Quantifying lifecourse drivers of international migration: A cross-national analysis of Mexico and the US.

Abstract

Background:

Evaluating the long-term health consequences of migration requires longitudinal data on migrants and non-migrants to facilitate adjustment for time-varying confounder–mediators of the effect of migration on health.

Methods:

We merged harmonized data on subjects aged 50+ from the US based Health and Retirement Study (HRS) and the Mexican Health and Aging Study (MHAS). Our exposed group includes MHAS return migrants (n=1555) and HRS Mexican-born migrants (n=924). Our unexposed group includes MHAS never migrants (n=16,954). We constructed a lifecourse data set from birth (age 0) until either age at migration to the US or age at study entry. To account for confounding via inverse probability of treatment weights (IPTW), we modeled the probability of migration at each year of life using time-varying pre-migration characteristics. We then evaluated the effect of migration on mortality hazard estimated with and without IPTW.

Results:

Mexico to US migration was predicted by time-varying factors that occurred prior to migration. Using measured covariates at time of enrollment to account for selective migration, we estimated that, for women, migrating reduces mortality risk by 13%, although this estimate was imprecise and results were compatible with either large protective or deleterious associations (HR=0.87, 95%CI: 0.60,1.27). When instead using IPTWs, the estimated effect on mortality was similarly imprecise (HR=0.98, 95%CI: 0.77, 1.25). The relationship among men was similarly uncertain in both models.

Conclusions:

Although time-varying social factors predicted migration, IPTW weighting did not affect our estimates. Larger samples are needed to precisely estimate the health effects of migration.

INTRODUCTION

More than 252 million individuals worldwide live in a country other than their birth nation.¹ Currently in the United States, over 15 million adults ages 50 years or older are foreign-born, and the majority of those are of Mexican origin. A great many more individuals who once migrated to the US have now returned to their country of origin; for example, between 2009 and 2014, approximately 1 million Mexicans who once migrated to the US have subsequently returned to Mexico.² As international migration becomes more common³ -- whether due to climate change, asylum, or economic opportunities -- understanding the long-term health consequences of migration is a central population health issue.

Identifying the effects of migration on health outcomes presents substantial methodologic challenges. In particular, it requires a comparison population of non-migrants in the country of origin, but few data sources provide harmonized measures on similar populations from both sending and receiving countries. Further, migrants are not a random sample from their country of origin. Migration is a selective process such that, compared to non-migrants who stayed in the country of origin, migrants may be selected – either positively or negatively -- on markers of early life socioeconomic status (SES) (e.g. education), health status (e.g. taller stature), or other lifecourse experiences.^4-10 Those pre-migration characteristics – also known as selection factors – predict migration and may also be important contributors to health outcomes later in life.^4,5 However, accounting for these selection factors is not straightforward; the act of migrating may itself modify the selection factors over time. For example, employment or marital history may vary during the time leading up to migration,¹¹ and migration in itself can change migrants’ opportunities and trajectories in life.^10,12,13 Similarly, education may influence chances of migrating,¹⁴ and migrating in turn may alter the educational opportunities available to the individual¹⁵ (see Figure 1). Thus, appropriately evaluating the health effects of migration requires closely harmonized cross-national data sets with measures of selection factors and statistical methods to account for time-varying confounder–mediators (i.e., factors that predict migration and are in turn modified by the act of migrating), to minimize confounding bias.

Figure 1a: Schematic Diagram for the relationship between migration and health in older age, including the role of time-invariant and time-variant confounders.

Note that time-varying characteristics (e.g. education, marital status) not only affect the decision to migrate (i.e. migration at T₀), but are also in turn affected by migration. The lifecourse nature of our study design (see Figure 2) enables us to accurately adjust for the time-varying values of those characteristics occurring up to the event of migration (i.e. the pre-migration characteristics displayed in the box occurring at T₋₁). The box indicates conditioning/accounting for pre-treatment covariates.

