Many different properties of objects and their names have been implicated as determinants of aphasic naming accuracy. Many of these variables are intercorrelated, making it very difficult for the experimenter to devise sets of items which differ on one property while being matched on all the others. Faced with this situation, experimenters have been tempted to employ the techniques of regression analysis on aphasic naming data, seeking statistical rather than experimental control over the variables. We report a study in which six aphasic patients attempted to name 139 object pictures on three separate occasions. Multiple regression analyses were carried out on the data for the whole group and the combined data for each patient, and logistic regression analyses on the data from individual administrations of the picture set. The different analyses produce disconcertingly different solutions, casting doubt upon the reliability of the outcomes of these analyses, particularly for single administrations of item sets where the 'significant' predictors only exert small effects. We conclude with a discussion of when, and how, regression analyses might be of use in the analysis of aphasic data. Copyright © 1996 Elsevier Science Ltd.