The purpose of this article is to point out a problem when using percentages for subgroups over time, or for members in a larger group, where the size of the denominator varies and probabilities are being estimated. Also to introduce a solution: adjusted p-chart scores (APC), a new way to score or compute percentages (e.g., in the p-chart setting).
ADVERTISEMENT |
Regular percentages are not a fair comparison when the denominators are different. Although 75 percent corresponds to three out of four, and 75 out of 100, and 3,000 out of 4,000, shooting three out of four free throws, for example, does not necessarily correspond to the same consistency as someone able to shoot 75 out of 100. Some degree of luck, or random error, is involved. The amount is directly related to the size of the denominator. Someone shooting three out of four free throws one time may shoot only one out of four (25%) the next time, while having the same underlying level of performance or ability. There is more range of error related to a small denominator, less with a larger denominator.
…
Comments
P' Charts
Some might be interested in the article "Improved Control Charts for Attributes" by David Laney (Quality Engineering, 2002), in which he deals with varying sample sizes in attribute control charts.
Not ready for prime-time
A few observations:
Author's reply
I appreciate your feedback, thanks. My comments: 1. APC charts are used when k = t or time. APC scores apply when k is more general, the binomial distribution is assumed. Which has but one parameter to estimate. I understand ANOM charts to estimate both mean and variance separately, which does not apply. Although ANOM does use the idea of charting over non-temporal variables, my article does not.
2. I had several PH.D. Mathematicians and Statisticians review the article before it was published. Quality Digest had a well known quality expert provide feedback as well.
3. I understand the Laney chart is used when the binomial distribution does NOT apply -- when there is overdispersion or underdispersion, i.e., when the width of the control limits is too tight or too loose. This is not the problem I address. I identify and offer a solution to the bias problem demonstrated with regular percentages when comparing subgroups over time (or members across a group).
4. Perhaps the Z chart transformed back to the P' chart scale is what you mean? Please let me know if this solves the sorting problem I present in my article. I found one example online where the P' chart had WIDER limits when subgroup size varied, but the limits varied. The APC scores adjust the actual "percentage" computation, not the width of the limits compared to binomial determined ones. Further, I am working on a more Robust APC method, which will certainly have advantages over the Laney tool.
Subgrouping
I enjoyed the article. I have data which has grouping as shown below. Do you know a control chart that takes account of the grouping as well? Thanks
group, id, number of items, number reviewed
1, 1, 567, 1000
1, 2, 543, 1020
1, 3, 456, 999
2, 4, 569, 1007
2, 5, 545, 1040
2, 6, 486, 1300
I'm not sure what you mean by
I'm not sure what you mean by "grouping." If the data are time-ordered samples from a process then a standard p-chart would work. A Laney p'chart might be better http://bit.ly/2hFuvwl.
Add new comment