What is a Latent Variable?

A latent variable is a construct that cannot be observed or measured directly but is inferred from observable indicators assumed to reflect it. Perceived usability is latent; the ten items of the SUS are its indicators. Customer loyalty is latent; repeat purchase, referral, and stated intent are indicators. The relationship is the heart of a measurement model: each indicator is treated as latent variable plus measurement error, so that several imperfect indicators, combined, estimate the underlying construct more faithfully than any one. The whole apparatus of psychometrics exists because the interesting variables in human research are latent, and getting from what people do and say to what those things mean is measurement's central problem.

From Constructs to Indicators

The bridge is operationalisation: defining a construct precisely enough to choose indicators, then choosing indicators that reflect it. Three failure modes recur. Single-indicator thinking treats one item as the construct ("satisfaction is the rating on question 4"), importing that item's specific quirks and error wholesale. Construct drift lets an indicator stand for something it only partly reflects (time-in-app as "engagement", capturing confusion as readily as delight). And proxy worship forgets the proxy was a proxy at all, optimising the indicator until it decouples from the construct, the mechanism behind Goodhart's law and most gamed KPIs. Construct validity is the discipline of checking the bridge: do the indicators converge, do they distinguish this construct from neighbours, do they predict what the construct should predict?

The Statistical Machinery

Factor analysis is the workhorse: it infers latent dimensions from the correlations among indicators, tells you how many constructs a set of items really measures, and quantifies each item's loading. Structural equation modelling extends the idea to relationships among latent variables (does latent trust predict latent loyalty, measurement error accounted for?), and latent class analysis finds hidden groups of people rather than hidden dimensions. Composite scores (averaging validated items) are the everyday approximation, honest when the items were shown to hang together and dishonest when someone averaged whatever was on the survey.

Latent Thinking in Product Research

1. Name the construct before the metric.
"We want to measure onboarding confidence" precedes "we'll ask two questions and watch completion", and the sentence in between (why these indicators reflect that construct) is the measurement rationale readouts should state.

2. Prefer validated multi-item measures for anything tracked.
Standardised instruments come with evidence that their indicators reflect a coherent latent construct, the pedigree a home-grown item lacks.

3. Triangulate indicators across modes.
A latent construct like frustration shows up in behaviour (errors, backtracking), self-report (ratings), and expression (what people say on video); mixed instruments that capture all three in one study (the design of a Ballpark task with a recorded answer) estimate the construct from independent angles, the strongest everyday defence against a single proxy's error.

4. Watch for decoupling.
When a metric moves and the construct visibly doesn't (scores up, users no happier), the indicator has slipped its moorings; revalidate rather than celebrate.

The Takeaway

Latent variables are what research is actually about, and indicators are all it can touch. Operationalise deliberately, measure with several indicators rather than one, validate that they reflect the construct, and treat every metric as a proxy that can drift. The number on the dashboard is not satisfaction; it is a shadow satisfaction casts, and good measurement is the art of reading shadows honestly.