Study population
We obtained 2017–2019 vital statistics data (birth certificate) from the Florida Department of Health’s Office of Vital Statistics. Eligible births were based on FEMA disaster declarations [17] to mothers whose county of residence received public and individual assistance (most affected areas). The analysis covered births delivered in the year before and after Hurricane Michael (before: Oct.6th, 2017 to Oct. 5th; after: Oct. 6th to Oct. 6th, 2019). Supplementary Table 1 and Supplementary Fig. 1 shows the numbers of births in each county.
Damage assessment
We used high resolution aerial imagery to estimate residential building damage in the affected counties. The imagery is available from the Land Boundary Information System which is sponsored by The Florida Department of Environmental Protection (FDEP), Division of State Lands, Bureau of Survey and Mapping. Images were available for 11 of the 12 counties designated for both individual and public assistance by FEMA after Hurricane Michael. The only county without images available was Leon County.
The aerial images contained a red, blue, and green visible color bands and 0.5 × 0.5 ft spatial resolution. Figure 1 shows the tarp distribution in the study area, a focus in one populated area (Panama City, FL), and a true color aerial image which covers one square mile within Panama City, FL. Features reflect visible light differently which can be used to distinguish objects on the ground in the images. The images were collected 2–3 months after Hurricane Michael’s landfall. We used blue tarps as an indicator for building damage. Based on the availability of high resolution aerial imagery, the analysis presumes that roofs were undamaged/not covered with blue tarps prior to Hurricane Michael’s landfall. Previous studies have demonstrated the use of blue tarps as post-hurricane indicators of damage and recovery, and the tarps are distinguishable using optical imagery [18, 19]. The data were classified in Python 3 with the sci-kit learn package [20]. A Support Vector Machine model classified the images first as a multiclass classification with seven classes: blue tarp, pool water, impervious surface, vegetation, bare soil, roof, and natural waterbody. This was then simplified to a binary classification (tarp or no tarp present). The model had an overall accuracy of 85.3% with a sensitivity of 74% and a specificity of 96.7%.
We complemented the aerial imagery with a publicly available “windshield” damage assessment in Leon County, which contains the state capital (Tallahassee) and one of the larger affected populations. Hurricane Michael caused notable damage in Leon County (which qualified for individual FEMA assistance) but was less catastrophic there than counties located closer to where the storm made landfall [21]. The county experienced 5851 births, which was 39.49% of the total study area births. Leon County completed a rapid, vehicle based “windshield” damage assessment within three days of the hurricane’s landfall. Damage assessors drove across the county and documented damage with photos and completed a standardized damage survey in ArcGIS collector.
The Florida Vital Statistics Office included geocoded maternal addresses with the requested records. Data were stored in secure, password-protected files on central servers at Tulane and FSU and accessed through encrypted computers. All staff with access to data received training in human subjects research, including confidentiality and data protection. We used tax parcel data from the University of Florida’s GeoPlan Center to match the geocoded births to tarp presence in a Geographic Information System. Damaged parcels were defined as a residential property with a blue tarp > 18.6 m^2 while all other properties were considered undamaged. However, the birth record geocoding process contained geographic positional error where addresses did not directly overlap with household property boundaries. Previous studies have demonstrated that geocoding positional error varies by method used and whether the location is in an urban or rural setting. However, for the majority of cases, the error is less than 100 m with an average of approximately 50 m [22]. To estimate infrastructure damage exposure more accurately, we used the proportion of parcels around a geocoded birth that had blue tarp present. We created 25 m, 50 m, and 100 m buffers around the geocoded birth record location and calculated the percentage of blue tarp damaged parcels.
Outcomes-perinatal outcomes and access to PNC services
Perinatal outcomes and access to PNC services were assessed using vital statistics. Perinatal outcomes include LBW, preterm birth (PTB), SGA, Caesarean section, breastfeeding, and access to PNC services. LBW was defined as a birthweight of an infant of 2,500 g or less, regardless of gestational age. PTB was defined as a birth before 37 weeks of gestation. Based on national standards [23], SGA was defined by birthweight below the 10th percentile for gestational age. Mode of delivery was defined as Caesarean section vs. other methods. Breastfeeding was based on the indicator for infants being breastfed between birth and discharge. Access to PNC services were evaluated by whether pregnant women had any PNC visits before delivery, the month of first PNC visit, and the Kotelchuck Index. Kotelchuck Index used information (PNC initiation time and number of PNC visits) from birth certificate to assess the PNC utilization adequacy. There are four adequacy categories in the Kotelchuck Index: adequate plus, adequate, intermediate, and inadequate [24].
