[Replying to Louis directly, using both e-mail addresses supplied, and to the edstat list] Your purpose(s) is/are unclear, at least to me. One infers that you are interested in estimating something about the 33 categories, not about the individual bugs; at least, the data you want to analyze (if I understand your description) contains no information about individual bugs. Detailed comments embedded in the original post (OP), copied below:
On 31 Jul 2003, Louis T wrote in part: > I'm doing an internship on software testing and I have to classify > reported bugs. The bug list is important so I decided to make a > sampling. Did you sample from (a) individual reports; (b) a list of individual bugs reported (which might not be the same as (a)); (c) some time frame, e.g. randomly chosen half-hours from a time series of some sort; (d) something else (and if so, what?)? > I know that by doing a sampling I may introduce an error when I will > generalize my result to the whole population, ... Perhaps. Since your "whole population" must include not only all the reports so far obtained but also all the reports that will ensue in the future, AND all the reports that might have been made but weren't, because certain particular combinations of parameter values have not been observed, there is inevitably some "error" in your estimates. The mere fact of sampling is ordinarily unlikely to increase that error. (Although this may depend on what you want to mean by "error". What I had in mind was what we would call an "error variance", a measure of the uncertainty of the estimate being produced.) > But I want to be able to measure the trust I can have in my work. I > feel I have to use the "chi square" test. Is it true and are there > other tests ? I take this to mean that you want to estimate the precision of whatever measure you're trying to produce. How one might do this depends on the nature of the intended outcome. If I understand your description correctly, your data permit you to estimate at least: (a) average number of total bugs (which is meaningful only if the total is with respect to some range of observations: e.g., for 100 hours of software use, or for 1000 instances of using the software); (b) average rate of encountering bugs (e.g., number of errors per 100 hours of software use, ...); (c) number of categories of bugs encountered; (d) average rate at which categories are encountered; (e) whether numbers (or rates) of bugs per category are about the same for all categories (or are consistent with some other distribution of numbers -- "equal values per category" may not be an interesting hypothesis to test). > The detail of my classification will depend on the amount of work I > have to make as my time is limited (no big news in this !). Are the classifications already known from previous work in the field? Or do you have to invent them as you go? > In my case, I have ten categories with a sub-division in 33 > categories (not for each category but for all of them). The number > of reported bugs should be around 3000. (I could get a more precise > number on this, but haven't searched yet). What should be the > correct size of my sample if I use the ten categories? What are you sampling from? A list of incidents of invoking the software (so some observations will contain zero bugs, some will contain 1, some will contain more....)? A log of time spent using the software, in which bugs are observed to occur from time to time? A list of the individual bugs encountered, with each bug classified into one of the categories of interest? A list of the numbers of bugs encountered in each category (seems unlikely, but one never knows...)? Etc. Amd (as remarked above) what question(s) do you want to ask? There's no such thing as a "correct size of sample", but a useful size of sample will depend on (a) what quantity(ies) you choose to estimate, (b) what degree of uncertainty you can tolerate in the estimate(s), (c) what population you are sampling from (times, incidents, etc.: see above). > And if I use the 32 subcategories? For openers, if you found a desired sample size per category for the ten categories, the same sample size per category might be applied to the 32 subcategories (for the same degree of precision in the estimates). Other answers depend on more detailed information about what you're trying to do. > I have some knowledge of statistics so I guess I could apply, and > hopefully understand, any formula as long as I could find one. If you're likely to be doing much stuff like this, you might consider talking your employer into supplying a good statistical package, such as MINITAB. This would permit you to concentrate on the logic of your problems, without worrying about how to apply formulas (which I take to mean, writing programs for such purposes). > Louis Tillier. Louis Tillier <[EMAIL PROTECTED]> ----------------------------------------------------------------------- Donald F. Burrill [EMAIL PROTECTED] 56 Sebbins Pond Drive, Bedford, NH 03110 (603) 626-0816 . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
