Internal migration following adverse life events: Assessing the likelihood of return migration and migration toward family

There are well-documented associations between life course changes and migration; yet, the occurrence, order, and timing of reasons for migrating are growing increasingly diverse. Migration following adverse life events, such as a divorce or an involuntary job loss, may be qualitatively distinct from migration undertaken for other reasons. Moves, especially long-distance moves, following adverse life events, may be defined more by seeking family and familiar locations. Moreover, a heightened probability of migration may occur not only immediately after an adverse life event but also in the years after. We explore these questions in the US context with longitudinal data from the 1983 to 2019 waves of the Panel Study of Income Dynamics, which provides information on residential locations, locations of family members, and adverse life events for individuals over time. We focus on five specific events: divorce, the death of a spouse, involuntary job loss, the onset of a chronic physical health condition, and the onset of a chronic mental health condition. Using multivariate regression, we find that divorce and job loss induce long-distance moves, especially return moves and moves towards family. Chronic physical conditions deter moving in general but increase the chances of return moves (after a period of time) and moves towards family. These results have implications for understanding migration as a response to adverse life events, both immediately and over time.

marriage, the birth of children, or retirement frequently trigger moves, including longdistance moves (Clark & Davies Withers, 2007;Kulu & Milewski, 2007).Often, these moves are made by people seeking to raise their standard of living via employment or educational opportunities or find an environment offering better amenities (Bernard & Kalemba, 2022).Migration is also prevalent following difficult and disruptive events, such as marital dissolutions (Cooke et al., 2016), housing crises (Stawarz et al., 2021), climate disasters (Warner et al., 2009), or crime victimisation (Xie & McDowall, 2014).Often, these are return moves, back to a parental home (South & Lei, 2015), or closer to family support networks (Spring et al., 2021;Wall & Von Reichert, 2013).Thus, prior research suggests that moves following adverse life events may be qualitatively distinct from moves that occur for other reasons.Moves, and especially long-distance moves, following adverse life events, may be defined by seeking support, seeking family, and seeking familiar locations.
In this paper, we utilise a life-course framework to explore migration as a response to various adverse life events.Extending prior research on migration following union dissolution (Cooke et al., 2016;Spring et al., 2021), we investigate whether other adverse life events, like the onset of health problems or involuntary job losses, also trigger migration.We are interested in whether migrations that follow these adverse events are especially likely to be return moves or moves towards family.We explore these questions in the US context with longitudinal data from the Panel Study of Income Dynamics, which provides information on residential locations, locations of family members, and adverse life events for individuals over our study period from 1983 to 2019.

| Migration after adverse life events
The life course perspective is a useful framework for understanding migration behaviour.Life course changes frequently trigger both short-and long-distance moves, and some life course stages, like young adulthood, are strongly associated with a higher migration potential (Clark & Davies Withers, 2007;Kulu & Milewski, 2007).Yet the life course is hardly standardised; there is immense and increasing diversity in the occurrence, order, and timing of reasons for migrating (Bernard & Kalemba, 2022).Among the motivations for migrating, a new job or an educational opportunity have remained among the top-cited reasons over the last 30+ years (Spring et al., 2016).These broad statistics paint a picture of migration as a key means to access opportunities and a better life.At the same time, migration can also be a reactive response to disruptive life events, like job loss or divorce (Clark, 2016).A deeper investigation into moves following a variety of disruptive events can help to highlight the way diverse life course events can differentially impact individuals' migration behaviour and outcomes.
Prior research has established that people are more likely to move, including long distances, following union dissolution (Spring et al., 2021) and the death of a spouse (Thomas & Dommermuth, 2020).Results have been mixed for some other adverse events, including health problems and involuntary job losses.For example, research on older adults using Swedish register data reported higher propensities to migrate after health crises (Artamonova et al., 2020).However, results from a younger population (aged 18-64) indicated that poor health reduced the likelihood of migration between England and Scotland (Wallace & Kulu, 2014).In a study of young adults in the Netherlands, job loss was associated with decreased migration propensities (Zorlu & Kooiman, 2019).However, using Norwegian register data, another study linked job loss to an increased likelihood of migration (Huttunen et al., 2018).Differences in the characteristics of the population under investigation may explain these mixed results.Our comparative analysis of adverse events occurring within the same sample may help clarify these prior results.

