Effectiveness and safety of intrauterine insemination vs. assisted reproductive technology: emulating a target trial using an observational database of administrative claims.

We emulated the above target trial using observational data from the MarketScan Commercial Claims and Encounters Database (MarketScan). MarketScan contains individual-level, de-identified healthcare claims data including clinical utilization, insurance enrollment/plan benefit for inpatient, outpatient, and prescription drug for individuals and their dependents (about 100 million individuals between 2011 and 2015) with employer-provided commercial insurance in the United States ( 30 , 31 ). The types of insurance plans included are preferred and exclusive provider organizations (PPOs and EPOs), comprehensive (COMP) plans, point-of-service (POS) plans, health maintenance organizations (HMOs), consumer-directed health plans (CDHP) and high deductible health plans (HDHP). Women were linked to their male partners and liveborn infants through a family ID variable. We used International Classification of Disease (ICD)-9 codes to identify female infertility diagnosis, PCOS, tobacco use disorder, alcohol abuse or dependence, obesity or overweight, and male infertility diagnosis ( Supplemental Table 1 ). We identified IUI and ART by current procedural terminology (CPT) codes in the inpatient and outpatient files ( 32 ) ( Supplemental Table 2 ). Cancellations of ART cycles were identified by the procedure codes “in vitro fertilization procedure canceled before aspiration”, where the outcome was coded as no live birth. Because cancellation of IUI procedures was not recorded in the claims data, the study did not include canceled IUI cycles (See Supplementary Materials for details). We identified abortion (spontaneous, elective, unspecified), stillbirth, and live births, using validated algorithms ( 33 ) based on ICD-9 code, CPT, and Healthcare Common Procedure Coding System, and Diagnosis Related Group codes. For women who disenrolled from MarketScan within 280 days after initiation of treatment, pregnancy outcomes were coded as missing. For each treatment cycle, we searched for pregnancies that ended before the next treatment or a pre-specified window (<45 weeks if the outcome was live births/stillbirths and <20 weeks if the outcome was abortions), whichever occurred earliest. Pregnancy outcomes not linked to any ART or IUI cycles were considered as pregnancies from natural conception. We estimated the gestational age at birth as {delivery date – (date of IUI or date of first ART procedure) + 14 days} or (delivery date – date of embryo transfer + 17 days) when the date of embryo transfer was available ( Supplemental Figure 1 ) ( 34 ). We identified maternal and neonatal outcomes using validated algorithms ( 35 – 37 ). Preeclampsia and gestational hypertension were defined as the presence of at least 2 inpatient ICD-9 codes after 140 gestational days and within 30 days after the delivery date ( 35 , 36 , 38 ) Gestational diabetes was defined as 1) at least one inpatient or outpatient ICD-9 code for gestational diabetes or CPT code for glucose tolerance test, 2) at least one diabetes code after 140 gestational days and the delivery date, and 3) an absence of diabetes codes or non-metformin antidiabetics before 140 gestational days. SGA and LGA were defined as the presence of at least one of the ICD-9 codes for poor fetal growth or excess fetal growth, respectively, in maternal or infant claims from delivery until 30 days after delivery ( 35 – 37 ). NICU admission was identified using CPT codes in maternal and infant claims within 30 days of birth. A non-chromosomal structural major malformation was defined by at least two inpatient or outpatient ICD-9 codes indicating a birth defect, or a diagnostic code and a corrective surgical procedure in infant claims during the 90 days after birth or maternal claims during 30 days after delivery ( 35 , 37 ). Women in the database who met the eligibility criteria of the target trial were classified into the treatment strategy with which their data were compatible at baseline. Because pregnancy test results are not systematically recorded, we defined pregnancy as the presence of codes for either a completed or terminated pregnancy, e.g., spontaneous or therapeutic abortion within 20 weeks or delivery within 45 weeks following ART or IUI. We assumed that the treatment strategies were randomly assigned within levels of the baseline covariates age, calendar year of cycle initiation, infertility diagnosis, PCOS, overweight or obesity, and region of residence. To estimate the observational analogs of the intention-to-treat effects and the per-protocol effects, we conducted analyses identical to the ones described for the target trial with one modification: In the target trial, we would study all outcomes in both the intention-to-treat and per-protocol analyses. In this emulation, however, we restricted the intention-to-treat analysis to pregnancy outcomes because the high nonadherence to the strategies assigned at baseline would make it hard to interpret intention-to-treat effect estimates for maternal and neonatal outcomes. The study received ethics board approval from the Harvard T.H. Chan School of Public Health and the MassGeneral Brigham Healthcare System Institutional Review Boards (Boston, Massachusetts).

