|
||||||||||||||
|
An Overview of Exact Statistical Inference with StatXact and LogXact
Timberlake Consultants Ltd have recently become the distributors of the Cytel Inc products in the About Exact Statistical Inference Nonparametric methods are indispensable for data analysis today. The primary reason is that they can be applied to a variety of problems without requiring restrictive (and often questionable) assumptions about the distributions that generate the data. Most popular statistical software packages provide a wide range of nonparametric tests: however results such as p-values are almost always approximations calculated by relying on normal theory assumptions! StatXact is the only software package that computes exact p-values, exact confidence intervals and exact power for a comprehensive array of parametric methods. These methods are applicable to both contingency tables and to continuous data. The first half of this seminar will outline the key ideas underlying the theory and computational algorithms for exact nonparametric inference. We will present examples illustrating exact inference for widely used nonparametric methods, such as Pearson’s Chi-square, Wilcoxon-Mann-Whitney, and Kruskal-Wallis tests, as well as more recently developed tests to detect clustering in binary data and for trends in multivariate binary data that will be available in the forthcoming version of StatXact. Logistic regression is the most popular data analytic method for regression modelling of binary response data. The standard method for testing and estimation (and the only method provided by virtually all statistical packages) is the maximum likelihood procedure. This method provides approximate tests and inferences based on large sample theory. These approximations will often be inaccurate for data sets of moderate size and even for large data sets with rare outcomes or imbalanced covariates. In fact there are situations where the standard procedure fails to converge! LogXact is based on award-winning algorithms created at Cytel to implement the exact method based on permutational distributions of sufficient statistics. The theory of this method was developed by Sir David Cox to overcome the limitations of maximum likelihood inference. LogXact provides exact procedures for logistic regression, stratified logistic regression and for other generalized linear models such as polytomous and Poisson regression. We will present examples using real data to illustrate how exact permutational inference overcomes the shortcomings of the maximum likelihood method. We will also describe a procedure proposed by David Firth that uses a penalized likelihood approach to inference. The method was designed to reduce the bias that can occur in maximum likelihood estimates even in large data sets. It also overcomes the problem of non-convergence of the maximum likelihood method for logistic regression. If time permits we will describe a method due to Joseph Ibrahim for missing data in categorical covariates in logistic regression. Both Firth’s and Ibrahim’s methods will be available in the forthcoming version of LogXact. Nitin Patel, Ph.D. - Founder and Co-Chairman, Cytel Inc.. Visiting Professor, MIT Dr. Patel has been a member of the faculty at MIT's
The Seminar The seminar includes lunch and refreshments. Computers will be available during breaks for you to try the software. You are invited to send us your own data, in Excel format, 1 week prior the seminar and you will be able to analyse it then. The number of delegates is restricted. Please register early to guarantee your place. Further instructions will be sent with the joining instructions. If you need assistance in locating hotel accommodation in the area, request the help of our Training Department. Timetable The seminar is FREE of charge. 10H00 – 10H30 Registration and coffee 10H30 – 12H30 StatXact 12H30 – 14H00 Lunch 14H00 – 16H00 LogXact 16H00 – Close and Tea |
||||||||||||||
|
||||||||||||||
