The Hispanic paradox and the ecological fallacy: explanatory hypotheses and what the evidence shows

Contents
  1. The Hispanic Paradox and Differences Across Indicators
  2. Testing the Healthy Migrant Effect and Salmon Bias
  3. Ethnic Misclassification on Death Certificates
  4. Differences in Smoking Behavior and Lifestyle Contributions
  5. Diversity Within the Population and Changes During the Pandemic
  6. Cautions When Reading Group-Level Data
  7. References

Consider two populations: one with lower incomes and a higher proportion of uninsured people, and one without these disadvantages. Which group would you expect to live longer? Intuitively, the socioeconomically advantaged group would seem more likely to stay healthy and live longer. Yet US epidemiological studies have repeatedly reported a reversal: Hispanic people, despite being disadvantaged in income and health insurance coverage, have overall mortality rates that are similar to or even lower than those of people classified in US statistics as non-Hispanic White. Hereafter, we refer to this reference population as the "comparison group." In a 2013 meta-analysis of 58 longitudinal studies (4,615,747 participants in total), Ruiz et al. showed that Hispanic people had a 17.5% lower risk of all-cause mortality compared with other racial and ethnic groups (odds ratio 0.825, 95% confidence interval 0.75–0.91)1. This phenomenon is known as the Hispanic paradox. In a 1986 review, Markides and Coreil summarized the health status of Hispanics in the southwestern United States and termed this an "epidemiologic paradox"2.

Odds ratio for all-cause mortality among Hispanics (pooled from 58 longitudinal studies)
0.70.80.91.0 0.825 Nodifference ← Lower mortality riskHigher →

Ruiz et al. (2013). Dot shows the odds ratio; horizontal line shows the 95% confidence interval.

This issue is not limited to Hispanic populations in the United States; it also arises when reading regional mortality maps or aggregated data at the municipality or county level. For example, an aggregate finding that "areas with a higher proportion of Hispanic residents have lower overall mortality" does not, by itself, mean that Hispanic individuals have lower mortality. Mistaking a group-level correlation for an individual-level characteristic is called the ecological fallacy3. To determine whether a numerical difference reflects a difference in individual risk or a combination of migration, recording practices, lifestyle habits, and regional characteristics, we need to examine how the data were generated.

A map divided into four regions containing many small people. A person standing beside it holds a magnifying glass over one person in one of the regions
Regional aggregates and the individuals living in those regions

The Hispanic Paradox and Differences Across Indicators

The lower mortality rate among Hispanics does not apply to every disease or every comparison group. According to Markides and Coreil's 1986 review, Hispanics in the southwestern United States (mostly Mexican Americans), despite being socioeconomically disadvantaged, had levels similar to the comparison group on major health indicators such as infant mortality, life expectancy, mortality from cardiovascular disease and major cancers, and measures of physical functioning2. On the other hand, they were reported to be clearly disadvantaged compared with the comparison group for diabetes, infectious diseases, and parasitic diseases2.

Differences by indicator (southwestern United States)
Similar to the comparison group
Infant mortality, life expectancy, mortality from cardiovascular disease and major cancers, physical functioning
Worse than the comparison group
Diabetes, infectious and parasitic diseases

Review by Markides and Coreil (1986)

Results also vary depending on which group is used for comparison and which age group is examined. In the meta-analysis by Ruiz et al., which pooled studies from January 1990 to July 2010, Hispanics had a lower risk of all-cause mortality than the comparison group, but a higher risk than Asian Americans1. Furthermore, the difference in mortality risk was more pronounced in older populations, and an advantage was observed both in initially healthy populations and in populations with cardiovascular disease1.

A 2014 scientific advisory published by the American Heart Association (AHA) also summarized the state of cardiovascular disease and stroke among Hispanics, the largest minority population in the United States4. To avoid overgeneralization, the advisory did not lump all Hispanics together and instead specified which subgroups were studied wherever possible4. The Hispanic paradox is not a uniform phenomenon across all diseases or all populations.

Testing the Healthy Migrant Effect and Salmon Bias

As reasons why a socioeconomically disadvantaged group would have lower mortality, selective migration was the first explanation proposed. The two leading hypotheses are as follows.