| Return moves and moves towards family
The life course perspective has highlighted the principle of linked lives (Elder, 1994) implying that life courses develop in coordination with and under the influence of the life courses of others.In line with this notion, return moves and moves towards family members (or in with family members) are especially likely under difficult circumstances.
For example, studies have found an elevated propensity to migrate long distances following union dissolution in the United States, Australian, and British contexts (Clark & Davies Withers, 2007;Clark & Huang, 2004;Clark, 2013;Flowerdew & Al-Hamad, 2004), but the location of family tempers this effect.Having nonresident family, and especially children, in the origin location substantially deters long-distance migration following separation or divorce (Cooke et al., 2016;Mulder & Malmberg, 2011;Spring et al., 2021).Among movers, returning to the location where they grew up is especially likely compared to migrating elsewhere, particularly if the return location also includes family members (Spring et al., 2021).Return migration and migration towards family are elevated at the time of divorce or separation, but also in the years after (Spring et al., 2021).Similarly, moves following the death of a spouse are especially likely to be towards family (Artamonova & Gillespie, 2022;Thomas & Dommermuth, 2020).Widowed parents are especially likely to migrate towards an adult child rather than the child towards the parent, suggesting people generally migrate to receive, rather than provide, care (Thomas & Dommermuth, 2020).Studies of migration in response to health issues have found a similar dynamic.Severe health problems among older adults elevate the likelihood of their moving towards adult children or into institutions, but not the likelihood of adult children relocating closer to their parents (Artamonova et al., 2020).Among older parents and adult children already living nearby, the onset of caregiving needs substantially deters either from moving (Artamonova & Syse, 2021).Among Swedish young adults, unemployment was associated with return migration (Mulder et al., 2020b).Huttunen et al. (2018) identified an increased likelihood of migration among the recently unemployed.Their findings also highlighted the importance of family ties-displaced workers with nearby families were less likely to move away after a job loss than those without local ties to kin.
The type of kin one has access to may also matter for migration.The bulk of research on migration towards family focuses on parents and adult children, but increasingly, migration scholarship incorporates a wider view of family, which includes siblings (Artamonova & Gillespie, 2022;Mulder et al., 2020a), nonresident young children (Van der Wiel et al., 2023-this issue), and extended kin (Spring et al., 2023).Individuals are generally more tied to the geographic location of nonresident nuclear kin like parents, children, and siblings than to the locations of extended kin like grandparents, aunts/uncles, and cousins (Spring et al., 2017).In part, this may be because individuals generally report stronger obligations towards nonresident nuclear than extended kin (Rossi & Rossi, 1990).But nuclear and extended family dynamics vary greatly across individuals and within individuals over the life course (Sarkisian & Gerstel, 2012;White, 2001).For these reasons, scholars have called to diversify the types of family members included in analyses (Bengtson, 2001;Cooke, 2008;Mulder, 2007).

| Motivations and outcomes of moves following adverse events
Much of this research suggests that moves following adverse events, particularly towards family members and familiar locations, may be crucial for accessing needed resources and support.For example, among unemployed individuals in Sweden, migration to be closer to family was associated with a higher likelihood of being employed after migrating (Gillespie et al., 2021).However, these "support-seeking moves" may be less-than-ideal for employment or educational opportunities in other circumstances.The same study of labour market outcomes of migration in Sweden found that if proximity to family was an important motivation for migration, employed individuals suffered worse labour market outcomes than those who reported job-related reasons, a combination of work and family motives, or other reasons for moving.For return migration in particular, a separate study using the same data found that individuals were more likely to experience either labour market deterioration or improvement (vs.staying the same) if they reported that family was an important reason for moving (Gillespie et al., 2022).In addition, moves following adverse life events are often unplanned (De Groot et al., 2011).Union dissolution, in particular, frequently triggers moves among people who did not have an initial intention to move (Clark, 2016;De Groot et al., 2011).Few studies delve into the emotional considerations of moves following adverse life events.One exception is Wall and Von Reichert (2013), who examined return rural migration following divorce.Qualitative interview data demonstrated that "attachment to a place that provides a sense of continuity, stability, and safety during a stressful life event" was equally important to functional and emotional support networks for shaping return migration choices (Wall & Von Reichert, 2013, p. 353).

| Confounding factors
Sociodemographic and locational characteristics, including income, education, employment, race/ethnicity, gender, age, parental status, and region of residence may confound relationships between migration and adverse events.Those with children, older individuals, and women are more likely to report moving toward family (Gillespie & Mulder, 2020).Low income and low education are major risk factors for divorce (Amato, 2010;Härkönen & Dronkers, 2006).Social class and racial/ethnic disparities in health are also prominent (Barr, 2014).As a result, marginalised social groups experience more stress across multiple domains of social life compared to non-marginalised groups (Boen, 2020).These same factors influence migration in general and specific types of migration (e.g., return migration, migration towards family).High-income and highly educated people are more likely to migrate towards family than are people with low incomes and less education (Thomas & Dommermuth, 2020).Lower-income and non-White groups are also more likely to live near family in the first place (Spring et al., 2023), perhaps reducing propensities for subsequent migration.In the US context, people from the West, South, and Midwest regions move longer distances, and so intergenerational distance between family members is greater in those regions than in the Northeast (Lin & Rogerson, 1995).
Other characteristics, like personality traits, are harder to measure but may also confound relationships between adverse events and migration.A recent study of interregional migration in Australia found that individuals with higher levels of extraversion and openness to experience were more likely to migrate, while individuals with low levels of agreeableness and emotional stability were more likely to migrate multiple times (Crown et al., 2020, see also Jokela, 2009).Personality traits have also been implicated in labour market outcomes (Caliendo et al., 2015;Cobb-Clark, 2015), marital stability (Lundberg, 2012), and diet and exercise choices (Cobb-Clark et al., 2014).In prior migration research, as in our study, personality traits are often unobserved.While we cannot avoid the limitation of unobserved confounding completely, we examine the robustness of our main results (which utilise clustered standard errors) to results that incorporate individual fixed effects, which effectively remove the influence of time-constant individual characteristics, including unobserved characteristics.

