Statisical methods for Customer Service Quality Assurance

Hello all,

I work in a customer service center and we are attempting to decide on the right sampling plan for our Quality Assurance program. I am not quite sure which one is appropriate or how to justify the method once chosen.

Basically, we have several different touchpoints (web based chat, email, phone) and low volume but many clients. We want to give a 95% confidence level, with 4% margin of error on both the client (external) and rep (internal) levels.

Any advice and/or directions to resources would be much appreciated.

Thanks!

Thank you for the statistical information - it helps out a lot. Well personally, I wanted to go with a straight 10% but my boss wanted to use sampling.

What other methods would you suggest we use?

Thanks again for the information

[Marc]

Rick! Thank you for helping out with this fellow's question and the others you have helped with as well. Your help is greatly appreciated!

[Rick Goodson]

The question is difficult to give a simple answer to. Statistically based samples are always more reliable than straight percentages. If you develop the Operational Characteristic (OC) curve for the sampling plan with a straight percentage you will see that as the population size decreases, the protection afforded by the sample gets worse. If you do not have a strong background in statistics I would recommend you look at a basic text on sampling or a basic quality control book with a good section on sampling before proceeding. My favorite is rather old but does have a current revision:

Quality Control by Dale Besterfield, ISBN 0-13-745232-2

Getting back to the original issue. If you are sampling as part of an ISO internal audit, you can not economically use statistical based samples unless you have an army of internal auditors. I am sure that Marc and some of the other contributors would agree. When I teach internal auditing I explain that there is not enough resource (auditors) to go around and therefore you have to take samples that are extremely small and not statistically significant. However, if you find a problem from a sample that is extremely small, it is usually an indicating of a problem worth investigating as the odds of your finding that problem were also extremely small.

Rick, thanks for your reply.

I have done some research into statistics and I admit that some of it goes way over my head.

I agree that as population decreases, the company will have to sample a LOT more to get the sort of accuracy we want.

Because our population is fairly low, I decided to use a longer time frame. As I stretch it out over a month and longer, the sessions to be reviewed decreases to a point where it is feasible to QC. Of course, we can't QC at the end of the time frame so we're using historical data to forecast and break up the QC on a daily basis.

That's been my solution so far, what do you think?

Thanks again for your input.

[Marc]

I teach my course the same way. I explain that there are so many specifics you have to set a sample but that they should understand that it would be overwelming to even begin to define what a statistically valid sample would be. As an example, for documents I tell them I suggest a typical sample size of 3 to 5 documents. No problems, go on your way. If you find 1 'reject', increase your sample size by 5. If you find another problem you are at a point where you will have to decide on the spot - can't be predicted as it will be a judgement call based on the specifics at hand. I emphasize that the sample size I recommend does NOT have a true statistical foundation.

[Rick Goodson]

Whether the use of statistical sampling is the right thing to do in this situation is a matter for debate. Never the less, the statistical side is relatively straight forward.

The formula for finding the sample size (assuming a continuous process that is in control) is:

n = sample size.
p = estimate of process percent defective.
Zcl= Z score of confidence level (remember
that this is a two sided estimate so
you have to use half the confidence
factor on each side, for 95% confidence
level you use 97.5% to find the Z
score).
E = Allowable error of the process percent
defective.

2
Zcl x p x (1-p)
n = ---------------
2
E

In essence, the sample size is equal to the Z score squared times the percent defective time 1 minus the percent defective divided by the allowable error squared.

References:

Process Quality Control by Ellis Ott, ISBN 0-07-047923-2, pages 76 - 79.

Modern Methods for Quality Control and improvement by Wadsworth, Stephens, and Godfrey, ISBN 0-471-87695-X, pages 173 - 176.

Hope this helps.

[Rick Goodson]

I hope you can make out the formula I just posted. Unfortunately the software makes it hard to accurately post a formula. Hopefully the text makes it clear.