What is LPA? Why Your Organization is Segmenting Wrong and What to Do About It

Organizational analysts often need to provide leaders with recommendations for actions or interventions. Actions are easier to take if you can aim at a reliable target, you can see and measure. So, you need to find a way to group, cluster, your employees into valid categories or personas.

Most minds go to existing groups as identified from other HR modalities such as compensation or learning and development. You probably have segments like “high performers,” “flight risks,” “disengaged,” or maybe demographic buckets: “early career,” “leadership pipeline,” or “tenured mid-level leaders. You set arbitrary thresholds. The boundaries are clean, easy to communicate, and completely made up.

The real patterns in your data are messier. Your people don’t often fall into one group or category. Humans are complex. This is where latent profile analysis comes in, and why it matters if you are trying to understand your workforce.

What LPA Does

Latent profile analysis is a statistical method that identifies hidden groups of people based on patterns in their survey responses. It is called “latent” because these groups aren’t labeled in your data. You don’t manually create them. The method is called “profile analysis” because it’s person-centered, not variable-centered. You’re asking, “what kinds of people do we have” instead of “how does engagement correlate with retention.”

LPA uses something called a Gaussian Mixture Model under the hood, which means it’s fitting multiple statistical distributions to your data simultaneously. It tests whether the data cluster into 2 groups, 3 groups, 4 groups, and so on, and asking which number of groups makes the most statistical sense. But you don’t need to know the math to use it. You need to understand what it’s answering: how many meaningfully different types of people do your survey and other indicators reveal? (Note: LPA works with numerical data only. Demographics should be validated against profiles afterward, not in the model.)

Why This Matters for IO Psychology

If you have an IO psychology background or you’re working in people analytics, you already think in profile.

It is important to note how different it is from K-means clustering. K-means is fast, it’s simple, and it’s very common. But K-means doesn’t assess whether clusters, or profiles, exist. It just forces your data into K groups, whatever K you pick. If you say, “I want 3 clusters,” it will give you 3 clusters, even if your data naturally has 2 or 5 or would be better described as a single continuum. It’s clustering by decree, not by what your data supports. LPA asks the question differently. It tests multiple solutions and uses statistical criteria to choose the one with the strongest evidence. This is the difference between finding the segments your workforce contains versus imposing segments and hoping they stick.

How LPA Chooses the Right Number of Groups

LPA doesn’t just fit one model. It fits many. It tests 1 group (null model, baseline), 2 groups, 3, 4, 5, …nth and so on up to some computationally considerate upper limit. Then it uses multiple statistical criteria to decide which solution is best.

The most common criterion is BIC, Bayesian Information Criterion. It penalizes model complexity, so it asks: ‘am I getting meaningfully better fit by adding another group, or am I just overfitting?’ SABIC (Sample-Size Adjusted BIC) and AIC (Akaike Information Criterion) do similar work, though AIC penalizes complexity less aggressively than BIC, favoring model fit over simplicity. Entropy measures how cleanly people separate into their assigned groups (you want entropy close to 1, meaning sharp boundaries). The BLRT, Bootstrapped Likelihood Ratio Test, compares adjacent models to see if K groups are significantly better than K-1.

These criteria prioritize differently. BIC is conservative as it penalizes complexity and works best at large sample sizes (N>300). At smaller N, it can underfit. SABIC and AIC are more permissive and favor model fit over parsimony. Entropy captures classification clarity directly: high entropy means each person clearly belongs to one group; low entropy means many ambiguous cases. BLRT is the gold standard for statistical significance but computationally expensive.

In practice: Use BIC and entropy together for most organizations. If N<200, weight entropy more heavily. If N>500, BIC becomes more trustworthy. BLRT is a tiebreaker when the first two disagree.