The evidence on the health effects of migration from Mexico to the US exemplifies these challenges. The US-Mexico migration corridor has been one of the busiest in modern history, and as of 2017 Mexican Americans account for more than 20% (or 3 million) of the foreign-born older adults living in the US.¹⁶ Strong selective migration is acknowledged as a source of confounding bias when evaluating the health effects of migrating from Mexico to the US; for example, such selection is typically invoked to explain the ‘Hispanic paradox’ of unexpectedly good health profiles among Mexican immigrants, who have low average socio-economic status compared to their US-born counterparts.¹⁷ While the field has increasingly recognized the methodologic challenge of finding the appropriate comparison population,^5,18-25 without cross-national data, the best evidence on the health effects of migration is often indirectly inferred from comparisons of the health of migrants residing in the US with their US-born counterparts (e.g., 2^nd generation immigrants) or with non-Hispanic whites and blacks.^26-31

Using harmonized cross-national data from the US and Mexico, we build on prior literature by quantifying selection factors for migration from Mexico to the US within a lifecourse approach that accounts for the time-varying nature of those factors. We do so by creating a propensity-to-migrate score updated at each year (age) of life up until the event of migration. We then apply those time-updated propensity scores as inverse probability of treatment weights (IPTWs) to evaluate the effect of migration from Mexico to the US on all-cause mortality, comparing migrants to never migrants.

METHODS

Data sources and study populations:

We merged harmonized data from two nationally representative datasets – the US based Health and Retirement Study (HRS) and the Mexican Health and Aging Study (MHAS).

The Health and Retirement Study (HRS 2000-2012) is a multi-stage, longitudinal survey of the US population over age 50 and their spouses. Enrollment was staggered by birth cohort with enrollments in 1992, 1993, 1998, 2004, and 2010. Since 1998, new enrollments occur every 6 years to include birth cohorts that have “aged into” the 50+ group. For our analyses, we restricted to HRS participants and their spouses who self-reported birth in Mexico and who were ages 50+ as of the 2000 interview wave, in order to be most comparable to the sister study, MHAS, 2001 baseline assessment (described below). We also included new HRS participants and their spouses who met the above criteria with enrollments in 2004 and 2010 as part of the refresher cohorts – again to be comparable with MHAS follow-up assessments. The University of Michigan’s Institutional Review Board (IRB) approved the HRS⁷. In this analysis, the HRS provides those participants who are Mexican-born migrants residing in the US.

The Mexican Health and Aging Study (MHAS 2001, 2003, 2012) is an HRS sister study in Mexico and, as such, was designed to largely mirror the structure and measures of HRS in order to facilitate harmonized analyses. MHAS enrolled a multistage, stratified probability sample of the Mexican population ages 50+ and their spouses in 2001. Follow-up survey visits that took place in 2003 and 2012 were included in this analysis. For our analyses, similar to HRS, we restricted to MHAS participants and their spouses who were ages 50+ in 2001. We also included new MHAS participants and their spouses who were ages 50+ with enrollments in 2003 and 2012 as part of the refresher cohorts. A detailed description of MHAS is available elsewhere^32,33. The IRBs of the University of Pennsylvania, the University of Maryland, and the University of Texas Medical Branch⁸ approved the MHAS. The IRBs of Columbia University and the University of California Sn Francisco approved the current analysis. Of relevance to this analysis, MHAS provides data on Mexicans residing in Mexico who either never migrated to the US or who migrated at some point in their life but returned to Mexico. A diagram describing the selection of HRS and MHAS participants for this analysis is included in eFigure 1.

Data structure and setup:

We constructed a lifecourse data set in which each participant contributed an observation for each year of their life beginning with birth (age 0) through either age at migration to the US (applicable to MHAS-return migrants and HRS- US residing migrants) or age at enrollment i.e. at first study visit (applicable to MHAS-never migrants). For example, an HRS-migrant who migrated at the age of 20 years had 21 rows of data, a row for each year of age since birth. Schematic Figure 2 is an illustration of the lifecourse data structure, using a simplistic scenario with only four age periods. While at age period 0 (i.e. birth), all participants are in Mexico, at age period 1 (i.e. 1 year old), a person could either still be in Mexico or could have migrated to the US. After the year that the individual migrated to the US (based on self-reported first year of migration), that person’s probability of being a migrant is remained unchanged. Similar to the lifecourse nature of migration, covariates such as years of education, age of life events (marriage, divorce or widowhood, entry into the labor force, and smoking initiation) may also vary and as such each participant’s covariate history was updated at each year leading up to migration (details in the covariate section). Such data set-up ensures that we only adjust for the values of confounders (pre-treatment covariates) that occurred prior to the event of migration or study entry.

Figure 2: Schematic illustrating the lifecourse nature of the event of migration and the longitudinal design of our study^a. For simplicity, displayed is a scenario with four age periods.

^aIn our longitudinal study design, migration status as well as the values of time-varying covariates are updated at each age period (year of life) up to the event of migration for ‘migrants’ or study entry for ‘non-migrants’. This data set up ensures that we only account for the values of time-varying confounders that occurred prior to migration or study entry.

