What is with all these different correlation thresholds?

Abstract

Aims: The concept of correlation was first introduced in the mid-1800s and is one of the most widely used measures of association. Since then, there has been widespread use of correlation coefficients across disciplines. While correlations are widely used, the interpretation or labeling of the correlation strength often creates confusion among researchers. For instance, labels such as ‘strong’, ‘very strong’, ‘large’, and ‘very large’ have been used for correlation coefficients ranging from 0.40 to 1.00. In the absence of context on how these coefficients are applied, there is a need to address such discrepancies. The overall objective is to present the context, type, and varying thresholds of correlation coefficients used in the literature. Methods: We took a pragmatic approach to search literature using a critical review methodology with the aim to identify publications presenting correlation thresholds guidance. Authors searched Ovid, MEDLINE, Embase, PsycInfo, and Google Scholar. Data focusing on the range of correlation coefficients, labels, thresholds, context, and measurement scale of the variables were extracted. Results: The search identified a total of 25 different sets of thresholds. Pearson’s correlation was most commonly referenced 40% (n = 10), followed by Spearman’s 20% (n = 5). Other coefficients (i.e., phi, tetra- and poly-choric, bi- and poly-serial) were rarely reported. Thresholds differed depending on the context of the research (i.e., measurement/ scale development, and/or research field). For instance, within the context of measurement development, thresholds typically had three levels: B 0.40 labelled as very weak, weak, or low; between 0.40 to 0.60 labelled as moderate; and > 0.60 labelled as strong, high, or very high. Within the behavioral and social sciences, coefficients seemed to be lower and values ranging from 0.30 to 0.40 were characterized as moderate to high. Empirically derived thresholds tended to be lower compared to those proposed theoretically. Conclusion: Given the abundance of how correlation coefficients are reported in the literature, there is an urgent need to develop field specific thresholds and reporting guidelines in an effort to stop asking what is up with all these correlation thresholds. This SIG will contribute to providing specific guidelines within the field of patient-reported data. On behalf of the ISOQOL Psychometrics Special Interest Group.