SEM: find what really drives NPS, and how

Cloud4Feed Research Team ·Updated 22 Oct 2026 ·7 min read

Structural equation modelling is a statistical method that tests how a set of variables influence one another, not just whether they move together. In VoC analysis it maps the routes by which a touchpoint, such as staff attentiveness or checkout waiting time, reaches an outcome such as NPS, satisfaction or loyalty.

The key idea is mediation. Touchpoints rarely change recommendation directly; they change a perception first, such as perceived service quality or emotional attachment, and that perception moves the score. Knowing how customers feel (the NPS) is only half the diagnosis. To act, you need to map the structural paths that create that feeling.

Each arrow carries a standardised coefficient (β). The effect of X through M is the product of the two coefficients on the way:

A mediation analysis tests four research hypotheses in sequence.

1. 1H1, direct effect (X → Y). Does the touchpoint affect satisfaction or NPS directly?
2. 2H2, mediator formation (X → M). Does the touchpoint significantly predict the perception?
3. 3H3, mediator effect (M → Y). Does the perception have a significant effect on the outcome?
4. 4H4, indirect effect (X → M → Y). How large is the real effect that travels through the mediating path?

A retailer measured five touchpoints and asked one question: which investments really increase the intention to recommend? The model grouped the five touchpoints into two perception mechanisms. Staff attentiveness and knowledge work through service satisfaction; store layout, checkout waiting time and discounts work through the shopping experience.

Staff attentiveness

Staff knowledge

Store layout

Discounts

Checkout waiting time

Service satisfaction

Shopping experience

NPS

.42.38.35.31.29.54.41−.18

People are the biggest lever. Staff attentiveness (.227) and knowledge (.205) together add up to .432, about half of the total indirect effect of all five touchpoints (.822). Investment in staff, through service satisfaction, is the strongest route to a higher NPS in this model.

Waiting time is a hidden risk. Checkout waiting time has a weak positive path through the shopping experience (.29 × .41 = .119) and a negative direct path to NPS (−.18). Its total effect is negative (−.061). When the direct and indirect paths point in opposite directions, the touchpoint is affecting recommendation through something the mediator does not capture. A simple correlation would show a small, harmless-looking number and hide both effects.

Enter the standardised coefficients from your model for each driver: X → M, M → Y and, if there is one, the direct path X → Y. The ranking updates as you type. It starts with the case-study values.

Rank the touchpoints by total effect and act in that order, but treat any touchpoint with a negative direct effect as urgent whatever its rank. In the case study:

• Staff attentiveness (.227). Coaching, staffing levels and service rituals that make customers feel noticed.
• Staff knowledge (.205). Product training and quick access to information on the shop floor.
• Store layout (.144). Wayfinding and tidiness in the areas customers use most.
• Discounts (.127). Clear, credible offers rather than more of them.

Then measure again. Track the touchpoints with transactional surveys, and check the overall effect with relational NPS; use MaxDiff when you also need customers’ stated priorities.

• Start from a theory. Draw the paths you expect before you look at the data. SEM tests a model; it does not discover causes on its own, and the arrows are only as causal as your design and theory allow.
• Measure perceptions with several items. Latent constructs such as service satisfaction are more reliable with three or more survey items each. Check reliability and the measurement model first.
• Collect enough responses. A few hundred complete responses is a common minimum for a model of this size; larger models need more.
• Check model fit. Report fit indices such as CFI, RMSEA and SRMR, and do not interpret coefficients from a poorly fitting model.
• Test indirect effects properly. Use bootstrapped confidence intervals for indirect effects rather than judging each path separately.