  • Healthy migrant effect: The hypothesis that healthier people are more likely to migrate to the United States, keeping the mortality rate of Hispanics living in the United States low as a result.
  • Salmon bias: The hypothesis that migrants who become ill or elderly return to their country of origin, so their subsequent deaths are not recorded in US statistics, causing statistical mortality rates to be underestimated.
A dotted path connects two houses on the left and right. A person carrying a pink suitcase walks toward the house on the right, while an older person with a cane walks back toward the house on the left
Migration of healthy individuals (left to right) and return migration due to illness or old age (right to left)

If these two hypotheses were the primary causes, what should we see in the data? If salmon bias were driving down mortality rates, the advantage should disappear in groups for whom return migration is difficult, or in groups whose deaths are still recorded in US statistics even if they return to their place of origin. Likewise, if the healthy migrant effect were the main cause, no advantage would be expected among US-born Hispanics who have never migrated.

This prediction was tested directly in a 1999 study by Abraído-Lanza et al.5 Using data from the National Longitudinal Mortality Study, the researchers analyzed 301,718 people in the comparison group and 17,375 Hispanics aged 25 years and older5. As groups unlikely to be affected by salmon bias, they examined Cuban Americans, for whom political barriers made return migration difficult; Puerto Ricans, whose deaths in Puerto Rico are also counted in US statistics; and US-born Hispanics, who are unaffected by migration5. Contrary to the hypotheses' predictions, both Cuban Americans and Puerto Ricans had lower mortality than the comparison group, and even US-born Hispanics had lower mortality than US-born members of the comparison group5.

Three groups unlikely to be affected by salmon bias or the healthy migrant effect
GroupResult
Cuban AmericansLower mortality than the comparison groupReason selected: return migration is difficult
Puerto RicansLower mortality than the comparison groupReason selected: deaths in Puerto Rico are also included in US statistics
US-born HispanicsLower mortality than US-born members of the comparison groupReason selected: neither migrated nor returned

Abraído-Lanza et al. (1999)

Furthermore, in 2008, Turra and Elo used US Social Security Administration beneficiary records and vital statistics to measure the actual magnitude of salmon bias among Social Security beneficiaries aged 65 and older6. They confirmed that while salmon bias does exist, its magnitude is small and it cannot be the main factor explaining the lower mortality of Hispanics6.

Ethnic Misclassification on Death Certificates

Because migration alone could not explain the paradox, researchers next examined whether the cause lay in how statistical data are recorded. This hypothesis, known as ethnic misclassification, proposes that people identified as Hispanic in population statistics such as the census (the denominator of the mortality rate) are erroneously recorded as non-Hispanic on death certificates (the numerator), making the apparent mortality rate look lower than it is.

If discrepancies in statistical classification were the cause, what would happen when survey records and death certificates are linked at the individual level to correct for misclassification? Hispanic mortality rates should rise substantially, and the gap with the comparison group should narrow or disappear.

A 2010 study by Arias et al. tested this point7. The researchers linked national survey records from 1979 to 1998 with death certificates to check agreement on Hispanic classification7. They found that the accuracy of Hispanic reporting on US death certificates was generally good, with only 5% more people recorded as Hispanic in survey records than on death certificates7. Moreover, even after recalculating age-adjusted mortality rates to correct for this misclassification, Hispanic mortality remained about 20% lower than that of the comparison group7. They concluded that statistical misclassification is not the primary cause of the paradox7.

How misclassification can make mortality appear lower
Denominator: population or survey records
Counted as Hispanic
Numerator: death certificates
If recorded as non-Hispanic, not counted among Hispanic deaths

Verification resultThe number of people recorded as Hispanic was only 5% higher in survey records than on death certificates. Even after correction, mortality remained about 20% lower than in the comparison group

Arias et al. (2010)

Differences in Smoking Behavior and Lifestyle Contributions

As candidates for what migration and statistical bias could not explain, Markides and Coreil had pointed to cultural practices, family support, and diet2. One factor examined in subsequent research is differences in smoking behavior. If lower smoking rates were the reason, what patterns would appear in life expectancy differences and in behavior before and after migration? If low smoking rates are driving down mortality, fewer smoking-attributable deaths should account for a large share of the life expectancy gap. Also, if only non-smokers were migrating under the healthy migrant effect, migrants would be expected to have lower smoking rates than people who remained in their country of origin.

