19. SEMinR Lecture Series - When to Delete or Not to Delete a Formative Indicator

Formative measurement models require a careful, content-driven deletion decision: non-significant indicator weights alone are not enough to justify removing an indicator, because doing so can break the construct’s domain coverage. The core rule is to treat formative indicators as non-interchangeable “building blocks” defined during conceptualization—so deletion must be justified by a combination of statistical evidence and relevance checks, while protecting content validity.

After checking collinearity and examining indicator weights/loadings, the decision process starts with indicator weight significance. If an indicator’s weight is statistically significant, it can usually be retained without further concern. If the weight is insignificant, the next gate is the indicator’s loading: when loadings are above 0.5, the indicator is generally kept. When loadings fall below 0.5, the indicator may be considered for deletion—but only if that deletion does not harm content validity.

This is where the lecture draws a sharp distinction between formative and reflective measurement. In formative models, indicators are not interchangeable; removing one can make the construct incomplete. The example uses a CSR construct defined as economic, legal, ethical, and philanthropic dimensions. Dropping the “economic” dimension would sacrifice content validity even if its statistical weight is not significant, because the conceptual domain would no longer be fully represented. In short: formative indicators with non-significant weights should not be automatically removed.

Beyond weight significance and loadings, the lecture adds a relevance check based on bootstrapped indicator weight values. Indicator weights are standardized between −1 and +1, and relevance is interpreted by magnitude: values closer to +1 or −1 indicate a stronger positive or negative relationship with the construct, while values closer to 0 indicate a weaker relationship. The closer the relevance is to one (in absolute terms), the more defensible the indicator’s retention.

A consolidated “rule of thumb” ties these diagnostics together. For redundancy/convergent validity, the path coefficient should exceed 0.708. Collinearity is evaluated with VIF-like thresholds: values above 5 are problematic, while 3 to 5 are usually not critical, and below 5 is generally acceptable. For formative indicators, weight significance is required, relevance should be closer to ±1 for stronger contribution, and loadings should be at least 0.5 and statistically significant. If loading is below 0.5 and not statistically significant, deletion becomes more defensible.

Finally, the lecture outlines an implementation workflow (with code available on the instructor’s website): load data, specify the measurement model (including formative weights for formative constructs), estimate the structural model, then bootstrap to obtain weights/loadings and inspect collinearity, significance, and relevance. The takeaway is procedural and practical: validate formative models using a layered set of thresholds, but let conceptual domain coverage—content validity—override simplistic “delete if non-significant” rules.