基于倾向性评分估计因果效应的方法研究及其在上市后药品不良反应评价中的应用
发布时间：2018-08-07 20:58
【摘要】：一、研究背景和研究目的 药品安全是一项关乎民生的基础工程，对药品安全性问题的研究一直是公共卫生研究的重点工作。虽然药品上市前已经经历了大量的动物实验和临床试验，但是由于动物实验的结果不足以预测人类应用药物的安全性；而临床试验，由于试验时间短、样本含量小、并且有严格的入选标准，其药物应用的条件与实际临床实践有一定的差异。因此，开展药品上市后不良反应监测工作尤显重要。在所有的药物警戒工具中，自发呈报系统提供的数据最多，所花的代价也越小，是目前世界上药品不良反应监测的最主要的手段。由于缺乏对整个人群中的事件背景发生率、暴露于研究药品的病例数及漏报率等的了解，无法计算出可靠的药品-不良事件期望发生数，目前常用的解决方案是进行数据挖掘。 然而这些数据挖掘方法主要侧重药品与不良事件之间的关联，没有尝试从因果方面解释药品与不良事件的关系。另一方面，没有考虑混杂因素如年龄、性别、病种、体重、合并用药等对检测结果影响，得到的检测结果存在一定程度的不准确性，包括由混杂因素导致的假阳性结果，以及遗漏了被混杂因素所掩盖的真实信息。由于自发呈报系统自身数据的特点，目前常用的控制混杂因素的方法，如分层分析法及Logistic回归模型并不能很好地控制混杂因素。 因此，本研究期望引入一种新的方法，对自发呈报系统中药品不良反应进行评价，既考虑到混杂因素对药品和不良事件的影响，又从因果概念上来分析药品和不良事件的关系，以期在发现信号后，进行专家评价、药物流行病学调查或专题研究确证因果关系前，对发现的信号进行进一步的确证，为药品风险管理、评价及决策提供依据。 二、研究方法 将Rubin因果模型框架引入自发呈报系统数据分析中，根据自发呈报系统中数据特点，构建Rubin因果模型，明确分析总体及潜在结果的定义。综合回顾利用倾向性评分估计因果效应的各种方法的理论，并侧重介绍当结果变量为二分类变量时各种方法的性质。 利用蒙特卡罗模拟方法，设置符合自发呈报系统数据特点的参数，考察利用分层、加权和匹配三大类倾向性评分估计因果效应方法的性质。分别构建正确与错误的倾向性评分模型，设置不同强度的协变量与分组变量之间的关系以及设置不同强度的协变量与结果变量之间的关系，计算各种设置下因果效应估计值的偏倚率、标准误和误差均方，以考察各种方法在不同情况下估计值的准确度和效率。另外，设置两种样本量相对较小的情况，模拟比较利用贝叶斯倾向性评分与传统倾向性评分估计因果效应值的估计值、标准误及置信区间。 对美国FDA自发呈报系统（FAERS）2011年及2012年两年的数据进行规范整理，利用常规数据挖掘方法，结合报告数和数据挖掘方法检测发现“可疑”的组合，确定目标研究药品。根据目标研究药品的适应症定义研究总体，选择所有可能服用该药品的人群，利用几种倾向性评分方法进行因果效应估计，最终得到因果效应估计值，以考察这些方法在实际数据中的可应用性。 三、结果 建立了自发呈报系统数据中的不良反应分析的因果模型，以所有可能服用目标药品的人群为研究总体；定义潜在结果为服用目标药品和不服用目标药品发生目标不良事件的可能性；定义系统中可能与是否服用目标药品有关及可能与是否发生不良事件有关的变量为协变量；定义潜在结果差值的平均值为因果效应。 模拟结果显示，目前使用比较广泛的基于对层内均数差值进行估算的倾向性评分分层法估计因果效应值时会造成偏倚。并且当样本量较大时，偏倚会随之增大。另外一种分层法利用层内回归估计代替了直接计算处理差值，这一方法显著地减少了估计偏倚，并且对倾向性评分模型多纳入变量不敏感，与其他方法相比，估计的效率也比较高。但是当结果变量与协变量呈非线性关系时，估计的方差将很难求得。利用倾向性评分加权法估计因果效应通常能得到一个无偏估计，模拟结果也显示，使用最广泛的利用固定差值计算因果效应的估计效率比较低，没有充分利用到样本的信息。另外，由于双稳健法的特殊性质，可以使在回归模型和倾向性评分模型任一构建正确的情况下得到一个无偏的估计。在本研究模拟设置条件下，利用贝叶斯倾向倾向性评分估计二分类结果的因果效应值与利用传统倾向性评分估计因果效应值区别不大，在小样本设置下，固定分为5层进行分析的两种分层法估计偏倚较大，效率也较低。 分析2011年及2012年FAERS报告的数据显示，使用双膦酸盐使骨折的发生率比不使用双膦酸盐骨折的发生率高，各种方法的效应差估计值分为IPW1：0.1083(0.0028,0.2138)； IPW2：0.1086(0.0049,0.2123)； DR：0.1065(0.0028,0.2102)； S：0.0711(-0.0544,0.1966)； SR：0.1123(0.0068,0.2178)。 四、结论 在自发呈报系统数据中引入因果模型概念，可以使结果解释更为直观。基于倾向性评分估计因果效应的方法，适用于自发呈报系统中药品不良反应的评价，可以克服以往的方法考虑报告数不考虑“混杂”的缺陷，使得到的结果更为可信，并为传统单药信号数据挖掘及确证最终因果关系之间提供了一个分析的方法。实例分析显示，双膦酸盐对骨折的发生可能有因果关系，，提示我们有必要对这一组合进行深入研究，如Meta分析、大规模的药物流行病学调查、专题研究等。
[Abstract]:First, research background and research purpose
Drug safety is a basic project related to the livelihood of the people. Research on the safety of drugs has been the focus of public health research. Although a large number of animal experiments and clinical trials have been experienced before the drug listing, the results of animal experiments are not enough to predict the safety of human applications. The test time is short, the sample content is small, and there is a strict admission standard. There is a certain difference between the conditions of the drug application and the actual clinical practice. Therefore, it is very important to carry out the monitoring of adverse reactions after the drug listing. At present, the most important means of adverse drug reaction monitoring in the world. Due to the lack of understanding of the incidence of events in the whole population, the number of cases and the rate of missing reports, the number of reliable drug adverse event expectations can not be calculated, and the current solution is to carry out data mining.
