This study leveraged a population-based state-wide sample of all births in California between 2005 and 2017, using birth certificate data from the Department of Health Care Access and Information. This dataset included information on the characteristics of birthing people and their infants: health and sociodemographic factors, perinatal outcomes, and address at the time of delivery. Addresses were geocoded to link to census tract identifiers, which enabled linkage to neighborhood-level gentrification variables.

From a total sample of 6,738,539 births, we excluded births if they could not be linked to a census tract, were missing, or had implausible gestational age (< 22 weeks or > 45 weeks), implausible pregnant person age (< 10 or > 60 years old), implausible birth weight (< 100 g or > 9000 g), or were plural births. We also excluded births that were missing complete exposure, outcome, or covariate information, and births where the birthing person’s race did not meet inclusion criteria (Supplemental Fig. 1). The final dataset included 5,116,131 births in 7575 census tracts; the mean number of births per tract was 675.4, with a minimum of 1 and a maximum of 6386. Study protocols were approved by the California Committee for the Protection of Human Subjects and the Institutional Review Boards of UC Berkeley (Protocol number: 13–05-1231).

Preterm birth (PTB) was defined as births after 24 weeks and before 37 weeks of gestation, and very preterm birth (VPTB) was defined as births after 24 weeks and before 32 weeks of gestation. Small-for-gestational-age (SGA) births had a birth weight less than the United States sex-specific tenth percentile of weight for each week of gestation [22]. Lastly, we assessed low birth weight (LBW) cases as infants born weighing less than 2500 g. As a sensitivity analysis, we also examined birth weight continuously, using birth weight z-scores for all infants and term birth weight for infants born between 37 and 44 weeks of gestation [22].

Neighborhoods were defined as census tracts. Metropolitan and Micropolitan Statistical Areas, defined by the Office of Management and Budget, were used as the regional boundaries, which we linked to census tracts using the Federal Information Processing System codes. Tract characteristics were compared with the corresponding regional characteristics.

We measured changes across two 10-year periods: 2000–2010, characterized using the 2000 Decennial Census and the 2008–2012 American Community Survey (ACS) 5-Year Estimates; 2007–2017, measured using the 2005–2009 and 2015–2019 ACS 5-Year Estimates [23]. Births were linked to their respective periods based on year, with a 5-year lag between the start of the gentrification period and the birth year to maximize the likelihood that the neighborhood was experiencing gentrification when the birth occurred (Supplemental Table 1).

Using census data, the Freeman method classified gentrification based on socioeconomic indicators [14, 19]. Tracts were classified as eligible for gentrification if 50% of the census blocks in the tract were urban, and the median household income and proportion of housing built in the prior two decades were lower than or equal to the regional median. Otherwise, the tract was determined to be ineligible for gentrification, or “excluded.” Among the tracts eligible for gentrification, those that saw an increase in median home value and percentage of residents with a bachelor’s degree that was larger than the corresponding regional change in these two characteristics during the respective period were classified as gentrifying, and the rest were classified as not gentrifying. In summary, this method classified census tracts as eligible for gentrification and gentrifying, eligible for gentrification and not gentrifying, and ineligible for gentrification. Tracts classified as “eligible for gentrification and not gentrifying” were used as the referent group to be comparable to existing findings on gentrification [18]. We also conducted a sensitivity analysis with “ineligible for gentrification” or excluded as the referent group to enable comparison to conceptually similar referent group using the Displacement and Gentrification Typology.

The D&G Typology leveraged census data and the Zillow Home Value Index to classify neighborhoods into nine categories based on community income and housing affordability. The specific criteria for each category are described in Supplemental Table 2. For this analysis, we further collapsed the nine categories into three broad stages of neighborhood change: (1) displacement, which included the categories “Low Income/Susceptible to displacement” and “Ongoing displacement of low-income households”; (2) gentrification, which included the categories “At risk of gentrification,” “Early ongoing gentrification,” and “Advanced gentrification”; and (3) exclusive, which included the categories “Stable moderate/Mixed income,” “At risk of becoming exclusive,” “Becoming exclusive,” and “Stable/Advance exclusive.” Tracts classified as “Exclusive” were used as the referent group. We made a small modification to the “Becoming Exclusive” category by excluding the criterion on the in-migration rate due to the lack of data availability. Materials on this measure can be found at https://github.com/urban-displacement/displacement-typologies.

Comparing the two measures, the Freeman method and its variations have been the most commonly used in epidemiologic studies [24, 25]. This method evaluates multiple socioeconomic and housing features of the neighborhood. The D&G Typology, on the other hand, was developed as a Neighborhood Early Warning System to identify patterns of investment and sociodemographic changes. It emphasizes housing affordability for low- and middle-income families and considers spatial proximity to increasing housing costs. It also measures displacement, allowing a more nuanced definition of neighborhood changes in addition to gentrification [26].

The sociodemographic covariates from birth certificate data included the pregnant person’s age (years) and the principal source of payment at delivery (private, public, uninsured or other). Pregnancy-related factors included parity (any or no prior live births) and receiving adequate prenatal care, which was assessed using the Kotelchuk index (inadequate or intermediate versus adequate or adequate +) [27].

We used self-reported information on birth certificates to determine the pregnant person’s race and ethnicity. The categories were non-Hispanic (NH) Black, NH Asian/Pacific Islander (API), NH American Indian/Alaska Native (AIAN), NH White, and Hispanic. We did not include pregnant people whose race was reported as “Other” or mixed race due to small sample sizes which may be insufficient for stratified analyses. This analysis conceptualized the variable of race and ethnicity as a proxy measure for exposure to past and present social marginalization that racialized people experience, which may influence how they experience gentrification [21].

Descriptive analysis assessed the prevalence of birth outcomes by neighborhood gentrification status and individual sociodemographic characteristics. We also examined the distribution of gentrification status and participant characteristics overall and by race and ethnicity.

We used mixed-effects logistic regression models, with a random intercept to account for individuals clustering within neighborhoods, to assess associations between gentrification and birth outcomes. Model 1 adjusted for sociodemographic factors (age and insurance type), which may be confounders by influencing people’s residential location and birth outcomes. Model 2 additionally adjusted for pregnancy-related factors (parity and prenatal care), which may be confounders by influencing neighborhood selection or mediators through which gentrification affects birth outcomes. Sensitivity analysis used mixed-effects linear models to assess associations with birth weight z-score and term birth weight.

Based on prior knowledge that gentrification may affect groups differently based on social marginalization and from assessing interaction terms between exposure and race/ethnicity (P-value < 0.001 for all birth outcomes), we used race and ethnicity-stratified models to investigate whether the influence of gentrification varied across racial and ethnic groups.