| Hypotheses
Our hypotheses follow from the notion that adverse life events may lead to migration, which we base on the previous literature on migration following separation and situations of need.We anticipate that adverse life events are associated with specific types of moves-moves that are more dominated by returning to a place where one has lived before and moves towards family.
Hypothesis 1. Adverse life events are associated with a greater likelihood of migration, net of individual, family, and locational factors.
Hypothesis 2. Among those who migrate, adverse life events are associated with a greater likelihood of return migration versus migration elsewhere, net of individual, family, and locational factors.
Hypothesis 3.Among those who migrate, adverse life events are associated with a greater reduction in the distance between individuals and their geographically closest kin, net of individual, family, and locational factors.
A main contribution of our analysis is to examine whether these three hypotheses hold for different types of events, including divorce/separation, the death of a spouse, involuntary job loss, and the onset of physical and mental health conditions.In addition, we differentiate adverse life events based on when they occurred, by comparing the likelihood of migration immediately after the event versus at a later point in time.Finally, we also examine whether Hypothesis 3 holds for nuclear and extended kin, given differing dynamics for these family groups.
Our data came from the Panel Study of Income Dynamics (PSID), a nationallyrepresentative, longitudinal study of US families that began in 1968 and has continued to follow individuals from the core sample and their descendants over time.Members of the initial 1968 panel of approximately 5000 families (approximately 18,000 individuals) were interviewed annually from 1968 to 1997 and biennially thereafter.New families were added to the panel as children and other members of original panel families formed their own households.The PSID maintained a reinterview rate between 95% and 98% across virtually all survey waves (PSID, 2019).While the PSID originally overrepresented Black and White families (a function of having been drawn in 1968), several panel refreshers helped to redress the underrepresentation of Latino/a and Asian populations.Although we refer to individuals in our sample as "respondents," the survey measures used here were completed by a designated household reference person on behalf of all others in the household.
We drew individual and household characteristics from the main PSID family files .We began our observations in 1983 because it was the first survey wave that identified permanent cohabiting relationships in addition to legal spouses.We then appended the geographic locations of respondents using data from the restricted geo-coded files of the PSID, which link PSID respondents to their census tracts at each survey wave.We determined return migration by attaching county identifiers for where respondents grew up, obtained from the restricted County Born/Grew Up file .Family members' locations were based on creating a database of respondents' kin members using data from the 2019 Parent Identification file and then attaching census tract identifiers for those kin members.We further describe the process for creating the database of kin members below.We appended census tracts' latitude and longitude (for respondents and their kin) to facilitate calculating geographic distances.
We constructed the data as a series of person-intervals nested within respondents.At each survey wave, we constructed indicators of adverse life events that happened during the 1or 2-year interval since the prior survey.We also constructed an indicator of whether the respondent migrated during the interval since the prior survey.
Respondents had to meet several criteria to be included in our population of interest (see Appendix Table A1).First, we selected household reference people or their spouses because our measures of adverse life events were mostly only available for these respondents and not for other family members (N = 247,380).Second, because we are interested in internal migration, respondents had to be born in the United States (this resulted in omitting 7060 person-intervals for foreign-born respondents).Third, because we are interested in return migration, respondents had to live outside of the county where they were born at the beginning of the survey interval (this resulted in omitting 101,271 person-intervals for respondents who were living in their county of birth).We excluded 15,948 person-intervals because they were missing information on the county of birth; 26,297 person-intervals because they were not asked questions on chronic health conditions; and 4719 personintervals because they were missing information for one or more of the covariates used in our analysis.Our final analytical sample totalled 92,085 person-intervals, reflecting 10,382 unique respondents.This sample reflects US-born adults who had relocated outside their county of birth, a somewhat specific group.They may be more likely to migrate again, having already migrated, compared to the general population.They are also better educated, have higher incomes, and are less likely to have children compared to adults in the full PSID sample.

