The common interpretation of statistical interactions of global measures of social support and stress in their effects on depression (the "buffer effect" of social support) as reflecting interactive processes within individuals was examined. It was argued that its adequacy depends on incorrect assumptions about the distribution of depressive symptoms within individuals over time and/or different levels of psychosocial adversity. With hypothetical data sets generated under the assumption of a threshold effect in the development of depression and of additive, rather than interactive, effects of social support and stress, it was demonstrated that spurious "buffer effects" of social support can appear in analyses of cross-sectional data. Methodological and substantive implications are discussed, and strategies for study design and data analysis are proposed that allow to assess whether empirical deviations from linearity are more adequately interpreted as a buffer or as a threshold effect.