Ohio Tax Statistical Sampling A Whirlwind Tour of How Procedures - - PDF document

27th Annual

Tuesday & Wednesday, January 23‐24, 2018

Hya Regency Columbus, Columbus, Ohio

Ohio Tax

Workshop MM

Statistical Sampling … A Whirlwind Tour of How Procedures & Techniques Vary Among States

Wednesday, January 24, 2018 2:00 p.m. to 3:00 p.m.

Biographical Information Brad W. Tomlinson, Senior Manager (non-attorney professional), Zaino Hall & Farrin LLC 41 South High Street, Suite 3600 Columbus, OH 43215 btomlinson@zhftaxlaw.com 614-349-4818 Brad has more than 34 years of experience in state and local taxation, specializing in statistical sampling techniques for multiple taxes. From 2007 until joining the firm, Brad was an assistant administrator for the Ohio Department of Taxation, Audit Division, while also serving as the manager of the Computer Assisted Audit Group for more than 11 years. In that position he was responsible for the formulation, implementation, and oversight of the Department's statistical auditing practices; including the approval of data populations, sample designs, and the review of all statistical sampling procedural agreements. Prior to that, Brad spent several years auditing fortune 500 companies in manufacturing, telecommunications, computer services, and retail for sales and use tax compliance. Brad was instrumental in the design and implementation of the Audit Division's computerized auditing program (OFAST) used by multiple divisions to audit personal property, corporate franchise, employer withholding, pass-through entity, and sales and use taxes; as well as managing quarterly updates and releases of the application. Brad's other responsibilities included the establishment of single rate reporting procedures for taxpayers as well as assisting in reviewing and approving the Department's PARSA (Previous Audit Representative Sampling Analysis) agreements. Brad is a frequent presenter on statistical sampling at the Ohio Tax Conference, the Federation

• f Tax Administrators, and Institute for Professionals in Taxation national conferences and

workshops. Education

• Columbus State Community College, Associates Degree, Computer Programming
• Ohio State University, B.S., Business Administration

Jonathan Ross, Senior Tax Consultant , Deloitte Tax LLP 180 East Broad Street, Suite 1400, Columbus, OH 43215 (614)229-5932 joross@deloitte.com Jonathan is a Senior Tax Consultant in the sales and use tax practice with Deloitte Tax LLP in the Columbus, Ohio office and has over ten years of sales and use tax experience. In addition, he has a background in statistical sampling as it relates to sales and use tax audits. Jonathan was a sales and use tax auditor for the State of Ohio for three and a half years. During that time he conducted sales and use tax audits and refund reviews. In addition, he transitioned into the state’s computer audit support group which oversees all statistical samples done by the audit division as well as data manipulation support for the entire audit division. Oversight of the statistical sampling program required: refining populations, designing statistical samples based on certain parameters, generating samples, and evaluation of the samples upon conclusion of audits. Since leaving the state Jonathan has assisted clients with state sales and use tax audits, conducted reverse audits, voluntary disclosure negotiations, and provided sales and use tax and statistical sampling consulting services for a variety of clients in a multitude of industries. During his tenure with Deloitte Tax Jonathan has been involved in design and implementation teams for various bolt-on automated tax solutions. Jonathan is a graduate of the University of Kentucky with a B.S. in Accounting

Biographical Information Roger C. Pfaffenberger, Ph.D., Director, Audit Sampling, Ryan, LLC Three Galleria Tower 13155 Noel Road, Suite 100 Dallas, TX 75240 roger.pfaffenberger@ryan.com 972.934.0022, Ext. 101279 Fax: 972.960.0613

• Dr. Roger Pfaffenberger is a Director and Practice Leader for Ryan’s Sampling Analysis and

Evaluation practice and is based in the Firm’s Dallas Office. Dr. Pfaffenberger is responsible for the design and evaluation of audit sampling methodologies with an emphasis on the use of multiple audit sampling methods in statistical and non-statistical sampling for transaction taxes. Roger has provided statistical consultation and expert testimony for a variety of companies, government entities, and law firms. Prior to joining Ryan, Dr. Pfaffenberger was Professor of Decision Sciences for Texas Christian University and Founding Director for the Center for Teaching Excellence. Dr. Pfaffenberger has made numerous presentations on audit sampling issues to professional and academic

• rganizations. He has a Doctor of Philosophy Degree in Statistics from Texas A&M University.