Ascertainment of migration status

Our exposure of interest is having ever migrated from Mexico to the US, and we study the first migration. Therefore, our ‘exposed’ group includes MHAS participants who once migrated to the US and later returned back to Mexico (Mexico return migrants, n=1555) and HRS participants who are Mexican-born immigrants residing in the US (US-residing migrants, n=924). Our ‘unexposed’ group includes MHAS participants who never migrated to the US (never migrants, n=16,954).

Migration status – ever-migrant (Mexico-return migrant or US residing migrant) vs never-migrant – and reported age at first migration were used to create a binary time-varying migration status indicator. For never migrants (i.e. unexposed group), migration status was always set to 0 (not migrated) from age 0 (birth) until their age at study enrollment. For return migrants and migrants (i.e. exposed group), migration status was coded as 0 before age at first migration, and coded as 1 at age at first migration.

Ascertainment of all-cause mortality

Mortality follow-up is through the last date of interview (04/30/2013 for HRS and 02/28/2013 for MHAS). In HRS, death dates were derived via a combination of the national death index (NDI) dates and the exit interview/proxy reported dates. If both sources were available, the NDI date was used. In MHAS, death dates were based on questions asked as part of the Next-of-Kin interview at each wave whenever the subject was reported to have died. A total of 493 MHAS subjects who were absent at study follow-up but had unknown mortality status and no next-of-kin interview were treated as alive but lost to follow-up.

Ascertainment of time invariant and time variant predictors of migration

To account for potential confounding bias due to time-variant covariates, we estimated exposure models in which migration was the outcome, and predictors were pre-migration characteristics at each year of life, detailed below. This model yielded a migration propensity score at each year of life which was then used in inverse probability weighting, described further below. In outcome models, i.e. the models estimating the effect of migration on all cause-mortality, we adjusted for time-invariant (constant) covariates in addition to applying inverse probability weights.

Time-invariant (constant) covariates – Characteristics included in both the exposure model (i.e. predicting migration) and the outcome model (i.e. predicting mortality): As potential predictors of both migration and mortality, we included the following covariates which could not plausibly have been affected by migration (i.e. time-constant): baseline age (only in outcome model), gender, birth cohort, and parental education. We calculated birth cohort based on reported age and year of interview. In MHAS, participants reported their parents’ education as none, some primary, primary, and more than primary. In HRS, parents’ education was reported in years and was recoded to match MHAS: 0 years of education as ‘none’, 1 to 5 years as ‘some primary’, 6 years as ‘primary’, and more than 6 years as ‘more than primary’. We recoded missing parental education into its own category.

Time-variant covariates – Characteristics included only in the exposure model (i.e. predicting migration): based on prior literature,^4-15 we additionally considered time-varying factors that were plausibly confounders of the migration–mortality association, but might also have been influenced by migration, including years of own education, height, marital status, smoking status, and working status measured at each year of life either until first migration (for migrants) or until entry into the study (for non-migrants). The coding of these covariates followed the lifecourse of participants.

• Own years of education: before 6 years of age, we set years of education to 0. At 6 years of age, participant’s education increased by 1 year for each year of age until reaching participant’s reported years of education. For example, if a participant reported 3 years of education, then we set their years of education to 0 from age 0 to age 5 years, 1 year at age 6, 2 years at age 7, and 3 years at age 8 and above.

• Height was self-reported at the baseline interview. Height is not established until late adolescence, but information on height is only available at study enrollment. Therefore, we considered height as missing/unknown until age 17, at which point we considered it as equal to adult height. We recoded structurally missing height before age 17 into its own category.

• Marital status: to describe participants’ marital history, we extracted all information related to marital status at each wave. In HRS, participants reported age of the beginning and end (due to divorce or widowhood) of up to four marriages. For example, a participant who reported three marriages and two divorces, would be assigned marital status of 0 (not married) when their age was smaller than their age at first marriage, between age at first divorce and age at second marriage, and between age at second divorce and age at third marriage. We applied the same approach to creating marital status in MHAS, using extracted information related to number of marriages, and ages at first and last marriage and divorce.

• Smoking status: We constructed smoking status using information on participant’s reported ever smoking (yes or no) and age at first smoking initiation. For those who never smoked, we set the participant’s smoking status to 0 (not smoking) for all time. For those who reported that they ever smoked, we set their smoking status to 0 (not smoking) before age at first smoking, and set it to 1 at age at first smoking.

