Introduction
PAR conjecture
The predictive adaptive response (PAR) conjecture tested in this paper is a cornerstone of the Developmental Origins of Adult Health and Disease (DOHaD). Although under a different name, its foundation was first laid down in the work of Barker and colleagues. Reference Barker, Gluckman, Godfrey, Harding, Owens and Robinson1–Reference Lucas, Fewtrell and Cole7 The original idea is that the embryonic and fetal developmental plan is modulated and adjusted in response to cues in the placental environment triggered by maternal conditions. When confronted with adverse nutritional challenges and stress signals, the program is fine-tuned, and energy supply to different physiological systems is rerouted with a strong bias toward growth of the heart and brain. Although this may result in constrained growth of some organs and increased risks of adult chronic conditions, it does optimize short-term chances of the organism’s survival.
More recently, a rapidly growing body of research spearheaded by Gluckman, Hanson and colleagues Reference Gluckman and Hanson8,Reference Gluckman and Hanson9 builds on the original idea of fetal programming and incorporates recent advances in developmental biology, epigenetics, and evolutionary biology. Reference Bateson and Gluckman10–Reference Godfrey, Lillycrop, Burdge, Gluckman and Hanson14 According to this line of thought, humans and other mammals evolved strategies other than homeostasis and fetal programming to maximize fitness under changing environmental conditions. There is significant empirical evidence of more nuanced responses whereby early developmental adjustments are made not just in response to cues from current environmental factors but as plastic accommodations of developmental plans to assessments of future conditions. In mammals, the maternal and placental environments generate cues that operate as sensors and allow the fetus to “read” conditions that might be encountered postnatally. Shifts in those readings promote the selection of alternative developmental paths. When the fidelity of the sensing mechanism is high, the resulting phenotype will possess enhanced fitness. When the prediction is inaccurate, the organism’s fitness will be compromised. Of interest is a special class of mechanisms that modulate fetal growth and developmental plans in response to external signals that predict post-birth environments characterized by poor nutrient supply or stressful conditions that may threaten survival. These are referred to as PARs. Reference Bateson, Gluckman and Hanson15–Reference Bateson and Gluckman16
Although PAR-related responses to nutritional (or other) stresses are more plastic and efficacious than homeostatic adjustments, most are irreversible and can backfire when predictions of future environments are incongruous with those encountered after birth. This type of mismatch is common in modern populations of low- and middle-income countries (LMICs), where ancestral environments of parents (and grandparents) are characterized by scarcity and harsh conditions while their offspring are born into a world of caloric abundance and obesogenic conditions. In these cases, future environments are incorrectly predicted in utero but individuals who survive to adulthood carry with them inappropriate adjustments that increase the risks of adult chronic conditions. The best documented cases of mismatches in humans involve pathways between fetal nutritional environments, adult obesity, type 2 diabetes (T2D), metabolic syndrome, hypertension, and cardiovascular disease. Reference Jaenisch and Bird17–Reference Gluckman and Hanson20
Importantly, there is strong empirical evidence from animal studies and from human populations exposed to extreme conditions, that epigenetic processes are an important component of the suite of responses to exposures during early stages of development. Reference Jaenisch and Bird17
Empirical evidence for PAR
Despite its theoretical soundness and appeal, it has been difficult to find empirical evidence that falsifies hypotheses from DOHaD in humans. A robust test of the PAR conjecture requires randomized control trials that compare the prevalence of adult chronic conditions in two, otherwise identical, populations that differ only in the presence (absence) of mismatches between prenatal and postnatal environment. The conditions for such a randomized control trial are, however, strict, and unlikely to be satisfied anywhere except in animal experiments. The bulk of empirical evidence for PAR and related mechanisms among humans is drawn from studies of adult health outcomes in individuals who survive to adulthood after being exposed early in life to episodic or long-lasting exogenous shocks that caused deprivation and stress such as wars, Reference Ramirez and Haas21,Reference Haas and Ramirez22 natural disasters, Reference Palloni, McEniry, Huangfu and Beltran-Sanchez23 famines, Reference Li and Lumey24 pandemics, Reference Almond25 and economic crisis. Reference Schmitz and Duque26 Much of the empirical evidence gathered thus far shows effects, albeit small, consistent with PAR. An important drawback that limits the inferential power of these studies is that they are vulnerable to sample selection, measurement problems, and omitted variable biases.