However, these data mining methods mainly focus on the association between drugs and adverse events, and there is no attempt to explain the relationship between drugs and adverse events from the causal side. On the other hand, there is no consideration of the effects of confounding factors such as age, sex, disease species, weight, and combination of drugs on the results of the test, and the results obtained are inaccurate to a certain extent. Sex, including false positive results caused by confounding factors, and the omission of real information concealed by confounding factors. Due to the spontaneous reporting of the system's own data, the commonly used methods of controlling confounding, such as stratification analysis and Logistic regression model, do not control confounding factors well.
Therefore, this study expects to introduce a new method to evaluate the adverse drug reactions in the spontaneous reporting system, taking into account the effects of confounding factors on drugs and adverse events, and from the causal concept of the relationship between drugs and adverse events, with a view to conducting expert evaluation, epidemiological investigation, or special topics after the discovery of the signal. Before confirming the causal relationship, further confirmation of the detected signals was carried out to provide evidence for drug risk management, evaluation and decision-making.
Two, research methods
The Rubin causality model framework is introduced into the data analysis of the spontaneous reporting system. According to the characteristics of the data in the spontaneous reporting system, the Rubin causality model is constructed, and the definition of the overall and potential results is clearly analyzed. The theory of various methods of estimating the effect effect by the tendency score is reviewed, and the two classification variables are introduced when the result variable is the result variable. The nature of various methods.
Using the Monte Carlo simulation method, we set up the parameters that conform to the characteristics of the data of the spontaneous reporting system, and examine the properties of the method of estimating causality effect by three major categories of layering, weighting and matching, and constructing the correct and wrong tendency grading model respectively, setting up the relationship between the covariate and the grouping variables of different intensity and setting up the relationship and setting up. The relationship between the covariate and the result variable of different intensities is used to calculate the bias rate of the estimated value of the causality effect under various settings, the standard error and the error mean square, in order to investigate the accuracy and efficiency of the estimated values of various methods under different circumstances. In addition, a relatively small sample of two samples is set up and the Bayesian tendency score is used for simulation and comparison. Estimation of the causal effect value, standard error and confidence interval with traditional tendentious score.