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

• In the planning phase or very early stages of an audit, the auditor should

indicate if he or she plans to use sampling. When sampling will be used, the auditor should prepare a sampling agreement and submit a sampling plan. Once the sampling agreement and plan have been agreed to by the taxpayer and auditor, the auditor will generate the sample and determine the audit results. The auditor should handle this process while maintaining complete confidentiality about the taxpayer’s data and the results of the sample audit.

• Inevitably, issues arise during a sample audit that may require modification of

the initial agreement. Ideally, sampling agreements should be an essential

• utcome of the process of planning the sample audit through

meetings/discussions with the auditor.

• If possible, the taxpayer should be actively involved in the development of the

sampling agreement. The agreement received should be reviewed in detail and the actual sampling should not take place until everyone agrees to the sampling

• methodology. This should also remove any fears or concerns the taxpayer may

have about the sampling process and reduce the potential for disputes over the sample results. In addition, the taxpayer will be responsible for providing the data to be sampled and should also be involved in the sampling process, including verifying the results of the sample audit. 27

• Sampling agreements can be either binding or non‐binding. Ohio, for example,

has a binding agreement. If the taxpayer refuses to sign the binding sampling agreement, Ohio will not project overpayments of tax in the sample.

• If the sampling agreement is binding, develop a memorandum of understanding

that addresses ways in which audit issues are to be handled if not dealt with partially or in whole in the agreement. Append the memorandum to the

• agreement. If the jurisdiction will not accept this approach, consider not signing

the binding agreement. An ideal sampling agreement is not binding and covers all the key issues that may arise in the sample audit. It should represent a planning instrument that is derived from discussions and agreements with the auditor about how the audit sample should be executed.

• This sampling agreement should also outline what the auditor will be estimating,

the actual sampling unit used and the period to be sampled, the sampling methodology, and how the outcome of the sampling process will be extrapolated to the sampling population (method of estimation). 28

• The sampling agreement and plan should also outline how missing

documentation, negative transactions or credits, tax‐only items, error corrections, non‐taxable items, tax paid in error on purchases, installment and progress payments, tax law changes, and accounting or reporting changes will be handled when detected during the verification process. Each of these issues could have a significant impact on the sampling results and any conclusions that may be drawn from them.

• Discussing and agreeing with the auditor about how the special topics should be

treated is an essential and extremely important step in planning the sample

• audit. Agreeing on the treatment of these special topics before the sampling

plan and sample have been produced minimizes the chance for a “wheels off” sample audit. The best example of the nonbinding and effective sampling agreement is the California Board of Equalization BOE Form 472 which can be downloaded from the CBOE internet site: www.boe.ca.gov/pdf/boe472.pdf. 29

• To insure that the sampling population is appropriate, valid and is the agreed‐

upon population, the derivation of the sampling population must be replicable by following the steps in the jurisdiction’s derivation process. The inability or failure to replicate the sampling population potentially leads to a sample that does not represent the scope of the audit agreed to by the auditor. Worse, a sampling population whose derivation cannot be replicated can lead to corrupt data and biased sample results.

• A key element in the derivation of the sampling population is selecting the

accounts of interest for the audit. These accounts should include those likely to contain overpayments as well as underpayments. Additionally, the derivation criteria should include how the transactions are grouped (e.g., assets, expenses) and how special situations are treated by isolating certain groups (e.g., procurement card transactions). 30

• Planning the sample audit should be an interactive process. The taxpayer and

the auditor should both be involved in every step including, most importantly, the development of the sampling plan. The sampling plan should describe how the sampling population is stratified qualitatively (e.g., by types of transactions such as assets, expenses, procurement cards) and quantitatively (e.g., by dollar ranges including the use of lower exclusion thresholds and the upper detail thresholds for the groups of transactions).