In a 2013 study, Fenelon tested this question using nationally representative survey data, focusing particularly on Mexican Americans8. The analysis showed that Mexican Americans—both US-born and foreign-born—had longer life expectancy than the comparison group, and more than 60% of that advantage could be explained by fewer smoking-related deaths due to lower smoking rates8. Meanwhile, smoking rates among Mexicans who migrated to the United States were similar to those among people who remained in Mexico, and both were substantially lower than in the US comparison group8. In other words, no evidence was found that only healthy individuals were selected for migration with respect to smoking habits8.

The impact of smoking has also changed over time. In 2015, Fenelon and Blue analyzed trends in life expectancy at age 50 from 1990 to 20109. The Hispanic life expectancy advantage over the comparison group widened from 2.14 years (95% CI 1.99–2.30) to 3.53 years (3.42–3.64) among women, and from 2.27 years (2.14–2.41) to 2.91 years (2.81–3.01) among men9. More than 40% of the widening advantage among women was due to a growing gap in smoking-attributable deaths, whereas smoking contributed little among men9.

Years by which Hispanic life expectancy at age 50 exceeded the comparison group
Women19902.14 yrs20103.53 yrsMen19902.27 yrs20102.91 yrs

Contribution of smokingMore than 40% of the widening among women was due to a growing gap in smoking-attributable deaths. Among men, the contribution of smoking was small

Fenelon and Blue (2015)

Diversity Within the Population and Changes During the Pandemic

Including the effects of smoking, the Hispanic advantage is not seen equally across all subgroups. In Fenelon's 2013 study, while US-born Mexican Americans retained a longevity advantage over the comparison group, among non-Mexican Hispanics, an advantage was seen only among the foreign-born, whereas the US-born fared worse than the comparison group8.

In a 2015 study of older adults aged 65 and over, Lariscy et al. linked survey data from 1989 to 2006 to analyze cause-specific mortality10. Although Hispanic mortality was lower than in the comparison group, the advantage was particularly pronounced among foreign-born Hispanics and for smoking-related causes of death10. At the same time, the authors noted that if Hispanics did not face socioeconomic disadvantages, their mortality levels would be even more favorable10.

During the 2020 COVID-19 pandemic, the drop in life expectancy varied widely by race and ethnicity. According to estimates published by Andrasfay and Goldman in 2021, overall US life expectancy at birth in 2020 fell by a median estimate of 1.13 years to 77.48 years (and by 0.87 years at age 65)11. By race and ethnicity, the comparison group experienced a drop of 0.68 years, whereas Hispanics experienced a sharp decline of 3.05 years11. As a result, the longstanding Hispanic life expectancy advantage of more than 3 years over the comparison group was estimated to shrink to less than 1 year11. The authors explained that this disproportionate burden on Hispanics reflected persistent structural inequities that increase both the risk of exposure to infection and the risk of death once infected11.

Reduction in life expectancy at birth in 2020 (medium estimate)
Overall1.13 yrsComparisongroup0.68 yrsHispanic3.05 yrs

Plotted from estimates by Andrasfay and Goldman (2021)

What we have learned from this body of research is as follows. The portion explainable by salmon bias or death certificate misclassification is small, and among Mexican Americans, lower smoking rates explained most of the life expectancy advantage. That advantage varies by country of origin, nativity, and cause of death. In 2020 estimates, the life expectancy advantage over the comparison group was projected to narrow to less than 1 year.

Explanatory hypotheses and verification results
HypothesisVerification result
Healthy migrant effectUS-born Hispanics also have lower mortalityAbraído-Lanza et al. (1999)
Salmon biasCuban Americans and Puerto Ricans, for whom return migration is difficult, also have lower mortality. Salmon bias exists, but its magnitude is smallAbraído-Lanza et al. (1999), Turra and Elo (2008)
Death certificate misclassificationEven after correction, mortality remains about 20% lower than in the comparison groupArias et al. (2010)
Lower smoking ratesExplains more than 60% of the life expectancy advantage among Mexican AmericansFenelon (2013)

Cautions When Reading Group-Level Data

In studies examining regional health disparities, aggregated data at the county or state level are sometimes used instead of following individuals one by one. Suppose, for instance, that an analysis of regional data finds a correlation: "the higher the proportion of Hispanic residents in an area, the lower the overall mortality rate of that area." From this result alone, we cannot conclude that "regional mortality is lower because Hispanic individuals have lower mortality."

