Chapter 5: A Theoretical Overview of GDA Methods
5.4 The CA Biplot
5.4.1 CA Biplot Variants
The various types of CA biplots are discussed in great detail in Chapter 7 of Understanding Biplots (Gower et al., 2011:289). Provided below is an explanation of the several variants of the independence model that can be approximated in the CA biplot.
5.4.1.1 Pearsonβs chi-squared:
By making use of the weighting the independence model, the contributions to the Pearsonβs chi- squared statistic can be approximated. Thus following from equation 5-17 the model which must be minimised is:
βπΉβΒ½(πΏ βπΉππβ²πͺ π ) πͺ
βΒ½β πΏΜβ2, (5-26)
where πΏΜ is the π-dimensional approximation of πΉβΒ½(πΏ βπΉππβ²πͺ π ) πͺ
βΒ½ (Gower et al., 2011:291). Note that the SVD of πΏΜ, the π-dimensional approximation of πΏ, is written as πΌπ±πΊπ½β².The π-dimensional approximation of the π2 contributions requires plotting of the first π columns of πΌπΊΒ½ and π½πΊΒ½. Such biplots provide a display from which it is possible to identify the elements of πΏΜ which contribute most to the total inertia as well as which element diverge from the assumption of independence.
5.4.1.2 Approximating the deviations from independence:
This method seeks to approximate the deviations from independence, πΏ β π¬, which are weighted by the inverse square roots of the row and column totals. This translates to the minimization of:
βπΉβΒ½{(πΏ βπΉππβ²πͺ
π ) β πΏΜ} πͺβΒ½β
2
(5-27)
(Gower et al., 2011:291). Equation 5-27 is minimized by πΉβΒ½πΏΜπͺβΒ½= πΌπΊπ±π½β². Solving for the π- dimensional approximation of πΏ gives πΏΜ = πΉΒ½πΌπΊπ±π½β²πͺΒ½. By plotting the first π columns of πΉΒ½πΌπΊΒ½ and πͺΒ½ππΊΒ½ a biplot which represents the deviations away from independence is achieved.
5.4.1.3 Approximation to the contingency ratio:
The contingency ratio, as defined in Correspondence Analysis in Practice (Greenacre, 2007:101) is the ratio of observed proportions of a contingency table to the expected proportions. A CA biplot which approximates the contingency ratios is defined as the minimization of:
βπΉβΒ½{πΉβ1(πΏ βπΉππβ²πͺ
π ) πͺβ1β πΏΜ} πͺβΒ½β
2 (5-28)
(Gower et al., 2011:292). In equation 5-28 the matrix πΏΜ approximates πΉβ1(πΏ βπΉππβ²πͺ π ) πͺ
β1= πΉβΒ½πΌπΊπ½β²πͺβΒ½. Note that a contingency ratio of unity is desirable, and the closer the value is to unity the closer the data are to independence (Greenacre, 2007:101). By plotting the first π columns of πΉβΒ½πΌπΊΒ½ and πͺβΒ½π½πΊΒ½ a biplot which depicts the deviation from the unity contingency ratio is obtained. Note that the biplot does not provide a direct deviation from unity but rather the deviation divided by π. This characteristic of the biplot has no material influence on the biplot but must be accounted for when calibrating the axes.
5.4.1.4 Approximation to the Chi-squared distance:
The independence assumption is evaluated by the π2 statistic. For any two-way table, the assumption can be tested via the π2 distance between rows or the π2 distance between the columns of the table. Therefore, there are two possible solutions for when approximating the Chi-squared distances in a contingency table.