• Working status: We constructed working status using information on participant’s reported ever worked (yes or no) and age at first job. For those who never worked, we set the participant’s working status to 0 (not working) at each year of age. For those who reported that they ever worked, we set their working status to 0 (not working) before age at first job, and set it to 1 at age of first job.

The value of each pre-migration characteristic was updated at each year of life (age) from birth until migration (for return migrants and US residing migrants) or until study entry (for never migrants).

Statistical analysis:

The analysis consisted of two steps. In step 1, we quantified selection factors that might predict first migration and could also affect mortality by computing the probability an individual experienced the migration history they actually did (either migrating at their actual age of first migration or never migrating), conditional on lifetime covariates. This was summarized into a propensity score. In step 2, we examined the disparity in all-cause mortality between migrants and non-migrants, first without and then with accounting for selective migration, by weighting analyses by the inverse of the probability each individual experienced the migration history he or she actually did.

Step 1 – Calculation of the propensity to migrate score:

To estimate the migration propensity score, we fit a pooled model predicting migration at each year of life. Informed by the observed probability of migrating in the sample, we modeled age with a combination of binary (indicating fairly stable annual probability of migration when ages 0 to 14 years, ages 19–23, and ages 52+) and linear (indicating linear trends in probability of migration from ages 15–18 years and ages 24–51 years) indicators. From a list of candidate confounders defined above plus interactions of age and sex, age and education, and sex and education, we applied backward selection to define a reduced final propensity-to-migrate model. Based on a priori hypothesis, we forced main effects of all measures and age by sex interactions to remain in the model. The model also included missing indicators for most covariates. Using the selected model, we calculated the probability of migration given the individuals’ past and current covariate for each individual in the cohort for each age from birth to the maximum of either age of migration or age at study entry in the cohort. The migration probability in each year was accumulated to calculate each individual’s probability of experiencing the migration history she or he actually experienced. For individuals who never migrated, this probability was:

$$1-{\prod }_{j=0}^{\max age}\text{Pr}(Migrateatagej\mid Covariatesatagej,Migrateatagej-1=0)$$ (1)

Where the maximum age was the age at study entry (i.e. at enrollment). For individuals who migrated, this probability was:

$$\begin{gathered} 1-{\prod }_{j=0}^{\text{Migration}age-1}\text{Pr}(Migrateatagej\mid Covariatesatagej,Migrateatagej-1=0))\ast \\ \text{Pr}(Migrateatmigrationage) \end{gathered}$$ (2)

We calculated weights as the inverse of the probability of receiving the treatment that the person actually experienced (migrated or did not migrate). To reduce the influence of extreme weights, weights were stabilized by calculating a similar probability using only covariates that were included in the outcome model (e.g., time constant covariates) and using this probability in the numerator of the weight calculations. Weights were trimmed at the 1^st and 99^th percentile before use in the outcome model.

To confirm the plausibility of the migration probabilities estimated in step 1, we evaluated the observed and predicted probabilities of migration in our cohort as a function of age and sex. We also compared the probabilities of migration at each age in our cross-national cohorts to the probability of migration at each age implied by age-of-migration reported by current US residents born in Mexico, from the US Census microsample.³⁴ We also presented the final selection model predicting migration (i.e. the exposure model) and evaluated the common support for the propensity score, comparing migrants and non-migrants.

Step 2 – Estimating the effect of migration on mortality:

Participants in this study contributed observed time at risk (in years) beginning their age at enrollment (registration date) and ending at date of death or censoring at last observed visit. We defined follow-up time (observed time at risk) as: (1) for deceased participants: the difference between death date and their registration date, (2) for living participants: the difference between last day of follow-up (28 February 2013 for HRS and 30 April 2013 for MHAS) and their registration date.

We first compared migrants and non-migrants on baseline characteristics, by gender. Second, we compared migrants to non-migrants using Cox proportional hazards models for incident mortality through 2012, overall and by gender. We adjusted our initial models for sex (in the non-stratified models), baseline age, year of birth, and parental education. An additional set of models accounted for the time-varying values of the predictors of migration (own education, height, ages at first marriage, smoked and job) using the IPTWs. We also estimated the Cox models including adjustment for the predictors of migration as measured at study enrollment, although this model is expected to be biased by controlling for factors that were potentially influenced by migration. We confirmed that the association of our predictor of interest, migration, satisfied the proportional hazards assumption by showing parallel lines when plotting the graphs of the survival function as well as the graph of the log(-log(survival)). Finally, we note that the outcome models did not include the survey weights because (1) they are already included in the construction of the IPTWs and (2) the populations we wish to weight back to are the Mexican birth cohorts; the HRS survey weights, however, are meant to weight back to the US population.