It is not uncommon for a correlation observed in group-level aggregates to disagree with individual-level behavior or risk. In his classic 1950 paper, Robinson argued that group-level correlations (ecological correlations) and individual-level correlations are distinct, and the former cannot be used as a substitute for the latter3. The example given at the beginning is an instance of this ecological fallacy. The proportion of Hispanic residents in an area may be serving as a marker for other regional characteristics, such as whether the area is urban, has well-developed healthcare resources, or has a large population size.

The same caution is needed when interpreting regional mortality indicators through maps and aggregate figures, as in Maps of 35 causes of death, one click apart. When looking at regional-level studies or maps, check the following three points.

Three points to check when reading regional-level studies or maps
  1. Compare with individual-level dataDoes the regional-level association point in the same direction in individual-level follow-up data?
  2. Suspect other regional characteristicsIs it acting as a marker for urbanization, well-developed healthcare resources, or large population size?
  3. Check for adjustmentsAre regional socioeconomic conditions and healthcare access adjusted for in the analysis?

References

  1. Ruiz JM, Steffen P, Smith TB. Hispanic mortality paradox: a systematic review and meta-analysis of the longitudinal literature. Am J Public Health. 2013;103(3):e52-e60. https://doi.org/10.2105/AJPH.2012.301103
  2. Markides KS, Coreil J. The health of Hispanics in the southwestern United States: an epidemiologic paradox. Public Health Rep. 1986;101(3):253-265. PMID: 3086917
  3. Robinson WS. Ecological correlations and the behavior of individuals. Am Sociol Rev. 1950;15(3):351-357. https://doi.org/10.2307/2087176
  4. Rodriguez CJ, Allison M, Daviglus ML, et al. Status of cardiovascular disease and stroke in Hispanics/Latinos in the United States: a science advisory from the American Heart Association. Circulation. 2014;130(7):593-625. https://doi.org/10.1161/CIR.0000000000000071
  5. Abraído-Lanza AF, Dohrenwend BP, Ng-Mak DS, Turner JB. The Latino mortality paradox: a test of the "salmon bias" and healthy migrant hypotheses. Am J Public Health. 1999;89(10):1543-1548. https://doi.org/10.2105/AJPH.89.10.1543
  6. Turra CM, Elo IT. The impact of salmon bias on the Hispanic mortality advantage: new evidence from Social Security data. Popul Res Policy Rev. 2008;27(5):515-530. https://doi.org/10.1007/s11113-008-9087-4
  7. Arias E, Eschbach K, Schauman WS, Backlund EL, Sorlie PD. The Hispanic mortality advantage and ethnic misclassification on US death certificates. Am J Public Health. 2010;100(Suppl 1):S171-S177. https://doi.org/10.2105/AJPH.2008.135863
  8. Fenelon A. Revisiting the Hispanic mortality advantage in the United States: the role of smoking. Soc Sci Med. 2013;82:1-9. https://doi.org/10.1016/j.socscimed.2012.12.028
  9. Fenelon A, Blue L. Widening life expectancy advantage of Hispanics in the United States: 1990-2010. J Immigr Minor Health. 2015;17(4):1130-1137. https://doi.org/10.1007/s10903-014-0043-6
  10. Lariscy JT, Hummer RA, Hayward MD. Hispanic older adult mortality in the United States: new estimates and an assessment of factors shaping the Hispanic paradox. Demography. 2015;52(1):1-14. https://doi.org/10.1007/s13524-014-0357-y
  11. Andrasfay T, Goldman N. Reductions in 2020 US life expectancy due to COVID-19 and the disproportionate impact on the Black and Latino populations. Proc Natl Acad Sci U S A. 2021;118(5):e2014746118. https://doi.org/10.1073/pnas.2014746118
English translation of the Japanese article.