The problem of economic interpretation of sentiment indices

Modern text-processing methods substantially expand the range of sources through which economists can study unobservable economic phenomena. Researchers use large corpora of textual data to construct sentiment indices, which are then given an economic interpretation. Public texts, however, are a highly specific source of information. They are embedded in a broader social context that cannot always be identified from the text itself, yet may shape the meaning of the statements it contains. For this reason, the technical accuracy of classifying statements, for example the successful detection of irony, is not in itself a sufficient evidence of a link with an economic phenomenon. This transition requires specific justification. The article examines the relationship between an index and an economic phenomenon through the concept of measurement validity. Using sentiment indices of inflation expectations and consumer sentiment as examples, it shows that validation requirements depend on the researcher's interpretation of the index. An index may be used as an applied predictor, as a proxy for an existing indicator, or as a measure of economic phenomenon. In each case, a type of evidence is needed to justify the relationship between the indicator and the measured object. The strongest requirements arise when an index is interpreted as a measure of a phenomenon, especially if the phenomenon itself has an uncertain conceptual status.