| Adverse life events-
The PSID asked about several adverse life events (Table 1), although their list is admittedly not exhaustive.In this study, we investigated divorce/ separation, the death of a spouse, involuntary job loss, the onset of a chronic physical health condition, and the onset of a chronic mental health condition.We describe the measurement of each below.The PSID also asked about home foreclosure, but not often enough to be useful for our analysis, and did not ask about other events we would have liked to include, like evictions or crime victimisations.Since respondents entered our sample at different ages and remained in the sample for different lengths of time, all models adjusted for their age at their first observation and the number of survey waves they were observed.
We coded divorce/separation as a binary indicator of whether the respondent experienced at least one divorce or separation since the previous interview.We discerned divorces/ separations by comparing the respondent's spouse's ID at the beginning and end of the survey interval and combining this with information on the spouse.Specifically, if the spouse was no longer present at the end of the survey interval or the spouse's ID changed, we combined this with information on why the original spouse was no longer in the household.When the spouse left the household for reasons other than death or relocating to an institution (a health care facility, educational institution, jail/prison, or armed forces), we coded the respondent as having divorced/separated from this spouse.An alternative strategy would have been to measure a transition in current marital status from "married" to "divorced/separated" during the survey interval, but re-partnership complicates this strategy.Because survey intervals are 1-2 years long, respondents can potentially separate and re-partner before the next observation.Our strategy accounts for re-partnership; however, there are still some limitations.If a respondent separated and re-partnered multiple times during the survey interval, we could only identify one divorce/separation.If the respondent separated and re-partnered with the same spouse, we could not identify the divorce/separation.A benefit of our measure of divorce/separation is that it reflects marital dissolution and separation from an unmarried partner since the PSID's definition of "married" included couples who were permanently cohabiting.
We coded the death of a spouse as a binary indicator reflecting whether the respondent experienced at least one spousal death since the previous interview.We discerned spousal death in a similar way to divorces/separations. We compared the spouse's ID at the beginning and end of the survey interval.When spouse IDs did not match, we added information on why the original spouse was no longer in the household; a possible response is that the spouse died.A benefit of our measure is that we capture spousal deaths even if the respondent re-partnered.Spouses are again inclusive of cohabiting relationships.
We coded involuntary job loss as a binary indicator representing whether the respondent experienced at least one job loss since the previous interview.Due to data limitations, there are some job losses we likely failed to capture.We captured as many job losses as possible by utilising multiple criteria.First, we identified a transition from "employed" to "not currently working" during the survey interval combined with the reason for the job loss.We coded responses of being involuntarily laid off, fired, or the company folding as involuntary job losses (as opposed to "voluntary" job losses like quitting or retiring, which we did not analyze).Since some respondents may lose a job and be re-employed within the survey interval, we also used respondents' reports of how much work they lost in the prior year due to lay-offs.We coded respondents who were employed at the start of the survey interval but in the following interview reported missing work in the prior year due to being temporarily laid off as having an involuntary job loss.In the case of 2-year survey intervals (1997 and later), this means we missed job losses that occurred in the first year of the survey interval for respondents who were re-employed by the next interview.Our measure of involuntary job losses is, therefore, an undercount.To reduce bias brought on by this data limitation, we controlled for predictors of job loss, including age, education, gender, and race/ethnicity (Montenovo et al., 2020), as well as the number of years since the prior interview.
We measured physical health crises as the onset of a new chronic physical health condition during the prior survey interval.We determined the onset of a chronic physical health condition from questions asking respondents the age they were first diagnosed with arthritis, asthma, high blood pressure, cancer, diabetes, heart attack, heart disease, lung disease, stroke, and "other" chronic conditions.The PSID survey asked the age of diagnoses beginning in 2005, but we back-coded responses to prior survey waves.However, we limited our analytical sample to those present in the sample sometime between 2005 and 2019, so that they were asked the questions on chronic conditions at least once.We then included all prior survey waves for those respondents.Reliability is a potential concern since the household reference person answered questions about health on behalf of their spouse.However, age-specific rates of health conditions in the PSID are consistent with the most widely used nationally representative health survey, the National Health Interview Survey (NHIS) (Andreski et al., 2009).Many respondents were asked the age of first diagnoses in more than one interview.Occasionally respondents reported inconsistent ages, meaning they reported one age at diagnosis in one interview and a different age for the same condition in a subsequent interview.Our strategy for age inconsistencies was to use the first age ever reported.Since the measures of chronic health conditions reflect the first onset, this means we do not capture subsequent episodes of the same condition.For example, if a respondent had multiple strokes over a period of time, we only knew the timing of the first stroke.However, we captured onsets of different conditions since the age of diagnosis was measured separately by condition.We summarised the various conditions into one summary measure, the onset of any new chronic physical health condition, which can occur more than once over the respondents' life if they experienced distinct new conditions.
We measured mental health conditions as the onset of a new psychiatric condition (of any type) during the prior survey interval.We determined onset from a question asking respondents the age of diagnosis for any psychiatric condition, which included depression, bipolar disorder, schizophrenia, anxiety, phobias, alcoholism, drug addiction, obsessive-compulsive disorder, and "other."Unfortunately, these conditions were grouped together in the question, and ages were not asked separately for each specific condition.We followed the same procedure as above to code (and backcode) the age of diagnosis.Since our measure of mental health reflects only the onset of the first psychiatric condition of any type, we did not capture subsequent diagnoses.
At each person-wave, each adverse event was coded as 0 = never having occurred, 1 = occurred since the previous interview, or 2 = occurred before the previous interview.We summarise the coding of the adverse event measures in Table 1.

| Migration
-Prior research does not provide a consistent definition of "migration".We are interested in long-distance moves that would implicate a substantial change in employment opportunities and local social networks.The exact distance threshold equating to a long-distance move is uncertain and arbitrary.Prior research relying on distance-based definitions of migration has often used a threshold of around 50 km, which we adopted for this paper (Ermisch & Mulder, 2019;Spring et al., 2021).We coded respondents who moved 50 km or more since the previous interview as having migrated, and tested the sensitivity of our results to other thresholds.Sensitivity results are reported below after the main results.

| Return migration-Our analytic sample consisted of person-intervals
where the respondent lived outside of the county where they grew up.Return migration was determined by whether the respondent had migrated back to the county where they grew up by the time of the next survey.Respondents who moved back to the county where they grew up but moved less than 50 kilometres were not coded as return migrants since they moved only a short distance.
Counties where respondents grew up were obtained from survey questions retrospectively asking respondents where they grew up or where they were born.Some respondents were asked both.We used county where grew up as the primary measure because the PSID asked it in more years.But, if values for the county where grew up were missing or were reported inconsistently within the same respondent over time (i.e., respondents reported different counties at different points in time), then we filled in with county born.If county born was also missing or reported inconsistently within the same respondent over time, then county where born/grew up was coded as missing for that respondent.We used respondents' county born/grew up identifiers from one wave, if available, to fill in missing data for other waves for the same individual.In the end, we were still missing county where born/grew up for 14.7% of respondents, who we excluded from the analysis because we could not assess return migration for this group.We also excluded respondents who grew up in a foreign country, who made up 6.3% of respondents.