We conducted two sensitivity analyses. In the first sensitivity analysis, we excluded the 493 MHAS participants with missing vital status, who we otherwise assumed were alive in the original analysis. In the second sensitivity analysis, we examined, overall and by sex, the effect of age-specific migration (instead of migration at any age) on all-cause mortality using the following cutoffs: migration prior to age 18 years old, migration between ages 18 and 34 years old, and migration at age 35+ years old.

Finally, in order to maintain the national representation of HRS and MHAS, we applied appropriate sampling weights from both studies. We also clustered standard errors at the household level, since both HRS and MHAS include eligible participants and their partners. All analyses were conducted using SAS version 9.4 and STATA 15. A co-author was not the primary coder reviewed our statistical code for the propensity score model, IPTWs, and outcome model and we posted it on the following Github page: https://github.com/AdinaZekiAlHazzouri/Selection-Into-Migration-SAS-Code.

RESULTS

Mexico to US migration for these cohorts varied in a non-monotonic pattern with age. Overall, the probability of migration was lower among females than among males; the probability of migration peaked between 20 to 30 years of age. Migration was predicted by all time-constant and time-varying covariates except smoking (Table 1). For example, the association between education and migration was gender specific and varied between primary, secondary, and post-secondary education. Taller height also predicted higher odds of migrating, whereas being unmarried or being employed predicted lower odds of migrating.

Table 1:

Predictors of migration at each year of life^a using a pooled logistic regression model, Health and Retirement Study (2000-2012) and Mexican Health and Aging Study (2001, 2003, 2012)

	OR (95% CI)
Gender (male vs. female)	1.4 (1.1, 1.7)
Age (vs. migrated at age <14 years)
While age <14 years	Ref
Per additional year of age when 14-18 yearsb	1.9 (1.6, 2.2)
While ages 19-23c	1.0 (0.75, 1.4)
Per additional year when 24-51 yearsb	0.97 (0.95, 0.98)
While ages 52+c	0.34 (0.23, 0.52)
Height (per 5cm)	1.1 (1.1, 1.2)
Own education (years)
Per year of education 0 to 5d	1.2 (1.1, 1.2)
Per year of education 6-11d	0.87 (0.81, 0.94)
Per year of education 12+d	1.0 (0.94, 1.1)
Mother’s education
None	Ref
Some primary	0.96 (082, 1.1)
Primary	1.3 (1.0, 1.6)
More than primary	2.4 (1.9, 3.0)
Father’s education
None	Ref
Some primary	0.91 (0.77, 1.1)
Primary	0.79 (0.63, 1.0)
More than primary	1.0 (0.81, 1.3)
Marital status (Unmarried vs. married)	0.80 (0.74, 0.88)
Job status (Unemployed vs. Joined the labor market)	1.4 (1.3, 1.6)
Smoking status (non-smoker vs. Initiating smoking)	0.93 (0.85, 1.0)
Male x Age (years)
Male x age 14-18 (per year of age increase)	1.1 (1.0, 1.2)
Male x age 19-23	0.98 (0.70, 1.4)
Male x age 24-51 (per year increase)	0.98 (0.97, 0.99)
Male x age 52+	0.86 (0.57, 1.3)
Age (years) x Education (years)
Age 24-51 x education <6 (per year increase)	0.99 (0.99, 1.0)
Age 24-51 x education 6-11 (per year increase)	1.0 (1.0, 1.0)
Male x Education (years)
Male x education <6 (per year increase)	0.95 (0.91, 1.0)
Male x education 6-11 (per year increase)	0.93 (0.88, 0.98)

^aBeginning from birth till age at migration (for migrants) or age at last observed interview (for non-migrants).

^bLinear terms (i.e. per additional year of age) reflecting the year-to-year trend over that age period.

^cIndicator variables reflecting the constant probability of migration over that age period.

^dLinear terms (i.e. per one additional year of education)

^eThe table shows the variables that were selected into the propensity score model predicting migration.

^fAll models account for HRS and MHAS sampling weights. For interaction terms, the estimates represent a ratio of odds ratios.

The predicted probabilities of migration calculated from the selection model described in Table 1 mirrored the observed probabilities of migration in our cohorts (Figure 3a) and the age of migration reported by Mexican migrants in the US Census data (Figure 3b). Finally, eFigure 2 suggests that there is good common support (i.e. overlap) in our estimated probability of migrating from Mexico to the US, across migrants and non-migrants.

