An Investigation of Software Metrics Affect on Cobol Program Reliability

The purpose of this research was to predict a COBOL program's reliability from software characteristics that are found in the program's source code. The first step was to select factors based on the human information processing model that are associated with changes in computer program reliability. Then these factors (software metrics) were quantitatively studied to determine which factors affect COBOL program reliability. Then a statistical model was developed that predicts COBOL program reliability. Reliability was selected because the reliability of computer programs can be used by systems professionals and auditors to make decisions. Using the Human Information Processing Model to study the act of creating a computer program, several hypotheses were derived about program characteristics and reliability. These hypotheses were categorized as size, structure, and temporal hypotheses. These characteristics were then used to test several prediction models for the reliability of COBOL programs. Program characteristics were measured by a program called METRICS. METRICS was written by the author using the Pascal programming language. It accepts COBOL programs as input and produces as output seventeen measures of complexity. Actual programs and related data were then gathered from a large insurance company over the course of one year. The data were used to test the hypotheses and to find a model for predicting the reliability of COBOL programs. The operational definition for reliability was the probability of a program executing without abending. The size of a program, its cyclomatic complexity, and the number of times a program has been executed were used to predict reliability. A regression model was developed that predicted the reliability of a COBOL program from a program's characteristics. The model had a prediction error of 9.3%, a R2 of 15%, and an adjusted R2 of 13%. The most important thing learned from the research is that increasing the size of a program's modules, not the total size of a program, is associated with decreased reliability.