Methods
Study population . This retrospective cohort study was conducted
at Kaiser Permanente, a US healthcare system providing healthcare and
insurance coverage. Participating regions were Washington, Southern
California, and Northern California, which together serve about 8
million people generally representative of the surrounding
communities.17 Data came from EHRs and linked birth
certificate data. These data have been used in many pregnancy
studies,18-21 and important variables and methods have
been validated.22-25 Study procedures were approved by
the regions’ institutional review boards and those of Washington State
and California, with a waiver of consent.
The population was women age 15-49 years with a singleton live or
stillbirth from 2005 through 2014. Women were required to be enrolled in
Kaiser Permanente from 16 weeks’ gestation through delivery, to have at
least one blood pressure (BP) measured before 20 weeks, and to have
chronic or gestational hypertension (defined from BP values, diagnosis
codes and medication fills; our algorithm is shown in Table S1 and has
been published26). We included both chronic or
gestational hypertension because in clinical practice, it can be
difficult to determine which type of hypertension is present and because
these conditions may represent different points on a continuum of
disease.
Women had to have filled at least one antihypertensive medication before
36 weeks gestation, to be on monotherapy, and to have been enrolled in
Kaiser Permanente for at least 150 days before their qualifying fill.
They could contribute more than one pregnancy to these analyses. We
excluded deliveries exposed to teratogenic medications or certain
high-risk maternal medical conditions (see Table S1 for more
information). The sample size was determined by the number of eligible
births.
Exposures. From pharmacy data, we identified fills for labetalol,
methyldopa, nifedipine and other β-blockers (Table S1). We considered
labetalol separately from other β-blockers because it is a combined α
and β-blocker and unlike other β-blockers, it is recommended as
first-line in US guidelines.6 Exposure was defined
based on the earliest fill after the first prenatal visit (typically at
8-10 weeks’ gestation) or, if the visit date was unknown, at ≥ 10 weeks
gestation; we called this the ‘index fill’. Using intent to treat
principles, women’s exposure status was fixed rather than time-varying,
because subsequent medication switches could be affected by the initial
medication choice.
Outcomes. Outcomes included small for gestational age (SGA),
preterm delivery, neonatal intensive care unit (NICU) admission,
preeclampsia, maternal ICU admission, and stillbirth or termination at
> 20 weeks. SGA was defined using sex and race-specific US
birthweight curves.27 The primary outcome was
birthweight <10th percentile for gestational
age and a secondary outcome < 3rdpercentile. Deliveries missing birthweight (n=32) were excluded from SGA
analyses. We defined preterm delivery using gestational age from the EHR
(preferentially) or birth certificate data, with the primary outcome
being delivery before 37 weeks gestation and a secondary outcome,
delivery before 34 weeks. We
considered preterm delivery a potential measure of medication
effectiveness, because less effective medications could lead to higher
risk of uncontrolled maternal hypertension or fetal growth restriction
(a potential consequence of severe hypertension) and via these pathways,
to indicated preterm delivery. The automated data available to us do not
reliably distinguish spontaneous vs. indicated preterm births. We
identified ICU admissions using EHR data. Preeclampsia was identified
from inpatient diagnosis codes after 20 weeks’ gestation, an approach
with a positive predictive value (PPV) of 90%.28 We
reviewed 45 charts meeting those criteria and found a PPV of 93%. We
identified preeclampsia cases with “severe features” using modified
criteria from the American College of Obstetricians &
Gynecologists,29 drawing on BP values, laboratory
results and diagnosis codes (Table S1).
Potential stillbirths and terminations after 20 weeks’ gestation were
identified using EHR data; we included as outcomes the 76% of potential
cases validated through medical record review or linkage to fetal death
certificates. We grouped together stillbirths and terminations for
several reasons. Terminations after 20 weeks might be done for fetal
anomalies, which could in theory be affected by medication choice, as
there is no definitive evidence about birth defect risk for some widely
used antihypertensive medications. Also, the decision to terminate might
be influenced by severe uncontrolled maternal hypertension, which could
be a consequence of the initial medication choice. Finally, we
hypothesized that variation in coding might lead to similar clinical
scenarios being classified as either a stillbirth or termination in
different instances.