Covariates
Known risk factors from the Vital Statistics for outcomes were assessed as potential confounders, including maternal age, race, ethnicity, education, smoking during pregnancy, alcohol use, pre-pregnancy BMI, and whether enrolled in the U.S. Department of Agriculture’s Supplemental Nutrition Program for Women, Infants, and Children (WIC) program. A causal diagram was created to identify confounders for each association we assessed [25]. The confounders were selected if these represented risk factors for the outcome of interest, were associated with the exposure (damage) but were not intermediate variables in the causal pathway between exposure and outcome.
Statistical analyses
Complete case analysis was conducted to assess the association between the percent of parcels damaged within a given radius and perinatal outcomes and access to PNC services. Outcome and covariate missing data were minimal. Missing data were categorized as follows: LBW (0.19%), preterm (0%), SGA (0%), time of first PNC visit (13.01%), PNC overall use (0.48%), Kotelchuck Index (16.74%); maternal age (0%), education (2.98%), race/ethnicity (1.80%), pre-pregnancy BMI (7.55%), WIC enrollment (1.77%), smoking during pregnancy (0.88%), and alcohol drinking during pregnancy (6.64%).
Exposure was categorized in three ways for the purpose of presenting demographics data and exploring different thresholds as we explored a new method to measure associations between hurricane and pregnancy outcomes (Fig. 2): First, it was dichotomized into areas of high (> 25% of parcels within a 25 m radius of the woman’s residence) and low damage (≤ 25%) to present the demographics of women giving birth before and after Hurricane Michael in the most affected areas. Because damages were measured after Hurricane Michael only, those areas were defined as areas of high (> 25%) and low (≤ 25%) risk for births happened before Hurricane Michael. Secondly, we further categorized the percent of parcels damaged into four categories, no damage, < = 33rd percentile, 33rd-67th percentile, > 67th percentile, using a radius of 25 m. Sensitivity analysis repeated the aforementioned analyses using a radius of 50 m and 100 m (Supplementary Table 2). Lastly, threshold regression was used to decide outcome-specific cutoffs of exposure (using the percent of parcels damaged within a 25 m radius) for each outcome of interest. Threshold models iteratively tested potential breaks points identify cutoffs of the proportion of blue tarp through Bayesian information criteria (BIC), Akaike information criterion (AIC), or Hannnan-Quinn information criterion (HQIC) [26], choosing the best cut-off to maximize model fit. The damage variable was categorized into three categories: 0%, 0% ~ outcome-specific cutoff, > = outcome-specific cutoff. Two sets of analyses were performed to assess the association between the extent of residential building damage and perinatal outcomes/access to PNC services. In the first scenario, births in the year before Hurricane Michael were coded as zero (no damage). In the second scenario, all births before Hurricane Michael were excluded.
Log-binomial regression models were used to compare the populations giving birth and living in and outside of high-risk/damage areas before and after the hurricane, respectively. Interaction terms between hurricane (before and after) and high-risk/damage areas (> 25% damage) in log-binomial regression models were used to evaluate the changes in population giving birth after Hurricane Michael. Log-binomial regression was used to estimate the associations between damage after Michael and binary outcomes (whether developed LBW, PTB, SGA, or had Caesarean section, breastfeeding, any PNC visits, intermediate/inadequate PNC). A semi-parametric linear model was used to evaluate the association between damage and time of first PNC visit. All estimates were compared unadjusted and after adjusting for potential confounders. Model performance was assessed by using the Value/DF of deviance. The Value/DF of deviance for models above ranged from 0.85 to 1.25. The main analyses (presented in the main text) (included 11 counties) excluded Leon County to have a consistent measure of damage from aerial imagery. Sensitivity analyses included Leon County’s complementary windshield survey (Supplementary Tables 3 and 4). Threshold regression was conducted in the software Stata 15 (College Station, TX: StataCorp LLC), and other statistical analyses were performed using the software SAS 9.4 (SAS Inc., Cary, NC).
This analysis was approved by the Institutional Review Boards of Tulane University (2019–529-TUHSC), Florida State University, and the Florida Department of Health.