Trajectories of Latent Vulnerability and Distress: Identifying Social and Spatial Fringes of the Swedish Population

This analysis assumed that latent vulnerability in register data is present in unobserved subgroups of individuals who experience short or long periods of unemployment, illness, low income, poverty, or have a need for social allowance or other welfare support, or live with a partner who requires such support (Bask, 2010). In addition to the above indicators, we included type of housing, whether equalized disposable income was larger than median, whether the household has children and whether they were living in a couple, to fully describe trajectories and for latent classes to relate to one another, beyond just indicators of poverty and vulnerability. To extract such trajectories from the data, we used LCA, MLR in the software Mplus (Henry & Muthén, 2010) for cohorts born between 1945 to 1980 for the years 1990 to 2019.

We used longitudinal, individual, registry data accessed on-line in MONA (Microdata Online Access) at the Statistics Sweden, SCB. The specific collection of registers and variables used is part of a project called Geographical context, which covers data for all residents in Sweden, from 1990 to 2019 and has been approved by the Swedish authority for ethical vetting (approval no. 2020–05521). Data is gathered by Swedish authorities and delivered to SCB, and SCB is keeping the longitudinal microdata available for researchers. All data is anonymized and has an individual reference number unique to the Geographical Context project. The attrition rate for each of the life phases was low, about 150 000 individuals per life phase was deleted due to not living in Sweden more than 9 years out of 15 years (1.8 million per life phase, Table 2). Furthermore, certain individuals were removed due to incomplete data concerning income, geo-coordinates, or other relevant indicators. Note that the attrition observed was not systematic.

LCA is a valuable method for reducing large numbers of trajectories to a smaller number that captures the variation in timing and sequence in life course transitions (Amato et al., 2008; Virtanen et al., 2011). Transitions during the life course occur in different order and at different ages, which can give rise to many different types of trajectories. With finite mixture modelling (Vermunt & Magdison, 2004), longitudinal data across different life domains can be used to identify typical, latent life course trajectories which individuals follow.

In a first step, the individual’s life trajectory was represented by indicator variables that specified for each year whether the individual had received e.g., social allowance (Ilmakunnas & Moisio, 2019) and so on for all indicator variables. With data covering the period from 1990 to 2019, individuals were followed for 30 years, but the analysis focused on shorter periods of 15 year age-windows, with a 5 year overlap, in order to follow ages in three life phases. In Table 1 the ages of individuals born in specific years are shown (we thus protect our analysis from cohort specific patterns) for the period 1990 to 2019 for which we have the data. The three life phases are marked in Table 1. With indicator variables representing each individual year, the individual’s life trajectory was captured by about 180 different variables (15 years × 12 indicators). (We do not assume independence over time, because the LCA allows for correlations between all observed variables (over time, as well as both within and between different indicators). Indeed, this is one of the strengths of our approach, namely that there are no assumptions about temporal relationships (e.g. Markov)).

In order to arrive at a number of five latent classes, LCs, we analysed models with three to eight different LCs and analysed reductions in the Bayesian information criteria (BIC), which indicated that there is less to be gained from increasing the number of classes above eight² concerning BIC, but also less to gain for reasons of presenting the analysis. Figure 1 shows the 3, 4, 5, 6, 7 and 8 classes solutions for the three life phases and the BIC value. The process of arriving at a number of classes balances between theoretical and statistical trade-offs (Nylund-Gibson et al., 2022, p. 8). Apart from collecting and tabulating fit for different models (BIC) and study patterns (compare different number of classes) we also took into account the possibility to interpret the results and also to make usable maps when deciding on the number of classes. The aim of this study was to identifie, define and further explore vulnerability in a spatial perspective, which is not always the case. For that reason we were interested to first, single out latent vulnerability (rather coarsely with the 5 LCs), and second, to analyse their geographical locations, and in a third step use the distilled two LCs to spatially explore distressed LCs. With register data follows a large population, here we have 1.8 million per life phase included in the LCA and presumably a large possible number of latent classes with the number of indicators we included. A large population, as available to us, and a large number of indicators (12) do seem to lead to good performance of LCA and stability in class solutions (Wurpts & Geiser, 2014). For our initial purpose of finding vulnerable life courses to further analyse, we believe we captured those life courses of interest with the number of latent classes we have chosen.

