Two-stage and sequential sampling for estimation and testing with prescribed precision

Abstract

Statistical data analysis includes several phases. First, there is the phase of data collection. Second, there is the phase of analysis and inference. The two phases are interconnected. There are two types of data analysis. One type is called parametric and the other type is nonparametric. In the present paper, we discuss parametric inference. In parametric inference, we model the results of a given experiment as realization of random variables having a particular distribution, which is specified by its parameters. A random sample is a sequence of independent and identically distributed (i.i.d.) random variables. Statistics are functions of the data in the sample, which do not involve unknown parameters. A statistical inference is based on statistics of a given sample. We discuss two kinds of parametric inference. Estimating the values of parameters, or testing hypotheses concerning the parameters in either kind of inference, we are concerned with the accuracy and precision of the results. In estimation of parameters, the results are precise if, with high probability, they belong to a specified neighborhoods of the parameters. In testing hypotheses, one has to decide which one of two or several hypotheses should be accepted. Hypotheses which are not accepted are rejected. We distinguish between two types of errors. Type I error is the one committed by rejecting a correct hypothesis. Type II is that of accepting a wrong hypothesis. It is desired that both types of errors will occur simultaneously with small probabilities. Both precision in estimation or small error probabilities in testing depend on the statistics used (estimators or test functions) and on the sample size. In this paper, we present sampling procedures that attain the desired objectives. In Sec. 2, we discuss estimation of the parameters of a binomial distribution. In Sec. 3, more general results about estimation of expected values are presented. In Sec. 4, we discuss the Wald Sequential Probability Ratio Test (SPRT), which has optimal properties for testing two simple hypotheses.