。
本研究將以實際資料探討孕婦在懷孕時期服用某類常用止咳藥物(以下簡稱M藥)對孕婦及胎兒的不良結果影響為例進行敏感度分析。受限於資料，本研究以孕婦生產時是否有因已知或疑似胎兒異常影響對母親之處置為主要的不良結果事件，以下簡稱孕婦及胎兒的不良結果。統計分析策略以資料配對搭配迴歸模式校正干擾因子，並以Rubin 多重插補法(Multiple imputation)的概念來處理研究中未觀測干擾因子，模擬不同的干擾程度進行敏感度分析，探討未觀測干擾因子對模式參數估計結果的影響程度。
研究分兩部分，第一部分利用了傳統配對或傾向分數配對兩種方法進行事前配對，再搭配傳統邏輯斯迴歸或條件式邏輯斯迴歸兩種統計模型來校正可觀測干擾因子進行分析，並比較其結果差異。第二部分使用多重插補法的概念，利用輔助的迴歸模型，隨機創造一個無法觀測到的虛擬二元變項，將這無法觀測的變項視為全部皆為缺失值，根據預設各種不同的干擾嚴重程度模式，進行插補。本研究在每一種嚴重程度皆各插補十組缺失變項，並以多重插補法加權平均校正此未觀察干擾因子後的估計值，以敏感度分析來探討未觀測干擾因子對結果的影響程度，檢查結果的穩健性。
研究結果顯示，在不考慮未觀測干擾因子時，無論資料是否經過配對，搭配傳統邏輯斯迴歸或條件式邏輯斯迴歸兩種統計模式，服用M藥與孕婦及胎兒不良結果的相關均未達統計顯著意義。若進一步對年齡進行分層，結果也相似。表示孕婦在懷孕時期服用M藥對孕婦及胎兒不良結果並沒有統計上顯著相關。
敏感度分析結果顯示，在未觀測干擾因子存在時，當未觀測干擾因子與懷孕期間是否服用M藥為負相關，且與孕婦及胎兒不良結果為正相關時，未觀測干擾因子對結果的影響最大。在未配對資料中不分層的情況下，未觀察到的干擾因子須分別與服用M藥和孕婦及胎兒不良結果的關係都達到均2倍(e^0.7)相關強度才有機會推翻結論，也就是從統計上不顯著變為顯著。當年齡分層為16~25歲，則須分別與服用M藥和孕婦及胎兒不良結果的關係都達到7.4倍(e^2=7.4)相關強度才有機會推翻結論；年齡分層為26~35歲，則須達到 5倍(e^1.6=5)相關強度才有機會推翻結論；年齡分層為36~55歲，則為所有組合均無法推翻結論。而在傾向分數配對中，不論有無年齡分層，在所有不同干擾程度的組合中均無法改變結論。
綜合以上結果，在校正所有可觀察到的干擾因子後，孕婦在懷孕期間服用止嗽藥M藥與孕婦及胎兒不良結果無統計上顯著關係。敏感度分析顯示，若資料未經配對，則可能因年齡不同，分別在相關強度達2倍、5倍及7.4倍時推翻結論。而當資料經過傾向分數配對後，除非存在著一個非常強的未觀察干擾因子，亦即此因子與M藥及與不良結果之間相關都達OR為7.4倍以上，才可能改變本次分析結果。然而現實生活中，即使是傳統上對孕婦不良影響最重要的干擾因子例如年齡或糖尿病，與M藥與不良結果的關係也未達此強度，因此即使未考慮未觀察干擾因子的存在，分析結果仍是穩健的。

Abstract
In medical and public health research, randomized controlled trials (RCT) are often infeasible. Controlling the confounder becomes an important issue in observational studies. Two commonly used approaches of managing confounders in observational studies are matching at design and statistical adjustment at analyses. These methods are only valid when all confounders are observed. When unmeasured confounders exist, one may correct the effects of the unmeasured confounder through an external auxiliary dataset that quantified the relationship between the unmeasured confounders and the outcomes. However, when the external validation data are not available, researches can perform a sensitivity analysis to quantify the effect of the main predictor on outcome after taking into account the unmeasured confounders.
The objectives of this study are to conduct a sensitivity analysis using an numerical example which explored the effect of using an antitussive medication (Drug “M”) during pregnancy on the potential adverse outcome related to fetus among women. Limited by the data, the adverse outcome in our research is only restricted to whether the medical treatment causing by known or suspected fetal abnormality as a surrogate adverse event.
The study is divided into two parts. In the first part, we analyzed data using several matching and statistical models to adjust the observed confounders. We considered two matching methods: the traditional matching method and propensity score matching method (PS). Then a traditional logistical regression model or a conditional logistic regression model were applied. In the second part, we conduct a sensitivity analysis for adjusting an unmeasured confounder, extending the concept of Rubin’s multiple imputation in dealing with missing values. . We simulated and imputed a binary unobserved covariate based on an auxiliary regression model according to different degrees of severity in the unmeasured confounders. We imputed 10 times for each scenario and obtain the estimate of the unmeasured confounder by multiple imputation weighted average. We used this imputation-based sensitivity analysis to assess the impact of results due to potential unmeasured confounder, and examined the robustness of our conclusions.
Without considering the unmeasured confounder, our results showed that whether the data were matched or not using conventional logistic regression or conditional logistic regression, the association between taking Drug M during pregnancy and the adverse reaction were not statistically significant. Stratifying the age further, we got the similar results. It implied that taking Drug M during pregnancy does not significantly affect the fetal abnormality based on the given data.
When the unmeasured confounder was considered, the sensitivity analysis
revealed that if the unmeasured confounder was negatively associated with taking Drug M and positively associated with fetal abnormality, the effect of unmeasured confounder would be the most obvious. In the unmatched data without stratification, the unmeasured confounder needed 2 times associated with taking Drug M and the fetal abnormality respectively to overturn the conclusion. When age was ranged in 16 to 25 or 26 to 35, the unmeasured confounder needed 7.4 and 5 times to overturn the conclusion, respectively. However, when the age ranged in 36 to 55, adjusting for the unmeasured confounder could not change the conclusion. In PS-matched data, the conclusion remained the same in all combinations of different levels of the severity in unmeasured confounder.
In summary, after adjusting for the observable confounders, the association between taking Drug M during pregnancy and the adverse outcomes were not statistically significant. Sensitivity analysis revealed that results might be overturned (becomes significant) in some age stratifications if the unmeasured confounders have a 2, 5, or 7.4 times of correlations with Drug M and/or outcome respectively when the data were unmatched. When the data were matched by PS, results might be changed if there was a strong unmeasured confounder, over 7.4 times associated with Drug M and/or adverse reactions. In real world, however, even important confounders such like age or diabetes mellitus do not have such a strong association with Drug M or with adverse reactions. Out study suggests that if data are matched ahead and adjusted for potential measured confounders by regression models, the results were still robust even if the unmeasured confounder was not taken into consideration.

