Nonparametric Inference, Auditing, and Litigation

Short course at XXIX International Forum on Statistics

Puebla, Mexico, 29 September–3 October 2014

Philip B. Stark

Department of Statistics, University of California, Berkeley

These materials are available at https://github.com/pbstark/MX14

Abstract: Many problems that arise in financial and election auditing, civil litigation, and causal inference can be reduced to statistical inferences about the mean of a nonnegative or bounded finite population. A variety of sampling plans can be combined with common probability inequalities to test hypotheses or make confidence intervals in these applications, in a fully nonparametric, conservative way. I will illustrate these methods with real and cartoon examples from election auditing, healthcare auditing, intellectual property litigation, wage and hour litigation, and online advertising. An especially useful class of methods can be derived from Wald's (1945) sequential probability ratio test (SPRT), which hinges on a generalization of the problem of gambler's ruin. Methods based on Wald's SPRT allow samples to be drawn incrementally and adaptively, often reducing the cost of financial and electoral audits, litigation discovery, and experiments without incurring any penalty from multiple testing.

Description of the materials

This course comprises a series of IPython notebooks.

To view them, you can either clone this github repository onto your own machine (which would then let you change parameters, modify the scripts, etc.) or navigate to http://nbviewer.ipython.org/github/pbstark/MX14/tree/master/ to see a statically rendered version.

The chapters are:

1. Canonical examples of real-world problems we will consider
2. Why not use the normal approximation?
3. The duality between confidence sets and hypothesis tests
4. Confidence bounds for the mean of a bounded population: Binomial and Hypergeometric
5. Confidence bounds from the Chebychev and Hoeffding Inequalities
6. Lower confidence bounds for the mean of a nonnegative population: Markov's Inequality & methods based on the empirical distribution
7. Wald's Sequential Probability Ratio Test
8. The Kaplan-Wald Confidence Bound for a Nonnegative Mean
9. Dollar-unit sampling and taint
10. Penny Sampling and Continuous Penny Sampling
11. Method shootout
12. Bibliography

If you discover errors or bugs, please let me know.

If you find these materials useful, please let me know.