Fig. 6: Spaghetti and scatter plots for the consequent one-stage IPD models, that is, PA subsequent to AWB.

a,c,e,g,i, The within-person associations of AWB and PA for positive affective states (a) (P = 0.0001, i = 140,170), negative affective states (c) (i = 150,118), valence (e) (i = 44,251), energetic arousal (g) (P = 0.00005, i = 43,270) and calmness (i) (i = 41,561). The individual regression lines (thin lines) of participants are shown in red (negative slope) and blue (positive slope). The bold line and associated interval represent the conditional fixed effect of person-mean-centred AWB with a 99.2% CI. b,d,f,h,j, The between-person associations of AWB and PA for positive affective states (b) (P = 0.0000000000007, n = 3,864), negative affective states (d) (P = 0.00006, n = 4,179), valence (f) (n = 1,354), energetic arousal (h) (n = 1,271) and calmness (j) (n = 1,161). Each dot represents one participant’s average AWB (x axis) and predicted average PA (y axis). The bold line and associated interval represent the conditional fixed effect of person-mean AWB with a 99.2% CI. Of note, in k = 5 studies included, the intercept of the outcome is systematically larger owing to the limited variability or the higher average of the outcome, leading to larger values on the standardized PA score. This can be seen in all panels where several lines/dots are clustered above the fitted lines. Removing these data points (see Supplementary Information Section 4.4 for sensitivity analysis) did not alter the conclusions drawn, suggesting that these between-study differences in mean outcome levels did not bias our results. PA was square-root transformed and regressed on person-mean-centred and person-mean AWB. Person-mean-centred AWB was nested in individuals, which were nested in studies. Multilevel models were controlled for gender/sex and age, and regression coefficients were standardized. Significance was derived using two-sided t-tests of regression coefficients, which were adjusted for multiple comparisons (P < 0.008). n indicates the number of participants and i the number of ratings. For visualization, predictions were back-transformed to the original response scale (instead of square-root transformed scale). Standard errors remain on the square-root-transformed scale.