Introduction

Alexander Pope wrote, "To err is human, to forgive Divine." Today we have scientific evidence that both of his ideas were true. First, to err truly is human. Everyone errs, and error rates are comparable across a broad range of human cognitive activities. We now have theories that show why error is so universal and constant. Basically, human cognitive processes are wired to make us fast and flexible, but speed and flexibility come at a price. The very way we think makes a small rate of error inevitable. This was not a problem when we were hunter-gatherers, but in activities such as software development where we must be precise in a large number of events, the very way we think makes such precision impossible.

Regarding the idea that to forgive is Divine, we too should be forgiving. Because human error is not avoidable, it should not be considered blameful. Fagan [1976] emphasized that blaming makes people defensive and harms error detection and correction. The correct attitude is that error is inevitable and that only a stance that rejects blamefulness is likely to be effective.

One page at this website is a minipaper on general human error theory. We would urge readers to study Baars [1992] and Reason [1990] for more detail on the inevitability of human error.

This website contains a number of tables organized in the following way:

Basic Error Rates. This table shows error rates for a wide variety of generally simple actions. The focus is on mechanical errors after any error correction done by the subject. In some cases, however, figures are given before subject corrections.

Errors in Multi-Step Tasks. Many tasks involve a series of smaller actions. A good example of this is commands typed at a command line. When there are many small tasks, error rates multiply. In addition, as the number of steps grows, the possibility of short term memory overload increases.

Spreadsheet Errors. This page links you to the Spreadsheet Research (SSR) website which contains such data.

For programming, we have especially rich error data thanks both to experiments and to code inspection studies that have reported the error rates they found. In programming, errors are called faults, and they are measured in terms of faults per thousand lines of (non-comment) source code, faults/KLOC. We have converted these figures into percentages for easier comparison with other data at the website.

Normal Development: This table shows error rates for normal traditional development. Data from thousands of code inspections are included.

Formal Development: New formal method try to prevent errors from happening in the first place. Their error rates are comparable to those in normal traditional development, after code inspection.

Late Inspection: This table shows error rates after individual modules have been tested and then combined into a full system. This table shows error rates found just as the system is about to go out the door.

Test-Retest: One study had students do procedures twice and studied how errors and correct performance on the first part were related to errors and correct performance in the second.

Error Detection. Finding errors can be very difficult. We only have 90% success rates in error detection in the simplest tasks.

Proofreading. There is a large literature on proofreading. Unfortunately, it does not focus on finding syntax errors, which are very difficult. Instead, it looks only at spelling errors. It differentiates between spelling errors that make another correct word (word errors) and spelling errors that lead to a text string that is not a real word (nonword errors). Word errors are much more difficult to find.

General Error Detection. This table looks at error detection in other contexts.

Experts versus Novices. Novices make more errors than experts, although not by as large a margin as most people think. In other tables, subjects often are highly experienced. Experts do make errors. They also tend to deny it.

Bibliography. Here is where to find all those citations listed in [square brackets].

Introduction

Copyright 1997-2008 Panko