The significance of P300 in investigation of the cognitive changes characteristic of aging and dementia is well established. But some controversies about sensitivity and specificity of P300 latency prolongation in detection of dementia still exist, and its predictive value and specificity in "real" clinical situations (i.e., in mixed population of neurological patients, both demented and nondemented but potentially cognitively impaired) were seldom estimated. In order to elucidate these questions, we recorded auditory event-related potentials ("oddball" paradigm) in 40 demented patients, 58 nondemented neurological patients, and 39 healthy subjects aged < or = 65 years. In addition, for the qualitative analysis of the data, we calculated three normality ranges of different width (i.e., control group's mean +/- 3 SD, 2.5 SD, and 2 SD, respectively). Our results showed that P300 latency was significantly longer in demented patients as compared to both controls and nondemented patients. Sensitivity of the P300 latency prolongation in detection of dementia depends on the width of the selected normality range, and is greatest for the narrowest range (70%), and diminishes with its widening. Specificity of this parameter, when only demented patients and controls were considered (approach used in the majority of the previous studies) was 100%, regardless of the range used. However, when the assessment was done in a mixed population of neurological patients, the P300 latency prolongation showed smaller but still very high specificity (from 86.2 to 100%) and the predictive value (from 77.8 to 100%). Depending on the width of the normality range selected, the rise of sensitivity was associated with fall of specificity and predictive value (and vice versa).