似然比法在物证理化检验结果评估中的应用（英文）

An Introduction to Evaluation of Physicochemical Data for Forensic Purposes with Application of Likelihood Ratio Test

新型犯罪日趋复杂,案件审判人员对高标准技术工作的需求不断增加,这些都要求研究新的证据评价方法,从而对各种微量物证理化检测数据的证据价值进行评估。证据评估方法能够反映法庭科学家在案件审理中的作用。这意味着上述数据(证据)应当在案件诉讼中控辩双方提出了相互对立的假设H1和H2的情况下接受评估,贝叶斯模型适用于这种情况下的证据评估。本文描述了在比较和分类(其实分类也是以比较为基础)问题中使用似然比方法(LR)对被观察的理化数据进行评估的原理。LR模型允许在一次计算中将所有重要的因素都包括在内,以此实现对相关理化数据的评估。这些因素包括,被比较样本间被比对理化数据的相似性,被测理化数据在有关总体中的稀有性,以及可能的误差来源(样本之内、之间的差异性)等。作为统计工具,LR模型只能用于仅以几个变量描述的数据库,而事实上大多数理化数据都是高度多维的(比如光谱),因此,需要使用缩维手段比如图形模型或适当的化学计量工具作降维处理,本文对此举例说明。需要指出,LR模型只应作为一种支持性(非决定性!)的工具,其结果(论)要接受严格的分析判断。换言之,统计方法并不能传达绝对的真相,采用的分析技术会有各自的不确定度,各种可能的错误答案也是统计方法的构成部分。因此,应当进行灵敏度检验亦即对所用分析方法作处理验证,从而确定其表现优劣。基于此,本文采用经验交叉熵方法举例说明如何对LR模型作校验。关于来源水平的理化数据评估,这涉及到比较样品是否源自于同一物体即是否具同一性的问题。通常,办案人员(法官、检察官或警察)会对被发现的取自于身体、衣服或鞋子的微量物证(显示与对照样本类似)是否发生了转移并留存下来的活动感兴趣,这就是所谓的活动水平检验,本文对此已有讨论分析。

The increasing complexity from new forms of crime and the need by those who administer justice for higher standards of scientific work require the development of new approaches for measuring the evidential value of physicochemical data obtained by application of numerous analytical methods during the analysis of various kinds of trace evidence. The methods used for evaluation of these data should reveal the role of the forensic experts in the administration of justice. This means that such data（evidence） should be evaluated in the context of two competing propositions H1 and H2 formulated by two opposite sides in the legal proceeding, i.e. prosecution and defence. Bayesian models have been proposed for the evaluation of evidence in such contexts. This paper describes the principle of likelihood ratio（LR） approach for evaluation of physicochemical data in so-called comparison and classification（in fact, classification is also based on comparison） problems. The LR models allow including all of important factors in one calculation run where evidential value of physicochemical data is to evaluate. These factors are the similarity of observed physicochemical data in compared samples, the rarity of determined physicochemical data in relevant population, and the possible sources of errors（within- and inter-sample variability）. The LR models, as statistical tools, can be only proposed for databases described by a few variables. However, most of physicochemical data are highly dimensional data（e.g. spectra）. Therefore, it is necessary to apply methods of dimensionality reduction like graphical models or suitable chemometrics＇ tools, with examples presented in the paper. The LR models should be always treated as a supportive（not the decisive!） tool and their results subjected to critical analysis. In other words, the statistical methods do not deliver the absolute truth as the levels of possible false answers are an integral part of these methods, in the same way like uncertainty related to the applied analytical techniques. Therefore, sensitivity convergence, an equivalent of the validation process for analytical methods, should be conducted in order to determine their performance. Thus, how to validate LR models is addressed in this paper by the example of application of Empirical Cross Entropy approach. There is the so-called source-level evaluation for physicochemical data as it helps to answer the question whether the compared samples are originated from the same object. Usually, the fact finders（judge, prosecutor, or police） are interested in recognizing the activity that made transferred and persisted of the recovered microtraces（which reveal similarity to control sample） from body, clothes or shoes. This is the so-called activity-level analysis, also discussed in the paper.