2.1 Respondents and Procedure
This study included 11,487 adult respondents (age 18
+) of the ongoing Dutch crowdsourcing study HowNutsAreTheDutch (HND; Dutch: HoeGekIsNL), who reported on the fulfilment of their social needs.
2 HND allows respondents to investigate their mental health dimensions and compare themselves to others by visiting the internet platform
hoegekis.nl. All procedural details are provided elsewhere (Van der Krieke et al.
2016). Respondents first completed a sociodemographic questionnaire (n = 14,489), after which they completed one or more topical modules (e.g., living situation, mood, or well-being). Respondents were in majority women (67%), on average 45 years old (
SD = 15), and higher educated (76%). Sampling biases were intended to be accounted for by weighting our results to match demographic strata in the general Dutch population.
The HND study protocol was assessed by the Medical Ethical Committee of the University Medical Centre Groningen. The committee judged the protocol to be exempted from review by the Medical Research Involving Human Subjects Act (in Dutch: WMO) because it concerned a non-randomized open study targeted at anonymous volunteers in the general public (registration number M13.147422 and M14.160855).
2.3 Analytic Plan
All analyses were performed in R (R Core Team
2018), unless specified otherwise. R scripts are provided at
https://osf.io/njwuy/. Descriptive statistics of the sample and age groups were derived first, including group means and their standard deviation (SD), and standardized mean difference
d between the age groups, heuristically interpreted as small from .20 to .40, medium from .41 to .79, and large after .80 (Borenstein
2009). The association between the social needs was examined using correlations and scatterplots. We classified correlations (
r) as very weak if between .00 and .29, weak between .30 and .49, moderate between .50 and .69, strong between .70 and .89, and very strong from .90 onwards (Mukaka
2012). In this study, only estimates with alpha levels below .001 will be interpreted to avoid type I errors.
Distinct social need fulfilment profiles were explored with Latent Class Analyses (LCA) in LatentGOLD (Vermunt and Magidson
2005). LCA is particularly useful for cluster analyses with a small number of items and a relatively large sample size (Eid et al.
2003). As only 16 respondents fell in the oldest-old group, this group was excluded from all cluster analyses. We fitted various types of LCA models, each with 1 up to 10 clusters. For each model, the number of clusters was identified by selecting the solution with the lowest BIC value. First, to examine cluster solutions for the general population, the three ordinal social need variables were clustered in the total sample and the weighted sample. Second, the same cluster analyses were performed for the five age groups separately. Third, three multi-group LCA’s were performed to examine the differences between age groups in more detail.
The first multi-group LCA has an equal number of clusters across all age groups, but the model parameters may differ across the age groups. This analysis gives insight into differences in cluster solutions between the different age groups. The second multi-group LCA is constrained insofar that the association between the latent classes and the social need variables is equal for all age groups, while the cluster sizes may differ across the age groups. In this way, the interpretation of the clusters is equal across the age groups, and we could examine if certain clusters are only present in specific age groups. The last multi-group LCA was most constrained because the equality constraints of the second model were applied extended with the constraint that the cluster sizes are equal across age groups. For this model, we took the same number of clusters as we had identified in the second model. A deviance test was applied to test whether the cluster sizes were equal across age groups. The multi-group LCA’s were performed using both the weighted and unweighted sample.
The differences in social need fulfilment between the age groups were examined using descriptive statistics, ANOVA analyses and pairwise t-tests. To examine whether comparisons of average social need fulfilment levels between age groups are valid, we fit measurement invariance tests (Supplement 5), which supported configural, metric, scalar, and strict invariance (CFI > .96, RMSEA < .06; Putnick and Bornstein
2016), in line with previous reports in five independent samples (Steverink et al.
2019).
Finally, age group differences in the associations between the social needs and happiness were examined in a series of regression analyses. As the relative benefit of a social need for happiness might depend on the level of need fulfilment, it is important to control in these analyses for the average levels of need fulfilment in each age group. For example, based on social emotional selectivity theory, we expect that status is stronger associated with happiness in younger adults compared to older adults. Simultaneously, younger adults are expected to have lower levels of status fulfilment than older adults as they still have to develop skills and a career. To ensure that our interaction effects do not reflect diminishing marginal returns (the higher the need fulfilment, the lower the relative benefit), our regression models are adjusted with quadratic terms of the social needs (see Nieboer and Lindenberg
2002).
In total, four regression analyses were performed. First, the association between social need fulfilment and happiness is examined (model 1), adjusted for gender, education level, and diminishing marginal returns (model 2). Second, the association between social need fulfilment and happiness is compared between age groups via their interaction terms (age group*social need, see model 3), which was also estimated adjusted for gender, education level and diminishing marginal returns (in model 4). All regression models were repeated using sample weights to derive more generalizable conclusions.
In general, conclusions from weighted and unweighted analyses were very similar. Therefore, we mainly discuss the results from the weighted analyses. Only when the results differ, we will also discuss the results from the unweighted analyses.