Abstract
Statistical tests are valuable tools that allow researchers to analyze data and evaluate hypotheses objectively. They are used to investigate correlations, identify disparities, and reduce data dimensionality (Casula et al., 2021). This paper analyzes three studies that used statistical testing. The researchers revealed where statistical tests are used. This will support professional activities, as the knowledge gained will contribute to the discipline of organizational behavior.
Introduction
Statistical tests are powerful tools that allow researchers to assess data and test hypotheses objectively. They are used to analyze relationships, infer differences, and reduce data dimensionality.
Objectives
The paper “Radiographic Severity Index in COVID-19 Pneumonia: Relationship to Age and Sex in 783 Italian Patients” used statistical techniques. To improve risk categorization for patients infected with severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), the diagnostic imaging department deployed a prototype chest X-ray (CXR) device to evaluate lung abnormalities. The goal of this study was to examine the correlation between CXR findings and age or gender in SARS-CoV-2-infected Italian patients.
The paper “Machine learning-based intrusion detection systems for IoT applications” is an example of a study that uses statistical testing to find discrepancies. The authors of this study intend to investigate the viability of using machine learning classification algorithms to protect the Internet of Things against DoS attacks (Verma and Ranga, 2020). A thorough examination of classifiers for creating anomaly-based intrusion detection systems (IDSs) is conducted.
Statistically significant differences between classifiers are analyzed using the Friedman and Nemenyi criteria. The authors also discuss how to select the best classifier based on application requirements (Verma and Ranga, 2020). Thus, the key aims of this study are to encourage Internet of Things security researchers to develop IDSs using ensemble learning and to provide adequate approaches for statistical evaluation of classifier performance.
The paper “Benchmark for filter methods for feature selection in high-dimensional classification data” is the third study that uses statistical tests to reduce data dimensionality. The analysis’s purpose is to examine how various filtering algorithms work, assess their usefulness in terms of execution time and prediction accuracy, and make application recommendations (Bommert et al., 2020). The authors use multivariate classification datasets to assess the filtering approach’s runtime and accuracy when combined with the classification technique.
Hypotheses
The first study hypothesizes that using an experimental chest X-ray (CXR) scoring system to quantify the severity and progression of pulmonary impairment in COVID-19-induced pneumonia improves risk stratification and enables clinicians to provide “individualized” care to high-risk patients. This CXR grading approach (also known as the Brixia scale) was developed exclusively for the semi-quantitative assessment of COVID-19 lung illness (Borghesi et al., 2020).
In the second study, the authors argue that the features of the Internet of Things make it simpler to use in real-world applications while also exposing it to security dangers. One of the most damaging attacks against the Internet of Things is denial-of-service (DoS). The null hypothesis (H0) states that there is no difference in classifier performance (Verma & Ranga, 2020). The alternative hypothesis (H1) asserts that at least one classifier performs significantly differently from at least one other classifier.
According to the authors of the third work, feature selection is one of the most fundamental issues in machine learning, and it is becoming increasingly popular as multidimensional datasets from other disciplines, including bioinformatics, become accessible. They also believe that filtering techniques play an essential role in feature selection (Bommert et al. 2020). The authors suggest that they can be used with any machine learning model to reduce the execution time of machine learning algorithms significantly.
Statistical Analysis
In the first study, all CXR reports using the new scoring technique were identified through a retrospective search of the department’s Radiology Information System (RIS)/Picture Archiving and Communication System (PACS). This study exclusively covered Italian patients hospitalized at the clinic with SARS-CoV-2 (Borghesi et al., 2020). Patients under 20 years of age were not eligible to participate in the experiment.
The study considered each patient’s age, gender, and the highest score on the CXR report. A radiologist (AB with 15 years of experience in thoracic imaging) examined the X-ray images associated with these CXR findings (Borghesi et al., 2020). The patients were divided into seven age groups (Borghesi et al., 2020). Since the patients’ ages and CXR scores were not normally distributed, data are presented as numbers, medians, and interquartile ranges.
The Kruskal-Wallis test was also used to see whether there were any significant differences in CXR scores across age groups. The statistical analysis was carried out by using readily available commercial tools. Statistical significance was determined as p < 0.05 (Borghesi et al., 2020). This statistical analysis method is excellent for this study since statistics are vital in medical investigations (Di Leo and Sardanelli, 2020). This strategy yielded more accurate and relevant data.
Two statistical tests of significance are appropriately used to compare the classifiers in the second work. The Friedman and Nemenyi tests were chosen for this purpose. Significance tests are used to determine whether classifiers differ significantly from one another (Verma and Ranga, 2020). The Friedman test was chosen primarily because it is a practical statistical approach for comparing several groups. The Friedman test is used to determine whether at least one classifier outperforms the others across all datasets (Verma and Ranga, 2020).
