MerQuR+Statistics Glossary

Statistics Glossary → Correspondence Analysis (CA / MCA)

Correspondence Analysis (CA / MCA)analysis

A method that visually summarises the association in a crosstab of two (or more) categorical variables by placing the row and column categories on a low-dimensional map according to their similarity.

Concept

In a large crosstab (e.g. 6 occupational groups × 5 holiday types), the chi-square test only says 'there is an association'; it is hard to see from the numbers which categories resemble each other or tend to occur together. Correspondence analysis turns the table into a map: rows with similar profiles lie close together, and rows that often occur with a particular column are drawn towards it.

The distance between each row category's profile (its row percentages) and the average profile of all the data is measured with the chi-square distance, and likewise for the columns. The total variation in the table is called the inertia and equals χ²/n. The analysis finds the axes that explain the most inertia; the first two axes usually carry most of it, and the two-dimensional map is drawn from them. For each point it computes the mass (the category's relative frequency), the contribution (its share in determining an axis) and cos² (how well the point is represented on the map). For three or more variables, multiple correspondence analysis (MCA) is used.

Correspondence analysis is descriptive and exploratory; it is not a hypothesis test in itself. Distances on the map are read differently depending on the normalisation chosen: distances between row points show how similar the row profiles are, but the distance between a row point and a column point cannot be interpreted directly under every normalisation. Points with low quality (small cos²) may sit in misleading positions. Rare categories can have too much influence on the map.

When to use it

  • To see on a map the structure of the association between two variables with many categories, and which categories are similar or linked (market research, ecology, survey analysis).
  • If the only question is whether an association exists, use the chi-square test of independence; if the association needs to be modelled and hypotheses tested, a log-linear model is more suitable.
  • To study three or more categorical variables together use multiple correspondence analysis (MCA); for continuous variables use principal component analysis.

An everyday example

A travel agency tabulates the occupational group (student, teacher, health worker, shopkeeper, retiree) and preferred holiday type (beach, hiking, cultural tour, city break) of 800 customers. On the correspondence map students may sit near city breaks, retirees near cultural tours and teachers near hiking, while shopkeepers appear together with beach holidays. If the first axis explains, say, 70% of the total inertia, the agency sees that its customers are largely separated along a single 'active-restful holiday' axis and can target its campaigns accordingly.

Parameters in MerQur

Row category and Column category

Concept: The row and column variables of the crosstab to be mapped. Because correspondence analysis treats the two variables symmetrically, swapping them does not change the solution; only which categories are shown as 'row' and which as 'column' points changes.

Everyday example: In the agency example the row variable is occupational group and the column variable holiday type. Swapping them shows the same structure on the map; only the kind of label on the points changes.

What it does: Selects the row and column variables of the contingency table.

Effect on the result: Swapping the direction does not change the eigenvalues, χ² and point statistics; the row and column roles are exchanged.

On the validation data: smoking x rank and rank x smoking give the same eigenvalues and χ².

Number of dimensions

Concept: Sets how many axes the map is summarised with. Each axis explains part of the total inertia; usually the first two are enough, but if they explain only a small share, more dimensions should be examined. The maximum number of dimensions is one less than the smaller of the numbers of row and column categories.

Everyday example: A table of 5 occupations × 4 holiday types has at most 3 dimensions. If the first two axes explain 92% of the inertia, the two-dimensional map carries nearly all the information in the table.

What it does: Sets in how many dimensions the coordinates and point diagnostics (quality, contribution, cos²) are given; usually 2, increased until the cumulative inertia reaches 80-90%.

Effect on the result: Adds a coordinate column per dimension; quality increases because it is the sum of cos² over the displayed dimensions. The eigenvalue list does not change. The upper limit is min(I, J) - 1 (3 in a 5 x 4 table); a 2 x 3 table gives a single dimension, a one-axis plot and a note line.

On the validation data: In smoke, from 2 -> 3 dimensions the quality of Senior_mngr goes 0.8926 -> 1.0000 (saturated solution).

→ Inertia, singular values and profiles

Normalization (Principal (rows and columns), Symmetrical (SPSS, √σ), Row principal, Column principal)

Concept: Sets how row and column points are scaled on the map. Under principal normalisation, distances among row points and among column points directly show profile similarity; row- or column-principal normalisation is better suited to interpreting the position of one side's categories relative to the other side's (a biplot). Inertia and contributions are unaffected by this choice.

Everyday example: If the agency asks 'which occupational groups are similar in their holiday preferences?', principal normalisation is appropriate. If it asks 'which holiday type do teachers lean towards?', a row-principal biplot gives a sounder reading.

What it does: Sets the scale of the map coordinates: principal (both sides with σ), symmetrical (both sides with √σ), row principal, column principal. For the row-column relationship (biplot) symmetrical or row/column principal, for category similarity principal (Greenacre, 2007).

Effect on the result: Only the "Coordinates" section, the biplot and the title change; inertia, χ², mass, contribution and cos² are independent of the normalization.

