between facility and UCSF study staff. Study Outcomes Unlike a superiority analysis, a noninferiority study design determines whether the effect of a new treatment is not worse than that of an active control by more than a specified clinically acceptable margin.29---32 We selected a noninferiority design because we were seeking not to replace physicians as abortion providers or to determine whether NPs, CNMs, and PAs were better than current providers of care but to identify additional, comparably safe providers to supplement the provider pool. Because NPs, CNMs, and PAs who are newly trained in aspiration abortion have less experience, we expected to find a statistically significant higher rate of complications among this group than among more experienced physicians. However, we also anticipated a low overall incidence of complications from procedures across both groups. Therefore, a noninferiority design provided a more clinically relevant analysis. Given a low expected complication rate in both provider groups, we prespecified the margin of noninferiority as a change of 2%, which was determined before the start of the study by a panel of researchers and clinicians and approved by the Data and Clinical Safety Monitoring Committee, who considered ethical and clinical issues and previous US-based studies, which showed abortion-related complication rates ranging from 1.3% to 4.4%.21,22,33---38 The primary outcome was the difference in incidence of complications within 4 weeks of the aspiration abortion between NPs, CNMs, and PAs and physicians. Complications were categorized as immediate (occurring before leaving the facility) and delayed (occurring £ 4 weeks after the procedure). Additionally, complications were classified as major if the patient required hospital admission, surgery, or a blood transfusion and minor if they were treated at home or in an outpatient setting. This classification schema is consistent with that used in other studies of abortion-related morbidity.34---37 Statistical Analysis We based sample size calculations for this study on an expected complication rate of 2.5%, which was based on mean complication rates cited in the published literature21,22,33---38 and powered at 90% to detect a 1.0% or greater difference in complication incidence between groups (a = .025, 1-tailed test). The study was powered specifically for a noninferiority analysis. Although we set a clinically acceptable margin of difference at 2.0%, we took a conservative approach and powered the study to detect an even smaller difference. We then further increased the sample size per group by 15% to adjust for clustering effects at the provider and clinic levels. We compared sociodemographic characteristics of patients seen by NPs, CNMs, and PAs and those seen by physicians using mixedeffects logistic regression for dichotomous variables, mixed-effects multinomial logistic regression for categorical variables, and mixedeffects linear regression for continuous variables, all of which included random effects for facility. Incidence of a complication was coded as a dichotomous variable. Complication incidence was calculated by provider group. We fit a mixed-effects logistic regression model with crossed random effects to obtain odds ratios that account for the lack of independence between abortions performed by the same clinician and within the same facility and cross-classification of providers across facilities. We included variables associated with complications in bivariate analyses at P < .05 in the multivariate model in addition to other clinically relevant covariates to adjust for potential confounders. To mitigate selection bias resulting from the lack of randomization, we replicated the analysis in a propensity score---matched sample, a method used to achieve balance between study groups in observational or nonrandomized studies using the predicted probability of group membership (NP, CNM, or PA vs physician group) on the basis of observed predictors.39---41 We used the Stata module pscore to develop the propensity scores based on a logistic regression model that included patient characteristics that potentially influenced to which provider type the patient was assigned (age, race/ethnicity, insurance type, gestational age, parity, history of cesarean delivery, history of miscarriages, history of abortions, screening for sexually transmitted infections, positive test for a sexually transmitted infection, selection of a clinical contraceptive method, and presence of risk factors). Patients with similar propensity scores in the 2 provider groups were matched using nearest neighbor matching. After testing that the balancing property of the propensity score was satisfied, we selected a matched sample composed of 78.3% of the original sample, among which we replicated our mixed-effects analysis. We used predictive probabilities to calculate risk differences and 95% confidence intervals (CIs) for all models. We used STATA version 12 (StataCorp LP, College Station, TX) for all analyses. RESULTS A total of 21 095 women were screened for eligibility. Of these, 3837 did not meet the eligibility criteria, most commonly because of patient age and gestational age. Among the 17 258 eligible women, 13 807 agreed to participate in the study. Of these, 2320 had procedures performed by NPs, CNMs, and PAs during their