If a suitable classifier is identified, a post hoc Newman-Keuls test for pairwise multiple comparisons is performed. I believe it is necessary to perform a post hoc test to detect changes in classifier performance. Friedman’s test verifies that there is a significant difference between the classifiers under consideration, but Nemenyi’s test specifies the source of this difference, thereby validating their use.
The third research employed univariate statistical analysis. Univariate statistical tests have only one dependent variable. This filter performs a variance analysis for each trait that explains the class variable. A score is calculated from the resulting statistic (Bommert et al., 2020). When the mean values of the related feature deviate between classes, the F statistic increases (Bommert et al., 2020). I believe the statistical analysis technique is appropriate for this study, as such tests are adequate for describing the operation and efficacy of specific processes.
Results
The initial study found 783 confirmed SARS-CoV-2 infections among Italian citizens. The median age was 65, with the interquartile range being 55-74 years (Borghesi et al., 2020). Each age group has its own set of CXR scores and patient gender information. CXR rates were significantly higher in men in groups D, E, F, and G compared to groups A, B, and C.
Female CXR frequencies were significantly higher in groups E, F, and G compared to A and B (Borghesi et al., 2020). The CXR rates were also significantly higher in Group G than in Groups C, D, and E. CXR rates were significantly higher in Group F than in Group C (Borghesi et al., 2020). To aid comprehension, all information is presented in figures and tables. Furthermore, the obtained data verified the theory.
According to the second study’s findings, RF outperforms other classifiers in terms of accuracy and specificity. In terms of sensitivity, GBM performed best (99.53%). XGB achieved the most excellent AUC of 98.76% (Verma & Ranga, 2020). MLP has the lowest accuracy (82.76%), whereas AB has the worst specificity (86.72%) and sensitivity (97.94%). CART has the lowest AUC (94.01%).
In terms of FPR, RF is the best classifier, while AB is the worst, with 8.89% and 13.26%, respectively. To facilitate comprehension and simplify the study, the findings are summarized and visually displayed in tables and figures. The acquired data were used to validate the study’s hypothesis.
The authors of the third research found that the performance statistics may be interpreted as unbiased estimates of the performance of new data generated during the data-generating operation that produced the dataset in issue. Some filters performed much better in the ina_agnostic dataset than others (Bommert et al., 2020). The JMIM, Permutation, Impurity, and Forest Importance filters perform well, whereas the DISR, MRMR, and variance filters do poorly. Overall, all filters outperform no filtering in terms of predictive ability. The data is presented in a tabular fashion to assist readers in understanding the information (Bommert et al., 2020). The results support the study’s premise.
Findings
According to the authors of the first study, males aged 50 and older, as well as women aged 80 and older, are most at risk of developing significant lung disease. These findings might help identify those at highest risk and those who require specific treatment choices (Borghesi et al., 2020). Based on the performance findings and statistical tests from the second study, the authors conclude that classification and regression trees, as well as the extreme gradient boosting classifier, offer the best balance between performance and response time (Verma & Ranga, 2020). As a result, they are both feasible solutions for creating IDSs based on IoT-specific anomalies. Based on this, the researchers want to develop IDSs to defend against routing attacks in IoT networks.
The authors of the third study found that no subset of filtering methods outperforms others across all datasets. Instead, the most effective filtering algorithms varied across datasets. On average, all filtering algorithms beat no filtering (Bommert et al., 2020). The authors interpret their findings as follows: if computational resources allow testing only one filtering strategy, they propose permutation (Bommert et al., 2020). This filtering method accurately fits classification models to the bulk of datasets (Bommert et al., 2020). If computing resources are available to identify a good filtering approach, we propose experimenting with permutation, impurity, and symmetry—uncert, limma, and JMIM.
Conclusion
Through my research, I have learned where statistical tests are used. This knowledge will support my professional work by contributing to my understanding and practice of organizational behavior.
References
Bommert, A., Sun, X., Bischl, B., Rahnenführer, J., & Lang, M. (2020). Benchmark for filter methods for feature selection in high-dimensional classification data. Computational Statistics & Data Analysis, 143.
Borghesi, A., Zigliani, A., Masciullo, R., Golemi, S., Maculotti, P., Farina, D., & Maroldi, R. (2020). Radiographic severity index in COVID-19 pneumonia: Relationship to age and sex in 783 Italian patients. La Radiologia Medica, 125, 461-464.
Casula, M., Rangarajan, N., & Shields, P. (2021). The potential of working hypotheses for deductive exploratory research. Quality & Quantity, 55(5), 1703-1725.
Di Leo, G., & Sardanelli, F. (2020). Statistical significance: Pvalue, 0.05 threshold, and applications to radiomics—Reasons for a conservative approach. European Radiology Experimental, 4(1), 1-8.
Verma, A., & Ranga, V. (2020). Machine learning based intrusion detection systems for IoT applications. Wireless Personal Communications, 111, 2287-2310.