To further compare patterns and decide on how many classes best describes the data, we evaluated the 5 class model and the 7-class model exploratively for all ages and with more descriptive variables. For example, when we compared 5-classes and 7 classes model, we detected a divide in the latent classes that could be characterized by early, respective, late family formation if choosing 7 LCs instead of 5 LCs for the youngest life phase. Thus, for our purposes, we concluded that the content of the 7 class solution did not add any important or meaningful information (description of classes and class sizes in results). Furthermore, assignments of individuals to a class in the models for the life phases were done with a high probability, ranging from 0.975 to 0.991 (shown in Table 1 Supplementary information). As the average latent class probability were high we interpreted the models to predict class membership well. Also, Entropy values greater than 0.80 indicate a good separation of the identified groups and our estimates were above 0.97 for the three models. In a second step, to identify vulnerable groups, we proceeded with using the vulnerable groups only in an LCA of distress applying the same indicators and the same procedure.

Among indicators of vulnerability, one important variable was ‘at risk of poverty’ (AROP), based on equalized disposable household income, coded 1 if income was less than 60 percent of the national median disposable income of the total adult population. Moreover, we included an indicator that we called in stress, which combined receipt of sick leave compensation, unemployment benefit, or early retirement (long term sickness benefit for unemployed) benefit for an individual. Social allowance is sign of poverty, as it is an allowance that is means tested by the local social authority according to national regulations. The housing allowance indicator shows if costs for housing is supported, most often the support is received by households with children but also other types of households may be eligible. The housing allowance is directed to households irrespective of their housing tenure form. The indicator alive, only for individuals above the age of 54 that is, only for the oldest in the oldest life phase, is a clear indicator of vulnerability and of health, in the life phase under investigation.

We also included two variables of the economic situation of the partner. These were firstly; based on partner unemployment, partner sickness, partner early retirement and secondly; whether the partner had low earned income. The first translated into partner in stress.

The indicator of being in a couple or not (Perelli-Harris et al., 2016) was based on family status—while divorced and never married can be single. A second demographic variable indicated presence of children in the household. It was based on family type, Table 2. Being in a couple or living single and having a child in the household is not part of the vulnerability examination, it is rather a characteristic to differentiate between types of vulnerability. The first income variable was based on earned income and was coded 1 if earned income was above 50 percent of the median, that is, a fairly good income in relation to the group. We also used the indicators of low earned income as well as high earned income (earned income included primarily labour income, but also entrepreneurial income). Residing in a single family house as opposed to living in an apartment is an indicator of life course circumstances included to differentiate between latent classes. Single family housing is privately owned and specify financial resources and to some extent geographical location.

In order to distinguish between several latent classes of vulnerability some variables included are not necessarily indicators of vulnerability as; living in a couple, having children in the household, an equalized disposable income larger than median or living in a single-family house. We argue however that these indicators give a fuller picture of the latent classes. For example, these types of indicators might help us to distinguish between two types latent classes both showing high proportions of vulnerability, but that differ due to their household characteristics. We argue that individuals in these types of situation, where one is living with a partner and have children in their household and the other, living single and no children in the household, have taken different paths in their life course trajectories. This also implies that their experiences of vulnerability will be different and also solutions to overcome it. We argue these individuals have taken different life course trajectories. In this way, the non-vulnerability indicators are there to distinguish between multiple latent classes.

Although important for vulnerability, we were not able to capture whether unemployment or social allowance receipt were caused by substance abuse, mental illness, physical disabilities or intellectual disabilities since our data did not include these strands of indicators. We believe that these conditions are not excluded but rather hidden (to a certain degree) in sickness benefit, unemployment and early retirement (long term sickness among unemployed) indicators. Another limitation when it comes to measuring vulnerability with Swedish register data is that we could only capture individuals who were registered in Sweden; thus, unregistered individuals (expelled individuals staying in the country, undocumented refugees, victims of trafficking and more) who often experience vulnerability were not included in our study.

