A New View of Statistics Go to: Next · Previous · Contents · Search · Home

Summarizing Data:
DIMENSION REDUCTION

Dimension reduction is a way of devising one or two variables to summarize the information contained in a whole lot of other variables. The three methods of dimension reduction are principal components analysis, factor analysis, and cluster analysis

PRINCIPAL COMPONENTS ANALYSIS

When you take lots of different measurements in a study, you sometimes want to combine them in some way to derive just one or two measures that summarize some aspect of the data. For example, you might have five different measures of body size, but would like to have simply one or two summary measure that combine the five. The summary measures are provided by principal components analysis.

All you do is tell the stats program what variables you want it to analyze. It comes up with a linear combination of the variables that somehow captures the biggest amount of common variation in all of them. It then goes on to produce another linear combination that captures the biggest amount of variation in what's left, and so. If you start with three variables, you'll get three principal components. The nice thing about them is, they are not correlated with each other, so they represent three totally independent measures. Exactly what they represent in reality has to be decided by looking at the weighting factors that the stats program derives to make the principal components. Sometimes it's not obvious that they represent anything meaningful, and you might have to abandon this approach.

FACTOR ANALYSIS

Here you want combinations of variables with equal weighting, and you're generally not concerned if the resulting composites are correlated. This method is used by psychologists (or their statisticians) to derive distinct dimensions of the psyche from subsets of items in multi-item questions. Factor analysis divides the items into subsets such that items correlate well within each subset but not so well between subsets. Each subject then gets a mean score for the items in each subset . The researcher has to decide what to call the mean scores by looking at the wording of the items.

It's a few years since I did factor analysis, which is why this section is so short! If there is a demand for it, I will include the detail on such things as promax rotation and deciding where to draw the line for inclusion of variables in a factor.

CLUSTER ANALYSIS

This is a particularly severe form of dimension reduction that reduces all variables and data down to one variable with only a few values (e.g. group A, group B, and group C). It's easiest to understand from the example in the figure, which shows heights and weights for a bunch of people who obviously fall into three groups or "clusters":

You can let the stats program decide on the number of clusters, or you can force it to find as many as you like. The program decides which observations belong to which cluster by minimizing the distances between points in each cluster. You are not restricted to two variables, of course. It's impossible to imagine clusters for more than three variables (unless you are an Einstein), but the stats program handles it without any problem.

Cluster analysis is used in market research, where you want to identify a few major target groups in a population. And it's a cool way of identifying groups in the population with particular lifestyles. Variables used in the cluster analysis might be age, sex, socio-economic status, level of physical activity, measures of diet, and so on.

Go to: Next · Previous · Contents · Search · Home