06.06.2013 Views

Theory of Statistics - George Mason University

Theory of Statistics - George Mason University

Theory of Statistics - George Mason University

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

3.1 Inferential Information in <strong>Statistics</strong> 211<br />

An additive model has the advantage <strong>of</strong> separability <strong>of</strong> first order expectations<br />

<strong>of</strong> the two components no matter what assumptions are made about<br />

joint probability distributions <strong>of</strong> elements <strong>of</strong> the one component and those <strong>of</strong><br />

the other. Note a questionable requirement for this separability: the variance<br />

<strong>of</strong> the residual component must be constant no matter what the magnitude<br />

<strong>of</strong> the expectation <strong>of</strong> the systematic component. Despite these issues, in the<br />

following, we will concentrate on models with additive random effects.<br />

In the case <strong>of</strong> the black-box model (3.8), both the systematic and random<br />

components are embedded in the box. The objectives <strong>of</strong> statistical analysis<br />

may be to identify the individual components or, more <strong>of</strong>ten, to determine<br />

“average” or “most likely” output Y for given input x.<br />

Because the functional form f <strong>of</strong> the relationship between Y and x may<br />

contain a parameter, we may write the equation in the model as<br />

Y = f(x; θ) + E, (3.12)<br />

where θ is a parameter whose value determines a specific relationship within<br />

the family specified by f. In most cases, θ is a vector. In the usual linear regression<br />

model, for example, the parameter is a vector with two more elements<br />

than the number <strong>of</strong> elements in x,<br />

where θ = (β0, β, σ 2 ).<br />

3.1.1 Statistical Inference: Point Estimation<br />

Y = β0 + x T β + E, (3.13)<br />

Statistical inference is a process <strong>of</strong> identifying a family <strong>of</strong> distributions that<br />

generated a given set <strong>of</strong> observations. The process begins with an assumed<br />

family <strong>of</strong> distributions P. This family may be very large; for example, it may<br />

be the family <strong>of</strong> distributions with probability measures dominated by the<br />

counting measure. Often the assumed family is narrowly defined; for example,<br />

it may be the family <strong>of</strong> univariate normal distributions. In any event, the<br />

objective <strong>of</strong> statistical inference is to identify a subfamily, PH ⊆ P, that<br />

contains the population from which the data arose.<br />

Types <strong>of</strong> Statistical Inference<br />

There are various types <strong>of</strong> inference related to the problem <strong>of</strong> determining<br />

the specific Pθ ∈ P or else some specific element <strong>of</strong> θ. Some <strong>of</strong> these are<br />

point estimation, hypothesis tests, confidence sets, and function estimation<br />

(estimation <strong>of</strong> the PDF, for example). Hypothesis tests and confidence sets<br />

are associated with probability statements that depend on Pθ.<br />

Beginning on page 12 and later in Chapter 2, we have distinguished families<br />

<strong>of</strong> probability distributions as either parametric or nonparametric. In<br />

<strong>Theory</strong> <strong>of</strong> <strong>Statistics</strong> c○2000–2013 James E. Gentle

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!