Methods
Cohort
The cohort was drawn from the Medicaid Analytic eXtract, which contains information on Medicaid beneficiaries. Medicaid is the joint state and federal health insurance program for people who are on a low income; it covers approximately 40% of all births in the United States. The Medicaid Analytic eXtract is a healthcare utilization database that records demographic and Medicaid enrollment information on beneficiaries, as well as healthcare utilization claims, including all recorded diagnoses and procedures associated with inpatient admissions and outpatient visits. It also contains claims for all filled outpatient drug prescriptions.
Using Medicaid Analytic eXtract data from 46 US states and the District of Columbia from 2000 to 2007, our group created a cohort for the study of drug safety in pregnancy, as previously described in detail. We identified women age 12 to 55 years with completed pregnancies and linked them to liveborn infants. Using a validated algorithm based on delivery date and diagnostic codes in the maternal and infant records, we estimated the last menstrual period before the pregnancy. We restricted the cohort to women who were continuously eligible for Medicaid from three months before the estimated last menstrual period month through one month post partum. To ensure complete ascertainment of claims throughout the entirety of pregnancy, we restricted our analysis to women without restricted benefits, private insurance, or certain capitated managed care programs that underreport claims to Medicaid Analytic eXtract. We also required that the linked infants met the same Medicaid eligibility criteria as their mothers for at least three months after birth, unless they died, in which case we allowed a shorter eligibility period. We excluded pregnancies in which the mother used known teratogenic drugs, including lithium, antineoplastic agents, retinoids, and thalidomide during the first trimester (n = 3528) and pregnancies in which the infant was diagnosed as having a chromosomal abnormality (n = 1175) (see Supplementary Figure S1).
(Enlarge Image)
Figure S1.
Flowchart showing selection of cohort.
Statin Use
In the primary analysis we defined statin use based on one or more claims for a dispensed statin from the last menstrual period through day 90 of pregnancy (first trimester), the causal relevant window of exposure for congenital malformations. We considered the following statins: simvastatin, lovastatin, pravastatin, fluvastatin, atorvastatin, cerivastatin, and rosuvastatin. The reference group for the primary analysis consisted of women who were not dispensed a statin during the first trimester. Statins are only available as prescription drugs in the United States (that is, they are not available over the counter).
Outcomes
The primary study outcome was the presence of a congenital malformation in the infant. We defined congenital malformations based on the diagnosis of one or more organ specific malformations. Organ specific malformations assessed included central nervous system malformations; eye, ear, neck, and face malformations; cardiac malformations; respiratory malformations; cleft palate or lip malformations; gastrointestinal malformations; genitourinary malformations; musculoskeletal malformations; or other malformations. These were identified based on the presence of international classification of diseases, ninth revision (ICD-9) diagnostic codes on two or more separate days in the infant inpatient or outpatient records during the first three months of life. We required diagnostic codes on separate days to increase the specificity of the definition, by excluding cases in which a single mention may be recorded to justify a diagnostic test to rule out that condition. Secondary outcomes included each of the nine organ specific categories of malformation.
Covariates
We considered four groups of potential confounders: maternal demographic characteristics, comorbid medical conditions, obstetric characteristics and conditions, and drugs dispensed to the mother. These covariates were selected because they are potential risk factors for malformations, or proxies for them. Demographic characteristics assessed included maternal age at delivery, race/ethnicity, geographic region, and year of delivery. Obstetric characteristics considered included multiparity and multiple gestations. Chronic comorbid medical conditions were defined during the baseline period (before the last menstrual period through to the end of the first trimester) and included pre-existing diabetes, dyslipidemia, pre-existing hypertension, chronic renal disease, obesity, and alcohol, tobacco, or illicit drug use/misuse. We included the number of distinct prescription drugs (other than statins) and physician visits in the three months before the last menstrual period as markers of general comorbidity. We also assessed the use of drugs during the baseline period, which may be markers for the presence or the severity of comorbid illness, including antihypertensives, insulin, and oral antidiabetes drugs, as well as the use of suspected teratogenic drugs during the first trimester (we excluded those pregnancies exposed to known teratogens).
