Which Text Matters: Handling Model Uncertainty in Text-as-Data Regressions, with an Application to Central Bank Communication

TEXT START: Social scientists routinely convert text into covariates like topic shares, ideal points or sentiment scores, and insert them into statistical models of interest.

The Dissection

The paper disciplines researcher discretion. It converts the arbitrary menu of text indicators and functional forms into a formal model-selection problem, then uses the ECB application to show that most candidate measures add no detectable information about professional inflation expectations.

Its deeper function is narrower and more political: it turns communication from an institutional act by a powerful unelected body into a set of selectable statistical features. The machinery improves inferential hygiene, but it does not explain what the communication causes or whether the institution’s language remains structurally important.

The Core Fallacy

The central error under the Discontinuity Thesis is scope inflation. Model uncertainty is real, but it is only one layer of uncertainty. Selecting variables more rigorously does not solve construct validity, causal direction, omitted variables, strategic communication, regime change, feedback effects, or the possibility that the chosen relationship is merely historical pattern-matching.

“No incremental information” means no detectable increment within this sample, outcome, candidate set, prior structure, and model class. It does not prove that nonlinear effects are absent, that communication is economically irrelevant, or that the result survives a changed monetary regime.

Most importantly, the paper refines an analytical pipeline without confronting its own automation. Under P1, constructing indicators, selecting models, testing nonlinearities, and explaining the output are precisely the kinds of cognitive operations AI can compress and eventually dominate.

Hidden Assumptions

• The 45 indicators contain the relevant dimensions of ECB communication.
• Monthly observations from late 2002 to mid-2025 provide a sufficiently stable data-generating process.
• Professional inflation expectations are a reliable and substantively adequate target.
• Sparse selection reflects genuine signal rather than collinearity, limited power, prior choices, or sample-specific regularization.
• The absence of selected nonlinear effects is evidence about reality rather than about the model’s ability to detect them.
• Textual associations remain useful when communication, policy regimes, and market expectations change.
• “Tracking” expectations is sufficiently valuable even without causal identification or demonstrated policy leverage.
• Human researchers retain a durable advantage in interpreting the result rather than merely operating an increasingly automated inference stack.

Social Function

Primary classification: partial truth. Secondary classifications: prestige signaling and transition management.

The paper addresses a genuine methodological weakness and offers a credible way to reduce arbitrary specification choices. But its technical sophistication also functions as a legitimacy ritual: mathematical machinery certifies that human text-based inference remains valuable. It manages uncertainty at the level of academic procedure while leaving the larger displacement question untouched.

This is not simple copium. It is a quality-control device attached to a task whose production chain is increasingly automatable.

The Verdict

Useful statistical refinement, no systemic rescue. The method may make ECB text regressions less arbitrary, but it does not establish causality, durability, or institutional importance. Under the Discontinuity Thesis, it is quality control around an automatable cognitive task—not a moat, not a reversal of P1, and not evidence that text analysts retain durable productive participation.