The data of the American FDA voluntary reporting system (FAERS) in 2011 and 2012 were standardized. The combination of the conventional data mining method, the number of reports and the data mining methods were used to detect the combination of "suspicious" and determine the target medicine. In order to investigate the applicability of these methods in the actual data, the population of the products is estimated by several tendencies.
Three, the result
A causal model for the analysis of adverse reactions in the spontaneous reporting system was established, and the population of all those who were likely to take the target drug was studied as a whole; the potential results were defined as the possibility of taking the target drugs and not taking the target drugs; the definition system could be related to the use of the target drugs and the possibility of taking the target drugs. The variables associated with adverse events are covariates, and the average value of the difference between potential outcomes is causal effect.
The simulation results show that biased rating stratification based on a relatively wide range of intra - level difference values estimates the bias when estimating the cause and effect value. And when the sample size is large, the bias will increase. Another method of stratification is used to replace the direct calculation of the difference. This method is significant. The estimation bias is reduced, and the tendency score model is insensitive to many variables. Compared with other methods, the estimation efficiency is also higher. However, when the result variable is nonlinear with the covariate, the estimated variance will be difficult. The simulation results also show that the estimation efficiency of using the most extensive use of fixed difference to calculate the causal effect is low and does not make full use of the information of the sample. In addition, due to the special properties of the dual robust method, an unbiased estimate can be obtained in the correct case of any construction of the regression model and the tendency scoring model. Under the condition of simulation setting, the causality effect value of the two classification results is estimated to be different from the traditional tendency score. Under the setting of small sample, the two stratification methods, which are divided into 5 layers, are more biased and less efficient.
Data from the 2011 and 2012 FAERS reports showed that bisphosphonates were used to make the incidence of fractures higher than those without bisphosphonates, and the estimated effects of various methods were divided into IPW1:0.1083 (0.0028,0.2138); IPW2:0.1086 (0.0049,0.2123); DR:0.1065 (0.0028,0.2102); S:0.0711 (-0.0544,0.1966); S. R:0.1123 (0.0068,0.2178).
Four. Conclusion
The introduction of the concept of causality model in the spontaneous reporting system can make the result interpretation more intuitive. The method of estimating the effect effect based on the tendency score is applicable to the evaluation of adverse drug reactions in the spontaneous reporting system, and can overcome the previous method of considering the number of reports without considering the defects of "confounding" and make the results more credible. The analysis shows that bisphosphonates may have causality in the occurrence of fractures, suggesting that it is necessary for us to study this combination, such as Meta analysis, large-scale pharmaco epidemiological investigations, and thematic studies.
【学位授予单位】：第二军医大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：R95
本文编号：2171324
[Abstract]:First, research background and research purpose
Drug safety is a basic project related to the livelihood of the people. Research on the safety of drugs has been the focus of public health research. Although a large number of animal experiments and clinical trials have been experienced before the drug listing, the results of animal experiments are not enough to predict the safety of human applications. The test time is short, the sample content is small, and there is a strict admission standard. There is a certain difference between the conditions of the drug application and the actual clinical practice. Therefore, it is very important to carry out the monitoring of adverse reactions after the drug listing. At present, the most important means of adverse drug reaction monitoring in the world. Due to the lack of understanding of the incidence of events in the whole population, the number of cases and the rate of missing reports, the number of reliable drug adverse event expectations can not be calculated, and the current solution is to carry out data mining.
However, these data mining methods mainly focus on the association between drugs and adverse events, and there is no attempt to explain the relationship between drugs and adverse events from the causal side. On the other hand, there is no consideration of the effects of confounding factors such as age, sex, disease species, weight, and combination of drugs on the results of the test, and the results obtained are inaccurate to a certain extent. Sex, including false positive results caused by confounding factors, and the omission of real information concealed by confounding factors. Due to the spontaneous reporting of the system's own data, the commonly used methods of controlling confounding, such as stratification analysis and Logistic regression model, do not control confounding factors well.
Therefore, this study expects to introduce a new method to evaluate the adverse drug reactions in the spontaneous reporting system, taking into account the effects of confounding factors on drugs and adverse events, and from the causal concept of the relationship between drugs and adverse events, with a view to conducting expert evaluation, epidemiological investigation, or special topics after the discovery of the signal. Before confirming the causal relationship, further confirmation of the detected signals was carried out to provide evidence for drug risk management, evaluation and decision-making.