| Migration towards family-
The structure of the PSID allowed us to locate PSID respondents within their broader kin network.The rules for follow by the PSID stipulate that sample members include individuals living in the original family unit at the time of the first interview in 1968 and all of their lineal descendants born after 1968.As a result, those born to a sample member automatically became sample members and were followed over time as they aged and formed their own households.Individuals who were not born into a study family but moved in later, such as a spouse marrying a sample member, also became sample members and were followable.As the networks of PSID kin branched out over time, they grew so that individual respondents could be linked not only to nuclear kin like parents, children, and siblings but extended kin like grandparents, aunts/ uncles, nieces/nephews, and cousins.We leveraged the PSID's built-in multigenerational structure by linking all members of the same extended family using the original 1968 Family Identification number they all share.We characterised types of kin using data from the Parent Identification file.Starting with the initial parent-child relationship, we inferred additional relationships, including siblings, grandparents, grandchildren, aunts/ uncles, nieces/nephews, cousins, step-parents, step-siblings, and other kin.Other studies have utilised similar kin-mapped data from the PSID, including Spring et al. (2017), Ackert et al. (2019), Daw et al. (2019), and Spring et al. (2021).
The availability of geographic identifiers for respondents and their nonresident kin at every survey wave allowed us to compare distance to the respondent's geographically closest nonresident kin member before and after a move.We calculated distance in kilometres based on the centroid of the respondent's census tract to the centroid of each kin member's tract and then looked for the closest distance.Respondents who co-resided with kin following migration were included, with distance coded as zero.We utilised two categories of kin: nonresident nuclear kin (parents, children, and siblings) and nonresident extended kin (everyone else) and calculated the spatial proximity measure within each category.We calculated the change in distance to the closest nuclear and extended kin by subtracting the distance at the end of the survey interval from the distance at the beginning of the survey interval.We truncated distance change at the 1st and 99th percentiles to adjust for skewness introduced by a small number of respondents who moved very far distances.As a sensitivity check, we truncated change in distance from family at the 5th and 95th percentiles and found substantively similar results.

| Covariates-All
models account for additional individual, family, and locational factors that are associated with migration and adverse life events based on the previous literature.These include an indicator of family at the origin location since prior research establishes that people are less likely to move when they have family nearby.We coded family at the origin as a binary indicator of whether the respondent had at least one family member (nuclear or extended) living within 50 km at the beginning of the survey interval before a potential move.We included a separate control variable indicating co-residence with kin at the origin.We also controlled for the respondent's age, sex, years of education, current marital status, race/ethnicity, family income, current employment status, children under 18 in the respondent's household, whether the family owns their home, region of residence, and survey year.Family income was standardised to 2010 dollars to adjust for inflation.We measured all covariates at the beginning of the survey interval before a potential move.To account for the PSID sampling structure, we also controlled for the respondent's age at the first observation, the number of survey waves they were observed, and the number of years since the previous interview.A small number of respondents were missing values for some covariates, including education (1.5% missing), homeownership (0.003% missing), race/ethnicity (2.3% missing), and employment status (3.10%).We used listwise deletion to remove respondents with missing data.This resulted in dropping 4.86% of person-intervals.We summarise covariate statistics for our sample in Table 2.

| Analytic methods
We are interested in how adverse life events affect the probability of migration in general, return migration, and migration towards family members.After examining descriptive statistics for our sample, we estimated three multivariate regression models: first, a binary logistic regression predicting migrating versus not migrating; second, a binary logistic regression predicting return migration versus migration elsewhere among only those who migrated; and third, ordinary least squares regressions predicting change in the distance to the closest nuclear and extended family member among only those who migrated.The main predictors of interest are adverse life events.We tested the impact of different events in separate models, controlling for all covariates listed above.We adjusted for the nonindependence of multiple observations of individuals over time and individuals from the same family by utilising standard errors clustered at the family level, as is common in prior migration research utilising the PSID (see e.g.Crowder et al., 2006;Crowder et al., 2011;Molloy et al., 2014).The descriptive and regression analyses apply sampling weights to adjust for differential probabilities of sample selection and retention.

| Descriptive statistics
The rate of events occurring since the previous wave ranged from 1% to 10% of our sample, depending on the event (see Table 3).The most common event to have just occurred was the onset of a new chronic physical health condition (10.4%), followed by a job loss (4.4%), a divorce/separation (2.1%), the onset of a mental health condition (1%), and the death of a spouse (0.7%).Another 4%-40% of our sample had experienced events in the past, with a prior chronic physical health condition diagnosis being the most common (39.7%), followed by a prior job loss (30.8%), a prior divorce/separation (16.3%), a prior chronic psychiatric condition diagnosis (8.5%), and the prior death of a spouse (3.9%).

