To date, only a few studies using scan-sampling intervals to record neighbours in groups of animals explain how the interval length was chosen. In this study, we investigated the effects of different scan-sampling intervals, behavioural context, and the definition of neighbours on the accuracy of data as compared to a quasi-continuous observation. The data was collected from dairy cows kept in cubicle housing systems. Dairy herds of 22-43 cows on 6 farms were observed once a minute with an automatic tracking system for 6 days, and both the nearest neighbour and all neighbours within a defined distance of each cow were recorded with references to the activity, feeding and lying area and stored in neighbourship matrices. In the analysis, we used data collected at intervals of 2, 3, 4, ..., 30 min, and correlated this simulated data with the data based on the 1-min interval using a specialised correlation coefficient for matrices τKr. This correlation coefficient was then used as the response variable in mixed-effects models. We found that the size of the correlation coefficients generally decreased as interval length increased. This decrease was less pronounced for all neighbours than for the nearest neighbour. Moreover, the decrease was greatest with data from the activity area and lowest with data from the lying area. We concluded that, even with the relatively slow dairy cows in a barn environment, neighbour recordings should be conducted at short scan-sampling intervals in order to achieve a minimum correlation of τKr=0.8 with the quasi-continuous data. Intervals of every 2, 8 and 17 min are recommended for observation of neighbours in dairy cows for the activity, feeding and lying areas, respectively, and species that move faster may well require even shorter sampling intervals.