The chi-squared distances between the πth and πβ²th rows of πΏ are defined as: πππ2β² = ( ππ π₯π.β ππβ² π₯πβ².) β² πͺβ1(ππ π₯π.β ππβ² π₯πβ².) (5-29)
(Gower et al., 2011:293). Equation 5-29 is of the form of weighted Euclidean distances between the rows of πΏ. The chi-squared distances are then calculated as the normal Euclidean distances between the rows of πΉβ1πΏπͺβΒ½. This can also be expressed as πΉβ1πΏπͺβΒ½βππβ²πͺΒ½
π , where ππβ²πͺΒ½
π is the translation term. The translation term has no material effect on the chi-squared measures between rows of πΏ but translates the points about the centroid in the biplot. The translated chi-squared distances can then be written as:
πΉβ1(πΏ βπΉππβ²πͺ
π ) πͺβΒ½= πΉβ1(πΏ β π¬)πͺβΒ½
(5-30)
(Gower et al., 2011:294). The biplot of the chi-squared distances between the rows of matrix πΏ is then based upon the minimization of:
βπΉΒ½{πΉβ1(πΏ βπΉππβ²πͺ π ) πͺ
βΒ½β πΏΜ}β2, (5-31)
βπΌπΊπ½β²β πΉΒ½πΏΜβ2. (5-32) The approximation, πΏΜ, is then obtained from the inner product of πΉβΒ½πΌπΊπ½β² and the π-dimensional biplot is obtained by plotting the first π columns of πΉβΒ½πΌπΊ for the row points and π½ for the column points (Gower et al., 2011:294). Note that π½ is an orthogonal matrix and has no effect on the distances between the row coordinates.
Similar arguments follow for the chi-squared distances between the columns of πΏ. The chi-squared distances between the πth and πβ²th columns of πΏ are defined as:
πππβ²2 = (ππ π₯.πβ ππβ² π₯.πβ²) β² πΉβ1(ππ π₯.πβ ππβ² π₯.πβ²). (5-33)
This leads to a biplot which attempts to minimise: β{πΉβΒ½(πΏ βπΉππβ²πͺ
π ) πͺβ1β πΏΜ} πͺΒ½β 2
, (5-34)
which simplifies to:
βπΌπΊπ½β²β πΏΜπͺΒ½β2. (5-35)
The approximation, πΏΜ, is then obtained from the inner product of πΌπΊπ½β²πͺβΒ½ and the π-dimensional biplot is created by plotting the first π columns of πͺβΒ½π½πΊ for the row points and πΌ for the column points. Similarly, πΌ is also an orthogonal matrix which has no effect on the distances between the points representing the columns of πΏΜ.
It is common practice to plot both the row and column chi-squared distances in one single plot, although the relationships between these two measures has no meaningful interpretation. This plot is simply achieved by plotting πΉβΒ½πΌπΊ and πͺβΒ½π½πΊ simultaneously as two sets of points.
5.4.1.5 Canonical Correlation Approximation:
Canonical correlation is associated with the optimal CA solution. The optimal solution can be viewed as a function of the correlation between the rows and columns of the two-way table. When the CA solution is aimed to maximize the correlation, then the name of the corresponding correlation for the optimal solution is known as canonical correlation. In this variant of the CA biplot there will be a search for the optimal row and column scores such that the correlation between rows and columns is maximised. Due to its lengthy derivation, the explanation of said method is not provided for. Rather, the reader is referred to Chapter 7 of Understanding Biplots for a detailed explanation (Gower et al., 2011:296).
5.4.1.6 Approximating the row profiles:
The final variant of the CA biplot is centred upon the row profiles, πΉβ1πΏ, of the two-way table. Similarly, the approximation of the column profiles can be achieved by simply replacing the input matrix with the transpose of πΏ. In this case the suggested biplot model is:
βπΉΒ½{πΉβ1(πΏ βπΉππβ²πͺ
π ) β πΏΜ} πͺ
βΒ½β2, (5-36)
and the approximation πΏΜ takes the form of πΉβΒ½πΌπΊπ½β²πͺΒ½ (Gower et al., 2011:298). The accompanying biplot is obtained by plotting the first π columns of πΉβΒ½πΌπΊ for the rows and πͺΒ½π½ for the columns. The distances expressed in this biplot are chi-squared distances, and the points for the columns provide axes for the row profile points. Here the row profiles are not the pure row profiles but rather the deviations from the marginal row profile, π
β²πͺ π .