In Which Direction Does Happiness Predict Subsequent Social Interactions? A Commentary on Quoidbach et al. (2019)

Method

The analysis reported by Quoidbach and colleagues was reproduced as closely as possible using the publicly available data from their study (see https://osf.io/bxgn4 for data from the original study). For this reanalysis, the following multilevel time-lagged logistic regression model was estimated by Quoidbach and colleagues (Equation 1, p. 1114):

logit P ( P t + 1 j ) = β 0 i j + β c i j H t + β day j H day + β d D + β T T + β p P t j ,

with β 0 i j = γ 0 j + u 0 i j and β c i j = γ c j + u c i j .

P t + 1 j represents individuals’ reports of being in a social interaction at the next measurement time point and is the dependent variable in this analysis (1 = in interaction, 0 = alone). The H t variable entails the previous measure of happiness (reported on a scale from 0 to 100), of which the fixed effect γ c j , representing the effect of happiness on subsequent interactions, is the coefficient of interest in this analysis. The H_day variable represents an individual’s average happiness on a given day, excluding the individual’s happiness at H_t. If there is one data point on a given day (66% of the data points), the H_day variable is identical to the happiness of the time point to be predicted (t + 1). On days when there are two observations (24% of the data points), H_day is computed using the average of H_t+1 and the other measure of that day. Also see Table S3 in the Supplemental Material available online for a data excerpt and example calculation of H_day. The variables D, T, and P t j control for the day of the week (weekday, Saturday, Sunday), time of day (in 2-hr bins starting at midnight), and whether the person was in a social interaction at the previous time point. The reference categories were specified as Saturday between 12:00 p.m. and 2:00 p.m.

The model is specified with random intercepts u 0 i j and random slopes u c i j , accounting for individual differences in social-interaction tendencies and individual differences in the effects of previous happiness measures (H_t) on social interactions. All βs and γs are subject to the estimation. The estimation was carried out with the glmr function of the lme4 package (Version 1.1-19; Bates et al., 2015) in R (Version 3.5.2; R Core Team, 2018). Data-preparation procedures and model estimation are documented in the R script available at https://osf.io/zk98q.

To better account for “low-frequency dynamics” of happiness that do not include future happiness values, I further included an H_pastweek variable and an H_pastday variable in the model, capturing the effects of the average happiness of an individual reported in the past week (excluding the current day) and the past 24 hr (excluding the measurements H t ), respectively. The calculation of H_pastweek and H_pastday is documented in the R script available at https://osf.io/zk98q. It is important to note that the variables H_pastweek and H_pastday were also calculated on the basis of only a few observations (M = 3.02, SD = 3.56, and M = 0.86, SD = 1.23, respectively). Moreover, they were computed using the publicly available data set, which was subdivided to contain only consecutive reports within a 12-hr time window. The nonsubdivided data might provide a more reliable estimate of the past week’s and past day’s average happiness.

Overall, this analysis focused on the comparison between model results based on different specifications (i.e., with and without the H_day variable and the alternative measures H_pastweek and H_pastday).

Results

Figure 1 shows the results of the reanalysis and the additional analyses conducted. The full model results are reported in Table S1 in the Supplemental Material. The red squares and error bars in Figure 1 represent the predicted values and 95% confidence intervals (CIs), respectively, of the odds ratios (ORs) of participants’ interacting with other people at time t + 1 depending on their level of happiness at time t. These values closely resemble the results reported by Quoidbach and colleagues, log OR = −0.003, 95% CI = [−0.003, −0.002], OR = 0.997, p < .001. The green triangles in Figure 1 represent predicted odds ratios of a model in which the H_day variable was removed. When the H_day variable was removed from the analyses, the association changed direction, log OR = 0.006, 95% CI = [0.005, 0.006], OR = 1.006, p < .001. Also, the raw data (i.e., means; gray dots in Fig. 1), which are not based on predicted values of a statistical model, suggest a positive association between happiness and the probability of being in a social interaction at the next time point.

OPEN IN VIEWER

Fig. 1.

Raw values (i.e., means) and predicted values of the odds ratios for participants’ reports of being with other people at time t + 1 (rather than being alone), depending on their previous happiness at time t. For predicted values, odds ratios are shown for models with the daily-average covariate H_day, without the daily-average covariate, with the daily-average covariate replaced by average happiness in the past week (H_pastweek), and controlling for average happiness in the past 24 hr (H_pastday). Full model results are reported in the Table S1 in the Supplemental Material available online. Error bars show 95% confidence intervals.

