- 130 Views
- Uploaded on
- Presentation posted in: General

Stratification Case study to illustrate alternative methods to stratify a sampling frame

Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author.While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server.

- - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - -

StratificationCase study to illustrate alternative methods to stratify a sampling frame

Dr. Will Yancey, CPA

This material is the property of the presenter and cannot be reproduced or used without the expressed

Written consent of the presenter.

VII-1

- Why stratify?
- Coefficient of Variation (CV)
- High and Low Thresholds
- Number of strata
- Strata Boundary Determination
Case study data for this presentation:

185,083 rows of purchase invoice line items.

VII-2

Parable of the Footballs and the Fish

- You are asked to determine the weight of 1,000 footballs. You know they are identical in weight. You can weigh only one football at a time. How many must you weigh?
- You are asked to determine the weight of 1,000 different fish taken from a lake. They are highly variable in weight. You can weigh only one fish at a time. How many must you weigh?

VII-3

- How could we organize the fish so we could get a reasonable estimate without weighing them all?
- What feature would we use to organize the fish?
- What features would probably not be useful for estimating total weight?
- How many piles should we have?

VII-4

Effective stratification: If possible, what we are measuring is similar within each stratum and different between strata.

Stratifying (grouping, categorization, segmenting, etc.)

- Grouping by account, type, division, or other attribute.
- Stratifying by dollar amount within group.
A sales and use tax audit goal is to estimate total tax dollar error.

Correlation of invoice line amounts with taxability or errors:

- If an error occurs, it is proportional to invoice line amount
- The relative frequency of error occurrence might or might not be correlated with invoice amount.

VII-5

185,083 rows of invoice line items

Range $0.01 to $26,763,476

$493 million total population base

4% of items with amount ≥ $10,000

VII-6

4% of items with amount ≥ $10,000 contain

$376 of the $493 million in population base = 76%

> $10K

VII-7

CV is a relative measure of the dispersion around the mean.

Dollar stratification results in lower CV within each stratum than in the combined unstratified sampling frame.

Caution: When the mean is close to zero, CV is very sensitive to small changes.

VII-8

Reducing CV usually improves precision.

(Remember Parable of Footballs and Fish.)

For each stratum compute the CV of the items’ invoice line amounts.

For a specific total sample size and stratified random sampling, the best precision usually occurs when the CV are relatively constant across the strata.

- Consider adjusting strata boundaries or adding more strata to adjust CV across the strata.

VII-9

VII-10

All items with dollar amount greater than High Threshold (H) will be detailed (actual basis exam) rather than sampled.

“This removal of the extremes from the main body of the population reduces the skewness and improves the normal approximation.” Cochran, Sampling Techniques, 3rd Edition, p. 44.

VII-11

- Approximately top 0.1% to 0.2% of items (or some other %).
- Greater than 3 standard deviations from the unstratified population mean.
As H decreases, the number of items in the detail stratum increases.

Items above H are from relatively few major vendors or major projects.

VII-12

Population Size = 185,083. Population Base = $492,953,742.

Exhibits in this presentation: H = $100,000.

VII-13

Accounting transaction data files have many small dollar items – particularly for purchases with invoice line items.

- Delivery charges, processing fees, etc.
Some sampling plans set a Low Threshold (L) such that every item below L is:

- Excluded (no change), or
- Minimum sample size, or
- Project results from other sampled strata onto the stratum below L.

VII-14

Policy for setting L depends on what will be done with items below L.

Possible criteria for setting a value for L

- Less than 1% or 2% of population dollars are below L (or some other %).
- Greater than 3 standard deviations below the unstratified population mean.
- Divide H by 1,000.

VII-15

Exhibits in this presentation: L = $100.

VII-16

- Adding more strata
- Reduces CV within stratum.
- Minimum sample size per stratum may result in total sample that exceeds budget.
- More than 6 strata probably does not improve precision [Neter and Loebbecke, Behavior of Major Statistical Estimators in Sampling Accounting Populations, (AICPA, 1975)].

- Pragmatic approach: Start with 3 strata and then add or delete strata as needed to achieve desired precision, CV, or other criteria.

VII-17

- Precision is a function of strata boundaries combined with other attributes in population and the sampling plan.
- Unless otherwise stated, the following case study shows:
Five strata = 3 sampled strata + Low + High

Low Threshold (L) = 100

High Threshold (H) = 100,000

VII-18

Observe: CV varies greatly across strata 2, 3, and 4.

VII-19

Observe: CV varies greatly across strata 2, 3, and 4.

VII-20

- Developed by Tore Dalenius, a Swedish statistician, in the 1950’s with the warning that it will not do well with all distributions.
- See numerical example in New York State CAA Manual, Publication 132, www.tax.state.ny.us/pdf/publications/sales/pub132_1001.pdf , pages 17-19.
- Cumulative square root method can be distorted when begin from zero and there are lots of small $ items (such as under $10).
- Mitigate by setting L threshold greater than zero.

VII-21

Observe: CV varies greatly across strata 1, 2, 3, and 4.

VII-22

Observe: CV is closer across strata 2, 3, and 4.

Setting an appropriate L has improved the stratification.

VII-23

- Developed by Will Yancey with co-authors Jane Horgan and Patricia Gunning at Dublin City University in Ireland in 2003.
- Assumes population distribution declines at a relatively constant rate.
- Requires setting thresholds L and H.
R = H / L = 100,000 / 100 = 1,000

For J=3 strata: r = R ^ (1/J) = 1,000 ^ (1/3) = 10.0

For J=4 strata: r = R ^ (1/J) = 1,000 ^ (1/4) = 5.623

VII-24

Observe: CV is relatively similar across strata 2, 3, and 4.

VII-25

Observe: Adding more strata lowers the CV.

VII-26

- Set a High Threshold (H).
- Set a Low Threshold (L).
- Choose number of strata.
- Set boundaries with a method.
- Compute CV in each stratum.
- Adjust by changing L, H, boundaries, adding or deleting strata.

VII-27