Figure 3a:

Observed and predicted probability to migrate as a function of age and sex, Health and Retirement Study (2000-2012) and Mexican Health and Aging Study (2001, 2003, 2012)

Figure 3b.

Reported age of migration among Mexican-born 2002 US Census respondents aged 50 years and older

For male and female participants, most characteristics varied by migration status (Table 2). For males, compared to non-migrants, migrants were slightly older, had fewer years of education, and were first married at a younger age. However, compared to male non-migrants, male migrants were taller, started working at a later age, and were more likely to have mothers and fathers with more than primary education. Among females, compared to non-migrants, migrants had more years of education, were more likely to have mothers and fathers with more than primary education, were taller, and started working at a later age but initiated smoking at an earlier age. Overall, mortality was higher among male migrants compared to male non-migrants (20% vs. 19%) but lower among female migrants compared to female non-migrants (12% vs. 17%).

Table 2:

Baseline^a characteristics comparing non-migrants and migrants, across sex, Health and Retirement Study (2000-2012) and Mexican Health and Aging Study (2001, 2003, 2012)

	Male		Female
	Non-migrants n=7,409 (81%)	Migrants n=1,694 (19%)	Non-migrants n=9,545 (92%)	Migrants n=785 (8%)
Baseline age (years), mean (SD)	60.9 (9.0)	62.4 (8.9)	60.7 (9.2)	60.3 (8.7)
Own education (years), mean (SD)	5.9 (5.2)	5.2 (4.7)	4.7 (4.3)	6.1 (4.4)
Mother education, n (%)
None	3,333 (45%)	759 (45%)	4,370 (46%)	256 (33%)
Some primary	1,900 (26%)	468 (28%)	2,637 (28%)	226 (29%)
Primary	638 (9%)	154 (9%)	822 (9%)	114 (15%)
More than primary	308 (4%)	104 (6%)	361 (4%)	105 (13%)
Father education, n (%)
None	2,873 (39%)	713 (42%)	3,660 (38%)	224 (29%)
Some primary	2,049 (28%)	473 (28%)	2,768 (29%)	199 (25%)
Primary	665 (9%)	156 (9%)	894 (9%)	99 (13%)
More than primary	470 (6%)	114 (7%)	592 (6%)	109 (14%)
Height (cm), mean (SD)	165.9 (8.8)	167.6 (8.0)	155.6 (7.9)	157.5 (8.0)
Age at first marriage (years), mean (SD)	26.6 (11.2)	25.2 (7.9)	22.2 (10.0)	22.7 (7.8)
Age at first smoked (years), mean (SD)	18.8 (7.9)	18.7 (9.3)	23.9 (11.3)	22.4 (10.5)
Age at first job (years), mean (SD)	14.6 (6.9)	16.0 (8.1)	20.0 (11.9)	25.3 (10.5)
Mortality – through 2012	1394 (19%)	339 (20%)	1571 (17%)	96 (12%)

^aExcept for mortality, all values are at the participant’s first observed/attended study visit

In age-, sex-, and birth year-adjusted models, we found no relationship between migration and the risk of mortality in the overall sample (HR = 0.97; 95% CI: 0.87, 1.1) (Table 3). The association was relatively unchanged when further adjusting for parental education (HR = 0.98; 95% CI: 0.88, 1.1) or for time-varying covariates using IPTWs (IPTW HR = 1.0, 95%CI: 0.87, 1.2). In models adjusted for the time-varying predictors of migration measured at study enrollment (i.e. incorrect adjustment of post-migration values), the HR was slightly below the null, although similar to the other models, this estimate had wide confidence intervals (HR = 0.94; 95%CI: 0.82, 1.1). These relationships in the overall sample were largely similar for men and women. For example, among women using measured covariates at time of enrollment to account for selective migration, migration was associated with an imprecisely estimated 13% lower risk of mortality (HR = 0.87, 95% CI: 0.60, 1.3). When instead using IPTWs to account for selective migration, the association of migration with mortality was almost exactly null and similarly imprecise (IPTW HR= 0.98; 95% CI: 0.77, 1.3).