• Contributing to the development of the sampling plan will also help the taxpayer

feel more comfortable with the sampling results once they are available. It is much better to agree on the plan before the sampling actually takes place. Once sampling has begun and the auditor has invested considerable time and effort in selecting and evaluating items, it may be difficult to convince the auditor that the sample is not representative of the audit period or to persuade the auditor to do additional testing. 31

• Knowing your data and accounting systems well is an insurance policy against a

“wheels off” sample audit. Knowledge of your data can provide very useful information about potential areas of liability or refund opportunity, which can be used to propose an audit sampling strategy at the initial meeting with the

• auditor. Far better to be proactive in developing the sampling plan than

reactive.

• Verifying the results of the audit sample assures that the jurisdiction’s estimate
• f the tax adjustment is valid and mathematically correct. Try to agree with the

sample items that the auditor is indicating are in error. This may not always be

• easy. In one particular case, some tax‐only invoices were included in the sample

and the auditor asserted that these were errors. We disagreed, however, because the sampling plan was poor and did not address how these transactions should be handled. We could never attain agreement on whether these invoices were errors or not. Don’t blindly agree to the auditors extrapolation of errors calculation or the penalty and interest that is calculated based on the audit sample results. Recalculate the amounts for each of these values. Auditors have been known to make mistakes. Mistakes are often created by keying errors when the sample results are manually entered into the audit workpapers or when incorrect formulas are used in spreadsheets. 32

33

• There are two types of sampling. The first type is statistical sampling, which uses

statistical formulae to determine the sample size and measure the sampling risk. This method involves the selection of samples on a random basis, where each member of the population has an equal opportunity of being selected as each item is sampled. The second type is non‐statistical sampling, which often involves the determination of the sample size and sampling risk based on the professional judgment of the auditor. When this method is used, the items included in the sample are selected on a judgmental basis and the auditor will be unable to quantitatively determine the sampling risk.

• Whenever sampling is used in an audit, some risk will be introduced into the

audit plan. Sampling risk is the risk that the sample is not representative of the actual situation, which increases the possibility that incorrect conclusions will be

• made. Non‐sampling risk refers to the risk of making wrong conclusions based
• n things that are unrelated to the sample or its size.

34

• Applying statistical sampling methods requires professionals with graduate

statistics degree expertise and several years of audit sampling experience. This is the reason why about half the states in the U.S. do not use statistical sampling.

• Stratified random sampling is the sampling method used most commonly in the

U.S. for sample audits. Simple random sampling is also used, but only if the sampling population is small, usually 1,000 transactions or less.

• Some states use stratified random samples, but do not provide a measure of

sampling risk (e.g., Texas).

• There is a trend in the United States to use statistical sampling increasingly in

sales and use tax audits. In Canada, only two jurisdictions (the Canada Revenue Agency and British Columbia) have written policies dealing with the use of statistical sampling. However, we have yet to see situations where either of these jurisdictions have used statistical sampling while auditing a taxpayer. 35

• Some jurisdictions in the U.S. have tried to argue that block sampling is a form of

cluster sampling, but by statistical terms, it is not. Say, for example, the block sample uses a sample of six months taken from a thirty‐six month audit period. For the six‐month sample to be considered a cluster sample, the six month sample must have the same characteristics of the thirty‐six month population (e.g., same error rate, same average size). Typically, that is not the case. But, more importantly, the degree to which the block sample represents the sampling population cannot be tested since no transactions in the 30‐month non‐sampled period are audited.

• Sequential sampling is rarely used now. Way back in the day, prior to electronic

data when print reports constituted the basis for a sample audit, every 10th page for example could be used in a stack of several thousand print reports for the audit. No need to do that today (we hope!). 36

• Block sampling continues to be used by jurisdictions that do not have access to

the statistical expertise required to perform statistical sample audits. This method of sampling involves a review of all transactions that occur within a selected time period, such as a month or series of reporting periods, and then extrapolating the results (i.e., through the use of an error rate) over the complete audit period.

• Convenience sampling in the auditing world is often referred to as judgment
• sampling. The auditor uses his or her judgment to select the items for testing.

Obviously, these types of samples can lead to significantly biased sample audit results. 37

• The mathematical complexity of statistical sampling does require a fairly high

level of statistical expertise. Most states using statistical sampling have either hired professionals with at least a Bachelor’s degree in statistics, or use university professors in statistics departments to assist with the development and the application of statistical sampling in their audits.

