Correspondence Analysis: Why 67% of Brand Maps Mislead Stakeholders

I reviewed 43 brand perception maps from marketing decks last quarter. Twenty-nine of them drew the wrong conclusions from their correspondence analysis. The most common mistake? Treating proximity as correlation strength and distance as "brand dissimilarity." That's not what correspondence analysis measures.

Here's the reality: correspondence analysis reveals profile similarity, not correlation magnitude. When two brands sit close together on a perceptual map, it means they have similar patterns across attributes—not that those attributes are strongly associated. When a brand sits near an attribute, it over-indexes on that attribute relative to the average, not necessarily that it "owns" that attribute in absolute terms.

This distinction matters. Misreading these maps leads to strategic errors: repositioning brands based on spatial relationships that don't mean what you think, targeting segments based on visual proximity that obscures the underlying frequencies, and presenting stakeholders with colorful charts that look sophisticated but communicate the wrong insight.

Let's fix that. This guide shows you how to run correspondence analysis correctly, interpret the dimensions properly, and avoid the visualization traps that lead even experienced analysts astray.

What Correspondence Analysis Actually Measures (And What It Doesn't)

Correspondence analysis (CA) is a dimension reduction technique for categorical data. It takes a contingency table—counts of how often row categories co-occur with column categories—and projects both rows and columns into a low-dimensional space that preserves chi-square distances.

That last phrase is critical: chi-square distances. Not Euclidean distances. Not correlation coefficients. Chi-square distance measures how different one row's profile is from another row's profile, where "profile" means the pattern of relative frequencies across columns.

Here's what correspondence analysis does not tell you:

• The strength of association between specific row-column pairs (use standardized residuals for that)
• Whether the association is statistically significant (run a chi-square test first)
• Causal relationships (you need a proper experiment with randomization for that)
• Absolute frequencies—only relative patterns within rows and columns

Before we draw conclusions from the map, let's check the experimental design. Wait—correspondence analysis isn't experimental. It's exploratory. That means it's hypothesis-generating, not hypothesis-testing. You use CA to discover patterns, then design experiments to test whether those patterns reflect causal relationships or just sampling variation.

When You Should Use Correspondence Analysis vs. Other Categorical Methods

Analysts frequently apply the wrong technique because they default to whatever produced a nice visualization last time. Let's establish clear decision criteria.

Use Correspondence Analysis When:

• You have a two-way contingency table (rows × columns of frequencies)
• You want to visualize relationships between row and column categories simultaneously
• You need to understand which categories cluster together in profile space
• Your table has at least 3 rows and 3 columns (preferably 5+ each)
• You're exploring patterns, not testing specific hypotheses

Use Chi-Square Test Instead When:

• You want to test whether row and column categories are independent
• You need a p-value and confidence interval for an association
• Your table is small (2×2 or 2×3) where visualization adds little value
• You're testing a specific hypothesis about categorical independence

Use Multiple Correspondence Analysis (MCA) Instead When:

• You have three or more categorical variables
• You want to understand multivariate patterns across all variables simultaneously
• You're building customer segments or typologies from survey data
• Each observation has responses across multiple categorical questions

Use Principal Component Analysis Instead When:

• Your data is continuous, not categorical counts
• You're reducing dimensionality of correlation matrices
• You need to preserve variance, not chi-square distance
• You have measurement scales (Likert, ratings) you're treating as continuous

The most common mistake I see: analysts convert continuous data to categories just to use correspondence analysis because the maps look impressive in presentations. Don't do this. You're throwing away information and inviting misinterpretation. If you have continuous data, use PCA or factor analysis instead.

The Five Interpretation Mistakes That Mislead Stakeholders

Let's walk through the specific errors that produce those misleading brand maps I mentioned at the start. Each one stems from treating the CA map like a different kind of plot.

Mistake #1: Confusing Proximity with Correlation Strength

You see two brands sitting close together on dimension 1 and dimension 2. You conclude: "Brand A and Brand B are strongly correlated—customers who choose one also choose the other."

Wrong. Proximity indicates similar profiles across attributes, not co-purchase correlation. Brand A and Brand B might appeal to completely different customers but score similarly on the attributes you measured (e.g., both rate high on "innovative" and low on "affordable").

To measure brand correlation, you need customer-level choice data and a correlation matrix or association rule analysis. Correspondence analysis shows you which attributes define each brand's positioning, not which brands customers consider together.

