We analysed health and accommodation data of 412 refugees and asylum seekers from up to 58 different accommodation facilities in the federal state of Baden-Württemberg in the south-west of Germany, collected in the RESPOND study [12]. For the latter, RESPOND researchers developed and validated an instrument, the “Small-area Housing Environment Deterioration” (SHED) index assessing different domains of housing and its physical environment [13]. The geographical location was specified at district level. A binary variable indicated population density; <150 inhabitants per km² was considered as rural, else urban, following the standard BBSR (Federal Institute for Research on Building, Urban Affairs and Spatial Development) definition. The distances to shops, medical and administrative services by foot and with public transport was estimated using Google Maps (GM) routes. Besides accommodation variables, the RESPOND study has also assessed a range of (subjective) physical and mental health outcomes, including a screening of GAD, depression symptoms, and the subjective general health status [12]. They were measured using the Generalized Anxiety Disorder GAD-2 [14], the Patient Health Questionnaire PHQ-2 for depression [15] as well as a general health indicator drawn from the European Health Interview Survey (EHIS), [16].

We performed a cluster analysis with the RESPOND variables that contained information about the accommodation and its physical context, i.e., the number of inhabitants, remoteness (walking distance to shops, medical or administrative services), urbanity (urban/rural distinction), and housing decay. For the latter, the overall SHED score as well as the normalized and z-standardized single scores were included in the analysis to identify housing differences. After exploring different cluster algorithms and methods, we chose a hierarchal, agglomerative clustering algorithm using Euclidean Distance and Ward’s method. We identified the number of clusters by visually assessing dendrogram and scatterplot, before adding the clusters as separate variables to the data set, representing different accommodation types. Additional thorough descriptive statistics were reported in previous articles [12, 17]. After bivariate analyses between clusters and health outcomes, we performed logistic regression analyses to obtain associations between the identified accommodation clusters and binary measures of general health, GAD, and depression adjusted for potential confounders. Both PHQ and GAD were dichotomised with scores above a cut-off value of 3 considered to be indicative of a depressive or anxiety disorder, respectively. General health was measured on a 5-point Likert scale and dichotomised at a cut-off value of 3 and upwards. The independent variables included age, gender, and the cluster variables. In subsequent analyses, we assessed how the distribution of GAD and depression differed between the clusters. We created three distribution variables (GAD or depression present, both present, both absent) and used these distribution variables as dependent variables in additional logistic regression analyses. All statistical analyses were performed with R (version 3.6.2).

Descriptive statistics of the variables included in the cluster analysis are provided in Table 1. Complete accommodation data of 411 refugees and asylum seekers were included in the cluster analysis. The resulting dendrogram and scatterplot pointed to either two or four clusters. To keep the within-group variance of the agglomerated clusters as high as possible, we decided to construct four clusters. A two-cluster solution would have led to a high level of dissimilarity. Cluster 1 and 3 were the smallest, containing 39 respondents of the RESPOND study each (from two and nine different accommodation centres, respectively). Cluster 4 contained 143 respondents from 29 different accommodation centres, and Cluster 2 190 respondents from 29 accommodation centres.

Descriptive statistics of the health outcomes are provided by clusters in Table 2. The four clusters showed differences in the health status of their respondents. In all three health variables, residents in Cluster 4 showed better health outcomes than in the other clusters, as Table 2 illustrates. For the general subjective health status, only 12.3% of Cluster 4 residents indicated poor general health, while Cluster 3 scored the highest proportion with 25.7%. The differences were not statistically significant, though. Symptoms of impaired mental health were more common than of impaired general health in all clusters. In terms of GAD, more than one in two residents of Cluster 2 indicated symptoms. For Clusters 1 and 3, proportions were also relatively high, with more than 42%. Cluster 4 residents scored lower, around one-third indicated symptoms here. The unadjusted differences in the frequencies of GAD symptoms between the clusters were statistically significant [x ² _{(df = 3)} = 13; p = 0.005]. The proportion of respondents with depression symptoms was lowest in Cluster 4 as well, with approximately 39%. In Clusters 1, 2 and 3, in contrast, around 1 in 2 people indicated symptoms of depression. Differences between the clusters were not statistically significant.

An overview of the results of the regression analyses for poor general health, GAD, and depression (PHQ-2) is provided in Table 3. In the age- and gender-adjusted Model 1, Cluster 1, 2 and 3 all showed increased odds for residents having a poor general health status than Cluster 4. Cluster 2 showed the highest odd ratios here and was nearly significant at the 0.05-significance level (OR = 2.05; 95%-CI [1.00–4.46], p = 0.058). The adjusted Model 2 showed a significant association of GAD with accommodation type in Cluster 2 (OR = 2.7; 95%-CI [1.55–4.79], p = 0.001) and Cluster 3 (OR = 2.4; 95%-CI [1.04–5.64], p = 0.042) relative to Cluster 4. In the adjusted model for depression (Model 3), all clusters showed odd ratios larger than 1 relative to Cluster 4 but no association was statistically significant (p-values between 0.13 and 0.62).