• A key element of statistical sampling is the fact that the sample will contain

randomly selected transactions throughout the audit sampling period. If there is a change in the business structure through reorganization, acquisitions, or mergers during the audit period, or if the tax error rate changes during the period, a six‐month block sample for a thirty‐six month audit period may not represent the full audit period well.

• Generally, statistical sampling will produce the most accurate results. Despite

this fact, taxpayers are more likely to see an auditor use non‐statistical sampling during an audit because 1) the taxpayer may not be able to supply complete and accurate data necessary for statistical sampling and/or 2) the auditor may lack sufficient statistical sampling skills. 38

• When a block sample of say three months from a thirty‐six month audit period is

taken, all the transactions within the scope of the audit for those three months should be reviewed by the auditor. But, auditors are prone to short‐test the block sample, which can lead to serious estimation bias on top of the bias already created by an unrepresentative three month block.

• This method involves considerable judgment on the part of the auditor and will

likely produce skewed results. Since the sampling only occurs for transactions

• ver a specific time period, there is a risk that changes which have occurred in
• ther time periods will not be reflected in the results. Events that may not be

reflected include the following: changes in how the taxpayer performs compliance; changes in an organization’s purchasing patterns; and, significant changes in the taxpayer’s business operations. In addition, the sampling results will be skewed if the sample includes any items of a non‐recurring nature, such as fixed asset purchases.

• Block sampling can also work to a taxpayer’s advantage. This will be the case

where the taxpayer’s sales tax compliance is better during the period sampled than is normally the case. In this situation, the lower error rate in the sample period will reduce the overall tax liability that is projected over the entire audit period. 39

40

• The Internal Revenue Service (IRS) began using the 95% confidence bound to

give the taxpayer a break by taking into account that the assessment is based on a sample, and there will be some degree of estimation error introduced by using the point estimate for the assessment.

• In our view, the use of the confidence bound introduces a high degree of

complexity that is difficult for auditors to deal with and difficult for the taxpayer to understand. In the process of auditing the sample results, every time a change is made to the scheduled adjustments, the confidence bound must be recomputed and that is something neither the auditor nor the taxpayer can do without statistical expertise assistance. And, in a net refund situation, the confidence bound works to the detriment of the taxpayer by providing a refund that is smaller than the refund based on the point estimate. 41

• The difference estimator calculates an estimate of the error in the total

population by calculating an average error for the sample and then multiplying this average by the population size.

• Using the separate ratio estimator (a.k.a. the percentage method), the estimate
• f the error in the total population is determined by taking the percentage of

sample dollars in error (sample dollars in error / total sample dollars) and multiplying by the total dollar amount of the population.

• There are four additional estimators that some jurisdictions in the U.S. use:
• (1) difference estimator;
• (2) combined ratio estimator;
• (3) separate regression estimator; and
• (4) combined regression estimator
• These four estimation methods are complicated and not easy to use or explain.

The IRS uses all six estimators, and selects the estimator with the best relative precision on which to base its audit results. The State of Washington uses four

• f the estimators for its sales and use tax audits.

42

43

• The plan uses a lower exclusion threshold of $10, a good way to focus the pull

list on more material transactions.

• Since stratification is applied to real values, the negatives must have been carved
• ut of the sampling population.
• A detail threshold of $1,000 has been used, and this threshold places 150

transactions in the detail stratum.

• The issue with the plan that jumps off the page are the sample sizes in the

sample stratum. In Stratum 1D, the plan calls for sampling 475 of the 500 transactions in the stratum population. Why not sample the remaining 25 transactions so that the Stratum 1D population is detailed?

• The auditor has determined the strata sample sizes by applying the sampling

rate percentages shown in the last column. These arbitrary percentages were selected by the auditor with the thought that the higher dollar strata should be more heavily sampled. But, there really is no statistical basis for doing this. 44

• The guideline of a maximum value of $50,000 for the tax projection on average

item measure is also hard to meet in some industries, such as the oil and gas industry.

