Nine of thirteen paths held. The pattern in the misses is the story.

Why PLS-SEM

The model has many constructs, a mix of reflective and formative measures, and higher-order constructs, and the aim was to explain and predict an outcome more than to confirm one established theory. Those are the conditions under which partial least squares is preferred over covariance-based SEM (Hair et al., 2011). It also makes no assumption that the data are normally distributed, which survey ratings rarely are.

Step one: are the measures sound?

Eight checks, each against its published threshold. Six pass cleanly. Two pass narrowly, and I show them that way because they affect how far the results can be pushed: several constructs in this study are closely related, which is worth knowing before reading the paths.

Item-level loadings and scale statistics are on the measurement page.

What each kind of value is made of

The stage-two weights show how much each dimension contributes to its kind of value. They are a finding in their own right, because they tell a product team where the value actually comes from.

Select an arrow to see the hypothesis, a plain-language reading and the statistics. Line weight follows the size of the effect.

Step two: the thirteen direct paths

Standardized coefficients from 95% bias-corrected bootstrap intervals, with six control variables in the model. The thesis reported two paths at the p < .10 level. Here every path is held to p < .05, so H2b is shown as not supported.

Of the control variables, individualism-collectivism had a positive effect on return and recommend (β = .330) and on browsing to discover (β = .153), and age had a small negative effect on searching to decide (β = −.109). The others were not significant.

Does immersion carry the effect of value?

Six mediation tests. The indirect effect is the part of a value’s influence that runs through immersion. Where it is significant and the direct effect is not, the value works through immersion only.

* direct effect significant at p < .05. The pattern is clean: usefulness acts on exploration directly, while enjoyment and social value act only by first drawing the shopper into immersion.

How much the model explains and predicts

R² is the share of variance explained in each outcome. Q² above zero indicates predictive relevance, and above .50 is considered large. In the PLSpredict check, the model produced lower prediction error than a plain linear regression benchmark on 14 of 18 indicators, which Shmueli et al. (2019) classify as medium predictive power.

R² values this high in a single-questionnaire study partly reflect that all measures share one method and one moment. I return to that on the last page.

What the numbers say

• Discovery is what brings shoppers back.

  Browsing to discover has about three times the effect of searching to decide on intent to return and recommend (.45 against .16).

• Usefulness is the only direct lever on exploration.

  It raises both kinds of exploration on its own, and most of its weight comes from personalization and navigation.

• Enjoyment and social features work through immersion.

  Neither moves exploration directly. Both raise immersion, and immersion raises exploration.

• Immersion deepens more than it widens.

  Its effect on searching to decide (.54) is more than twice its effect on browsing to discover (.24).