Methods
Mass General Brigham Institutional Review Board approval was obtained
for this study, and the need for informed consent was waived (Protocol
#2020P001079). Universal testing of all patients admitted for labor and
delivery at all Mass General Brigham hospitals began on April 19, 2020.
Mass General Brigham includes four large hospitals: two academic medical
centers and two community hospitals with a combined estimated annual
delivery volume of approximately 15,000 deliveries per year. All women
who were admitted for labor and delivery at all Mass General Brigham
hospitals between April 19, 2020 and May 16, 2020 and tested for
SARS-CoV-2 up to 48 hours before admission or upon admission were
included in this study. All SARS-CoV-2 testing was performed by
nasopharyngeal reverse transcription-PCR (RT-PCR) using assays approved
via United States Food and Drug Administration Emergency Use
Authorization. Electronic health records were reviewed for all patients
admitted to labor and delivery during the study period to abstract
SARS-CoV-2 test results, demographic data, and medical variables that
may be associated with SARS-CoV-2 infection.
We calculated the risk (and exact 95% confidence interval) of a
positive SARS-CoV-2 test up to 48 hours before or on admission to labor
and delivery among our study population over the entire study population
and by study week. Patients not tested for SARS-CoV-2 were excluded from
these analyses, as were patients who tested positive earlier in
pregnancy but negative at the time of admission for labor and delivery,
given the focus of the study was on the risk of SARS-CoV-2 infection at
the time of delivery. We calculated the percentage of women testing
positive for SARS-CoV-2 on admission who were asymptomatic, with
symptoms defined as fever, chills, cough, dyspnea, myalgia, headache,
anosmia, ageusia, sore throat, rhinorrhea, nausea or vomiting, abdominal
pain, and diarrhea. The average daily positive tests per 100 admissions
to labor and delivery were compared to statewide data from the
Massachusetts Department of Public Health per 100,000 residents aged
20-39 by study week.
Demographic, socio-economic and clinical factors evaluated for their
association with infection included maternal age, body mass index (BMI),
race, co-morbidities (gestational diabetes, preexisting diabetes,
asthma, smoking, opioid use disorder), zip code, known SARS-CoV-2
infection in a household member, parity (as a surrogate for number
children in the household), occupation, and insurance type (MassHealth
or Medicaid vs. commercial insurance). We identified the factors most
strongly associated with SARS-CoV-2 infection and determined the risk of
infection, stratified based on the number of factors associated with
infection present.
The zip code of patient residence was mapped to the corresponding
towns.10 The COVID-19 rate, defined as the number of
confirmed cases per capita provided from the Massachusetts Department of
Public Health,1 was recorded on May 13, 2020.
Occupation for each patient was classified into categories based on the
United States Bureau of Labor Statistics 2018 Standard Occupation
Classification System.11 Occupations were then
classified as essential workers vs. nonessential workers, with
healthcare workers being a subset of essential workers. Occupations were
determined to be essential based on the emergency order enacted by the
governor of Massachusetts on March 23, 2020;12 those
included as essential were: building and grounds cleaning and
maintenance occupations, food preparation and serving related
occupations, healthcare practitioners and technical occupations,
healthcare support occupations, installation, maintenance, and repair
occupations, military support occupations, protective service
occupations, and transportation and material moving occupations. The
medical records of all essential workers were manually searched for
documentation of whether the patient was working from home or not
working. If patients whose job fell into the essential workers category
were specifically noted to be working from home or not working for over
two weeks prior to delivery, they were not included in the essential
worker category.
Due to the limited number of
SARS-CoV-2 infections, estimating a full multivariable logistic
regression model including all covariates of interest was not
feasible.13 Therefore, a set of simple logistic
regression models was used to assess the univariate association between
each covariate of interest and the odds of SARS-CoV-2 infection. As an
exploratory analysis, multivariable logistic regression with the least
absolute shrinkage and selection operator (lasso)14was used to identify a small subset of predictors with the strongest
association with SARS-CoV-2 infection. Lasso is a penalized regression
method that constrains the sum of the magnitude of regression model
coefficients such that covariates that do not improve prediction of the
outcome are shrunk to zero, thus creating a more parsimonious
model.15 The degree of penalization, lambda, was
selected as the largest value that maintained 10-fold cross-validated
prediction error within 1 standard error of the
minimum.14 The risk of SARS-CoV-2 infection and
corresponding 95% confidence intervals were estimated amongst subgroups
of patients using Poisson regression with robust error variance. Missing
data on race (9.9%), occupation category (7.6%), and delivery BMI
(0.2%) were addressed using multiple imputation by fully conditional
specification, assuming that data were missing at random given observed
data.16 Specifically, predictive mean matching,
logistic regression, and the discriminant function method were used to
impute continuous variables, binary and ordinal categorical variables,
and nominal categorical variables, respectively, to create 20 complete
datasets. Imputation models included all predictors assessed for
univariate association with SARS-CoV-2 infection, as well as delivery
hospital and SARS-CoV-2 test result. Odds ratio and risk estimates with
corresponding standard errors were obtained from each of the 20 complete
datasets and combined using Rubin’s rules to produce pooled estimates
with 95% confidence intervals.17 Lasso selected the
same set of predictors across all imputations, and final coefficients
were obtained by averaging across imputations. A complete case analysis
was performed as a sensitivity analysis. Statistical analyses were
performed using SAS software version 9.4 (SAS Institute Inc, Cary, NC,
USA) and R software version 3.6.1 (R Foundation for Statistical
Computing, Vienna, Austria).