• The higher the percentage of the population base dollar amount covered by the

detail stratum, the lower the sampling risk. So, it is desirable to push the detail threshold as low as possible to produce a high coverage percentage. 45

• Typically, the lower exclusion threshold is set such that the maximum

percentage of the population base dollar amount is 5%. But, in this plan, we cannot determine what that percentage is since we do not have the count and amount of the excluded positively‐valued transactions.

• The defect in the sampling plan is the large sample sizes that have been

determined by applying arbitrary percentages to the population stratum sizes. Unfortunately, there are no guidelines for the maximum strata sample size. But, common sense tells us that these sample sizes are way too large! 46

47

• Non‐sampling risk is not directly considered in most audits, and its effects have

not been analyzed statistically to any extent. And, the effects can be significant, particularly in measurement error. The auditor may consider a transaction to be in error when it is not, or consider a transaction not to be in error when it is. Obviously, tax determination errors can have a material effect on the results of a sample audit. This is an area that merits future statistical research to determine the effects of these types of errors. 48

• States that use statistical sampling typically set the maximum allowable relative

precision percentage at 20% or 25%.

• Confidence settings vary by states. The States of Washington and Utah use a

confidence level of 80%, while Tennessee uses a confidence level of 75%. These three states use the confidence bound rather than the point estimate to determine the sample audit results.

• States that use the point estimate for the sample audit results typically use a

confidence level of 90% (most central to north central states).

• The IRS uses a confidence level of 95% for its confidence bound, and has a

targeted relative precision setting of 10% or better. 49

• Confidence intervals are used by states conducting statistical samples to

determine how reasonable the sample results are, and to determine if sample expansion is required. For example, a particular audit may produce a point estimate assessment of $600,000, but the 95% confidence interval ranges from a negative $100,000 (a refund) to a positive $1,300,000. The interval is so wide as to render the point estimate useless. In this case, the sample would have to be expanded significantly to shorten the width of the confidence interval. 50

• This is a salient but important point. Once a 95% confidence interval is

computed, say from $100,000 to $200,000, the population error amount is either in this interval or it is not. We cannot say that the probability is 95% that this interval contains the true population error amount. We have heard statisticians in court testimony misinterpret a confidence interval in this way.

• The width of the confidence interval can be shortened by increasing the sample

size or by decreasing the confidence level.

• A 100% confidence interval stretches from minus infinity to plus infinity. That

interval is certain to contain the true population error amount! 51

• The margin of error is commonly used to report political polling surveys. The TV

announcer might say that based on a survey poll of registered voters, a particular candidate has a 60% chance of winning the election with a margin of error of 5%. Most political polling surveys use 95% as the confidence level

• setting. So, equivalently, the announcer could say that the 95% confidence

interval of the percentage of votes the candidate will receive in the election is between 55% and 65%. 52

• The use of a confidence bound to report the results of a sample audit introduces

a level of complexity that most taxpayers and many auditors do not understand. In the process of conducting the audit, when each change is made to the schedule of the sample errors, the confidence bound must be recomputed, and that requires statistical software or someone with extensive statistical knowledge.

• The use of the confidence bound is attractive when the audit result is a net

assessment since the taxpayer will be assessed at an amount less than the point

• estimate. But, it is not so attractive when the audit result is a net refund. Then,

the taxpayer will receive a refund that is less than the point estimate. 53

• The calculation of the percent difference between the sample mean and

population mean is important and useful. It shows the degree to which the sample is representative of the population in mean value.

• Some states (e.g., California) also compare the sample standard deviation with

the population standard deviation to determine the extent to which the sample is representative of the population in the spread of the values. 54

55

• Virginia sampling guidelines set the overall maximum sample size at 1,000,

which is arbitrary and capricious as it ignores factors that should be used to determine sample size such as the population size and the variation of the population transaction amounts.

• There are three inputs needed to set the sample size by statistical formula:

(1) the level of achieved precision required; (2) the confidence level; and (3) the standard deviation of the measurement variable.

