使用mclust包Mclust函数时出现方差/σ无穷值异常结果求助

Hey there! I’ve run into this exact problem before when using Gaussian-based mixed models like Mclust on integer rating data (like your 1-7 scale columns). Let’s walk through why this is happening and how to fix it:

Why You’re Seeing Infinite Variance #

The diabetes vignette uses continuous physiological data (e.g., blood glucose levels) where Gaussian assumptions hold well. But your data is discrete, ordered rating data—and worse, some rows/columns have identical values (like row 3 in your sample, where every entry is 7). When Mclust tries to fit a Gaussian model to a subset of data where a variable has zero variance (all values the same), it can’t compute a valid covariance matrix (since division by zero leads to infinite values).

Step-by-Step Fixes #

1. First: Hunt for Zero-Variance Variables or Uniform Subgroups #

Start by checking which columns have no variation, or if any potential clusters would have uniform values:

# Check variance of each column
sapply(example, var)
# Check for rows with all identical values (like row 3 in your sample)
row_all_same <- apply(example, 1, function(x) length(unique(x)) == 1)
table(row_all_same)

• If you find columns with zero variance, drop them—they add no information and break the model.
• If you have rows with all identical values, consider removing them (if they’re outliers) or grouping them manually before fitting the model.

2. Switch to a Model Built for Discrete Data #

Mclust is designed for continuous data. For your ordered rating data, Latent Class Analysis (LCA) is a better fit. The poLCA package is made exactly for this:

library(poLCA)
# Convert all columns to factors (required for poLCA)
example_factors <- as.data.frame(lapply(example, as.factor))
# Define the LCA model (all variables as indicators of latent classes)
lca_formula <- cbind(HI01_01, HI01_02, HI01_03, HI01_07, HI01_08, HI01_09, HI01_10, HI01_12) ~ 1
# Fit a 3-class model (adjust nclass to test different numbers of classes)
lca_model <- poLCA(lca_formula, data = example_factors, nclass = 3)
# View results
summary(lca_model)

3. If You Must Use Mclust: Constrain the Variance Structure #

If you need to stick with mclust, avoid the default full covariance matrix ("VVV"), which is too flexible for discrete data. Instead, use a constrained variance structure:

library(mclust)
# Try spherical variance (all variables have same variance)
model_eii <- Mclust(example, modelNames = "EII", G = 2:5)
# Or diagonal variance (each variable has its own variance, no covariance)
model_vii <- Mclust(example, modelNames = "VII", G = 2:5)
# Check which model fits best
summary(model_eii)
summary(model_vii)

Constrained models require fewer parameters, so they’re more stable with discrete or small datasets.

4. Preprocess Your Data #

Standardizing your rating data can help stabilize parameter estimates:

# Standardize each column to mean 0, variance 1
example_scaled <- scale(example)
# Fit Mclust on scaled data
model_scaled <- Mclust(example_scaled, G = 2:5)

Just remember to interpret results relative to the scaled values, or reverse-transform if needed.

Quick Check: Sample Size per Cluster #

If your model is trying to fit a cluster with only 1-2 observations, variance estimates will be unstable. Use model$classification to check cluster sizes after fitting, and adjust the number of classes (G) if needed.

内容的提问来源于stack exchange，提问作者Saja01