On the validation data: Senior_empl and None Dim1: principal 0.3806 / 0.3933; symmetrical 0.7279 / 0.7522; row principal 0.3806 / 1.4385; column principal 1.3920 / 0.3933 (with the sign convention).

→ Inertia, singular values and profiles

Supplementary row categories and Supplementary column categories

Concept: Categories that take no part in computing the solution but whose profiles are placed on the map afterwards. They allow a comparison group or an external distribution to be shown in the same space without altering the map's axes.

Everyday example: The agency could add last year's customer profile, or the distribution of customers who booked holidays abroad, as a supplementary row. Where that point falls on the map shows which holiday types the group is close to relative to this year's main customer base.

What it does: The named categories do not determine the solution; their profiles are projected onto the map afterwards. It positions an external comparison distribution (national proportion, subgroup) without disturbing the solution.

Effect on the result: Observations in that category are removed from the active table; the supplementary points are scaled with the selected normalization and shown in a separate section and on the plot. A warning is written for a name not in the data.

On the validation data: National (supplementary row) principal (0.2584, 0.1176), Nondrink (0.1148, 0.3620); "None" as a supplementary column (-0.2916, -0.1873). In symmetrical normalization National (0.4941, 0.3719).

→ Supplementary (passive) points

Profiles and singular values

Concept: Row and column profiles are each category's own percentage distribution, and the map is built on comparing these profiles. Singular values and eigenvalues show how much inertia each axis carries, and the χ² share shows each axis's contribution to the total association.

Everyday example: The retirees' row profile might be '10% beach, 15% hiking, 55% cultural tour, 20% city break'; since they depart from the average profile most on cultural tours, they are drawn towards cultural tours on the map.

What it does: Row profiles (row percentages), column profiles; per dimension the singular value, eigenvalue and χ² share (n·σ²).

Effect on the result: Additional tables.

On the validation data: Senior_mngr row profile None 0.3636, Light 0.1818, Medium 0.2727, Heavy 0.1818; dimension χ² 14.429 / 1.933 / 0.080.

→ Inertia, singular values and profiles

MCA variables (2+)

Concept: Runs multiple correspondence analysis, which examines the shared structure of three or more categorical variables on a single map. Because raw inertia percentages understate the importance of the axes in MCA, Benzécri- or Greenacre-adjusted percentages are preferred for interpretation.

Everyday example: If the agency adds age group and accommodation preference (hotel, guesthouse, camping) to occupation and holiday type, MCA puts the categories of all four variables on the same map; a cluster such as 'young, student, camping, hiking' might gather in one corner.

What it does: Performs multiple correspondence analysis with the checked categorical variables (CA of the indicator matrix); it explores the joint structure of three or more categorical variables. Benzécri- or Greenacre-adjusted inertias are preferred in interpretation (Greenacre, 2007, chapter 19).

Effect on the result: A separate MCA section: inertia (λ), %λ/Σλ, Benzécri-adjusted inertia, Benzécri and Greenacre percentages, category principal coordinates. Missing values are deleted listwise; a single checked variable gives a warning.

On the validation data: ISSP A-D (n = 871): λ₁ = 0.45738; Benzécri 0.07646; Greenacre percentage 44.91 and 34.20; Benzécri percentage 50.73%. Cars 7 variables n = 334, American (-0.4035, 0.8129).

Example on real data

Data: Greenacre (1984) smoke data; Stata ca, R ca

Variables: row category rank (staff group), column category smoking (smoking level)

  • Question: How does the structure of the association between staff group (5 levels) and smoking level (4 levels) look on a two-dimensional map? (N = 193)
  • Independence: Pearson χ²(12) = 16.44, p = .172; the association is not statistically significant, so the map should be interpreted descriptively.
  • Inertia: Total inertia = χ²/N = 0.0852. The first dimension explains 87.76% of the inertia (λ1 = 0.0748), the second 11.76%; together the two dimensions 99.51%.
  • First dimension: Non-smokers (None, +0.393; contribution 0.654) and senior employees (Senior_empl, +0.381; contribution 0.512) are at one end, heavy and medium smokers with junior employees and junior managers at the other: the dimension is a "non-smoking - smoking" contrast.
  • Second dimension: Junior managers (-0.243) and senior managers (-0.194) are close to heavy smokers (-0.198); this dimension is shaped mainly by the manager groups.
  • Quality: All points are well represented in two dimensions (0.89 to 1.00).
  • APA: In the correspondence analysis the first two dimensions explained 99.5% of the total inertia (λ1 = 0.075, λ2 = 0.010); χ²(12, N = 193) = 16.44, p = .172.

Related concepts

Inertia, singular values and profilesSupplementary (passive) pointsEigenvalue and eigenvectorBiplotListwise and pairwise deletion

The numbers in the examples are MerQur's output on independent, publicly available validation data; see the MerQur Technical Note for details.