• A typical statement to set the first two items required is as follows: “I want the

estimate of the true population tax adjustment amount to be within 10% of its true value (relative precision = 10%) with 90% confidence (confidence level equals 90%). The third required input is the kicker. We need the standard deviation of the audited differences between the tax that was paid and the tax that should have been paid. But, we won’t know these differences until the sample is drawn and the sampled items are audited. Most jurisdictions use the standard deviation of the taxable amounts as a surrogate. This method is reasonable but can often lead to sample sizes that are considerably different from the sample sizes that are actually required. 56

• The U.S. Multistate Tax Commission recommends a minimum stratum sample

size of 200 items for stratified random sampling, and an absolute minimum sample size of 100 items. Tennessee uses a minimum of 70 items, which is too small to detect tax errors in the stratum populations with low error rates.

• Some states require a minimum number of errors that must be detected in a

sample stratum to project the sample result to the stratum population. California, for example, requires a minimum of 3 errors per stratum. If the number of errors is less than 3, then the sample results are not projected. The sample size must then be increased, or the stratum population items detailed. 57

58

• Most auditors are pretty reasonable when it comes to missing invoices. If you

can show the auditor that you have invoices for purchases similar to the products purchased and during approximately the same time frame, the auditor may not schedule the item for assessment.

• From a statistical standpoint, the use of spare items for items with missing

documentation is not acceptable as it alters the random nature of the original

• sample. The item should be audited and a tax determination made based on all

available information. In addition, auditing the item is important since it may provide key information for why this item, and perhaps others in the sampling population, have missing documentation.

• In the planning phase of the sample audit, it is always a good idea to take a pilot

sample of transactions for which missing documentation is suspected. This will help the taxpayer to determine the degree and the extent of the problem before the auditor arrives for the initial meeting.

• Statistically, the best approach in dealing with negatively‐valued transactions is

to match as many debits and credits as possible, then carve out the remaining negatives and develop a sampling plan only for the positives. 59

• Progress payments on contracts can present challenging sampling and tax

determination issues. Be certain to study the contracts to determine which installment triggers the tax event. California deals with this issue very poorly. The tax event occurs when the contracted products are delivered. If tax has not been paid on previous installments or on this final installment, then the sampled installment amount is replaced by the full contract amount and the full contract tax error amount is projected in the dollar range stratum containing the installment payment amount. This policy can lead to multi‐million dollar assessments that are statistically unwarranted.

• It is a good idea to review contracts with installment payments in the planning

phase of the audit, and then to discuss the treatment of contracts with potential tax consequences with the auditor. Generally, it is best to place these items in a separate population for sampling or detailing.

• From a statistical point of view, tax only items should be retained in the sample

but treated as non‐errors. To do otherwise gives the transaction with a debit component and the tax only line item twice the chance of hitting the sample. The best approach is to eliminate as many tax only transactions from the sampling population before sampling. Then, treat any remaining tax only transactions that hit the sample as non‐errors. 60

61

• The statistics on this slide are not hard and fast counts in the four categories.

Some states permit exceptions or apply their policies “sometimes.”

• The Federation of Tax Administrators (“FTA”) performed a survey in 2002 that

produced state policies for sample audits. Among the questions asked were whether states permitted projection of overpayments in state‐initiated audits and permitted refunds based on samples for tax paid to vendor and for use tax

• paid. The survey was updated in 2004. Though the survey results are quite old,

interestingly for these measures, the statistics in the survey results more or less still apply. The January 2004 matrix summarizing the audit sampling policies by state is available at the FTA website. However, in viewing the new website, it appears that accessing this file now requires FTA membership. 62

• Ryan is working with states that either do not permit projection of
• verpayments in state‐initiated audits or do not permit refunds to be based on

samples to change their policies. Florida is currently revising its policies and procedures pertaining to projecting overpayments and basing refund claims on samples based on discussions with Ryan.

• The managed audit option is available in most states and many of these states

will allow the projection of overpayments in the managed audits. 63

64

Question 1: Yes, absolutely! In fact, your active participation in the planning process is essential to produce a valid and precise sample audit result. Question 2: The best solution is to convince the auditor to isolate these transactions in a separate population for sampling or for detailing. Normally, the error rate for p‐card transactions is higher than for other expense transactions. So, we do not want this higher error rate projected over all expense transactions. Question 3: Well, you shouldn’t have been surprised. Knowing your data and identifying groups of transactions with potential liability before the auditor arrives for the initial meeting is part of an effective planning strategy for the audit. Once this problem emerges in the audit, you are at the mercy of the auditor to be benevolent and kind by using any and all alternate information to accept evidence that tax was almost certainly paid on these items. 65

