<International circulation>: Dr Rumsfeld， confounding a large scale epidemiology study can be problematic.  How do you minimize confounding in large data set analyses?

　　《国际循环》：混淆大规模流行病学研究可能会产生问题。您如何最大限度地减少大型数据集的分析混淆？
　　Dr Rumsfeld  :  Well this can be a very long answer but to keep it somewhat shorter， first is very important， it is very important to have a carefully phrased clinical study aim or question with a hypothesis you are trying to test.  That is if you go into large scale databases and just start looking for statistical associations， you will find them and you will not know if they are true or due to confounding as the answer.  It is much better to approach these analyses with a pre-formed study aim and hypothesis.  Second of all we do use advanced statistical methods， certainly risk adjustment.  We do a lot of stratification to looking into sub groups of patients and more and more we do things such as propensity matching to try to balance the characteristics between the two groups you are comparing.  At the end of the day though， I think it is widely known that while these things do help， there could still be unmeasured confounding.  We can never be sure in an observational study of a casual relationship.  Ultimately observational studies are very good to look at patient safety problems and they are very good for hypothesis generation but as a rule if possible， you should prove those hypotheses in subsequent clinical trials.　

　　Rumsfeld博士：第一点是非常重要的，有一个精心设计的临床研究目的或试图检验假设性问题是非常重要的。也就是说，如果你进入大型数据库只是寻找数据关联，你会找到这些关联，但你无法知道他们是否真实还是由于答案混淆。最好带着预先设定的研究目的和假设进行这些分析。第二，我们确实使用先进的统计方法，当然包括风险调整。我们做了很多分层寻找到患者亚组做越来越多的倾向匹配，尽量平衡比较两组的特点。虽然这些东西确实能够帮助我们，但有可能仍是不可测的干扰。我们永远不能肯定是一个因果关系的观察研究。最终的观测研究要适合研究病人安全问题，适合假设一代，但结果是如果可能的话，你应该在以后的临床试验中证明那些假说。