MULTIPLE CORRESPONDENCE ANALYSIS
Correspondence analysis and multiple correspondence analysis are techniques from multivariate analysis. You can use the techniques to find clusters in a data set. Correspondence and multiple correspondence analysis are similar to principal component analysis, in that the analysis attempts to reduce the dimensions (number of columns or rows) of a set of intercorrelated variables so that the smaller dimensioned (number of columns or rows) variables explain most of the variation in the original variables. However, correspondence and multiple correspondence analysis are for categorical variables rather than the numerical variables of principal component analysis. Correspondence analysis was developed by Jean Paul Benzecri and measures similarities of patterns in contingency tables.
The Mathematics Behind Correspondence Analysis
Correspondence analysis is used in the analysis of just two categorical variables. In correspondence analysis, the reduced variables are found by applying singular value decomposition to a transformation of the contingency table created from the two original variables. The transformation replaces the value in each cell of the contingency table by the original value minus the product of the row total and the column total divided by the overall total, with the difference divided by the square root of the product of the row total and the column total. The resulting cells then contain the signed square roots of the terms used in the calculation of the chi square test for independence for a contingency table, divided by the square root of the overall total.
The Mathematics Behind Multiple Correspondence Analysis
For multiple correspondence analysis, more than two categorical variables are reduced. The reduced set of variables is found by applying correspondence analysis to one of two matrices. The first matrix is a matrix made up of observations in the rows and indicator variables in the columns, where the indicator variables take on the value one if the observation has a quality measured by the variable and zero if the individual does not. For example, say there are three variables, ‘gender’, ‘hair color’, and ‘skin tone’. Say that the categories for gender are ‘female’, ‘male’, ‘prefer not to answer’; for hair color, ‘red’, ‘blond’, ‘brown’, ‘black’; and for skin tone, ‘light’, ‘medium’, and ‘dark’, then, there would be three columns associated with gender and, for a given person, only one would contain a one, the others would contain zeros; there would be four columns associated with hair color and, for a given person, only one would contain a one, others would contain zeros; and there would be three columns associated with skin tone and, for a given person, only one would contain a one, the others would contain zeros.
The second type of matrix is a Burt table. A Burt table is multiple sort of contingency table. The contingency tables between the variables make up blocks of the matrix. From the example above, the first block is the contingency table of gender by gender, and is made of up of a diagonal matrix with the counts of males, females, and those who did not want to answer on the diagonal. The second block, going horizontally, is the contingency table of gender by hair color. The third block, going horizontally, is the contingency table of gender by skin tone. The second block, going vertically, is the contingency table of hair color by gender. The rest of the blocks are found similarly.
Plotting for Clustering
Once the singular value decomposition is done and the reduced variables are found, the variables are usually plotted to look for clustering of attributes. (For the above example, some of the attributes are brown hair, male, red hair, light skin tone, each of which would be one point on the plot.) Usually just the first two dimensions of the reduced matrix are plotted, though more dimensions can be plotted. The dimensions are ordered with respect to how much of the variation in the input matrix the dimension explains, so the first two dimensions are the dimensions that explain the most variation. With correspondence analysis, the reduced dimensions are with respect to the contingency table. Both the attributes of the rows and the attributes of the columns are plotted on the same plot. For multiple correspondence analysis, the reduced dimensions are with respect to the matrix used in the calculation. If one uses the indicator variable matrix, one can plot just the attributes for the columns or one can also plot labels for the rows on the same plot (or plot just the labels for the rows). If one uses the Burt table, one can only plot attributes for the columns. For multiple correspondence analysis, the plots for the columns are the same by either method, thought the scaling may be different.
Interpreting the Plot
The interpretation of the row and column variables in correspondence anaylsis is done separately. Row variables are compared as to level down columns and column variables are compared as to level across rows. However if a row variable is near a column variable in the plot, then both are represented at similar relative levels in the contingency table.
The interpretation of the relationship between the variables in the indicator or Burt table is a bit subtle. Within an attribute, the levels of the attribute are compared to each other on the plot. Between attributes, points that are close together are seen at similar levels.
An Example Using the Deficit by Political Party Data Set
Below is a multiple correspondence plot for which only the column reduction is plotted. We look at the size of the deficit (-) / surplus (+) using the political affiliation of the President and the controlling parties of the Senate and the House of Representatives. The size of the onbudget budget deficit (-) / surplus(+) as a percentage of gross domestic product for the years 1947 to 2008 was classified into four classes. The largest deficit over the years was -6.04 percent and the largest surplus was 4.11 percent. The class breaks were -6.2, -4, -2, 0, and 4.2, which gives four classes. (There was only one observation with a surplus greater than 2, so years of surplus were all classed together.) Each of the President, Senate, and House variables were classified into two classes, either Democrat or Republican. Since the budget is created in the year before the budget is in effect, the deficit (-) / surplus (+) percentages are associated with the political parties in power the year before the end of the budget year.
The first principal axis appears to measure distance between the relative counts for the parties of the senate, and house. On the plot, the greatest distances on the first principal axis are between the Republican and Democratic senates and the Republican and Democratic houses. Looking at the tables below, the lowest counts were for the Republican senates and houses, which means the highest counts were for the Democratic senates and houses. For the deficit classes, classes (0,4.2] is the smallest class in the table and, like the Republican senates and houses, is to the left on the plot. Still there is not much difference between the deficit classes on the first principal axis (or the parties of the presidents).
The second principal axis appears to show how the differing levels of deficit (or surplus) are associated with the political parties of the presidents, senates, and houses. Looking at the senates and houses, there is not a lot of difference between the Democrats and the Republicans on the second principal axis, both cluster around the deficit class (-4,-2]. For the presidents, however, there is a great difference. Democratic presidents are near the deficit classes (-2,0] and (0,4.2] on the second principal axis while Republican presidents are between the deficit classes (-4,-2] and (-6.2,-4].
Below are two classifications of the data.
In this example, we have treated the deficit, which is a numeric variable, as a categorical variable. In the post before this post, an example of principal component analysis, the same data is analyzed, with all of the variables treated as numeric rather than categorical.