Based on the historical data of 5 past and similar projects, you have calculated these average numbers of defects detected in system testing: - for…

The analysis involves comparing two prediction models against the actual outcome.

1. Prediction based on LOC: The historical rate is 200 defects per 10,000 LOC. For the estimated 20,000 LOC, the expected number of defects is (20,000 / 10,000) 200 = 400 defects.

2. Prediction based on Effort: The historical rate is 49 defects per person-month. The project effort is 4 developers 4 months = 16 person-months. The expected number of defects is 16 49 = 784 defects.

The actual number of defects discovered was 797. This number is extremely close to the effort-based prediction (784) but significantly different from the LOC-based prediction (400). Therefore, the statement that is "certainly true" is that the number of defects is approximately proportional to the development effort for this project.

A. This is not certain. The higher defect count could be due to a larger code size than planned, not necessarily a higher defect density (poor quality).

C. This is a plausible inference to explain the high defect count, but it is not a certainty. The discrepancy could also be explained by a higher defect density.

D. This statement combines two unproven assumptions (more LOC and higher defect density). It is possible that only one of these factors, or neither, is true.

1. Kan, S. H. (2003). Metrics and Models in Software Quality Engineering (2nd ed.). Addison-Wesley Professional. Chapter 9, "In-Process Quality Metrics for Software Testing," discusses how defect density can be measured against both code size (e.g., KLOC) and effort. The scenario illustrates a case where effort is a more accurate predictor than the initial size estimate.

2. Fenton, N. E., & Neil, M. (1999). A critique of software defect prediction models. IEEE Transactions on Software Engineering, 25(5), 675-689. https://doi.org/10.1109/32.815326. This paper evaluates the reliability of different predictors for software defects, including size (LOC) and process-related metrics like effort. It supports the principle that the predictive power of different models must be empirically validated, as the question requires, showing one model (effort-based) can be more accurate than another (size-based) in a specific context.