Further variables were used after the LC analyses to descriptively analyse the different trajectories found, in Supplementary information. Furthermore, we also used maps to analyse the geographical patterns across LCs. For the maps, proportions of the LCs across neighbourhoods of 200 individuals (total population) based on east and north coordinates were used (Hennerdal, 2019).

The life phases: Young adulthood, 25 to 39 years; Early middle age, 35 to 49 years; and Late middle age, 45 to 59 years partly follows on the nomination by Slotkin (1954) in which he, based on earlier research, labels ages from 28 the stage of maturity. However, today ‘maturity’ is probably delayed compared with the 1950s, why we label the youngest cohort from 25 years, Young adulthood. Ages starting at 43 Slotkin labels “testing stage of early middle age”, and from age 48 an “indulgence in later middle age”, and finally, not included in this study, ages from 64 and up he calls the “completion stage of old age” (p. 171).

In each life phase, we found two latent classes that could be identified as vulnerable by our own examination of proportions of the indicators; in stress, AROP, social allowance, low earned income and low income, compared to the other latent classes, see Fig. 2 and Table 3. In Fig. 2 proportions of each indicator for Young adulthood, Early Middle age and Late Middle age are shown for an age span of 15 years. The diagrams are read as e.g. LC4, Affluent norm (25–39 years), where the indicators Child in household and living In couple is rising to over 90% and income indicators are high and, vulnerability indicators are low.³ The size of the LCs is fairly even. For Young adulthood the proportions of individuals in LCs 1 to 5 are, 0.25, 0.19, 0.12, 0.20 and 0.23. For Early middle age the proportions in LCs from number 1 to 5 are, 0.27, 0.16, 0.18, 0.15 and 0.24 and lastly for the Late middle age cohort the proportion in LCs 1 to 5 are, 0.13, 0.25, 0.25, 0.19 and 0.18.

The two top rows of five diagrams are accounted for in Table 3, descriptions of vulnerable life courses. The other latent classes in diagrams, that do not show vulnerability are also shown in Fig. 2, but since they are out of the scope for this paper, they will not be addressed further. In the third and final step of the analysis, we will return to the vulnerable LCs for the Late middle age, which are used for another LCA to distillate different distressed life course trajectories only.

In sum, the vulnerable latent classes could be distinguished from a total of five latent classes per life phase, all of which show more favourable development. The fact that life courses are described as singles and families/couples, we do not argue being a finding, it is just a part of the description of different types of vulnerability and type of life course. Typically, the latent classes with favourable developments consist of households that have attained higher incomes and earnings, where a majority enter the home-owning sector and few experience stress or low earned income.

For all the life phases, the total proportion of the two vulnerable LCs is about 31 percent. That is, we find that 31 percent of the individuals across the three life phases we have studied, are living in a vulnerable position in society. This might seem a large proportion but it actually corresponds to the proportion that has been found in previous research on the Swedish population, namely the proportion of the population having welfare problems/living in deprivation as measured by (Halleröd & Larsson, 2008).

In the second step, we investigated the geography of the vulnerable latent classes that we have described above. We did this by looking firstly at the distribution of vulnerable latent classes across municipality types (Gillingsjö & Ekholm, 2016) and secondly, across neighbourhoods classified on the basis of the distribution of the local population, nearest 1600 neighbours, income deciles contexts, Table 4, and thirdly, by analysing the maps in Fig. 3. Income decile context deserves some explanation. It is based on a variable of deciles of equalized disposable income for the population 15 years and older. It is measured among 1,600 nearest neighbours for the years 1990–2015, all years, pooled data for 26 years for 214,644 geo-coordinate pairs. A cluster analysis resulted in 12 clusters: Very high income 1, 2, High income 1, 2, Mixed 1, 2, 3, 4, Low income 1, 2, Very low income 1, 2, also used in Andersson et al., 2022. The Very high income 1 cluster has a high concentration of individuals in the highest income decile. The Very low income 2 cluster, conversely, has a high concentration of individuals in the lowest decile. In between these extremes, the clusters have progressively fewer individuals in the highest income deciles and progressively more individuals in the lowest income deciles. The most even distribution is found in the Mixed 2 cluster, where each decile is represented by close to 10% of the population.