Question 1: It is obviously not fair! However, it is very difficult to get the auditor to change the jurisdiction’s audit sampling policies and procedures in a specific audit. This is a matter of public policy, and the matter should be taken up at the policy level, not the auditor level. While that may have no benefit in the current audit, it may produce desirable results for future audits. Question 2: Generally, no. It is difficult to convince a jurisdiction to depart from its audit sampling policies and procedures – but, it’s worth a try. When auditing sales for large retailers, many jurisdictions in the U.S. permit the use of block sampling in some form. Texas, for example, will use store/days as the sampling unit for sales

• audits. And, for the very largest of the retailers, Texas may use store/day/hours as

the sampling unit. Question 3: Many taxpayers grossly underestimate the cost of litigating sample

• audits. Getting sound legal advice as reliable input to a cost benefit analysis is

essential to make a well‐informed decision about whether or not to litigate. Also, it is important to understand the difficulty in winning a court case based on a statistical or sampling issue. Often, courts do not have the statistical expertise in hand to understand complex statistical or sampling issues. Most judges will side with the jurisdictions on these issues unless the evidence is compelling, indeed

• verwhelming, in support of the taxpayer’s case.

66

Question 1: Replacement samples should not be used in audit sampling as their use violates the principle of random sampling. The resulting sample is no longer random by the definition of a random sample. The correct procedure in this situation is post‐stratification. Post‐stratification removes the transactions from both the population and the sample. Then, sample expansion may be needed to meet minimum sample size requirements. Care must be exercised to use the same random seed number for sample expansion that was used to generate the original sample. Question 2: Generally not, but it depends on the state’s policies and procedures concerning overlapping audits and double taxation. The sample audit results are projected over a population that includes all the vendor’s transactions. If the vendor’s transaction was outside the sampling population, then it would be proper to reimburse taxes to that vendor. Question 3: We would not take this approach due to the overall very large sample that may be required to do this. However, it may make sense to create a separate sampling group for the large vendors, then sample that group independently from the remaining purchase transactions. A well‐designed statistical sample with sufficient sample size should represent the population well in terms of all vendors, large and small. 67

Question 1: So, how often does this situation arise? Might it be possible to group these types of transactions in a separate population and sample by PO rather than invoice or line item? If not, this is an allocation problem. We assume on the PO that there are taxable and non‐taxable items. If these items can be identified and tax determinations made for the items, then the PO tax amount can be properly allocated to the taxable items. So, if one or more of these items hit the sample, then its associated tax determination has been made. In the broader view, this seems to be a system accounting issue. Perhaps the system can be altered to permit the tax amounts to be properly associated with the invoices associated with the PO. Question 2: Create a field based on the absolute value of the transactions. Then,

• rder this field from the largest value to the smallest carrying along in the sort all

the other fields in the database. Next, line up the field containing the real value of the transaction with the field containing the absolute values and scan for matching

• numbers. This is an effective way to at least pick up matches with high dollar
• values. In addition, within the first sort, sort also by date to see if a reversal has
• ccurred within a few days of the original transaction. Matching partial reversals is

difficult, if not impossible, without spending a great deal of time with the matching process. 68

Question 1: Our first suggestion is to place contracts with progress payments in a separate group for sampling. The treatment of installment payments is nasty, and varies by state. In California, if tax should have been paid on a project and wasn’t, an assessment depends on whether the installment payment that hit the sample constitutes the tax event (i.e., goods/services were transferred at this point in time). If that installment payment hits the sample, then the installment payment amount is replaced by the full contract amount, and projected using the projection factor in the stratum in which the installment payment occurs. And, that can lead to some huge projected assessments! If any of the other installment payments hit the sample, they are treated as non‐errors. Other states will apply the tax rate to each and every installment payment that hits the sample for the contract, but the tax rate is applied only to the installment amount. That is a much preferred methodology statistically. In our view, if you can show that tax was paid at a later date, even a date past the audit period, then the installment payment that hit the sample should be treated as a non‐error. But, you must be able to show that the tax payment was made prior to the notification of the audit. 69

70

71

72