Multivariable regression is a statistical procedure defining the relation between an outcome and a set of surrogate observations. The procedure is used either to identify the observations that provide independent information with respect to the outcome (descriptive modeling) or to estimate the likelihood of the outcome in a given patient from the set of observations particular to that patient (predictive modeling).

Bickell and colleagues [1] recently used a multivariable regression model to evaluate retrospectively the presence of gender bias in the management of patients with coronary artery disease. The male-female odds ratio for surgical referral was calculated for different risk strata using a previously validated logistic regression model. The calculated risk based on this model was then used as an unbiased estimate of true outcome.

Unfortunately, the logistic coefficients used in the prediction model and their associated standard deviations were not included in the article, nor have they been published. An otherwise excellent scientific paper is thereby reduced to the level of an internal memorandum, and its conclusions must be accepted as a matter of faith.

This deficiency may have been an oversight; however, some investigators justify such nondisclosures because others may misapply their models to populations that differ from those in which they were developed. Indeed, one of us has observed that "... a (diagnostic) test will perform ‘as advertised’ only if it is applied to patients similar to those in whom it was first assessed" [2], and the same goes for a prediction model. Despite this potential for abuse, disclosure is essential to allow others to attempt to replicate the findings and to test the methods to answer other questions. If the methods are not disclosed, however, no replication or further application can occur.

The journal Science states that "When a paper is accepted for publication ... it is understood that ... any materials and methods necessary to verify the conclusions of the experiments reported will be made available to other investigators under appropriate conditions" [3]. Such a policy of methodologic disclosure should be universal, and statistical prediction models (but not the raw data from which the models derive) should be covered by that policy.