Why study migrant populations?
An alternative research strategy for testing DOHaD hypotheses is to identify populations that approximately satisfy strict requirements of randomized trials. Under proper conditions, a comparison of migrant populations from LMIC and native populations in high-income countries (HICs) might fulfill, albeit imperfectly, some of the randomized trial criteriaFootnote a . Most migrants from LMIC experience sharp contrasts between their ancestral environments and those that prevail at destination and are, therefore, more likely to be exposed to mismatches that increase the risk of chronic conditions. Furthermore, offspring of the first cohort of migrants will also experience the clash between their ancestral (parental and grandparental) and current environments during critical periods of growth and development.
There is an additional, but mostly neglected, set of conditions that makes the study of modern migrants from LMIC a potentially powerful tool for testing PAR conjectures. The bulk of adult populations residing in LMIC countries experienced an unprecedentedly rapid mortality decline that began in earnest after 1945–1950. Unlike HICs, these improvements in survival were driven less by amelioration of standards of living than by the diffusion of new medical technology (sulfa, antibiotics, pesticides, etc). Reference Palloni and Wyrick27,Reference Preston28 Under this emerging mortality regime, children who experienced hardship and adverse early conditions that would have been lethal under the old mortality regime, are able to survive to adulthood. They will thus increase the fraction of modern adult populations in LMIC scarred by early experiences and primed to manifest delayed adult responses. When there is a strong correlation between the severity of early deprivation and child mortality levels, the size of the adult population at risk of expressing delayed effects is reduced. When these two phenomena are decoupled, as happened under mortality declines experienced by modern LMIC, delayed effects and the manifestation of PAR mechanisms are more likely to be observed.
Although a comparison between these modern migrant and native populations is very distant from a rigorous randomized control trial, it can generate useful empirical evidence. Over the past ten years or so, a growing number of studies have focused on the health status of migrant populations and in some cases compared them with those of the native populations. Although most of these do not directly address PAR conjectures, all include potentially valuable empirical evidence. To harness the latent power of this evidence, we carry out a rigorous meta-analysis of selected studies. It is well known that inferences from meta-analysis can be as strong or stronger that those from individual studies. Reference Borenstein, Hedges, Higgins and Rothstein29–Reference Harrer, Cuijpers, Furukawa and Ebert31
Hypotheses
To parse the meta-analysis into well-defined components and to organize the data analysis, it is most useful to formulate expected empirical findings in the form of four precisely stated hypotheses:
i. The prevalence of obesity, T2D, and hypertension is higher among adult migrants compared to the adult native populations.
ii. The prevalence of obesity is higher among children of adult migrants born either in the country of origin or of destination, compared to native children in the population at destination.
iii. There should be no differences between adult migrants and adult native populations in health outcomes not directly implicated by PAR.
iv. Differences between migrant populations and populations of stayers with similar ancestry in the country of origins should be at least as large as those in (i).
Method
Search strategy and selection criteria
The systematic review and meta-analysis were conducted using the PRISMA guidelines as reference (Table S1). Reference Page, McKenzie and Bossuyt32 Search was carried out in PUBMED and SCOPUS databases for studies published prior to May 2021, reporting health outcomes associated with DOHaD and/or migrants, without language restrictions. In addition, we used reference lists and relevant reviews to identify additional studies of interest. Terms used in the search are in Table S2. The initial selection was narrowed down to a reduced number of health outcomes and contrast groups identified in our hypotheses. The final database includes original observational (cohort and cross-sectional) studies that at least meet the following inclusion criteria (see full set of evaluation criteria in Supplementary Text S1): (1) reports estimated effects (and their standard errors) on either prevalence or odds of diabetes, hypertension, or obesity/overweightFootnote b ; (2) unambiguously defines the LMIC migrant group and the contrast reference group. The latter must be either the native HICs host population and/or the population at origin; (3) the sample sizes of migrants and reference populations are adequate, and 4) the models employed to estimate effects on prevalence or odds ratios, include full controls for age, sex, and SES/education. Controls for age and sex are essential for these studies to produce useful inferences. A control of socioeconomic status (SES) (and/or education) is necessary as socioeconomic condition is a potent confounding factor associated with both migrant status and health outcomes. All studies focused on well-defined subpopulations of first- and second-generation migrant children (up to 10 years of age) and adults. We excluded studies with coarse, broad, or ill-defined immigrant or ethnic groups, those that focused on health outcomes not included as our chosen targets, those that did not use a native population from HICs as a contrast, and those with very small and/or nonrepresentative samples. Finally, we excluded studies limited to maternal or perinatal outcomes as well as those that report estimated effects on continuously measured BMI, blood pressure, or glucose tolerance.
