MathematicalStatistics (1)

“Models of course, are never true but fortunately it is only necessary that they are useful” – Geoge Box

• We can conceptualize the data structure and our goals more precisely, begin this in the simple examples.
• We can derive methods of extracting userful information from data and give methods that assess the generalizability of experimental results.
• We can assess the effectiveness of the methods we proposed.
• We can decide if the models we propose are approximations to the mechanism generating the data adequate for our purposes.
• We can be guided to alternative or more general desciptions that might fit better. Hierarchies of models are discussed throughout.

Examples:

a. A population of N elements, $N\theta$ of which are defetive. To get information about $\theta$, a sample of n is drawn without replacement and inspected.

b. We want to study a physical or economic feature, for example, height or income, is distributed in a large population.

c. An experimenter makes n independent determinations of the value of a physical constant $\mu$.His or her measurements are subject to random fluctuations(error) and the data can be thought of as $\mu$ plus some random errors.

d. We want to compare the efficacy of two ways of doing something under similar conditions such as brewing coffee, reducing pollution,treating a disease,producing energy,learing a maze, and so on.Random variablity here would come primarily from differing responses among patients to the same drug but also from error in measurements and variation in the purity of the drugs.

Sampling Inspection:

The sample space consists of the numbers 0,1,…,n corresponding to the number of defective items found.

$$P[X=k]=\frac{\left(\begin{array}{c}N\theta \\ k\end{array}\right)\left(\begin{array}{c}N-N\theta \\ n-k\end{array}\right)}{\left(\begin{array}{c}N\\ n\end{array}\right)}$$

If $max(n-N(1-\theta ),0)\le k\le min(N\theta ,n)$

Thus X has an hypergeometirc $\mathcal{H}(N\theta ,N,n)$ distribution.

$N\theta$ is unknown, we cannot specify the probability completely, only give a family { H ( N θ , N , n ) } \{\mathcal{H}(N\theta,N,n)\} {H(Nθ,N,n)} of probability distributions. Any one of which could have generated the data actually observed.

Sample from a Population, One-Sample Models:

If the measurements are scalar, we observe $x_1,\cdots ,x_n$ , which are modeled as realizations of $X_1,\cdots ,X_n$ independent,identically distributed (i.i.d.) random variables with common unknown distribution function F.

$$X\sim F$$

The model is fully described by the set $\mathcal{F}$ of distributions that we specify.

$$X_i=\mu +{ϵ}_i,1\le i\le n$$

We postulate:

• The value of the error committed on one determination does not affect hte value of the error at other times. That is ${ϵ}_1,\cdots ,{ϵ}_n$ are independent.
• The distribution of the error at one determination is the same as that at another.Thus ${ϵ}_1,\cdots ,{ϵ}_n$ are identically distributed.
• The distribution of $ϵ$ is independent of $\mu$
• The common distribution of the errors is $\mathcal{N}(0,{\sigma }^2)$.

Two-Sample Models:

Drug A is a standard or placebo,we refer to the x’s as control observations. y’s denote the responses of subjects given a new drug or treatment that is being evaluated by comparing its effect with that of the placebo, that we call treatment observations.

Natural initial assumptions here are:

(1). The x’s and y’s are realizations of $X_1,\cdots ,X_m$ a sample from F, and $Y_1,\cdots ,Y_n$ a sample from G.The model is specified by the set of possible (F,G) pairs.

(2) Suppose that treatment A had been administered to a subject response x would have been obtained. Then if treatment B had been administered to the same subject instead of A, response $y=x+\Delta$ would be obtained where $\Delta$ does not depend on x.

(3) The control responses are normally distributed.