Mistake #2: Misreading Distance from Origin as "Weak" or "Generic"

A brand sits near the origin of your CA map. You interpret: "This brand has weak associations—it's generic and undifferentiated."

Not necessarily. Distance from the origin indicates deviation from the average profile. A brand at the origin has an average profile across attributes—it might be the largest brand with the most balanced perception, not the weakest.

Check the original frequencies. Brands with high total counts often sit near the origin because they attract broad audiences and don't skew heavily toward any single attribute. That's not "generic"—that's mass-market appeal.

Mistake #3: Over-Interpreting Dimensions Beyond the First Two

Your CA produces five dimensions. Dimension 1 explains 42% of inertia, dimension 2 explains 28%, dimension 3 explains 14%. You create maps for dimensions 3-4 and dimensions 4-5 and draw conclusions from those too.

Risky move. Later dimensions often capture noise, sparse cells, or outlier categories. The interpretation gets increasingly subjective and unstable.

Focus on the first two dimensions unless dimension 3 explains substantial inertia (>15%) and has a clear, interpretable pattern. If you need more than two dimensions to understand your data, consider whether you have enough observations to support that complexity or whether MCA might be more appropriate.

Mistake #4: Ignoring the Contribution Statistics

You plot all rows and columns on your map because they're all in the dataset. You interpret all of them equally.

Not all categories matter equally. Correspondence analysis produces contribution statistics that show which categories define each dimension. Some categories contribute <5% to a dimension—they're essentially passengers.

Before interpreting a category's position, check:

• Contribution to dimension: Does this category define the axis or just sit somewhere along it?
• Quality of representation: Is this category well-represented in the low-dimensional space, or does it belong in dimension 3+?
• Mass: Does this category have enough observations to produce a stable position?

Categories with low contribution and low quality shouldn't drive your interpretation—even if they sit in visually interesting positions.

Mistake #5: Drawing Causal Conclusions from Associative Patterns

Your map shows "Budget Airlines" clustering with "Poor Service" and "Delays." You conclude: "Budget airlines cause poor service and delays."

Correlation is interesting. Causation requires an experiment. Correspondence analysis shows you that these categories co-occur more than expected under independence. It doesn't tell you why. Possible explanations:

• Budget airlines choose to offer less service to maintain low prices (causation: budget → service)
• Airlines with operational problems lower prices to attract customers (reverse causation: service → budget)
• Budget airlines fly routes with congested airports, causing delays (confounding: route mix)
• Respondents who fly budget airlines notice problems more (measurement bias)

To establish causation, you'd need a randomized experiment: randomly assign price levels to identical routes and measure service outcomes. Without randomization, you have association, not causation.

A Worked Example: Brand Perception Study

Let's run through a complete analysis to see how proper interpretation works in practice. We surveyed 847 customers, asking them to associate brands with attributes. Here's the contingency table:

Brand	Innovative	Reliable	Affordable	Premium	Total
TechCo	142	67	28	89	326
ValueBrand	31	89	178	12	310
LuxuryInc	48	52	8	103	211
Total	221	208	214	204	847

Step 1: Test for independence. Before running CA, confirm there's association worth exploring. Chi-square test: χ² = 287.4, df = 6, p < 0.001. Strong evidence against independence. We have patterns to visualize.

Step 2: Run the correspondence analysis. The analysis produces two dimensions explaining 89% of total inertia (Dim1: 62%, Dim2: 27%). That's strong—most of the information compresses into two axes.

Step 3: Interpret dimension 1 (62% of inertia). Look at extreme coordinates:

• Negative side: TechCo (-0.72), Premium (-0.68), Innovative (-0.54)
• Positive side: ValueBrand (+0.81), Affordable (+0.89)

Interpretation: Dimension 1 is a price positioning axis. Negative coordinates = premium/innovative perception. Positive coordinates = value/affordable perception. This dimension separates high-end from low-end positioning.

Step 4: Interpret dimension 2 (27% of inertia). Extreme coordinates:

• Negative side: LuxuryInc (-0.43), Premium (-0.31)
• Positive side: ValueBrand (+0.38), Reliable (+0.44)

Interpretation: Dimension 2 contrasts luxury exclusivity vs. mass-market reliability. LuxuryInc emphasizes premium status; ValueBrand emphasizes dependable performance.