The database search resulted in 22,635 articles of which 22,324 were duplicates or considered not relevant based on title and abstract. An additional 48 studies were identified in citations, resulting in 359 articles which were categorized by outcome and migrant origin and destination to identify unique studies that approximately satisfy PAR conditions (29 reviews were excluded at this stage). One hundred ninety-nine full-text articles were assessed for eligibility (12 could not be retrieved), of which 38 met the inclusion criteria (Fig. 1). Three of these included minority populations as part of the host contrast group and were only used in sensitivity analyses. Table 1 summarizes the characteristics of included studies. A narrative synthesis is in Supplementary Material Table S3.
Study selection.
In studies reporting estimates of prevalence in the reference group but only estimated effects on odds ratios, an approximation was employed to compute effects on prevalence. Reference Zhang and Yu33 This allows pooling studies that only reported odds ratios with those that utilized outcomes’ prevalenceFootnote c . Altogether, 38 studies and 272 unique effects were included in the main analyses.
Data analysis
All analysis were completed using STATA (version 17.0) and R (version 4.1.2) packages METAFOR (version 4.1.3), DMETAR (version 0.0.9000), and META (version 5.2-0). Studies with prevalence estimates of conditions (obesity, T2D, hypertension, other) were pooled into separate groups to obtain estimates of effects associated with a migrant group on the prevalence rate of each of these health outcomesFootnote d . With a few exceptions noted in the text, analyses are confined to contrasts between a migrant group and the native population. The latter excludes migrants from Western Europe and/or the USA and Canada as well as minority populations residing in the place of migrants’ destination. We estimated random effects (RE) models by subgroups (and/or with moderators) and generated summary estimates of effects sizes and corresponding 95% CI’s. Observations to compute estimates of effect sizes for a single outcome include multiple estimates associated from the same studyFootnote e .
Between-study heterogeneity was assessed using the I 2 statistic. To partially account for between-study heterogeneity we use group-specific estimation, and differences by region of origin and/or destination and by baseline contrast group were evaluated. Parameter estimation is extended to models that include region of origin as a moderator. The statistic to assess heterogeneity, I 2, has well-known flaws Reference Borenstein, Higgins, Hedges and Rothstein34,Reference Migliavaca, Stein and Colpani35 and, to partially circumvent them, we used the range of study-specific estimates, a more informative measure of between-study heterogeneity. This is justified because in most cases the estimated effects from different studies are of the same sign and their ranges center quite a distance away from the null or no-effect values. Finally, several sensitivity tests were implemented to verify the robustness of findings. These are described in Supplementary Materials (Section II).
Results
Table 2 displays a global estimate of effects for migrant excess risk with host population as contrast regarding all outcomes. Tables 3–7 do the same according to region of origin. Finally, Tables 8–16 display results of meta-analyses in subgroups defined by region of origin and by combination of origin and destination. The first set of results pertains to estimation of effects of migrant status on the prevalence of T2D in studies in which the samples included population aged 18+ and the contrast group was the host population. Fig. 2 shows that study effects for T2D are consistently large and statistically significant. Effects sizes are, on average, of the order of logPR 0.40 (95% CI: 0.30-0.50), with fairly narrow confidence intervals. This implies that adult migrants experience T2D risks about 48% larger than the native populations. Subgroup analyses defined by region of origin suggests that African, Asian, and South American migrants fare worse than the contrast populations. The effect sizes range between logPR 0.37 and logPR 0.49, with African migrants exhibiting the worst profile. Significantly, and as expected, migrant groups originating in neither of these regions (mostly European migrants) do not experience excess risks.
Diabetes risk of migrants vs. host population (by place of origin).