Consistent with annual rates of US long-distance migration (Current Population Survey
[CPS], 2021), 7.8% of our sample had migrated more than 50 km since the previous interview (see Table 4).About 17% of these long-distance moves were return moves, whereas 83% were moves elsewhere.Of those who migrated and had nonresident nuclear kin members, the average distance moved was 69 km closer to nuclear kin.For extended kin, the average distance moved was 59 kilometres closer, among those who migrated and had nonresident extended kin.Some of these moves (9%) were moves into co-residence with kin.

| Predicted probability of migration after adverse life events
We graphed the predicted probability of migrating 50 km or more by the occurrence of adverse life events in Figure 1.The predicted probabilities were derived from logistic regression estimates of migration predicted by adverse event indicators and covariates.We report the tabular results in Appendix Table A2.In Figure 1, the shaded vertical line represents the predicted probability of migrating averaged across the full sample, taking into account covariates but no adverse event indicators (see Table A2, Model 1).This provided a baseline for comparison to the sample as a whole.We also graphed the 95% confidence intervals in Figure 1 to assess the statistical significance of the differences in migration propensities by adverse event occurrences.
In line with prior research (Cooke et al., 2016;Spring et al., 2021), we found strong evidence that migration was more likely following divorce.The probability of migrating was over three times greater for those who had just divorced (since the prior wave) compared to those who had never divorced.We also found that the impact on migration continued for a period of time following divorce.Those who were divorced in the past (before the last survey wave) had significantly higher chances of migration than those who had never been divorced.We also found strong evidence that migration was related to involuntary job loss and chronic health conditions.Migration was almost twice as likely for respondents who had lost a job since the prior survey wave compared to those who had never lost a job.Previous job losses, on the other hand, were not significantly related to migration.The onset of a new physical health condition since the prior survey decreased the probability of migration compared to those who had never had a chronic condition, by about 15%.Prior new conditions (diagnosed before the previous wave), also reduced the propensity to migrate.The results suggest that unlike divorce and job loss, chronic physical health conditions inhibited rather than encouraged migration.
We found minimal evidence that death of a spouse was associated with migration.An increased propensity to migrate following the death of a spouse was marginally significant (at p < 0.1), but did not reach the 95% confidence threshold.We also did not find evidence that the onset of chronic psychiatric conditions was associated with migration, either immediately or over time.We note that death of a spouse and mental health diagnoses were less common than other events in our sample, leading to a smaller sample size of people who had experienced these events, which may make it difficult to detect significant effects.
We thus find mixed support for our first hypothesis.While divorce and job loss both induced migration (including for an extended period of time after divorce), chronic physical health conditions reduced the chances of migration, and results were inconclusive for death of a spouse and chronic mental health conditions.

| Predicted probability of return migration
We also predicted that the types of moves migrants undertook would differ by adverse event experiences.We anticipated that respondents with recent or previous adverse events would be more likely to return migrate compared to other respondents.We report the predicted probability of return migration in Figure 2. Estimates were derived from logistic regression estimates of return migration versus migrating elsewhere among those who migrated.Tabular results are in Appendix Table A3.
Again, we found the largest effects for divorce/separation.Respondents who were just divorced were 1.7 times more likely to return migrate than respondents who had never been divorced, compared to migrating elsewhere.A previous divorce (compared to having never been divorced) was also associated with a greater propensity to return migrate.An involuntary job loss was also associated with return migration versus moving elsewhere, both immediately after the job loss and down the line.Those who experienced a job loss since the last wave were 1.4 times more likely to return migrate (vs.migrating elsewhere), and those who had a job loss before the last wave were 1.36 times more likely to return migrate (vs.migrating elsewhere), compared to those who had never lost a job.We also found that the onset of a chronic physical health condition was associated with return migration, but only over time.We did not find evidence that spousal death or the onset of psychiatric conditions were associated with returning versus migrating elsewhere.
These results generally support our second hypothesis, that adverse events are associated with a greater likelihood of return migration versus migration elsewhere.Results support this hypothesis for divorce/separation and job loss.They also support it for chronic physical conditions, but only several years after the onset of the condition.This suggests there is a temporal dimension to return migration following physical health problems.One potential explanation is that as the condition progresses, people are increasingly likely to return migrate to family, hometowns, or places with which they are familiar.

| Predicted change in distance towards family
Return migration often, but not always, entails moving to locations with family members.We predicted that moves towards family are another unique type of move that may be especially prevalent following adverse events.We divided our distance analysis into nuclear and extended kin, because prior research points to differing dynamics for these family groups.We report the predicted change in distance to the respondents' geographically closest nuclear family member (defined as nonresident parents, children, and siblings) in Figure 3. Values are in kilometres, with negative values indicating declines in distance to the nearest nuclear family member and positive values indicating increases in distance.The values in Figure 3 were derived from OLS estimates of change in the distance to nuclear kin before and after a move among those who migrated.Tabular results are in Appendix Table A4.
Moving closer to nuclear kin was most strongly associated with divorce, job loss, and the onset of a chronic physical health condition.Migrants who became divorced, lost a job, or had a new chronic physical condition since the last wave moved on average 92 to 129 kilometres closer to their nuclear kin compared to migrants who never experienced those events.The onset of a chronic physical condition in the past, before the previous wave, was associated with moving 64 kilometres closer to nuclear kin.We did not find evidence that the death of a spouse or the onset of psychiatric conditions were associated with moving towards nuclear family.
We also found associations between some adverse events and moving towards extended kin (see Figure 4).Migrants who were divorced or lost a job since the last survey wave moved on average 89-117 km closer to their nuclear kin compared to migrants who never experienced those events.Some events were marginally associated with moving towards extended kin, including death of a spouse and the diagnosis of a chronic physical health condition since the prior survey wave, but these events did not reach the 95% confidence threshold.And, migrants who were diagnosed with a chronic physical health condition in the past, before the previous wave, moved 50 km closer to extended kin than migrants who never had a chronic physical condition.We did not find evidence that psychiatric conditions were associated with moving towards extended family.
Taken together, the results for distance to family somewhat support our third hypothesis.Divorce, job loss, and physical health problems were associated with a reduction in the distance between individuals and their closest nuclear kin member.These same events were also associated with moving towards extended kin members, although for physical health conditions, results were stronger for previous rather than recent events.Our results for the death of a spouse were inconclusive but point to patterns worthy of additional investigation.

