Appendix: Monte Carlo simulation results We conducted a Monte Carlo simulation study addressing the following questions: (1) Are 3 observations per person sufficient in a multilevel model (MLM) to yield an unbiased estimate of the parameter addressing age-differences in the within-person relationship between happiness and sadness? (2) Is the standard error of the parameter unbiased in this case (i.e., are statistical significance tests accurate)? (3) How does the statistical power of our study compare to the power of prior research with more observations per person but lower sample size? The model analyzed in the Monte Carlo study corresponds to those used to examine “age differences in mixed emotions using the covariation approach” in the manuscript. Specifically, the following MLM was tested: Within-subject equation: Yij = β0j + β1jXij+ rij (1) Between-subject equations: β0j = γ00 + γ01Wj + u0j (2) β1j = γ10 + γ11Wj + u1j (3) In this model, Yij is the happiness rating of person j for episode i, Xij is the sadness rating of person j for episode I, and Wj is the age of person j. In the within-subject equation (1), β0j and β1j represent intercept and slope of the regression of sadness on happiness for subject j, and rij is a within-subject residual term. In the between-subject equation (2), a subject’s intercept is a function of a common (fixed) intercept for the population (γ00), a common (fixed) effect of age (γ01), and a residual intercept variance term (u0j). In the between-subject equation (3), a subject’s slope is a function of a common slope effect (γ10), a common (fixed) effect of age on the regression slope (γ11), and a residual slope variance term (u1j). For the simulations, the parameter values used in the data generation were taken from the results from the PATS sample reported in the manuscript. Analyses were conducted using the Mplus version 7.3 Monte Carlo simulation facility, using 1000 replications for each simulation analysis (Mplus example code is given below). Two conditions were varied in the simulation: (a) sample size and (b) number observations per person. Specifically, in one series of models, the sample size was held constant at N=180 with varying numbers of observations per person (3, 15, 25, 35, 45, 55, 67, 700); In a second series of models, the sample size was varied (N=180, 900, 1500, 2100, 2700, 3300, 4000) and the number of observations per person was held constant at 3. The total number of observations (N times number of observations) was matched in each series (i.e., 180*15 = 900*3; 180*25 = 1500*3, and so on). The combination of N =180 with 35 observations per person corresponds with the design by Carstensen et al. (2000), and the combination of N=4000 with 3 observations corresponds with the design of the PATS sample in the current study. Of particular interest in the present context is the performance of parameter γ11, which represents the estimated effect of age on the relationship between happiness and sadness. The results (% parameter estimates bias, % standard error bias, power) for this parameter are shown in the tables below. Bias of parameter estimates and standard errors: Percent bias of the parameter estimate and standard errors were well within acceptable limits (for focal parameters, bias should be below ±5%, (Wang, 2012)) for all tested conditions. For instance, for a sample size of 180 and 35 observations per person – corresponding with the design in Carstensen et al. (2000)-- parameter bias was -2.90% and SE bias was -2.96%. For a sample size of 4000 and 3 observations per person -- corresponding to the design for the PATS data in the current manuscript -- parameter bias was 0.65% and SE bias was 1.75% (see Table). Given that some prior studies in this literature relied on small samples with fewer than 50 participants (Ong & Bergeman, 2004; Ready, Carvalho, & Weinberger, 2008), we also wondered whether biases would increase for smaller samples. For a sample size of N=50 (with 35 observations per person), SE bias increased to -8.0%; for a sample size of N=25 (with 35 observations per person), SE bias increased to -11.8%. Thus, a small sample size at level 2 (i.e., a small number of subjects) yielded biased estimates, but a small number of observations (i.e., 3) combined with a large sample size yields unbiased parameters and standard errors. Statistical power: The statistical power to detect age differences in the relationship between happiness and sadness based on the MLM is also shown in the table below, and graphically illustrated in the figure below. As can be seen, for a given number of overall observations (sample size * observations per person), increasing the participant sample size (lower part of the table) has a much greater impact on power than increasing the number of observations per person (upper part of the table). Keeping the number of observations per person constant at 3, a sample size of 1500 is sufficient to detect an effect of the size found in the PATS sample with a power exceeding 90%. On the other hand, keeping the sample size at 180 individuals, the power to detect this same effect does not reach 80% even with a very large number (i.e., 700) of observations per person. Table: Parameter bias, standard error bias, and power for parameter γ11 Total number of observations 540 2700 4500 6300 8100 9900 12060 126000 Sample size = 180, varying number of observations per person Sample Observations % parameter % SE bias size per person bias Power 180 180 180 180 180 180 180 180 0.23 0.55 0.62 0.63 0.66 0.67 0.70 0.75 3 15 25 35 45 55 67 700 -0.97 +0.97 +0.65 -2.90 -2.26 -1.29 +0.65 +1.0 -0.75 -1.95 +0.00 -2.96 -0.78 +0.00 +2.48 -3.07 Total number of observations 540 2700 4500 6300 8100 9900 12000 Observations per person = 3, varying sample size Sample Observations % parameter % SE bias size per person bias Power 180 900 1500 2100 2700 3300 4000 0.23 0.72 0.91 0.98 0.99 0.997 0.999 3 3 3 3 3 3 3 -0.97 +0.65 +1.29 +0.65 +0.32 +0.32 +0.65 -0.75 -0.81 +1.06 +0.00 +1.43 +1.59 +1.75 Power for varying ns (subjects or observations) 1 2100*3 2700*3 3300*3 4000*3 Varying the number of subjects (keeping number of observations at 3) 180*67 Varying number of observations per subject (keeping number of subjects at 180) 1500*3 0.8 900*3 0.6 Power 180*25 180*35 180*45 180*55 180*15 0.4 0.2 180*3 0 0 2000 4000 6000 8000 10000 Total sample size (subjects * observations) Mplus syntax for the Monte Carlo simulation: MONTECARLO: NAMES ARE happy sad age; nobservations = 6300; ncsizes =1; csizes = 180 (35); seed = 5859; nreps = 1000; within = sad; between = age; ANALYSIS: 12000 type = twolevel random; MODEL POPULATION: %WITHIN% [sad@0]; sad@2; slope | happy on sad; happy*1.366; %BETWEEN% slope on age*0.031; happy on age*0.044; [happy*4.107]; [slope*-0.332]; happy*1.351; slope*0.08; happy with slope*0.032; [age@0]; age@3.2; MODEL: %WITHIN% slope | happy on sad; happy*1.366; %BETWEEN% slope on age*0.031; happy on age*0.044; [happy*4.107]; [slope*-0.332]; happy*1.351; slope*0.08; happy with slope*0.032; References Carstensen, L. L., Pasupathi, M., Mayr, U., & Nesselroade, J. R. (2000). Emotional experience in everyday life across the adult life span. Journal of Personality and Social Psychology, 79(4), 644-655. Ong, A. D., & Bergeman, C. S. (2004). The complexity of emotions in later life. Journals of Gerontology Series B-Psychological Sciences and Social Sciences, 59(3), P117-P122. Ready, R. E., Carvalho, J. O., & Weinberger, M. I. (2008). Emotional Complexity in Younger, Midlife, and Older Adults. Psychology and Aging, 23(4), 928-933. Wang, J. (2012). Wiley Series in Probability and Statistics : Structural Equation Modeling with MPlus : Methods and Applications (3rd Edition). Somerset, NJ, USA: John Wiley & Sons.