Figure 1

A schematic overview of the generalised additive models (GAMs) using one example pig and 14 days-period. At the top left, the probability (prob.) that a pig eats at every moment (panel c) is modelled by adding the probability to start eating across days (panel a) and the diurnal probability surrounding this daily trend for each of the six periods (panel b). On the right, the probability-corrected (pb-corr) intake of a pig (panel f) is modelled similarly, by adding the trend (panel d) and diurnal intake surrounding the trend of each of the six periods (panel e). The resulting probabilities to eat (panel c) and probability-correct intakes (panel f) are multiplied to obtain the final predicted intake of the six periods (panel g). Six features of diurnal patterns in feed intake were extracted from this final prediction: the height (h+), width (w), and timing (t+) of the highest peak, the height of the lowest intake (h−), the number of peaks (black dots), and the proportion of intake obtained at night (grey shading). From the hourly probability plot (panel c) the height of the lowest (c−) and highest (c+) probabilities of eating were extracted as an approximation of day-to-day consistency. Note that quantities (height, width and proportion of intake) and probabilities (lowest and highest) were corrected for the trend, to allow comparison between periods. The model in R code: gamlss(intake ~ ga(~ s(hour, bs = “cp”, k = 8, by = period) + s(day, bs = “ps”, k = 4)), nu.formula =  ~ ga(~ s(hour, bs = “cp”, k = 8, by = period) + s(day, bs = “ps”, k = 4)), sigma.formula =  ~ period—1, data = na.omit(hourly.intake.data), family = ZAGA).