## 3.2 Conditions of Application

### 3.2.1 Linearity and Symmetry

We have seen the importance of the correlation coefficient (or covariance) in PCA. We can actually present PCA as a visualization technique of a correlation matrix (or covariance matrix). The technique will excel when the correlation coefficient is a *good* measure of the association between variables. The ideal conditions to apply PCA are when the association among variables are linear and their distributions are symmetric (i.e. closer to the normal distribution).

Consequently, we need to be cautious when the distributions are extremely asymmetric or when the associations among variables are not linear.

A common case that can limit the applicability of PCA is when analyzing variables that are seemingly continuous, but that in reality are a hybrid of continuous and nominal scale. For example, this is the case of variables like *payed work time* of women: this is null for a woman that is a housewife, while the distribution is continuous for women that have a payed job.

Nonlinearity associations can also limit the applicability of PCA. This is illustrated with the relation betwen age and income: overall, income tends to increase with age during active working years, but when a person retires the income tend to decrease.

Phenomena of lack of symmetry and lack of linearity will affect the results of a PCA. If these issues are not identified, they can lead to wrong interpretations and conclusions. However, the presence of these phenomena will become apparent for the well trained eyes of an experienced analyst.

We should say that techniques such as Multiple Correspondence Analysis (MCA) can always be used after having encoded (categorized) the continuous variables. Compared to PCA, MCA has the advantage of being inherently non-linear, and thus can be used in situations when PCA is limited.

### 3.2.2 Balancing the content of active variables

More often than not, Principal Component Analysis is performed on variables having different units of measurement. In this case, the variances tend to be vary considerable in magnitude, and are not directly comparable. The typical solution to overcome this issue is to rescale the variables in standard units (i.e. mean of zero, unit variance). In this way, all variables wil be given the same importance, and we don’t have to worry about units of measurement anymore. In fact, this type of transformation has become the default solution in most PCA computer programs: to carry out a normalized PCA and work on the matrix of correlations. Keep in mind that this transformation modifies the shape of the cloud of points by providing the same spread among all directions in the space of origin.

Despite the usefulness of transforming variables into standardized scale, this transformation is not always the ideal solution to *balance* the variables. For example, if there is a subset of variables that are highly correlated among each other, this subset will dominate the first principal component, and therefore, will have a higher importance in the analysis.

Suppose that you have 5 variables that are measuring the same aspect of a certain phenomenon, and that the other aspects are covered each one by just one variable. You can think of the group of 5 variables as being just one variable but with a variance 5 times larger than the rest of the variables. Consequently, the first axis will be determined by the cumulative effect of the 5 highly correlated variables. In summary, we should pay attention to the effect produced by groups of variables that are highly correlated, and have a mechanism to balance the importance of each aspect in the studied phenomenon.