Of 29,021 eligible women ( Supplemental Figure 1 ), 18,495 initiated IUI and 10,526 initiated ART. Compared with IUI initiators, ART initiators were older, less likely to have a diagnosis of PCOS or anovulation, and more likely to have infertility of tubal origin ( Supplemental Table 3 ). Figure 1 shows treatment trajectories for the first 3 cycles in IUI and ART initiators. The proportions of women who adhered to the protocol was 35% for the IUI group and 50% for the ART group. The probabilities of pregnancy per cycle were about 13–15% for IUI and about 33–40% for ART ( Figure 1 ). The probability of live birth was 28.4% among IUI initiators (21.8% from the assigned IUI treatment, 0.9% from additional IUI cycles, 3.2% from switching to ART, and 2.6% from natural conception) and 41.5% among ART initiators (27.4% from the assigned ART, 11.8% from additional ART cycles, 0.03% from switching to IUI, and 2.3% from natural conception). The estimated probability of live birth was 26.3% under the IUI strategy and 27.4% under the ART strategy; risk difference 1.0% (95% CI: −0.1%, 2.2%). The risk difference was 4.3% (95% CI: 3.7%, 4.9%) for multiple births, 3.4% for preterm births (95% CI: 2.8%, 4.0%), and 1.5% (95% CI: 0.9%, 2.1%) for NICU admission. The absolute risk differences for other neonatal outcomes were less than 0.5% ( Table 2 ). The estimated probability of pregnancy lasting ≥ 20 weeks was 26.7% under the IUI strategy and 28.0% under the ART strategy; risk difference 1.3% (0.2% to 2.4%). The risk differences were 0.6% (95% CI: 0.2%, 1.0%) for gestational diabetes and less than 0.5% for preeclampsia and gestational hypertension ( Table 2 ). The differences of conditional risks among live births ( Table 2 ) and women with pregnancy lasting at least 20 weeks were in the same direction. The probability of live birth was similar under both strategies in women aged ≤ 40 years, but greater for ART (14.4%) than IUI (7.4%) in women aged 41–45 years ( Figure 2 ); the risk difference was 7.0% (95% CI: 4.6%, 9.1%). The risk differences for neonatal and maternal outcomes by age group were consistent with those in the main analysis, but with wider 95% confidence intervals ( Supplementary Tables 4 - 8 ). Risk differences were similar when restricting the analysis to women living in States with an insurance mandate ( Supplementary Table 9 ), women with unspecified infertility origin ( Supplementary Table 10 ), with at least 18 months of enrollment before baseline ( Supplementary Table 11 ), without loss to follow-up ( Supplementary Table 12 ), and when redefining consecutive cycles as cycles occurring within 35 days from each other ( Supplementary Table 13 ). Results were also similar in subgroup analyses by type of insurance plan ( Supplementary Tables 14 - 17 ). When natural pregnancies were permitted in a variation of the protocol, the probability of live birth was 29.8% under ART versus 31.0% under IUI; risk difference −1.3% (95% CI: −2.3%, −0.3%). The risk differences for neonatal and maternal outcomes were similar to those in the main analysis ( Supplementary Table 18 ).

The target trial is a (hypothetical) pragmatic randomized trial that would answer our question of interest ( 25 – 27 ). We can conceptualize observational analyses as an attempt to emulate a target trial. The approach has two steps: 1) specifying the protocol of the target trial, and 2) emulating the components of that protocol using the observational data. We now describe both steps ( 25 – 27 ).

We emulated a target trial of two fertility treatment strategies using observational data from 29,021 women in a large healthcare database. Had all participants adhered to the protocol, we estimated that three cycles of IUI would result in a similar probability of live birth than one cycle of ART (26.3% vs. 27.3%), but in lower risks of multiple births, preterm, NICU admission, and gestational diabetes. Results were similar in women with unspecified infertility origin and in women aged between 18 and 40 years. However, in women aged 41 to 45 years, a single ART cycle resulted in a higher probability of live birth than did 3 IUI cycles. Our estimates of probability of live birth are compatible with those from randomized trials that compared 3 cycles of stimulated IUI with one cycle of ART in women ≤ 38 years. The pilot of the INeS trial in the Netherlands (n=116) found 21% for IUI compared with 22% for ART ( 16 ) and a UK randomized trial (n=207) found 24.7% for IUI vs 31.1% for ART ( 13 ). Note that in our study, live birth rates were higher for ART when not adjusting for adherence to the 1-cycle strategy because the analysis then included all pregnancies within 4 months from additional ART or other fertility treatments. Our findings also support ASRM’s recommendation of ART for women over 38 years ( 10 ). This recommendation was based on the FORT-T trial (n=154), which showed that in women aged 38–42 years, two cycles of ART had a greater probability of live birth compared with two cycles of stimulated IUI (ART vs clomiphene-IUI vs gonadotropin-IUI group: 31.4% vs 13.5% vs 15.7%) ( 17 ). We found that compared with two cycles of IUI, one cycle of ART was as effective in women between 38 and 40 years (18.2% vs. 18.1% of live births; Figure 2 ) and more effective in women older than 40 (14.4% vs. 6.4% of live births; Figure 2 ). Further, because the probability of live birth decreases with age, the shorter time to pregnancy after a single cycle can make ART more attractive than 3 IUI cycles even for younger women when the time to a live birth is an important