1.1. Introduction
The subject of multivariate analysis deals with the statistical analysis of the data collected on more than one (response) variable. These variables may be correlated with each other, and their statistical dependence is often taken into account when analyzing such data. In fact, this consideration of statistical dependence makes multivariate analysis somewhat different in approach and considerably more complex than the corresponding univariate analysis, when there is only one response variable under consideration.
Response variables under consideration are often described as random variables and since their dependence is one of the things to be accounted for in the analyses, these response variables are often described by their joint probability distribution. This consideration makes the modeling issue relatively manageable and provides a convenient framework for scientific analysis of the data. Multivariate normal distribution is one of the most frequently made distributional assumptions for the analysis of multivariate data. However, if possible, any such consideration should ideally be dictated by the particular context. Also, in many cases, such as when the data are collected on a nominal or ordinal scales, multivariate normality may not be an appropriate or even viable assumption.
In the real world, most data collection schemes or designed experiments will result in multivariate data. A few examples of such situations are given below.
During a survey of households, several measurements on each household are taken. These measurements, being taken on the same household, will be dependent. For example, the education level of the head of the household and the annual income of the family are related.
During a production process, a number of different measurements such as the tensile strength, brittleness, diameter, etc. are taken on the same unit. Collectively such data are viewed as multivariate data.
On a sample of 100 cars, various measurements such as the average gas mileage, number of major repairs, noise level, etc. are taken. Also each car is followed for the first 50,000 miles and these measurements are taken after every 10,000 miles. Measurements taken on the same car at the same mileage and those taken at different mileage are going to be correlated. In fact, these data represent a very complex multivariate analysis problem.
An engineer wishes to set up a control chart to identify the instances when the production process may have gone out of control. Since an out of control process may produce an excessively large number of out of specification items, detection at an early stage is important. In order to do so, she may wish to monitor several process characteristics on the same units. However, since these characteristics are functions of process parameters (conditions), they are likely to be correlated leading to a set of multivariate data. Thus many times, it is appropriate to set up a single (or only a few) multivariate control chart(s) to detect the occurrence of any out of control conditions. On the other hand, if several univariate control charts are separately set up and individually monitored, one may witness too many false alarms, which is clearly an undesirable situation.
A new drug is to be compared with a control for its effectiveness. Two different groups of patients are assigned to each of the two treatments and they are observed weekly for next two months. The periodic measurements on the same patient will exhibit dependence and thus the basic problem is multivariate in nature. Additionally, if the measurements on various possible side-effects of the drugs are also considered, the subsequent analysis will have to be done under several carefully chosen models.
In a designed experiment conducted in a research and development center, various factors are set up at desired levels and a number of response variables are measured for each of these treatment combinations. The problem is to find a combination of the levels of these factors where all the responses are at their 'optimum'. Since a treatment combination which optimizes one response variable may not result in the optimum for the other response variable, one has a problem of conflicting objectives especially when the problem is treated as collection of several univariate optimization problems. Due to dependence among responses, it may be more meaningful to analyze response variables simultaneously.
In many situations, it is more economical to collect a large number of measurements on the same unit but such measurements are made only on a few units. Such a situation is quite common in many remote sensing data collection plans. Obviously, it is practically impossible to collectively interpret hundreds of univariate analyses to come up with some definite conclusions. A better approach may be that of data reduction by using some meaningful approach. One may eliminate some of the variables which are deemed redundant in the presence of others. Better yet, one may eliminate some of the linear combinations of all variables which contain little or no information and then concentrate only on a few important ones. Which linear combinations of the variables should be retained can be decided using certain multivariate methods such as principal component analysis. Such methods are not discussed in this book, however.
Most of the problems stated above require (at least for the convenience of modeling and for performing statistical tests) the assumption of multivariate normality. There are however, several other aspects of multivariate analysis such as factor analysis, cluster analysis, etc. which are largely distribution free in nature. In this volume, we will only consider the problems of the former class, where multivariate normality assumption may be needed. Therefore, in the next few sections, we will briefly review the theory of multivariate normal and other related distributions. This theory is essential for a proper understanding of various multivariate statistical techniques, notation, and nomenclature. The material presented here is meant to be only a refresher and is far from complete. A more complete discussion of this topic can be found in Kshirsagar (1972), Seber (1984) or Rencher (1995).