Dichotomizing

Dichotomizing converts a variable that has a range of values into two categories. It is used for predictors or outcomes to obtain a simple yes/no interpretation or to fit binary models. This typically involves choosing a threshold, with values on one side assigned to one category and those on the other side to the second category. Thresholds may be predefined (clinical cutoffs) or data-derived (for example, median splits or ROC-based thresholds).

Common approaches include using a clinical or policy cutoff, a median-based split, or a threshold optimized to discriminate outcomes. Dichotomization is popular for its simplicity and interpretability, but it can distort relationships and comparisons if the underlying variable is continuous.

Advantages include ease of interpretation and compatibility with binary methods such as logistic regression. Disadvantages include loss of information and statistical power, potential misclassification, and sensitivity to the chosen threshold, which can produce misleading or inconsistent results.

Alternatives and best practices emphasize retaining the variable in continuous form when possible or modeling nonlinearity with splines or fractional polynomials, or using multiple categories when appropriate. If dichotomization is used, preregister or justify the threshold, report the distribution around the cut point, and perform sensitivity analyses to assess robustness of conclusions.

Examples include dichotomizing a depression or anxiety scale at a clinical cutoff to identify cases or categorizing income as above or below a poverty line. Such decisions should be reported with clear rationale and accompanied by disclosures about potential limitations.