# random sampling invoices for errors?

June 30, 2016 4:09 PM Subscribe

To check for errors, what would be a good number to sample?

I'm helping a new biz owner catch up with the bookkeeping.

I've already found one error in an invoice from a sub-contractor. There are approximately 30 invoices each month. I think this particular error was likely to be a typo (the keys are adjacent).

She's very cost-conscious, but of course, if there are lots of errors, that may be why things are feeling tight.

I'm helping a new biz owner catch up with the bookkeeping.

I've already found one error in an invoice from a sub-contractor. There are approximately 30 invoices each month. I think this particular error was likely to be a typo (the keys are adjacent).

She's very cost-conscious, but of course, if there are lots of errors, that may be why things are feeling tight.

Yeah, this is one of those cases where a small population size (~30/mo) and a (hopefully!) low count of the thing you're looking for bites you. Assuming you want your estimate to be reasonably accurate (let's say ±10%, which itself implies a fairly high % of incorrect invoices), you'll still have to examine 24+ invoices. May as well check them all…

And calgirl makes an excellent point: know your aim. If it's just to assess the # of incorrect invoices then random sampling at minimal effort (which in this case is still a large % of the whole) is fine. If you want to see if the problem correlates to something (as per her examples of invoice type or vendor) then a different / more intensive sampling strategy might be required.

posted by Pinback at 6:37 PM on June 30, 2016

And calgirl makes an excellent point: know your aim. If it's just to assess the # of incorrect invoices then random sampling at minimal effort (which in this case is still a large % of the whole) is fine. If you want to see if the problem correlates to something (as per her examples of invoice type or vendor) then a different / more intensive sampling strategy might be required.

posted by Pinback at 6:37 PM on June 30, 2016

I wouldn't sample. You can do the following:

- set a lower bound threshold to disregard (eg, everything under $100). Set an upper bound where everything above is reviewed (eg $5000). Set a time period for review (eg 1 year). Start reviewing in descending order by size of invoice until you either reach your upper threshold or get tired of reviewing.

posted by crazycanuck at 7:18 PM on June 30, 2016 [1 favorite]

- set a lower bound threshold to disregard (eg, everything under $100). Set an upper bound where everything above is reviewed (eg $5000). Set a time period for review (eg 1 year). Start reviewing in descending order by size of invoice until you either reach your upper threshold or get tired of reviewing.

posted by crazycanuck at 7:18 PM on June 30, 2016 [1 favorite]

Thank you!

I don't think there's a big problem; but I feel like I need to check and confirm that these are simply errors and not a problematic vendor.

So setting the boundaries will help and checking all those for the last 3 months. That lines up with when the biz started to get busier, and being more pressed for time, errors may have escaped detection.

I appreciate it - I needed something effective but not costly!

posted by egk at 4:40 AM on July 1, 2016

I don't think there's a big problem; but I feel like I need to check and confirm that these are simply errors and not a problematic vendor.

So setting the boundaries will help and checking all those for the last 3 months. That lines up with when the biz started to get busier, and being more pressed for time, errors may have escaped detection.

I appreciate it - I needed something effective but not costly!

posted by egk at 4:40 AM on July 1, 2016

If this is math/statistics question - the sampling rate would depend on the error rate (the higher the error rate, the lower the sample would have to be to detect them).

A targeted sample is good from a business/accounting view. From a math/statistics point of view you should sample randomly until you have enough positive hits (ie errors) to calculate the error rate. If you get 8 errors in a random sample of 10, that enough to suggest a rate of 80%, If you only find 1 error in 100 samples theoretically you should sample more (maybe 500) before deducing the error rate is 1%.

posted by Xhris at 12:44 PM on July 1, 2016

A targeted sample is good from a business/accounting view. From a math/statistics point of view you should sample randomly until you have enough positive hits (ie errors) to calculate the error rate. If you get 8 errors in a random sample of 10, that enough to suggest a rate of 80%, If you only find 1 error in 100 samples theoretically you should sample more (maybe 500) before deducing the error rate is 1%.

posted by Xhris at 12:44 PM on July 1, 2016

« Older Things to do in Reykjavik when you're too early... | I want to travel but I don't want to go anywhere Newer »

This thread is closed to new comments.

i would check all of them for errors in calculation for a test period of about 3 months.

Then you can assess whether errors are frequent, are limited to a type of invoice or even a particular vendor.

posted by calgirl at 4:15 PM on June 30, 2016 [2 favorites]