In the sample, depression and anxiety disorders were correlated (r = 0.52). We additionally assessed the distribution of depression and GAD in the clusters (see Figure 1) and analysed how regression results changed when using these distribution variables as dependent variables (see Table 3). Higher proportions of residents in all clusters showed both symptoms than GAD symptoms without depression, or vice versa Only cluster differences found for both symptoms present were statistically significant [x ² _{(df = 3)} = 8; p = 0.05].

In the regression analyses, we found age- and gender-adjusted significant differences between Cluster 2 and 4 (OR = 2.3; 95%-CI [1.24–4.10]; p = 0.009) in the model for comorbid GAD and depression (Model 4, Table 3). A model for the absence of both conditions (Model 5) supported these results, again indicating significant effects for Cluster 2. Further, a model for the presence of either GAD or depression symptoms could not identify any statistically significant cluster effects.

We have addressed contextual variabilities of collective refugee accommodation centres in Germany to investigate associations between housing and refugee health in a comparative way. By using a cluster analysis approach, we identified four distinct types of refugee accommodation. This enabled us to go beyond regular multivariate analyses based on individual contextual characteristics and focus on contextual combination of different characteristics when assessing their associations with health and presenting health disparities.

Cluster size varied considerably, ranging from 39 cases to 190 respondents. Since one of the smaller clusters was consistent and repeatedly occurred in different pre-analyses with varying clustering approaches, we decided to accept the variability in cluster size (and the small number of accommodation centres cluster 1 was based on), being aware that the clusters can be compared with each other only to a limited extent.

In regression analyses, only few effects were found to be statistically significant. This may be due to the small sample sizes in some of the clusters. Especially Cluster 1 failed to yield significant associations relative to Cluster 4. Another statistical issue was missing values in the data for GAD and depression. Whereas the percentages of missings were relatively even in Cluster 1, 2 and 4, Cluster 3 had lower rates of missings (see Table 2). Since Cluster 4 contained the highest number of missings overall with 17.5% (GAD) and 18.3% (depression), associations found in this analysis could have been weaker and frequencies of residents showing GAD and depression symptoms could have been higher. By using PHQ2, we may have underestimated negative mental health effects besides depression. Future studies should further explore the concept of “demoralization” [31] among asylum seekers and refugees, especially given its relevance for suicidal ideation and behavior [32].

Further, we could identify few cluster differences regarding urbanity. Given our definition of urbanity, it is rather unsurprising that each accommodation type identified here was predominantly located in urban districts. For example, in Baden Württemberg, where data has been collected, overall population density was 311 inhabitants per km² in 2019 [33], far above the cut-off applied here. However, as refugees are distributed to districts according (among other factors) their population size, the majority of refugees will be accommodated in more urban areas.

Also, we could not investigate to which extent the residents of the larger accommodation types lived in overcrowded settings according to the Eurostat definition since the data did not provide information about room numbers of the accommodation centres assessed. However, with an average number of residents of up to 150, we can assume that the specifications of the Eurostat definition are not met, and that residents of these accommodations (clusters 1 and 2) share rooms. As the Eurostat definition applies to private households rather than to collective accommodation centres, a more appropriate concept should be developed.

Additionally, given the important role of work and workplace in potentially shaping health as well as facilitating societal and language integration, future analyses could be enriched by data on transient employment and workplace distance. At the same time, an overall such effect may be subsumed under our data on remoteness in terms of distance to shops, medical or administrative services.

The cross-sectional design of the data allowed us to identify only associations between the accommodation cluster and unfavourable health outcomes. Longitudinal approaches are needed to assess whether the accommodation types are causally associated with certain health outcomes. O’Brien et al. [24] for example argued that poor mental health may also result in more negative assessments of the physical environment. Since the SHED index was assessed by researchers, not by residents themselves, this form of same-source bias is unlikely here. It may be possible, though, that researchers would consider the accommodation as run-down while refugees might appreciate the conditions compared to other accommodation settings they have experienced during flight.

Future research should also account for the stage of the asylum application process of interviewees. For example, it is possible that Cluster 4 contained a lower proportion of residents whose asylum claim was still under process, which has been found to be a significant predictor of mental health problems [34]. Since asylum seekers live more often in deprived areas with unfavourable living conditions [35] like a run-down physical environment, the pathway of how exactly accommodation type impacts health warrants further research.

The context in which people live can affect their health in many ways. By conducting a cluster analysis of variables related to refugee accommodation in Germany, we have identified four different accommodation types reflecting contextual differences in the number of inhabitants, the degree of housing deterioration, and the distance to important amenities, thereby contributing to the development of a typology of refugee accommodation. The accommodation type with a moderate number of occupants, lowest levels of deterioration, and a central and urban location was associated with favourable health outcomes in terms of subjective general health status, depression, and GAD—the latter most strongly. The odds of unfavourable health increased for accommodation clusters that either showed a high state of decay and a large accommodation size (Cluster 2), or a small accommodation size remote from important amenities (Cluster 3). Effects were weakest for depression. Subsequent analyses showed, though, that when assessing comorbidity of GAD and depression, the larger, more deteriorated accommodation type (Cluster 2) yielded similar significant effects compared to the analysis of GAD alone. The findings from this study should be considered in political decisions on the layout as well as the choice of location of collective accommodation centres in Germany and comparable countries.