What is the difference between FIMIX-PLS and multi-group analysis?

Multi-group analysis compares groups you define in advance, such as country or gender. FIMIX-PLS discovers segments from the data when you do not know the grouping, which is why it addresses unobserved heterogeneity.

How many segments should I extract?

Run several class counts and compare information criteria (AIC, BIC, CAIC, MDL5) and normalized entropy, alongside segment size and interpretability. No single criterion decides it on its own.

What do I do after FIMIX finds segments?

Find an observable variable that explains the segment split, describe the segments, and ideally confirm the differences with a multi-group analysis on that variable.

PLS-SEM METHOD

FIMIX-PLS: finding hidden segments in your data

FIMIX-PLS uncovers unobserved heterogeneity: subgroups in your sample for which the structural relationships differ, even though you did not define the groups in advance. This page explains what it does, how to read the fit criteria and segment solutions, a worked example, and the common pitfalls.

What FIMIX-PLS does

A single PLS-SEM model assumes one set of path coefficients holds for everyone. That assumption can hide unobserved heterogeneity: distinct segments for whom the relationships differ. Unlike multi-group analysis, where you supply the grouping variable, FIMIX-PLS (Hahn et al. 2002) discovers the segments from the data itself.

It is a finite-mixture method: it assumes the sample is a mixture of a fixed number of latent classes, each with its own path coefficients, and estimates both the class memberships and the per-class model at once using an EM algorithm.

A worked example

You run FIMIX-PLS for two classes. The aggregate model estimated one coefficient of 0.34 for a key path, but the two-class solution splits the sample into Segment 1 (62% of cases, coefficient 0.55) and Segment 2 (38%, coefficient 0.09). The single aggregate number was hiding two very different groups.

The next step is to find an observable variable that explains the split, so the segments can be described and used.

An aggregate coefficient of 0.34 splits into 0.55 for Segment 1 (62% of the sample) and 0.09 for Segment 2 (38%).

How to choose the number of segments

FIMIX-PLS does not tell you the number of segments directly. You run it for several class counts and compare fit and separation criteria.

AIC, BIC, CAIC, MDL5: Information criteria across candidate solutions. Lower is better. When they disagree, AIC3 and CAIC together, or a consensus across criteria, are commonly used to pick the number of classes.
Entropy (EN) > 0.50: Normalized entropy measures how cleanly cases are assigned to classes. Above 0.50 indicates the segments are well separated; higher is better.
Segment size: Each segment should be large enough to estimate its own model. Tiny segments well below 10% of the sample are usually not usable.

A good solution has low information criteria, high entropy, and segments that are big enough and interpretable. Statistics alone do not settle it; the segments must make theoretical sense.

What a FIMIX run reports

Class sizes

The estimated relative size of each segment (its mixing proportion).

Class paths

The structural path coefficients estimated separately within each segment.

Fit criteria

AIC, AIC3, AIC4, BIC, CAIC, MDL5 and normalized entropy for the solution, used to compare different numbers of classes.

Segment membership

A most-likely class label for each case, which you can relate to observable variables afterwards.

Running FIMIX-PLS step by step

Estimate the aggregate PLS-SEM model first as a baseline.
Run FIMIX-PLS for a range of class counts, for example two to four.
Compare the fit criteria and entropy across solutions to choose the number of segments.
Inspect the per-segment path coefficients to see how the relationships differ.
Find observable variables that characterize the segments, then consider a multi-group analysis on that grouping to confirm.

Common pitfalls

Picking the number of segments from one criterion alone. The criteria often disagree; weigh several together with interpretability.
Keeping segments that are too small to estimate a stable model.
Treating the latent classes as final without finding an observable variable that explains them. Unexplained segments are hard to act on.
Running FIMIX on a sample that is too small; finite-mixture estimation needs a reasonable number of cases per class.

FIMIX-PLS in OpenPLS

OpenPLS includes a FIMIX-PLS panel. Choose the number of classes, and it reports the segment sizes, the per-segment path coefficients, the full set of fit criteria (AIC, AIC3, AIC4, BIC, CAIC, MDL5 and entropy), and a most-likely segment label for every case. Free and runs in the browser.

Look for hidden segments in your own data

Build your model in the visual editor and run FIMIX-PLS for several class counts in a few clicks. Free, no install.

Open OpenPLS What is PLS-SEM?

Frequently asked questions

What is the difference between FIMIX-PLS and multi-group analysis?: Multi-group analysis compares groups you define in advance, such as country or gender. FIMIX-PLS discovers segments from the data when you do not know the grouping, which is why it addresses unobserved heterogeneity.
How many segments should I extract?: Run several class counts and compare information criteria (AIC, BIC, CAIC, MDL5) and normalized entropy, alongside segment size and interpretability. No single criterion decides it on its own.
What do I do after FIMIX finds segments?: Find an observable variable that explains the segment split, describe the segments, and ideally confirm the differences with a multi-group analysis on that variable.