Covariates . Covariates included maternal age at delivery, Kaiser
Permanente region, delivery year, hypertension type (chronic or
gestational), BP values, race/ethnicity, parity, maternal education,
pregestational diabetes, depression, tobacco use, body mass index (BMI),
and prior use of certain medications (Table S1). Hypertension was
categorized as chronic if it was present prior to pregnancy or during
the first 20 weeks gestation and as gestational otherwise. To account
for hypertension severity, we identified the most recent BP value prior
to the index fill and also determined whether a woman experienced one or
more BPs ≥ 160/110 before pregnancy or during this pregnancy before the
index fill. We categorized history of antihypertensive medication use as
no use prior to the index fill, continuous use up to the index fill
(allowing for 80% adherence), or prior use with a gap. Other covariates
included prior use of angiotensin converting enzyme inhibitors,
angiotensin receptor blockers, thiazide diuretics, diabetes medications,
benzodiazepines, statins,
antidepressants or antiseizure medications.
Analyses . Descriptive analyses included counts and proportions
for categorical variables and means and standard deviations for
continuous variables. Primary analyses used logistic regression to model
study outcomes, with labetalol as the referent group. Inverse
probability of treatment weights (IPTW) were used to account for
confounding. We calculated weighted outcome prevalences for each
medication group and adjusted odds ratios (aORs) and 95% confidence
intervals (CIs). We used the bootstrap to account for multiple
pregnancies per woman and for the estimation of the
weights.30,31 Treatment weights were generated from
propensity scores calculated using a multinomial logistic regression
model including all covariates shown in Table 1 except for BMI,
education, parity, and timing of prenatal care. We omitted these
variables because they were well balanced before weighting and a small
proportion of deliveries had missing information for each of these
characteristics. Table S2 lists variables in the propensity score. To
improve statistical efficiency, we calculated stabilized weights
including some baseline covariates in both the outcome model and the
numerator of the weights.32,33These were Kaiser Permanente
region, race/ethnicity, diabetes, type of hypertension (chronic vs.
gestational), and gestational age at index fill.
For statistical modeling, we categorized delivery year as 2005-2008,
2009-2010, 2011-2012, and 2013-2014. We grouped together the four
earliest years because very few deliveries were included from 2005-2006,
when only one region had electronic BP values available. Maternal age
was categorized as < 30, 30-34, 35-39 or ≥ 40 years.
Gestational age at the index fill was modeled as a linear spline with
knots at 140 and 210 days. The systolic and diastolic BP values closest
to the index fill were modeled using linear splines, with knots at 140
mm Hg and 90 mm Hg respectively. Deliveries missing race/ethnicity
(0.5%) were grouped with those with “other” race/ethnicity and
treated as a category of race/ethnicity in statistical models.
To assess covariate balance, we calculated the average standardized mean
differences across all treatment groups before and after
IPTW.34,35
We excluded stillbirths/terminations from analyses of SGA, NICU and
preterm delivery because they are competing events. We used inverse
probability of censoring weights to account for possible bias due to
excluding stillbirths; Table S3 lists the variables used to model these
weights.
In sensitivity analyses, we restricted the analysis to women with
chronic hypertension (87% of the population) and excluded women with
pregestational diabetes. In subgroup analyses, we examined new users
separately from women with prior antihypertensive treatment. Analyses
were performed using R, version 3.5.
Funding. This study was funded by the US National Institute on
Child Health and Human Development grant R01HD082141. The grant proposal
underwent external peer review for scientific quality, and priority was
assessed by scientific staff and a scientific council at NICHD. The
Group Health Foundation funded Dr. Chen’s fellowship. The funders did
not play a role in conducting the research or writing the paper.
Patient involvement. There was no patient or public involvement
in the study.