Latent Class Analysis

Unlike hard clustering methods that force each observation into a single group, LCA assigns posterior membership probabilities for each individual across all latent classes.

• Each customer receives a probability vector (e.g., 0.82 for Class 1, 0.15 for Class 2, 0.03 for Class 3)
• Accounts for classification uncertainty rather than assuming perfect separation
• Enables soft segmentation where customers can exhibit partial membership in multiple behavioral archetypes
• Particularly valuable in CLV contexts where purchasing patterns may blend across segments

LCA operates under the principle that observed indicators are conditionally independent given latent class membership.

• Within each latent class, correlations among manifest variables are explained entirely by class membership
• This assumption is what makes the model identifiable and computationally tractable
• Violations can be diagnosed using bivariate residuals (BVR) between indicator pairs
• In customer segmentation, this means that within a true behavioral segment, purchase frequency and product category preference should be uncorrelated

Determining the optimal number of latent classes requires comparing models across multiple information criteria and likelihood-based tests.

• BIC (Bayesian Information Criterion): Most commonly used; penalizes complexity more heavily than AIC
• AIC (Akaike Information Criterion): Tends to favor more classes; useful for exploratory analysis
• Lo-Mendell-Rubin (LMR) Test: Compares k-class model against k-1 class model with a likelihood ratio test
• Bootstrap Likelihood Ratio Test (BLRT): Computationally intensive but preferred for definitive class enumeration
• Entropy: Measures classification clarity (0 to 1); values above 0.80 indicate good class separation

LCA can incorporate concomitant variables (covariates) to predict class membership without distorting the measurement model.

• The three-step approach prevents covariates from influencing class formation:
  1. Estimate the unconditional latent class model
  2. Assign individuals to classes based on modal posterior probability
  3. Regress class membership on covariates with correction for classification error
• Common CLV covariates: acquisition channel, demographic indicators, initial purchase category
• Avoids the pitfall of covariates driving segmentation rather than the behavioral indicators themselves

Latent Transition Analysis (LTA) extends LCA to repeated measures, modeling how customers move between latent states over time.

• Estimates a transition probability matrix showing the likelihood of moving from one latent class to another between time points
• Identifies common migration paths (e.g., high-value → at-risk → churned)
• Critical for dynamic CLV forecasting where customer segments are not static
• Can incorporate time-invariant and time-varying covariates to predict transitions
• Used to model the progression through customer lifecycle stages with probabilistic uncertainty

LCA flexibly handles mixed measurement scales through different indicator parameterizations within a unified framework.

• Binary indicators: Modeled with logistic regression; parameters represent item endorsement probabilities per class
• Ordinal indicators: Use adjacent-category or cumulative logit models respecting the ordered nature of Likert-scale responses
• Continuous indicators: Assume class-specific normal distributions with estimated means and variances
• Count indicators: Poisson or negative binomial distributions for transaction frequency variables
• Nominal indicators: Multinomial logit parameterization for categorical variables like preferred product category

A foundational Buy-Till-You-Die model that predicts future transactions by modeling purchase frequency with a Poisson process and dropout with a Beta-Geometric distribution. Unlike LCA, which identifies static latent classes, BG/NBD captures continuous heterogeneity in transaction rates and churn propensity across a customer base.

• Assumes customers become inactive after an unobserved dropout event
• Uses Gamma and Beta distributions to model population-level heterogeneity
• Ideal for non-contractual settings like e-commerce

A sub-model that estimates the average monetary value of a customer's transactions, independent of purchase frequency. It assumes transaction values follow a Gamma distribution with individual-level scale parameters that also follow a Gamma distribution.

• Accounts for spend heterogeneity across customers
• Assumes no relationship between frequency and monetary value
• Combined with BG/NBD to produce full CLV estimates

A temporal probabilistic model where a customer's unobserved state transitions over time according to a Markov process, with each state emitting observable purchase behaviors. HMMs provide a dynamic alternative to LCA by allowing customers to transition between latent states.

• States represent behavioral modes (e.g., active, at-risk, dormant)
• Transition probabilities capture state-switching dynamics
• Useful for modeling lifecycle stage progression

A statistical approach that estimates individual customer parameters by borrowing strength from the population distribution. This enables robust CLV predictions even with sparse individual data, addressing a key limitation of LCA when class assignments are uncertain.

• Uses partial pooling to shrink extreme estimates toward the mean
• Quantifies uncertainty through posterior distributions
• Naturally handles new customers with limited history

A validation technique that ranks customers by predicted CLV, divides them into ten equal groups, and compares predicted value against actual realized value for each decile. This is the standard method for evaluating whether LCA-based segmentation translates to actionable value differentiation.

• Reveals calibration quality across the value spectrum
• Identifies overestimation or underestimation in specific segments
• Complements Lift Curves and Gini Coefficient metrics

A statistical framework for analyzing the expected duration until a specific event occurs, such as customer churn. The Cox Proportional Hazards Model extends this by assessing how covariates influence the hazard rate without specifying a baseline distribution.

• Models the hazard function over time
• Handles censored data where churn hasn't yet occurred
• Pairs with LCA to identify class-specific churn trajectories