Instead of controlling for an individual’s daily happiness average, I created a further model that controlled for the individual’s average happiness over the past week. The predicted ORs based on the model including H_pastweek are reported as blue asterisks in Figure 1, log OR = 0.004, 95% CI = [0.003, 0.005], OR = 1.004, p < .001. The predicted ORs of a model containing the H_pastday variable are represented as blue squares in Figure 1, log OR = 0.003, 95% CI = [0.002, 0.004], OR = 1.003, p < .001. The models including H_pastweek or H_pastday suggest a positive association between happiness and subsequent social interactions. Further models using other alternative measures (e.g., past happiness measures of the current day, H_{earlier_in_day}; person-mean-centered happiness, H t PMC ) can be found in Table S2 in the Supplemental Material and the accompanying R script (https://osf.io/zk98q).

So why does only the H_day model reveal a negative association between happiness and subsequent social interactions? One reason might be the overrepresentation of H_t+1 in H_day, because a model with H_t+1 instead of H_day suggests almost identical results (see Table S2). After the model adjusted for the positive cross-sectional association (effect of H_t+1), log OR = 0.015, 95% CI = [0.014, 0.015], OR = 1.015, p < .001, previous happiness at time t was negatively associated with being in an interaction at t + 1, log OR = −0.003, 95% CI = [−0.004, −0.003], OR = 0.997, p < .001. In this case, the effect of H_t cannot be interpreted independently of the H_t+1 effect, as they are highly correlated, r(220292) = .70, p < .001, and part of the same model. Note that the H_t+1 effect is about 5 times larger than the effect of H_t. Hence, in order for the total contribution of happiness to be negative, H_t must be about 5 times larger than H_t+1, which is rarely the case in the observed data (2.5% of cases).

The nature of the interplay between H_t and H_t+1 can best be understood when looking at the change in happiness between t and t + 1 (i.e., using ΔH = H_t+1 − H_t). A model with ΔH instead of H_t+1 was estimated (see Table S2). This model essentially contains identical information to the H_t+1 model, but interpreting it is more straightforward. The results of that model suggest that if there is no happiness change between t and t + 1, the level of happiness at t predicts interactions at t + 1 positively, log OR = 0.012, 95% CI = [0.011, 0.012], OR = 1.012, p < .001. With each positive change in happiness, the likelihood of interactions increases, log OR = 0.015, 95% CI = [0.014, 0.015], OR = 1.015, p < .001. In the other direction, this also means that if someone becomes unhappy between t and t + 1, interactions become less likely. Hence, the residual negative effect of H_t on P_t+1 that emerged in the original analysis was likely driven by observations of individuals who were happy before (at t) and were not interacting at the subsequent time point (t + 1) because they had become less happy by the subsequent time point. The aspect of how future happiness (H_t+1)—as part of H_day—plays a role in the claimed association is not discussed in the original article.

Discussion

The results reported in this Commentary suggest that one of the main findings claimed by Quoidbach and colleagues is not robust. Quoidbach and colleagues reported a negative association between individual’s happiness (H_t) and the likelihood of future interactions (P_t+1). However, in my reanalysis, I found that the raw data and statistical models, in which the control variable H_day was removed, suggest that the association between happiness and the likelihood of future interactions is positive. Also, in models with alternative specifications with average happiness in the past week (H_pastweek) or the past 24 hr (H_pastday), the stated association was estimated to be positive.

The proclaimed negative effect of happiness on subsequent social interactions seems to be biased by the presence of unreported H_t+1 information in the model. The overrepresentation of H_t+1 in H_day is neither discussed nor considered in Quoidbach et al.’s interpretation of effects. This reanalysis suggests that if individuals were happy before and they were not interacting at the subsequent time point (the proclaimed negative effect of H_t), it is likely that they became unhappy between the two measurement time points. Although dynamic investigations of happiness and social interactions are novel, the finding that less happy individuals are less likely to socially interact has been frequently demonstrated (e.g., Elmer & Stadtfeld, 2020; Pavot et al., 1990; Sandstrom & Dunn, 2014). This reanalysis is in line with that literature.

A limitation of this reanalysis is that the alternative covariates (H_pastweek and H_pastday) were not computed on the complete data set, that is, before the data were reduced to only subsequent reports within 12 hr. The calculation of the alternative covariates was based on the reduced data used for the present reanalysis (see https://osf.io/bxgn4 for data).

The findings regarding the probability of participants’ interacting with certain interaction partners (e.g., best friends) when they were previously unhappy (see Quoidbach et al., Fig. 3a) was also affected by the H_day variable. See the Supplemental Material for additional analyses of these findings.

It is important to note that the intent of this Commentary is not to undermine the work of Quoidbach and colleagues. The importance and novelty of their research is without question. Nevertheless, the findings related to the prediction of social behavior (see Quoidbach et al., Figs. 2 and 3a) are not robust. Fortunately, Quoidbach and colleagues published their data online, which made this reanalysis possible. Beyond further encouraging open-science practices, I suggest that raw data and results of stepwise models should be more frequently reported in the psychological sciences.

Supplemental Material