目錄
中文摘要 i
Abstract iii
目錄 vi
表目錄 vii
第一章 緒論 1
第一節 研究背景與動機 1
第二節 研究目的 2
第二章 研究方法 4
第一節 配對方法 4
第二節 統計模式 6
第三節 多重插補之敏感度分析 8
第三章 實際案例與資料分析：不考慮未觀測干擾因子 12
第一節 資料來源 12
第二節 研究變項 12
第三節 資料處理 13
第四節 分析結果 22
第四章 實際案例進行敏感度分析：考慮未觀測干擾因子 25
第一節 參數設定 25
第二節 分析結果 25
第五章 結論與討論 37
參考文獻 40

表目錄
表1 各種包含 Dextromethorphan 之藥品代碼及內容 16
表 2 原始資料與配對後資料M藥與ICD9的次數分配表 17
表 3 原始資料與配對後資料中其他變項之敘述統計 18
表 4 原始資料與配對後資料中其他變項之敘述統計 16~25歲 19
表 5 原始資料與配對後資料中其他變項之敘述統計 26~35歲 20
表 6 原始資料與配對後資料中其他變項之敘述統計 36~55歲 21
表 7 傳統邏輯斯迴歸模型Medicon參數估計:未經配對資料 22
表 8 邏輯斯迴歸模型Medicon參數估計:傳統配對方法 23
表 9 邏輯斯迴歸模型Medicon參數估計:傾向分數配對方法 23
表 10 未經配對資料使用不同干擾程度之參數估計結果 28
表 11 以四種組合了解未觀測干擾因子之趨勢 29
表 12 未經配對資料使用不同干擾程度之參數估計結果 16~25歲 30
表 13 未經配對資料使用不同干擾程度之參數估計結果 26~35歲 31
表 14 未經配對資料使用不同干擾程度之參數估計結果 36~55歲 32
表 15 傾向分數配對資料使用不同干擾程度之參數估計結果 33
表 16 傾向分數配對資料使用不同干擾程度之參數估計結果 16~25歲 34
表 17 傾向分數配對資料使用不同干擾程度之參數估計結果 26~35歲 35
表 18 傾向分數配對資料使用不同干擾程度之參數估計結果 36~55歲 36

參考文獻
1. Rubin, D.B., Matching to remove bias in observational studies. Biometrics, 1973: p. 159-183.
2. Lin, H.-W. and Y.-H. Chen, Adjustment for missing confounders in studies based on observational databases: 2-stage calibration combining propensity scores from primary and validation data. American journal of epidemiology, 2014. 180(3): p. 308-317.
3. Rosenbaum, P.R. and D.B. Rubin, The central role of the propensity score in observational studies for causal effects. Biometrika, 1983: p. 41-55.
4. Hsieh, P.-Y., Matching Methods and Analysis Strategies in Time to Event Models. 2016.
5. Cox, D. and E. Snell, The analysis of binary data. London. UK, Methuen and Co, 1970.
6. Hosmer Jr, D.W., S. Lemeshow, and R.X. Sturdivant, Applied logistic regression. Vol. 398. 2013: John Wiley & Sons.
7. Breslow, N., Covariance adjustment of relative-risk estimates in matched studies. Biometrics, 1982: p. 661-672.
8. Rubin, D.B., Multiple Imputation for Nonresponse in Surveys (Wiley Series in Probability and Statistics). 1987.
9. Schafer, J.L., Multiple imputation: a primer. Statistical methods in medical research, 1999. 8(1): p. 3-15.
10. Shi, J., et al., Unmeasured Confounding in Observational Studies with Multiple Treatment Arms. Epidemiology, 2016. 27(5): p. 624-632.
11. Ichino, A., F. Mealli, and T. Nannicini, From temporary help jobs to permanent employment: What can we learn from matching estimators and their sensitivity? Journal of Applied Econometrics, 2008. 23(3): p. 305-327.
12. Lu, B., et al., Estimating the effect of premarital cohabitation on timing of marital disruption: using propensity score matching in event history analysis. Sociological Methods & Research, 2012. 41(3): p. 440-466.
13. Waller, J. L. Where’s The Match? Matching Cases and Controls after Data Collection