The final step in analysing vulnerable latent classes was to perform a second LCA including only LC1 and 5 for the Late middle age life phase 45 to 59 years. This explorative part of the analysis thus included 567 437 vulnerable individuals divided into five new latent classes, only for Late middle age, Fig. 4, hereafter called DC, distressed classes. The life phase chosen is advantageous to study exploratively in that the ages 45–59 are expected to be stable and established concerning education, employment and children, and we were also able to include a proxy for health, which is if the individual is alive or not at the end of the trajectory.

The five distressed latent classes, DCs are presented in Fig. 4 below. As we can see in the top left frame in Fig. 4, a special characteristic of DC1, named Less distressed couples, multifamily houses, is that individuals are predominantly in a couple (green line) and that they are living in an apartment (low proportion of individuals in a single family house). About 50 percent of individuals in this trajectory type are in stress, that is, and to repeat; they receive sick leave compensation, unemployment benefit, or early retirement/sickness benefit. Also, they have low earned income including about 40 percent of individuals. The stress and low earned income are a concern for their partner as well. Very few have high earnings, but a good proportion, about 30–50 percent, have above median for disposable household income, hence the name ‘less’ distressed.

A first characteristic of DC2, in the top right frame in Fig. 4, Increasingly distressed singles is their low earned income; that they are singles without children, but 35 percent are living in a house. Averaging over the life phase, 70 percent are in stress and with low earned income, which seems comparably high. Also, at risk of poverty increases over the life phase to almost 40 percent.

Distinguishing DC3, distressed low survival, is that only 60 percent are still alive at 55 years. One out of five has a partner in the early phase but later live as singles. When the life phase starts at 45 years of age the trajectory shows about 35 percent being in stress. Also, at 45 years, about 25 percent live in single family housing with a housing allowance; however, both proportions decline with age, thus apartment-living dominates.

Characterizing DC4 Extra distressed with social allowance, is again a large proportion of individuals with low earned income, but also at risk of poverty and receiving a social allowance. Half of the individuals following this life course are in a couple and starting the life phase with a child at home. The trajectory shows the largest proportion of at risk of poverty out of all distressed classes. The majority live in an apartment, but receive housing allowance. About half in this trajectory type are in stress. A proportion of 40 percent have a social allowance.

The DC5, Less distressed couples single family housing, are the ones with the largest majority in single family housing (> 90 percent) as well as living in a couple (comparable to DC1 except for housing). About half have low earned income and are experiencing stress, and some are at risk of poverty. Their partner is to a similar, but slightly lesser extent, with low earned income and experiencing stress as well.

The observations from the profiles in Fig. 4 are confirmed in that distressed classes 1 and 5 with ‘less distress’ include individuals that in this sample have a slightly higher education (19 and 18 percent respectively), higher mean disposable income and are to a lesser degree workers (81 and 82 percent respectively) (except DC3 at 75%).⁴ However, Halleröd and Larsson state precisely that income poverty measures do not always capture individuals living under poor conditions and suffering from welfare problems (2008). In this respect we use the term latent vulnerability.

The vulnerable latent classes found in the first step of analysis, did show high proportions of individuals born outside of Sweden. However here it is important to note that with the refinement of distress in several LCs only two classes come out with high percentages of people born in another country. This is for the life course DC1 (43%) that is concentrated to the metropolitan areas and cities; they are not dispersed over Sweden at large, and for the DC4 Extra distressed with social allowance, that have a high proportion of both born outside of Sweden (61%) and visible minorities (36%).⁵