Table 3:

Effect of migration on all-cause mortality through 2012, overall and by sex, Health and Retirement Study (2000-2012) and Mexican Health and Aging Study (2001, 2003, 2012)

	Overall HR (95%CI)	Male HR (95%CI)	Female HR (95%CI)
Model 1: adjusted for Sex, age, and birth cohort
Non-migrant	Ref	Ref	Ref
Migrant	0.97 (0.87, 1.1)	0.99 (0.87, 1.1)	0.93 (0.76, 1.2)
Model 2: + mother’s and father’s education
Non migrant	Ref	Ref	Ref
Migrant	0.98 (0.88, 1.1)	0.99 (0.88, 1.1)	0.98 (0.79, 1.2)
Model 3: model 2 + time-varying covariates’ values at study enrollment
Non migrant	Ref	Ref	Ref
Migrant	0.94 (0.82, 1.1)	0.95 (0.83, 1.1)	0.87 (0.60, 1.3)
Model 4: model 2 + IPTW to account for selection into migration
Non migrant	Ref	Ref	Ref
Migrant	1.00 (0.87, 1.2)	1.02 (0.87, 1.2)	0.98 (0.77, 1.3)

Abbreviations: HR: Hazard ratio; IPTW: inverse probability treatment weight

^aModel 1 adjusts for sex (only in overall model), baseline age, and birth cohort; Model 2 additionally adjusts for mother’s education and father’s education; Model 3 is model 2 additionally adjusted for the baseline values of the time-varying covariates (i.e. at study enrollment): years of education, height, age at first marriage, age at first job and age at first smoking; Model 4 is model 2 additionally adjusted for the time-varying covariates in the form of time-updated stabilized IPTWs. All models account for clustering at the household level

Our findings remained largely unchanged when we excluded the 493 MHAS participants with missing vital status, who were otherwise assumed alive in the original analysis (eTable 1). Finally, while migrating before age 18 was associated with higher risk of mortality, migrating at or after age 18 was associated with lower risk of mortality, though all estimates were imprecise. Similar to the original analysis using migration at any age, the estimates from our model 4, which accounts for selection into migration, were more similar to those from models 1 and 2 than to model 3, which we considered the ‘incorrect model’ (eTable 2). Gender stratified results were similar (eTable 3).

DISCUSSION

Identifying the effects of migration on health outcomes presents substantial methodologic challenges. In this work, we appropriately evaluated the health effects of migration using harmonized cross-national datasets to account for confounding due to time-varying pre-migration characteristics. Mexico–US migration for these cohorts – ages 50+ between 2000 and 2012 in Mexico and the US – was strongly patterned by time-varying factors that occurred prior to migration and which in turn are potentially influenced by migration itself. Social factors such as education had strong, non-linear, and gender-specific relationships with migration probabilities. Height -- a broad indicator of physical health in early life reflecting the effects of nutrition, infection, and psychosocial stressors – also positively predicted migration. These findings vindicate concerns about strong selection for health among migrants.

All models had wide confidence intervals, indicating the need for more data to precisely estimate mortality effects of migration. It is worth contrasting our results with evidence from prior work, however. Models that incorrectly controlled for time-varying confounder–mediators measured at time of enrollment ( i.e. post migration; incorrect model) suggest that, for women, migration is associated with a reduced risk of mortality, although this estimate is very imprecise and the data are compatible with both large protective and deleterious associations. When instead using IPTWs, the risk of mortality is similarly imprecise. In short, despite a sample size of over 10,000 women, these data do not provide conclusive evidence on the direction of the conditional association. Among men, we estimate a similarly uncertain relationship. Although the estimator for the effect of migration on mortality is statistically indistinguishable in these data whether using the conventional model or the IPTW model, migrant selection is likely to be an important potential bias given the influence of time-varying social risk factors on migration. Accounting for such time-varying confounders correctly necessitates careful modeling, but the impact of these corrections on the effect estimate may not be statistically detectable without very large sample sizes.

Previous work, using the appropriate comparison group as ours (i.e. non-migrants from the country of origin), has employed a number of methods to address and account for pre-migration characteristics in cross-national studies^5,18-23 as well as in prospective studies that follow non-migrants in country of origin and eventual migrants over time^7,24. Examples of such methods include regression stratification,^7,20,21,24 matching migrants and non-migrants on observed characteristics (e.g. exact matching)¹⁸ or on a propensity-to-migrate score generated using pre-migration characteristics.²⁵ However, many of these previous methods did not adequately account for the time-varying nature of many pre-migration characteristics. For example, educational attainment is a major predictor of mortality³⁵ and also has been shown to be a predictor of Mexico–US migration,^14,15 albeit in complex ways that depend on historical, political, and economic context, as well as individual demographic factors and social connections to other migrants. Educational attainment at time of study enrollment (i.e. post-migration) is therefore typically included in models of the effect of migration on health outcomes. However, for individuals who migrate in early-life or who otherwise seek out schooling after migration, total educational attainment may be the result of migration. Stratifying on such a post-migration variable could induce collider bias³⁶ (see Figure 1) or over-control for a key mechanism by which migration effects mortality. Considering the entire lifecourse history of migration and educational history – as well as other time-varying lifecourse drivers of migration and mortality – allows us to appropriately account for key confounders of our association of interest while avoiding conditioning on an outcome or mediator of the association. Employment or marital history are other example of such key confounders whose values may vary during the time leading up to migration,¹¹ and migration in itself can change migrants’ opportunities and trajectories in life.^10,12,13