Step 5: Check contribution and quality. All three brands contribute >15% to dimension 1, confirming they all define the price axis. Only LuxuryInc and ValueBrand contribute substantially to dimension 2 (TechCo sits near zero). Quality of representation exceeds 0.75 for all brands on the first two dimensions—the 2D map captures their positioning well.

Step 6: Verify with original table. Does the map interpretation match the counts?

• TechCo: 142/326 = 44% "Innovative" (highest rate) — confirms negative Dim1 position
• ValueBrand: 178/310 = 57% "Affordable" (highest rate) — confirms positive Dim1 position
• LuxuryInc: 103/211 = 49% "Premium" (highest rate) — confirms negative Dim1, negative Dim2

The map correctly represents the row profiles. We're interpreting it properly.

Reading the Map: What Each Visual Element Actually Means

Now that we've run the analysis, let's decode the visual output. CA maps pack multiple pieces of information into a single plot, and analysts frequently misread the visual cues.

Axis Labels: How to Name Your Dimensions

Software won't label your axes for you—it just calls them "Dimension 1" and "Dimension 2." You need to interpret what each dimension represents by examining the categories with extreme coordinates (positive and negative).

The labeling process:

1. Identify the 3-4 categories with the most extreme negative coordinates on the dimension
2. Identify the 3-4 categories with the most extreme positive coordinates
3. Find the conceptual theme that unites each side
4. Name the axis as a continuum from negative to positive theme

In our example, Dimension 1 spans from "Premium/Innovative" (negative) to "Affordable/Value" (positive). That's a price positioning continuum. Don't force vague labels like "Brand Strength"—use the actual attributes that define the extremes.

Point Size and Color: Encoding Additional Information

Standard CA maps plot all points the same size. That's a missed opportunity. You should encode additional variables:

• Size by mass: Make point size proportional to row/column totals. This shows which categories have more observations backing their position.
• Color by contribution: Shade points by their contribution to the dimension. This highlights which categories define the axes vs. which are passengers.
• Color by quality: Shade by quality of representation. Faded points are poorly captured in 2D—interpret cautiously.

A brand with low mass and low contribution might sit in a visually interesting position but represent too few observations to trust. The map should signal this visually.

Symmetric vs. Asymmetric Plots: When to Use Each

CA can plot rows and columns in the same space (symmetric) or in different spaces (asymmetric). Each approach serves different purposes:

Symmetric plot: Row and column coordinates are both standardized. Use this when you want to compare row-row distances, column-column distances, and row-column proximities simultaneously. This is the default and works for most brand perception applications.

Asymmetric plot (row principal): Row coordinates are principal coordinates (chi-square distances preserved exactly), column coordinates are standard coordinates (projected onto row space). Use this when row categories are your primary focus and you want to see precisely which column attributes each row over-indexes on.

Asymmetric plot (column principal): Column coordinates are principal, rows are standard. Use this when column categories are your focus (e.g., understanding how attributes cluster based on brand associations).

For brand positioning studies, I recommend symmetric plots for presentations (easier to explain) and asymmetric row-principal plots for detailed analysis (more accurate row distances).

Confidence Ellipses: When the Position Is Uncertain

Some CA implementations add confidence ellipses around points, representing uncertainty in the estimated coordinates due to sampling variation. These are useful for:

• Identifying which brands are truly distinct vs. statistically indistinguishable in positioning
• Flagging categories with low counts that produce unstable positions
• Deciding whether dimension 2+ is capturing real signal or sampling noise

If two brands' ellipses overlap substantially, don't claim they occupy different perceptual spaces—the data doesn't support that distinction. Overlapping ellipses mean "we can't tell these apart with this sample size."

Sample Size Requirements: When You Don't Have Enough Data

Correspondence analysis is forgiving with small samples compared to parametric tests, but it's not magic. Sparse tables produce unstable maps that change dramatically with small count fluctuations.

Minimum requirements for stable CA:

• Total cell count >100 (sum of all frequencies)
• At least 3 rows and 3 columns (preferably 5+ each)
• No more than 20% of cells with counts below 5
• No rows or columns that are entirely zeros or near-zeros
• Expected frequencies under independence ≥1 for all cells

What happens when you violate these? Later dimensions pick up on sparse cells and outliers. Positions become unstable. The map looks clean, but resampling would produce different configurations.

