Conclusions

Abstract

The basic hypothesis behind automated data analysis that text data can be used to detect psychological signals—is confirmed by the established statistical methods results presented in the chapter. Text data from discussion forums, interview transcripts and free text questionnaire sections can be subjected to automatic analysis to detect the psychological signals indicating the likelihood of certain syndromes, such as PTSD, in the object of the document, with the results accuracy approaching that of human experts. The chapter discusses the human-annotated data set used for the study, suggests several methods for evaluating the quality of the outcomes, presents the results obtained from validation data and from actual users of the environment, and outlines the directions for future work.

From the statistical analysis standpoint, this chapter presents results of PTSD and Control group difference analyses of variables of interest (e.g., memory, attention, processing speed). The simplest and informative standard characteristics of the variables’ (of their distributions) are the mean and standard deviation; the former measuring the central tendency while the latter the scattering (dispersion) around the mean. The properly normalized difference between the sample means, called the (Student) t-statistic, is a standard tool to decide if there is a real difference between the mean values of the variables under study in the two groups. This is done to distinguish from differences in the samples where the difference may be explained by random fluctuations of the variables. In many medical studies the following principle is used: if the value of the t-statistic calculated from the data is such that the probability of getting a larger value when there is no difference in the population means (this probability is called the p-value of the data) is less than 0.05, then the data are considered sufficiently strong evidence for the real difference. In other words, the probability of claiming that there is a real difference when actually there is none (the so called Type I error), is less than 0.05 (i.e., only less than five instances out of 100 will result in the wrong decision).