- 期刊
檢討科學數據統計分析之客觀性
Re-Examining the Objectivity of Statistical Analysis of Scientific Data
摘要
不管在自然科學或社會科學的研究中，欲由實驗數據獲致真正客觀的結論，長久以來一直採用標準統計法的「精確假設考驗」，此種統計考驗幾乎成為實徵研究的必要程序之一。假設考驗之基本邏輯如下：（一）將研究假設轉換為虛無假設，並確定所容許的第一類型誤差機率。（二）以統計考驗方法決定虛無假設為真之機率，並與事先設定的誤差機率比較。（三）如果所得到的機率小於事先設定的第一類型誤差機率，則所得結果即達「統計上之顯著水準」；在此情況下即可拒斥虛無假設，而使研究假設得到支持。標準統計分析法雖然相當合乎邏輯，但亦有不少缺點，本研究旨在以實例對其應用之結果加以質疑，從而介紹「貝氏統計分析法」，使用同一例子來分析比較。簡言之，欲使科學數據之統計分析，給吾人帶來真正有洞見的結論，在使用標準統計分析時，必須提防其虛偽的客觀主義，且以貝氏分析方法，考量主觀性在統計分析的角色與需求，方能使吾人更有彈性且更有效力地瞭解所分析的數據。
關鍵字
無資料並列摘要
Standard statistical analysis of testing hypothesis has been used to obtain authentic and valid conclusion from empirical data in both natural and social sciences. In spite of its crucial role in scientific investigations, it is not entirely excluded from potent limitations. The procedure of testing hypothesis is featured with the basic logic as follows: (a) Transforming any research hypothesis into a null hypothesis and assigning the desired probability P0 of Type I error;(b) Determining the probability P when null hypothesis is true by statistical testing, and comparing it with the assigned probability P(subscript 0);(c) If P ＜ P(subscript 0) , the result is concluded to be statistically significant, the null hypothesis is therefore rejected but the research hypothesis is supported instead. This paper tried to argue even standard approach of statistics depends upon the subjective intention of the investigator. Followed in the next section, Bayesian approach of statistical analysis was thus introduced. A hypothetical experiment was succeedingly used as an example to compare the standard statistical approach with the Bayesian approach. In conclusion, the adoption of standard approach in statistical analysis has helped improve the interpreting of experimental data. But common practice of standard approach can produce misleading inferences because of its exaggerated objective way of analyzing data. Acknowledging the role of subjectivity in the interpretation of scientific data could open the way for more accurate and flexible statistical judgments.