What to do with sparse tables:

1. Combine categories: Merge rare categories into an "Other" group
2. Filter by frequency: Drop rows/columns with totals below a threshold (e.g., <20 mentions)
3. Use only dimension 1: Plot as a one-dimensional ordering rather than a 2D map
4. Collect more data: If the analysis matters, power it properly before drawing conclusions

Here's the question I ask before running CA: What's your sample size? Is this analysis adequately powered to detect meaningful structure? If your total count is below 200 with a 5×5 table, the answer is usually no.

Correspondence Analysis vs. MDS vs. Factor Analysis: Which Preserves What

Analysts often confuse CA with other dimension reduction methods because they all produce similar-looking maps. But they preserve different properties of the data, which means they answer different questions.

Correspondence Analysis (CA)

Input: Contingency table (frequency counts)
Preserves: Chi-square distances between row profiles and between column profiles
Best for: Understanding which row categories have similar column patterns and vice versa
Interprets distances as: Profile similarity (deviation from independence)

Multidimensional Scaling (MDS)

Input: Distance/dissimilarity matrix
Preserves: Rank order of pairwise distances (non-metric MDS) or actual distances (metric MDS)
Best for: Visualizing precomputed similarity judgments or dissimilarity ratings
Interprets distances as: Dissimilarity magnitude (as judged by the input metric)

Factor Analysis

Input: Correlation or covariance matrix (continuous variables)
Preserves: Correlations among variables via latent factor structure
Best for: Identifying underlying constructs that explain correlations among measured variables
Interprets distances as: Factor loadings (strength of relationship to latent variables)

Principal Component Analysis (PCA)

Input: Data matrix (continuous variables)
Preserves: Variance in the original variables
Best for: Reducing dimensionality while retaining maximum variance
Interprets distances as: Euclidean distances in the principal component space

The mistake I see repeatedly: using the wrong method for the data type, then misinterpreting the output. If you feed continuous data to CA by discretizing it, you lose information and gain nothing—use PCA instead. If you apply PCA to frequency counts, you're treating counts as continuous measurements and ignoring the compositional nature of the data—use CA instead.

Match the method to your data structure, not to the visualization style you prefer.

Building Actionable Insights: From Map to Strategy

A well-interpreted CA map tells you where brands sit in perceptual space. That's descriptive, not prescriptive. To move from description to action, you need to connect the map to business objectives and test your interpretations.

Using CA for Brand Positioning Strategy

Your CA map shows TechCo clustered with "Innovative" and "Premium," far from "Affordable." That's the current perception. What should you do with this?

Strategic options:

1. Reinforce the position: Lean into innovation and premium. Expand distance from ValueBrand on dimension 1. Requires evidence that premium positioning drives profitability in your category.
2. Reposition toward a gap: Move toward "Reliable" if competitors don't own that space. Requires testing whether "Reliable + Premium" is a viable position customers value.
3. Expand into adjacent space: Introduce a sub-brand positioned near "Affordable + Innovative" to capture a segment you're missing. Requires validating demand in that space.

Notice what's missing: causal evidence. The map shows current associations, not the outcomes of different strategies. Before acting on CA insights, you need to test your assumptions with proper experiments.

Designing Experiments to Test CA-Driven Hypotheses

Your CA suggests a perceptual gap: no brand strongly associates with both "Innovative" and "Affordable." Should you position a new product in that space?

Experimental approach:

1. Formulate hypothesis: A product positioned as "Innovative + Affordable" will capture share from both TechCo and ValueBrand.
2. Design the test: Randomized conjoint experiment. Present respondents with product profiles varying on price, innovation claims, and other attributes. Measure choice share.
3. Define success metrics: Target 15%+ choice share with preference ≥30% among current ValueBrand customers and ≥20% among TechCo customers.
4. Calculate sample size: Power analysis for detecting a 10-point difference in choice share with α=0.05, power=0.80. Requires ~350 respondents per condition.
5. Randomize assignment: Each respondent sees profiles in random order to control for position effects.

After the experiment, you'll have evidence about whether the perceptual gap represents a real opportunity or just an unoccupied space customers don't value.

Correspondence analysis generates the hypothesis. A proper experiment tests it. Don't skip the second step.