Study characteristics
OW = overweight (includes obesity); OB = obesity; T2D = type II diabetes; Diabetes = unspecified type I or type II; HT = hypertension; Other = outcomes not associated with PAR (predictive adaptive responses): asthma, cancer, and mental disorders. LMICs = low- and middle-income countries; HICs = high-income countries; SEA = South-East Asia; BMI = body mass index; SES = socioeconomic status; Nr = not reported.
Original estimated effect.
a Odds ratio/beta.
b Odds ratio; without sufficient information for transforming to prevalence ratios.
c Hazard ratio.
d Inversed reference group.
* All include both genders except study by Simchoni (2020) (male sample only).
† Rounded to one full year or whole number. Range of mean age when results are presented separately for males/females or different age groups.
Estimates for migrant excess risk with host population as contrast
“Other” includes outcomes not directly invoked in DOHaD as a result of exposures to early conditions: asthma (n = 5), cancer (n = 3), and mental disorders (n = 2).
T2D by subgroup of origin (migrants vs. host)
Obesity by subgroup of origin (migrants vs. host)
Hypertension by subgroup of origin (migrants vs. host)
Child obesity by subgroup of origin (migrants vs. host)
Other by subgroup of origin (migrants vs. host)
SA + Mx = South America and Mexico.
“Other” includes outcomes not directly invoked in DOHaD as a result of exposures to early conditions: asthma (n = 5), cancer (n = 3), and mental disorders (n = 2).
T2D estimates with contrasts in origin population
T2D estimates with contrast origin population and ignoring size effects > 1
T2D estimates by origin and destination combination (migrant vs. host)
Obesity estimates with contrasts in origin population
Obesity estimates with contrast origin population and ignoring size effects>1
Obesity estimates by origin and destination combination (migrant vs. host)
Hypertension estimates with contrasts in origin population
Hypertension estimates by origin and destination combination (migrant vs. host)
Child obesity estimates by origin and destination combination (migrant vs. host)
SA + Mx = South America and Mexico; EU = Europe; NA = North America.
Tables 8–16 do not include “other” migrant as the combination of origin destination leads to an unfeasible large number of groups.
A table for child obesity with contrast to origin populations could not be estimated because there are no observations.
To address Hypothesis 4 requires knowing whether migrants from a region experience worse conditions than those who stayed behind with whom they share similar ancestry. We estimate models in which the contrast group is always the stayers’ population at origin. Fig. 3 shows that, on average, migrants experience excess risks of T2D close to twice (logPR 0.59 [95% CI: 0.34–0.85]) as large as their those residing in regions of origins, their ancestral populations. The bulk of the burden, however, is borne by African migrants who experience risks nearly three times as large. In contrast, the difference between Asian and South American (plus Mexican) migrants and their peers at origins are centered around 0 (Table 8). Alternative results drawn from a sample of studies that excludes all those in which the effect sizes exceeded 1 are presented in Table 9. Although the total effect becomes statistically insignificant, it is still the case that African migrants are at higher risk than their nonmigrant counterparts and by a large margin.
Diabetes risk of migrants vs. peers-in-origin (by place of origin).
We also investigate whether the combination of place of origin and destination matters as much or more than the place of origin. This is a more direct test of the idea that it is the degree of dissonance between migrants’ current and ancestral environments that matters. Table 10 displays estimates of models that shed some light on this conjecture. These models were estimated using results from studies in which effect sizes are measured for African and Asian migrants who migrated to North America and EuropeFootnote f . The results suggest that Asian migrants fair worse and by a large margin, pointing to excess T2D risks that top 58% (logPR 0.46 [95% CI: 0.17–0.76]) of those in the host populations, both in Western Europe and North America. Only African migrants to Europe (but not to North America) perform equally bad or worse, with excess risks of the order of 200% or more. An explanation for this result can be deduced from the PAR conjecture: if the effect of mismatches is equally powerful among migrants to either Western Europe or the USA, it should be more visible in the former as the overall levels of T2D in the native population are much lower than in North America. Another explanation is that there is stronger migrant selection into North America than to Europe and much less so among Asians in both places of destinationFootnote g .