It is good LPA practice to use multiple criteria to evaluate the sturdiness of its chosen K, to triangulate the true effect. It builds reliability and confidence in your outcomes. If all of them point to K=4, you have 4 groups. If they disagree (BIC says 3, entropy favors 4), that’s a signal that the solution is ambiguous, and you should be cautious about drawing strong conclusions. This is one place where LPA diverges from frequent practice. Most organizations pick one criterion or eyeball an elbow curve. LPA researchers have tested this (e.g., Nylund, Asparouhov, and Muthén, 2007) and found that different fit indices perform better under different conditions. The responsible approach is to look at all of them together.

In practice, here’s how to read disagreement:

• If all four indices (BIC, SABIC, entropy, BLRT) point to the same K, you have a robust solution. Act with confidence.
• If three agree and one diverges, the solution is still solid. The outlier usually reflects a different penalty structure (e.g., BIC penalizes complexity more than AIC). Proceed, but note it in your documentation.
• If they split 2–2 (BIC and SABIC say K=3, entropy and BLRT say K=4), the true K is genuinely ambiguous. Examine both solutions and choose based on interpretability and sample size. At N<200, entropy becomes less reliable; favor BIC. At N>500, entropy gains strength.
• If all four disagree, your solution space is unstable. Collect more data or reconsider your indicators.

When LPA Works Well and When It Doesn’t

LPA works well when you have:

• Enough respondents (roughly 200+ for stable solutions; 300+ is better)
• High individual response variability (if everyone gives the same response, nothing clusters)
• Items of quality (indicators that separate people meaningfully matter more than quantity; 6 strong items beat 15 weak ones)

LPA struggles when you have:

• Small samples (under 100 respondents, the solutions become unstable)
• Highly skewed or unusual distributions (LPA assumes roughly normal data within each group)
• Too many indicators (more than 12 or so makes model fitting difficult)
• Indicators that don’t cohere (if you mix unrelated constructs, you don’t get coherent profiles)
• Too much missing data

The Technical Depth (If You Want It)

If you’re going to use LPA in production or explain it to a technical audience, you should understand a few things:

LPA assumes multivariate normality within each group. If you have one indicator that’s bimodal (two humps instead of one bell), it can distort the whole solution.

The covariance structure matters. LPA can fit different models where the variance within groups is the same across profiles (more restrictive) or varies by profile (more flexible). Verifying which fits best is part of the enumeration process.

The outcome of an LPA for each respondent is a cluster assignment based on posterior probabilities. These probabilities tell you how sure the model is about each person’s group membership. You want most people in the 0.7–1.0 range, meaning they have a higher probability of belonging to one group. If someone has a posterior of 0.45 for Group 1 and 0.55 for Group 2, they’re ambiguous. They might belong to either group depending on sampling variation. Good practice flags these people separately instead of forcing a classification.

Sample size affects stability. Rules of thumb suggest at least 5–10 observations per estimated parameter. If you have 4 groups and 7 indicators with a full covariance structure, you’re estimating roughly 60–70 parameters. That means you want 300–700 respondents. Below 200, the fit indices become unreliable, and solutions tend to be unstable across repeated samples.

Why You Should Care Right Now

If your organization is doing people analytics, designing engagement interventions, segmenting your workforce for targeted HR programs, or trying to understand organizational change readiness, you probably need segments that exist in your data rather than segments that fit a spreadsheet.

The alternative — continuing with arbitrary cutoffs and K-means clustering — isn’t harmless. It distorts who you think you have. It leads to interventions designed for segments that don’t coherently exist. It burns credibility when stakeholders realize the “high performers” and “flight risks” don’t cluster the way you described them.

LPA takes longer to explain than “we cut people at the median on engagement.” It requires more statistical sophistication to execute properly. But it answers a fundamentally different question, and in most cases, the right question.

The next article explains how to make this methodology reliable and executable. Rather than leaving segmentation as analysis-only, Part II shows how to build a harnessed agent system that is scalable. [Next article in construction].

References

Nylund, K. L., Asparouhov, T., & Muthén, B. O. (2007). Deciding on the number of classes in latent class analysis and growth mixture modeling: A Monte Carlo simulation study. Structural Equation Modeling, 14(4), 535–569. https://doi.org/10.1080/10705510701575396