Statistical Analyses
We first determined the baseline characteristics of women in the cohort, stratified by statin use during the first trimester of pregnancy, and the frequency of malformations in the infants of women who did or did not use statins. This made it possible to calculate an unadjusted risk ratio and 95% confidence intervals for congenital malformations associated with statin use. Because pre-existing diabetes was expected to be an important confounder, we estimated the association between statin use and the primary outcome stratified on pre-existing diabetes using the Mantel-Haenszel method.
To account for all measured differences in baseline characteristics between women who did and did not use statins, we used propensity score based methods. The propensity score was determined using a logistic regression model that estimated the probability of being dispensed a statin in the first trimester based on all potential confounding variables mentioned, without further selection. Based on propensity score, we matched women who did and did not use statins in the first trimester in a fixed 1:3 ratio, using a nearest neighbor greedy matching algorithm with a maximum matching distance of 0.05; covariate distribution and the risk ratio and 95% confidence interval for malformations were described in the matched cohort. In an alternative approach (which should be interpreted as the primary result), intended to preserve the information content of the large group of women who did not use statins, we conducted propensity score stratified analyses using the Mantel-Haenszel method after the creation of 100 strata of equal propensity score width. We additionally used high dimensional propensity scores to adjust for empirical covariates in addition to the investigator specified covariates, as confirmatory analyses; high dimensional propensity scoring has been shown to further improve control of confounding in some circumstances. In this approach, the algorithm for high dimensional propensity scores evaluates thousands of inpatient and outpatient diagnoses and procedures as well as pharmacy claims, and prioritizes 100 covariates that may act as proxies for unmeasured confounders. These are then combined with investigator specified covariates for inclusion in the propensity scores model. The high dimensional propensity scores based analyses were conducted using stratification, as described.
Subgroup and Sensitivity Analyses
We performed several subgroup and sensitivity analyses to assess the robustness of our primary findings (any congenital malformation), again using stratification by propensity score. Because they more readily cross the placenta, lipophilic statins (simvastatin, lovastatin, fluvastatin, atorvastatin, and cerivastatin) have been postulated to have greater teratogenic potential than hydrophilic statins. We therefore separately assessed the association between use of lipophilic statins during the first trimester and malformations, with exposure defined based on the specific statin prescribed. To estimate the potential impact of exposure misclassification, we explored two alternative definitions of first trimester statin use: firstly, days supply of statin overlapping the first trimester, based on dispensings in the 90 days before the last menstrual period through to the end of the first trimester (for example, a woman would be considered to have used a statin if she was dispensed a statin before pregnancy, but with a supply that would be expected to extend into pregnancy); and secondly, two or more statin dispensings during the first trimester. To assess the potential impact of outcome misclassification we defined malformations based on a single diagnostic code in the infant inpatient or outpatient record, extended the follow-up of infants to one year, and defined malformations based on codes in either the mother or the infant claims, since in some circumstances in Medicaid Analytic eXtract, infants' conditions are applied to the mother's claims history for the first several months after birth. We also performed a probabilistic assessment to explore the effect of potential outcome misclassification across a range of potential sensitivities and specificities for the malformations, and an assessment of the potential impact of analysis based only on live births (see Supplementary Appendix 1 and 2). Because some women had more than one pregnancy included in the cohort, we also repeated the analysis restricting it to the first pregnancy recorded. We also considered the possible effect of the complex relation between preterm delivery and malformations, since infants with malformations are more likely to be born preterm and infants born preterm will sometimes have conditions that might be coded as malformations that would have spontaneously resolved had the infant been carried to term. We therefore performed a secondary analysis restricting the cohort to term deliveries.