Findings for adult obesity are less sharp than those for T2D. Column 2 of Table 2 shows that the migrant effect on obesity prevalence is negative and with a confidence interval whose upper bound is 0 (logPR −0.11 [95% CI: −0.21–0.01]. Furthermore, subgroup analysis in Fig. 4 suggests that the average effect is similar across all migrant groups. Note that the effect associated with Asian migrants is significantly different from 0 but negative, for example, Asian migrants experience less obesity than native populations. Although this result is consistent with empirical findings regarding Asian migrants in general, Reference Gong, Shi and Huang74 it is inconsistent with studies showing that Asians of Indian origin fare much worse than native populations. Reference Fernandez, Miranda and Everett75 Fig. 5 reveals a startling result: the migrants’ pooled sample experience twice as large a prevalence of obesity (logPR 0.75 [95% CI: 0.40−1.09]) as their peers left behind (Table 11). This excess is reduced to logPR 0.12 [95% CI: 0.00−0.25] when we exclude all studies that report size effects larger than 1 (see Table 12). Although this is consistent with the PAR conjecture, other explanations are possible. Thus, for example, strong migrant assimilation effects could produce similar patterns even in the absence of mismatches. This latter argument, however, relies on the assumption that African migrants are much more sensitive to assimilation effects than Asians, something we cannot verify with these data.
Adult obesity risk of migrants vs. host population (by place of origin).
Adult obesity risk of migrants vs. peers-in-origin (by place of origin).
As in the case of T2D, Table 13 reveals that African migrants exhibit higher levels of obesity that native populations but only if their region of destination is Western Europe, not North America. As before, this finding could be explained by invoking the PAR conjecture or could be accounted for the fact that the average prevalence of obesity is much larger in North America than in Europe. It can also be dismissed altogether as an outcome of differential migrant selection by place of destination.
Table 2, column 3 shows that the total effect of migrant status on prevalence of hypertension is centered on zero (logPR 0.07 [95% CI: −0.10–0.15] and it is only significant among migrants originating in Africa (Fig. 6). Note that migrants from “other” regions fare better than natives, as they are likely not at risk of PAR (as are Africans and Asians). Fig. 7 and Table 14 show that African and South American (plus Mexican) migrants are particularly prone to hypertension when compared with populations of origin, and that African and Asian migrants to Western Europe fair worse than native host populations. Only Asians suffer a higher risk with respect to North American host population (Table 15).
Hypertension risk of migrants vs. host population (by place of origin).
Hypertension risk of migrants vs. peers-in-origin (by place of origin).
The rationale for Hypothesis 2 is that, under conditions regulated by PAR, young children born in the country of origin or destination should be more likely to be exposed to contrasts between “current” and ancestral environments. It is, after all, the population of children that experiences the full blow of disharmony between ancestral and current conditions during a most sensitive period of growth and development, either in utero, infancy, early childhood, or combinations these.
Because the number of unique effects including children younger than 10 years of age is small (23 effects, from 7 studies of migrants to Western Europe), our inferences are tentative. Estimates in Table 2 (column 4) indicate that, on average, child obesity prevalence among migrants is about 22% higher than among natives (logPR 0.20 [95% CI: 0.04–0.36]. Fig. 8 and Table 16, however, suggest that not all migrant groups are equal: only the children of Asian migrants experience statistically significant excesses. Prevalence of child obesity among African migrants is higher as well, but the magnitude of the excess is not statistically significantFootnote h .
Risk of obesity among migrant children/first generation children vs. host population (by place of origin).
According to Hypothesis 3 migrant groups should not experience worse conditions than populations at destination when the health outcome is not one influenced by PAR conditions. Unfortunately, there are just 10 effects from only three studies that account for non-PAR-related health outcomes. From these, five effects referred to asthma, three to cancer, and two to mental disorders. They do not include observations of African or South American (plus Mexican) migrant populations. Estimates in Table 2 (column 5) and Fig. 9 show that migrants from other regions and from Asia experience risks that are between 50 and 60% percent lower than populations at destination. This is consistent with Hypothesis 3.
Risk of non-PAR-related health problems in migrants vs. host population.
