Table 1 reports descriptive statistics of the individual and neighbourhood characteristics in the baseline assessment. The average age is 44.67 years (SD = 9.67), with most participants being female, Dutch-born, and having middle to high education. Most participants cohabit with their partner and children, report a good social circle and are non-smokers. On average, participants report 31.23 paid working hours per week. The average home value at the neighbourhood level is €196,640. The sample shows a slightly higher proportion of Western (non-Dutch) than non-Western populations, with most participants living in rural areas. The average distance to a GP is 5.31 km (SD = 7.24), indicating variability in healthcare access. The reported percentages of MDD of different categories in the case of discrete characteristics also differ significantly from each other.

The results of the Bayesian comparison test, applied to the linear-in-means model without peer effects, correlated effects, or both, are reported in Table 2. With a posterior model probability of 0.9988 in the first wave and 0.9991 in the second wave, a linear-in-means model with contextual and correlated effects is 900 to 1250 times more likely than its counterpart with contextual and peer effects. Apart from a few studies on depression [20], there is also generally little evidence of peer effects on MDD in adults. One possible explanation is that MDD represents an extreme manifestation of depression, typically characterised by a high degree of isolation. Consequently, not only is the presence of MDD less frequently observed, but individuals diagnosed with MDD also tend to have less interpersonal contact, which in turn influences the likelihood of peer effects.

Table 3 presents the estimation results of the two-step approach applied to the linear-in-means model (Model 1), comprising Step 1 and Step 2. The covariates in this model include individual and neighbourhood characteristics, referred to as “Own characteristics”, as well as contextual and correlated effects, referred to as “Social interaction effects”. Step 1 reports the AMEs estimated using a probit model based on a pooled sample of two cross-sections of 19,700 participants in both waves. Step 2 reports parameter estimates that can be interpreted in a way similar to AMEs. These estimates were obtained from a linear regression model based on a pooled sample of 614 participants from the first wave and 689 participants from the second wave. For comparison, and to assess the implications of accounting for the right-skewed distribution of MDD, particularly differences between diagnosis and severity, we also report the results of a naïve specification (Model 2) that does not incorporate this distinction between both steps. As in Step 2 of Model 1, Model 2 provides parameter estimates from a linear regression model, and as in Step 1 of Model 1, these estimates are based on pooled data from two repeated cross-sections of 19,700 participants.

The first important finding is that the results differ not only between the naïve and two-step models, but also between Step 1 and Step 2. The former is driven by the observation that most characteristics have larger parameter estimates in the naïve model than in either step of the two-step model, indicating that the naïve model tends to overestimate their magnitude. The latter is driven by the observation that many individual characteristics significantly associated with the diagnosis of MDD (6 out of 12) are not significantly associated with its severity. Together, these results highlight the importance of distinguishing between the diagnosis and severity of MDD.

The second finding is that the number of significant social interaction effects is substantially higher than the number of significant neighbourhood characteristics. Of the seven neighbourhood characteristics considered, only the average home value appears to have a significant impact on the probability of being diagnosed with MDD (AME =  − 0.08, p < 0.05) and a weakly significant impact on its severity (− 1.64, p < 0.1). In contrast, five contextual effects significantly affect the probability of being diagnosed with MDD (Step 1, column “Social interactions effects”), and three affect its severity (Step 2, column “Social interactions effects”). The results of Step 1 further show that participants surrounded by people with secondary education (Education level: Middle) and/or more paid work hours per week are less likely to be diagnosed with an MDD episode (AME =  − 0.02, p < 0.1; AME =  − 0.08, p < 0.1, respectively). Conversely, participants surrounded by childless neighbours (AME = 0.02, p < 0.05) with poorer social ties (AME = 0.08, p < 0.05) and a history of depression are at greater risk of being diagnosed with MDD (AME = 0.06, p < 0.01). The results of Step 2, in turn, show that participants diagnosed with MDD surrounded by neighbours with a history of this depression (0.79, p < 0.05) are particularly vulnerable, unless they are male (− 0.84, p < 0.05) or do not cohabit with a partner (− 0.78, p < 0.1). Moreover, both steps of Model 1 show significant correlated effects between the error terms, which align with the posterior model probabilities from the Bayesian comparison test in Table 2. This suggests that shared unobserved characteristics within the neighbourhood, such as stigma or environmental stressors, influence both the diagnosis and the severity of MDD.

The third and last finding is that the individual characteristics remain the strongest determinants of MDD. Nine of them significantly contribute to the diagnosis of MDD, and four of these also affect its severity (Table 3, column “Own characteristics” of both Step 1 and Step 2). For example, people with a fair or poor social circle, rather than a good one, are at greater risk of being diagnosed with MDD (AME = 0.05, p < 0.01 and AME = 0.11, p < 0.01, respectively), as well as more severely affected (0.16, p < 0.05 and 0.52, p < 0.01, respectively).