Multivariable models in biobehavioral research.

OBJECTIVE: To review contemporary multivariable modeling and statistical reporting practices in psychosomatic and behavioral medicine research.
METHODS: A random sample of 40 original research articles involving multivariable models was obtained from the 2005 volumes of four of the leading psychosomatic and behavioral medicine research journals. A random comparison sample was obtained from the 2005 volumes of four of the leading general medical and psychiatric journals. Multivariable modeling and reporting practices were systematically coded. The evaluation focused primarily on issues raised in 2004 Statistical Corner article by Babyak.
RESULTS: Deficiencies were found in a large proportion of the articles published in psychosomatic and behavioral medicine journals. The single most common problem was a lack of clear information, or any information at all, about important aspects of the statistical methods. Other frequent problems included post hoc selection of variables, lack of clear rationales and well-specified roles for selected variables, inadequate information about models as a whole (e.g., goodness of fit), failure to test model assumptions, and lack of model validation. Overfitting of multivariable models was the exception rather than the rule, but still a significant problem.
CONCLUSIONS: There is room for improvement in the use and reporting of multivariable models in psychosomatic and behavioral medicine research journals. These problems can be overcome by adopting best statistical practices, such as those recommended by Psychosomatic Medicine's statistical guidelines and by authoritative guidebooks on statistical reporting practices.