Discussion
Results from the meta-analysis lead to five inferences. First, we find abundant support for the part of Hypotheses 1 and 4 that refers to T2D. All migrant groups, irrespective of origin or destination experience substantially higher risks of T2D than either native populations at destination or peer populations at origins. Second, less robust is the evidence for the part of Hypotheses 1 that refers to obesity and hypertension. Indeed, although we find either no or negative migrant effects (in the case of Asians) for obesity and hypertension, we uncover significant contrasts in the expected direction when comparing African (obesity and hypertension) and Asian (hypertension) migrants with the European host populations. Third, with respect to Hypothesis 4, migrants show substantially higher risks of obesity and hypertension in comparison with their populations at origin, but this effect is driven by African migrants. Fourth, although based on a smaller sample of studies, we find support for Hypothesis 2 as there are large excesses of child obesity among all migrants, particularly among those from Asian origins. Fifth, and as expected by Hypothesis 3, we found no migrant effects for other health outcomes unrelated to the PAR conjecture.
Despite satisfactory performance of multiple robust sensitivity tests to detect flaws in model estimation (see Supplementary Material, Section II), our study shares limitations inherent to all meta-analyses. First because we do not have access to the original data, we cannot always ensure that proper controls for confounding variables were always introduced or are comparable across studies. Our protocol only required controls for age, gender, and educational attainment or other measures of SES. Similarly, the studies’ statistics we used are selected and may not always coincide with the whole suite that could be employed to test the PAR conjecture. We only used those required by meta-analytic models that were available in the publications, for example, obesity rather than BMI. Second, some findings may be accounted for alternative explanations. A particularly important one is related to migrants’ assimilation and selection. Because most studies do not control for duration since migration, the influence of assimilation is unaccounted for. However, although this may play a role in comparisons involving migrants and population at origins, it is irrelevant to account for differences between migrants and population at destination. In fact, because duration distributions are likely to be left skewed and mismatches are more likely to be manifested at longer durations, our estimates might be biased downward, not upward. A limitation unique to our study is that we are not able to distinguish more sharply between migrants from one region and the chosen destination. Thus, our findings regarding obesity and associated excesses among migrants to Europe (but not to North America) is not unambiguous support for PAR as it could be accounted by differential migrant health selection.Footnote i Despite these limitations, the present study shows important empirical evidence consistent with selected PAR-like conjectures.
Two classes of implications can be drawn from our study. The first is substantive and relates to the strength of findings. Results of our meta-analyses are uniformly consistent with the most important hypotheses derived from the PAR conjecture. Patterns of contrasts between migrants and nonmigrants (at origin and destination) regarding T2D, hypertension, and child obesity constitute an evidentiary corpus superior to that embedded in each of the individual studies separately. Some of the effects sizes, such as for T2D, are substantial and, very likely, underestimated. The non-finding associated with health outcomes not implicated by PAR offers an important complement to the positive evidence for the other outcomes.
The second implication is for future research based on migrant populations and targeting DOHaD hypotheses. As anticipated by Gluckman and Hanson, Reference Gluckman and Hanson19 studies of migrant populations are highly valuable for they can produce empirical evidence that, albeit in a very limited way, has at least some of the merits of randomized trials. To maximize their inferential power, however, they should clearly identify migrants at origin and destination, not just the latter as is normally done. To maximize the robustness of each study and any meta-analysis including them, these studies should follow a common template guiding their sampling plan, study design, and content. Our paper demonstrates that inferences generated with pooled studies of this kind can be powerful, all the more so if they are articulated ex ante.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S2040174423000429.
Data availability statement
Data extracted from the published studies included in our analyses and the statistical code for our main model and select subgroup and sensitivity analyses will be made available online at https://github.com/palloni/metaanalysis.
Acknowledgments
We thank Bertie Lumey for useful suggestions and discussions. We are also grateful for the input from members of the ECHO research group.
Financial support
Guillermo Gomez and Clara Bueno were supported by the program of Plataformas Transversales Interdisciplinarias in Neuroaging supported by the Cajal International Neuroscience Center of the Spanish Consejo Superior de Investigaciones Científicas (CSIC) under the Next Generation EU program. Alberto Palloni’s research was supported by the National Institute on Aging via research project grants R01-AG016209, R03-AG015673, R01-AG018016, R37-AG025216, R01AG056608, and R01AG52030; by a Fogarty International Center award for Global Research Training in Population Health, D43-TW001586; by core grants to the Center for Demography and Ecology, R24-HD047873 and to the Center for Demography of Health and Aging, P30-AG017266, both at the University of Wisconsin–Madison; and by a European Research Council (ERC) grant under the European Union’s Horizon 2020 research and innovation program (grant agreement no. 788582).
Competing interests
None.
Open access