Two, research methods
The Rubin causality model framework is introduced into the data analysis of the spontaneous reporting system. According to the characteristics of the data in the spontaneous reporting system, the Rubin causality model is constructed, and the definition of the overall and potential results is clearly analyzed. The theory of various methods of estimating the effect effect by the tendency score is reviewed, and the two classification variables are introduced when the result variable is the result variable. The nature of various methods.
Using the Monte Carlo simulation method, we set up the parameters that conform to the characteristics of the data of the spontaneous reporting system, and examine the properties of the method of estimating causality effect by three major categories of layering, weighting and matching, and constructing the correct and wrong tendency grading model respectively, setting up the relationship between the covariate and the grouping variables of different intensity and setting up the relationship and setting up. The relationship between the covariate and the result variable of different intensities is used to calculate the bias rate of the estimated value of the causality effect under various settings, the standard error and the error mean square, in order to investigate the accuracy and efficiency of the estimated values of various methods under different circumstances. In addition, a relatively small sample of two samples is set up and the Bayesian tendency score is used for simulation and comparison. Estimation of the causal effect value, standard error and confidence interval with traditional tendentious score.
The data of the American FDA voluntary reporting system (FAERS) in 2011 and 2012 were standardized. The combination of the conventional data mining method, the number of reports and the data mining methods were used to detect the combination of "suspicious" and determine the target medicine. In order to investigate the applicability of these methods in the actual data, the population of the products is estimated by several tendencies.
Three, the result
A causal model for the analysis of adverse reactions in the spontaneous reporting system was established, and the population of all those who were likely to take the target drug was studied as a whole; the potential results were defined as the possibility of taking the target drugs and not taking the target drugs; the definition system could be related to the use of the target drugs and the possibility of taking the target drugs. The variables associated with adverse events are covariates, and the average value of the difference between potential outcomes is causal effect.
The simulation results show that biased rating stratification based on a relatively wide range of intra - level difference values estimates the bias when estimating the cause and effect value. And when the sample size is large, the bias will increase. Another method of stratification is used to replace the direct calculation of the difference. This method is significant. The estimation bias is reduced, and the tendency score model is insensitive to many variables. Compared with other methods, the estimation efficiency is also higher. However, when the result variable is nonlinear with the covariate, the estimated variance will be difficult. The simulation results also show that the estimation efficiency of using the most extensive use of fixed difference to calculate the causal effect is low and does not make full use of the information of the sample. In addition, due to the special properties of the dual robust method, an unbiased estimate can be obtained in the correct case of any construction of the regression model and the tendency scoring model. Under the condition of simulation setting, the causality effect value of the two classification results is estimated to be different from the traditional tendency score. Under the setting of small sample, the two stratification methods, which are divided into 5 layers, are more biased and less efficient.
Data from the 2011 and 2012 FAERS reports showed that bisphosphonates were used to make the incidence of fractures higher than those without bisphosphonates, and the estimated effects of various methods were divided into IPW1:0.1083 (0.0028,0.2138); IPW2:0.1086 (0.0049,0.2123); DR:0.1065 (0.0028,0.2102); S:0.0711 (-0.0544,0.1966); S. R:0.1123 (0.0068,0.2178).
Four. Conclusion
The introduction of the concept of causality model in the spontaneous reporting system can make the result interpretation more intuitive. The method of estimating the effect effect based on the tendency score is applicable to the evaluation of adverse drug reactions in the spontaneous reporting system, and can overcome the previous method of considering the number of reports without considering the defects of "confounding" and make the results more credible. The analysis shows that bisphosphonates may have causality in the occurrence of fractures, suggesting that it is necessary for us to study this combination, such as Meta analysis, large-scale pharmaco epidemiological investigations, and thematic studies.
【学位授予单位】：第二军医大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：R95
【参考文献】
相关期刊论文 前1条
1 周晓枫;刘青;蔡兵;Andrew Bate;;全球上市后药品主动监测系统概况[J];药物流行病学杂志;2012年07期
本文编号：2171324
本文链接：https://www.wllwen.com/yixuelunwen/yiyaoxuelunwen/2171324.html
最近更新
教材专著