As discussed throughout this manuscript, to appropriately evaluate the health effects of migration requires closely harmonized cross-national data. However, few data sources to our knowledge provide harmonized measures on similar populations from both sending and receiving countries. And without cross-national data, the best evidence on the health effects of migration has often been indirectly inferred from comparisons of the health of migrants residing in the US with their US-born counterparts (e.g., 2^nd generation immigrants) or with non-Hispanic whites and blacks.^26-31. For example, it is well-established in the literature that Mexican immigrants enjoy a mortality advantage relative to US-born Mexican Americans or non-Hispanic whites.^{17,26,28,29,31,37-41} And migrant selection, though not tested or quantified in these studies, has often been proposed as one of several mechanisms by which such an advantage might occur. While these studies provide invaluable information about health disparities in the US, they cannot directly estimate migrant selection or the effect of migration on health.)

Our study has some limitations and implications for future work that are worth noting. Constructing a pre-migration history of the time-varying covariates for each subject is time consuming and weights inflated standard errors relative to those estimated with conventional regression models, yielding imprecise estimates. Our estimates also depend on the correct specification of the selection model. Compared to non-migrants, the migrant sample was relatively small in size. However, to maintain their national representation, we applied appropriate survey weights from both studies when constructing the weights. We do not address potential bias due to selection into study enrollment. We also cannot rule out unmeasured confounding as we were also missing potentially important predictors of migration to the US. For example, migration decisions are likely influenced by region of origin in Mexico^42-44, yet our MHAS data did not include information on place or region of residence in early life, prior to migration to the US. Furthermore, we computed variables such as smoking status and working status using only age at first smoking and age at first job, and as such may not capture the constructs adequately, potentially resulting in model misspecification. The adjudication of mortality differed between HRS and MHAS, and was less rigorous in MHAS especially for subjects who dropped out of the study. However, our findings remained largely unchanged when we excluded the 493 MHAS participants with missing vital status, who were otherwise assumed alive in the original analysis. Our study evaluated the total effect of migration, which is the result of initial/first migration decisions (ever vs. never migrate), but many more aspects of migration may influence health, including length of residence in the US, the residing experience in the US and even in Mexico for the return migrants. Finally, our findings pertain to Mexico–US migration occurring largely pre-IRCA (Immigration Reform and Control Act) whose dynamics are likely very different than those of more recent migrant cohorts due to difference in migrant network, policy, and demographics.⁷

Notable strengths of this study include leveraging two well-established cross-national samples that were harmonized to enable pooled analyses. In particular, the two studies enabled a better characterization of the counterfactual contrast for US migrants – when often studies use US born Hispanics or non-Hispanic whites as the non-migrant comparators. With an increasing number of high quality population health data sets in countries of origin for large migrant populations, this type of design will become feasible for many other settings.^18,25 Further, our study also builds on an important literature regarding selective migration and international migration and is the first, to our knowledge, to quantify time-varying pre-migration selection forces in the form of time-updated IPTWs. As such, our work makes an important contribution to a literature that has mainly focused on pre-migration characteristics measured only at enrollment.

In conclusion, we predict that migration patterns will continue to have profound implications for population health in coming decades. Discussions about migration policies should incorporate rigorous evidence about the effects of migration for migrants; delivering this evidence is an urgent research priority.⁴⁵ Evaluating the health effects of migration entails research designs leveraging harmonized data sets from both the sending and receiving countries, with measures and analytic strategies that can appropriately account for complex and dynamic selection processes that may vary across the lifecourse. Our study demonstrates the challenges of this approach, including sample size and modeling considerations, but also confirms the overall feasibility with existing data sets. Similar evaluations of other major migration corridors may be feasible with other emerging sources of international data sets.

Figure 1b: Schematic Diagram for the relationship between migration and health in older age, including the role of time-invariant, time-variant confounders, and selection into study enrollment^a.

^aWhile Selection into study enrollment may result in potential bias, we do not quantify it or account for it in the analysis.