| Sensitivity checks
The results presented here are robust to several different measurement and model specifications.We re-estimated all models using a lower, 30-km threshold for migration, and a higher, 70-km threshold for migration.Results were substantively similar to results using the 50-km threshold, with a few exceptions.At the lower threshold, we did not find an association between the onset of a physical health condition and migration whereas in our main results the association was negative.The discrepancy could suggest that people who experience the onset of a physical health condition have lower propensity to migrate long distances compared to those who do not, but the risk of moving shorter distances is similar.
We also tested the sensitivity of our results for moving towards family in several ways.We used a categorical measure of change in distance from family among those who migrated, with the categories of (i) reduction in distance of more than 50 km; (ii) increase in distance of more than 50 km; and (iii) no change in distance or change less than 50 km.Results were generally consistent with those presented here for continuous distance changed.However, we also found an association between divorces in the past and decreasing distance from nuclear kin, relative to no change in distance.And we did not find an association between job losses and change in distance to extended kin (whereas in our main models the association was negative).As another check, we removed respondents who moved into co-residence with kin from the analysis and estimated results only for migrants who did not move in with kin.Results were similar, except that we did not find associations with physical health conditions and moving towards extended kin.
Finally, we re-estimated all models using respondent-fixed effects rather than clustered standard errors (see Appendix Table A6).Fixed effects have the benefit of adjusting for observable and unobservable characteristics of respondents that do not change over time.We found substantively similar results using fixed effects, with a few exceptions.We found that fixed effects estimation was more precise for the rarer events in our study-death of a spouse and the onset of a psychiatric health condition-compared with our main results.With fixed effects, we found evidence that widowed respondents were more likely to migrate and to move closer to extended kin compared to respondents who had never experienced the death of a spouse.With fixed effects, we also found that migrants who had psychiatric conditions diagnosed in the past, before the previous wave, were more likely to return migrate compared to migrants who had never been diagnosed with a psychiatric condition.
For other conditions-divorce, job loss, and physical health conditions-fixed effects estimates tended to be less precise than our main estimates.With fixed effects, we did not find evidence that previous divorce (before the last wave) was associated with migration, or that divorce was associated with moving towards nuclear family.And, we found weaker associations of job loss and chronic physical conditions with migration and return migration, and no associations with moving towards family.These differences between the main and fixed effects imply that fairly constant but unobserved characteristics, like personality traits, might explain some migration differences following particular events, like job losses and physical health problems.With these caveats in mind, our fixed effects estimates still point to the same general patterns as our main results: increased migration following divorce, job loss, and to a marginal extent, death of a spouse; the inconsistency of physical health's association with migration; and weak evidence of an association between migration and mental health.