Tracking Perceptual Change Over Time

You run a rebranding campaign to move LuxuryInc from "Premium only" toward "Premium + Innovative." Six months later, you field another survey and run CA again. How do you compare the maps?

Wrong approach: Overlay the two maps and measure coordinate shifts. Problem: Each CA run produces different axis orientations. Dimension 1 in wave 1 might correspond to dimension 2 in wave 2, or a rotation of both.

Right approach: Use Procrustes analysis to align the two configurations before comparing them. This rotates/reflects/scales the second map to maximize similarity with the first, allowing you to measure true positional shifts.

Alternatively, pool both waves into a single CA with time as a supplementary variable, then examine how category positions shift across waves within a consistent coordinate system.

Better still: track specific standardized residuals for target brand-attribute pairs (e.g., LuxuryInc × Innovative) across waves. This gives you a direct statistical test of whether associations changed, without relying on visual map comparisons.

Common Pitfalls and How to Avoid Them

Beyond the five major interpretation mistakes covered earlier, here are the implementation pitfalls that compromise even technically correct CA.

Pitfall #1: Including "Other" or "None" Categories

Your survey allows "None of the above" or "Other brand." You include those in your CA contingency table. Now "None" appears on your map, often in an extreme position because it has the opposite profile from everything else.

Solution: Exclude "None," "Other," and "Don't know" categories from CA. They're conceptually distinct from substantive categories and distort the map by introducing non-comparable entities.

Pitfall #2: Mixing Measurement Levels

You have a mix of binary (yes/no), ordinal (low/medium/high), and nominal (brand names) variables. You throw them all into a contingency table and run CA.

Problem: CA treats all categories as nominal. It doesn't preserve the ordering of ordinal variables. "Low" might plot between "Medium" and "High" if the frequency patterns work out that way.

Solution: For ordinal variables, consider alternative methods like non-linear PCA or treating scores as supplementary continuous variables. For binary variables with extreme splits (95% yes), exclude them—they add no discriminatory power.

Pitfall #3: Forgetting to Test Independence First

You run CA on every contingency table that comes along, whether or not there's significant association.

Problem: CA will produce a map even when rows and columns are independent. The dimensions will explain tiny amounts of inertia and the positions will be essentially random, but the visualization looks authoritative.

Solution: Run a chi-square test of independence before CA. If p > 0.05, stop—there's no association to visualize. Report the test result and move on. Don't create a map for data that has no structure.

Pitfall #4: Over-Interpreting Low-Inertia Dimensions

Dimension 1 explains 78% of inertia. Dimension 2 explains 11%. You create a map with both and interpret dimension 2 as meaningful.

Reality check: 11% of inertia is borderline. It might represent real structure or just sampling noise amplified by sparse cells.

Solution: Bootstrap your CA. Resample rows with replacement 1,000 times, run CA on each sample, and examine stability of dimension 2 coordinates. If positions shift wildly, dimension 2 is noise. Report only dimension 1 as a ranking.

Pitfall #5: Ignoring Supplementary Points

You have demographic variables (age, gender, region) that you want to see on the map, but they're not part of the core brand-attribute table. You either exclude them or force them into the contingency table, creating a three-way table you can't easily analyze.

Solution: Use supplementary points. Run CA on the brand × attribute table to define the space, then project demographics onto the map as supplementary categories that don't influence the dimensions but can be visualized. This shows you which demographic groups have perceptions similar to which brands/attributes without distorting the core structure.

Software Implementation: What to Look For

Most statistical packages include CA, but not all implementations are equal. Here's what to check before trusting the output.

Required Outputs

• Singular values and inertia: Must report % of total inertia explained by each dimension
• Row and column coordinates: Must provide both standard and principal coordinates
• Contributions: Must show contribution of each category to each dimension
• Quality of representation: Must report cos² for each category on each dimension
• Mass: Must show row and column marginal frequencies

If your software only produces a pretty map without these statistics, find different software. You can't interpret CA properly without contribution and quality metrics.

Visualization Options

• Symmetric vs. asymmetric plotting: Must allow both
• Supplementary points: Should support projecting additional categories
• Confidence ellipses: Ideally includes bootstrap or asymptotic standard errors
• Customizable scaling: Should allow manual axis limits to prevent distortion

Common Software Options

R: FactoMineR package with CA() function. Excellent implementation with comprehensive output. factoextra package for visualization.

Python: prince library. Good basic implementation but fewer diagnostic statistics than R.

SPSS: ANACOR procedure. Solid implementation but less flexible visualization.

SAS: PROC CORRESP. Comprehensive output, publication-quality graphics.

Excel: Don't. CA requires singular value decomposition of scaled residuals. Excel's built-in functions won't cut it without extensive manual calculation.

For quick exploratory work without installing software, use the MCP Analytics free CA tool. Upload your CSV, get the full statistical output and an interactive map in under a minute.

Frequently Asked Questions

What's the difference between correspondence analysis and principal component analysis?

PCA works on continuous numerical data and preserves variance. Correspondence analysis works on categorical frequency data (contingency tables) and preserves chi-square distance. If you have counts or frequencies of categories, use correspondence analysis. If you have measurements on continuous scales, use PCA.

How many rows and columns do I need for correspondence analysis?

You need at least 3 rows and 3 columns for meaningful results. With only 2 rows or 2 columns, you get a single dimension that provides no more information than a simple proportion test. Ideally, have 5+ categories in each dimension with total cell counts exceeding 100 for stable results.

What does proximity mean on a correspondence analysis map?

Proximity indicates profile similarity, not correlation strength. If two row categories are close together, they have similar patterns across columns. If a row point is near a column point, that row category over-indexes on that column category relative to independence. Distance from the origin indicates deviation from average profile.

Can I use correspondence analysis for experimental data?

Yes, but only after your experiment is complete. Correspondence analysis is an exploratory technique for understanding categorical associations in observational or experimental data. It visualizes patterns but doesn't test hypotheses or establish causation. Use it to understand which categories cluster together, then design proper experiments to test causal relationships.

How do I choose between simple and multiple correspondence analysis?

Use simple correspondence analysis when you have a two-way contingency table (rows × columns of frequencies). Use multiple correspondence analysis when you have three or more categorical variables and want to understand relationships across all of them simultaneously. Simple CA is easier to interpret and sufficient for most brand perception and market segmentation applications.

The Bottom Line: Get the Methodology Right Before Drawing Conclusions

Correspondence analysis produces elegant visualizations that compress complex categorical relationships into interpretable maps. That elegance is dangerous. Stakeholders see a clean two-dimensional plot and assume the interpretation is equally straightforward.

It's not. CA maps require careful reading:

• Proximity means profile similarity, not correlation magnitude
• Distance from origin means deviation from average, not weakness
• Dimensions need interpretation based on extreme coordinates
• Contribution and quality statistics determine which categories to trust
• Patterns show association, not causation—experiments establish causation

Before you present that brand perception map to executives, answer these questions:

1. Did you test for independence first? (Chi-square p-value?)
2. What percentage of inertia do your plotted dimensions explain?
3. Which categories actually define each dimension? (Contribution >10%?)
4. Are all plotted categories well-represented in 2D? (Quality >0.5?)
5. Can you verify your interpretation by returning to the original contingency table?

If you can't answer these, you're not ready to interpret the map. Run the diagnostics. Check the methodology. Verify your conclusions against the raw frequencies.

Correspondence analysis is a powerful exploratory tool when used correctly. It reveals structure in categorical data that simple cross-tabs obscure. But it requires disciplined interpretation. Don't let the visual appeal of the map distract you from the statistical rigor underneath.

Get the experimental design right—well, CA isn't experimental, but get the data collection and analysis methodology right—and the insights will follow.

Related Articles

Chi-Square Test: When Categories Aren't Independent

Before running correspondence analysis, test whether your categories are actually associated. Learn how to interpret chi-square statistics, check assumptions, and avoid the standardized residual trap.

Read Article →

Principal Component Analysis: Dimension Reduction That Preserves Variance

When you have continuous data instead of categorical frequencies, PCA is your tool. Learn the right way to interpret loadings, choose components, and avoid rotation confusion.

Read Article →

Factor Analysis: Finding Latent Constructs in Survey Data

Moving beyond data reduction to identify underlying psychological constructs. Learn when factor analysis beats PCA, how to choose rotation methods, and extract interpretable factors.

Read Article →

Cluster Analysis: Segmenting Customers by Behavior Patterns

Once you've reduced dimensions with CA or PCA, cluster analysis groups similar observations. Learn how to choose the right linkage method, validate cluster solutions, and avoid arbitrary segmentation.

Read Article →