| DISCUSSION AND CONCLUSION
This study extended prior research linking union dissolution to migration (Spring et al., 2021) to other types of adverse life events.We hypothesised that adverse events, including divorce/separation, the death of a spouse, job loss, and the onset of chronic health conditions, would induce migration in general and specific types of migration: return migration and migration towards family.We situated our analysis within the life course framework to build an understanding of migration as a response to life events.Results for migration in general, return migration, and migration towards family point to patterns consistent with our expectations.
Migration, especially return migration and migration towards family, is more likely after an adverse life event.However, the onset of physical health problems is positively associated with return migration and migration towards family but negatively associated with migration in general, and we did not find associations between mental health conditions and migration.The findings for physical health could indicate that health problems limit options to migrate for education, jobs, or amenities, but at the same time, induce migration towards one's social network.
With these exceptions related to health, elevated rates of migration occur immediately after the event and in some instances continue for a period of time.Divorces that happened in the past still contribute to a higher current likelihood of migration in general, but this result was not robust to fixed effects.Return migration is also predicted by past divorces, job losses, or chronic condition diagnoses, and past chronic conditions also predict moves towards nuclear family.One explanation is that in these instances, the stress of prior events is ongoing, resulting in more support-seeking moves down the line.Another explanation is that past events impact income, employment, and other predictors of current migration in the direction that makes migration more likely.We are unable to fully test between these explanations, but when we removed potential mediators from our models (income, employment, presence of children, and homeownership) we did not find markedly increased effects of past events, indicating that conditioning on the mediators is perhaps minimal.This is in line with research demonstrating that people tend to rebound quickly from many adverse events (Seery et al., 2010), yet there is also evidence to the contrary (Brand, 2015).These questions all relate to a broader limitation of our analysis: we are unable to ascertain why adverse events are associated with migration.This is due to limitations with our data and with adverse event checklists more broadly.Noting whether an event did or did not occur fails to capture the contextual and personal circumstances surrounding the event.Future research on migration following adverse events may find a risk and resilience framework useful.Researchers have used this framework to understand why some people maintain healthy levels of well-being and avoid negative outcomes during or following potentially negative events and others do not.For example, prior research identifies pre-existing depression as a key risk factor for negative health outcomes following divorce (Sbarra & Nietert, 2009;Sbarra et al., 2014).Their research suggests it is the context surrounding the divorce, and not the event of divorce itself, that is linked with subsequent negative outcomes.Migration is not a "negative outcome," but exploring the contextual circumstances surrounding adverse life events could prove even more fruitful for understanding who migrates following adverse life events and the types of moves made.
Our study provides some useful starting points for future research but also comes with some additional limitations.A major limitation is that our measures of adverse events were imprecise.Because of the 1-or 2-year interval between successive PSID interviews, we could not account for multiple events that occurred within the interval.For example, we could not account for multiple divorces and re-partnerships, or re-partnerships to the same spouse.We also did not capture all involuntary job losses, and our measures of chronic conditions rely on retrospective reports.Nonetheless, we were able to find significant differences between groups that had experienced these events and those that had not, even though the line between those two groups was blurred.We were also unable to account for adverse events that happened not to the respondent but to someone else in their family.A family member losing a job or getting diagnosed with a health condition might also prompt migration.Another limitation is that we did not have data on specific motivations for people's moves.We made assumptions about why adverse life events are associated with migration, but we did not know the underlying reasons and motivations.
Despite these limitations, our study provided important insights into how adverse life events impact migration.Extending the previous research on divorce, we found that some other life events operate similarly to induce particular types of migration (with health events being somewhat unique from other events in this regard).Future research could examine whether similar migration patterns exist for other adverse life events, such as home foreclosures, evictions, crime victimisations, or ongoing adverse events, such as family violence.Future research could also examine whether marginalised social groups, such as women and racially/ethnically minoritized people, are more likely to migrate after adverse events.Future research might also investigate the effects of cumulative events on migration.The added stress of multiple adverse events is underscored in the literature on mental health (Chen et al., 2022) and identified as a major contributor to racial health disparities (Boen, 2020).
But, cumulative events have not been linked to migration specifically.If the accumulation of multiple adverse events is particularly likely for marginalised social groups, the social and economic trade-offs of migration following adverse events is an important topic for further study.

TABLE A2
Logistic regressions predicting migrating 50+ km (reference group = did not migrate).A2.Predicted probability of return migration (vs.migrating elsewhere) by adverse event occurrence (with 95% confidence intervals).The shaded vertical line represents the average predicted probability of migration in the full sample with the 95% confidence interval.The predicted probabilities are based on marginal effects that are averaged across the sample and weighted by the PSID longitudinal sampling weights.Tabular results are in Appendix Table A3.Predicted change in distance to nuclear kin (in kilometres) by adverse event occurrence (with 95% confidence intervals).The shaded vertical line represents the average predicted distance change in the full sample with the 95% confidence interval.The predictions are based on marginal effects that are averaged across the sample and weighted by the PSID longitudinal sampling weights.Tabular results are in Appendix Table A4.Predicted change in distance to extended kin (in kilometres) by adverse event occurrence (with 95% confidence intervals).The shaded vertical line represents the average predicted distance change in the full sample with the 95% confidence interval.The predictions are based on marginal effects that are averaged across the sample and weighted by the PSID longitudinal sampling weights.Tabular results are in Appendix Table A5.

FIGURE 1 .
FIGURE 1.Predicted probability of migration (vs.not migrating) by adverse event occurrence (with 95% confidence intervals).The shaded vertical line represents the average predicted probability of migration in the full sample with the 95% confidence interval.The predicted probabilities are based on marginal effects that are averaged across the sample and weighted by the Panel Study of Income Dynamics (PSID) longitudinal sampling weights.Tabular results are in Appendix TableA2.

TABLE A4
OLS regressions predicting change in distance (in kilometres) to nuclear kin among those who migrated.

TABLE A5
OLS regressions predicting change in distance (in kilometres) to extended kin among those who migrated.

TABLE A6 Migration
, return migration, and family migration estimates with respondent-fixed effects.Migration (

Change in distance to nuclear kin (among those who migrated) 2 Change in distance to extended kin (among those who migrated) 2
Estimates are weighted by Panel Study of Income Dynamics (PSID) longitudinal sampling weights and adjust for time-varying respondent characteristics, including age, education, marital status, family income, employment status, presence of children, homeownership, region of residence, family within 50 km of origin, co-residence with family at origin, years since prior interview, and year of survey.Estimated with logistic regression with respondent-fixed effects (using the xtlogit command in STATA).2Estimated with ordinary least squares regression with respondent-fixed effects (using xtreg command in STATA).

TABLE 2
Weighted sample statistics.
Popul Space Place.Author manuscript; available in PMC 2024 May 02.

TABLE 3
Descriptive statistics of adverse life events.

event never occurred 1 = event occurred since the previous wave 2 = event happened before the previous wave
Weighted by Panel Study of Income Dynamics (PSID) longitudinal sampling weights.
Popul Space Place.Author manuscript; available in PMC 2024 May 02.