consideration. Our target trial differs from these trials in two aspects. First, our target trial evaluated the comparative safety of the two strategies while the sample size of most trials only allowed them to evaluate the probability of live birth ( 13 , 15 , 16 ). One exception is the INeS trial, which found similar probabilities of having a healthy infant for 3 cycles of ART with single embryo transfer versus 6 cycles of stimulated IUI, although the 95% confidence intervals were wide ( 12 ). Second, the proportion of ART procedures with live births that resulted in multiple births was 27.7% in our study compared with 6% to 8.3% in the trials ( 12 , 13 ). Our data thus reflects the higher use of multiple embryo transfers in the US, where the proportion of multiple births (from fresh non-donor ART cycles) was about 28.8% in 2011 and 22.7% in 2015 ( 3 , 7 ). Despite this higher use of multiple embryo transfers, we did not estimate ART to be more effective than 3 IUI consecutive cycles in women ≤ 40 years. The lower probability of multiple births in the ART group of the INes trial may explain that it did not find a higher risk of preterm delivery for ART (7%) compared with IUI (11%). In contrast, the probability of multiple births in the IUI group of our study (12.8%) and of trials (6% to 13%) was comparable ( 13 ). Our target trial emulation is not directly comparable with previous observational studies for three reasons. First, previous observational studies used comparators that are less relevant for decision making (e.g., natural conception from fertile couples, natural conception from infertile couples, non-ART groups including live births from natural or non-ART treatments) ( 18 – 24 , 39 ). Second, most studies did not consider sustained treatment strategies involving multiple attempts, which are more relevant for decision making. That is, children were classified according to the conception method while disregarding that women who delivered a child after the first treatment attempt likely differ from women who delivered after several failures. Third, most studies only reported associations between treatment and neonatal outcomes conditional on live birth, which can introduce selection bias ( 40 ). In contrast, we presented risks of live birth with and without an offspring event ( 29 ). Our study has several limitations. First, the outcome ascertainment was based on diagnostic and procedure codes and thus is subject to misclassification. However, most outcomes were identified through validated algorithms with good positive predictive value ( 33 , 35 – 38 ). Second, we cannot rule out some pregnancies that linked to treatments may have resulted from a natural conception, especially those deliveries after 41 weeks from a cycle. Nonetheless, such proportion was small ( Supplemental Figure 2 ). In addition, our estimate of pregnancy, miscarriage, and live birth percentages following ART treatments in Marketscan closely matched the national reports ( Supplemental Tables 19 - 20 ). Third, the information available on the actual procedures in claims allowed us to exclude individuals with codes for preimplantation genetic testing, but not those who might have paid out of pocket (thus no claims) for preimplantation genetic testing. Also, because the proportion of single embryo transfers in the U.S. increased from 25%−38% between 2012 and 2014 (our study period) to 64% in 2017 ( 4 , 6 , 9 ), the risk of multiple births and associated complications (e.g., preterm, NICU admission) in the ART group is expected to be lower today than during the study period. Fourth, we could not identify women who might have decided to cancel IUI after ovarian stimulation (in contrast, we could identify women who decided to cancel ART). However, this likely small proportion of women would have been excluded from our per-protocol analyses anyway (See Supplemental Materials ). Fifth, our population only included women with insurance coverage for fertility treatments, which may limit its generalizability to women who paid out of pocket if factors that affect the source of payment also affect the biological relations studied. However, the distributions of age, infertility diagnosis, and tobacco use of our study population are comparable to those of ART patients nationwide ( Supplemental Tables 19 and 21 ) and previous studies suggested that, these and other factors (e.g., age, ethnicity, types of subfertility, duration of subfertility, and body mass index) do not predict who would benefit from immediate ART over IUI ( 14 , 41 ). Therefore, our findings are probably transportable to other women seeking fertility treatments in the US. Last, as in any observational analyses, confounding is always a possibility However, we adjusted for known predictors of the outcomes and restricted the cohort to women without records of IUI or ART for at least one year. Potential residual confounding by aspects not recorded in the data (e.g., infertility duration; lifetime treatment history; reasons for infertility), if any, may channel more severe infertility to ART group, which would bias the results towards more favorable findings for IUI. In summary, our findings support the effectiveness and safety of three IUI cycles compared with one ART cycle as a first-line treatment for women ≤ 40 years. In women > 40 years, one ART cycle had a higher probability of live birth than three IUI cycles. Our target trial emulation replicates the findings from randomized trials, expands the inference to maternal and neonatal complications, and provides a blueprint for ongoing evaluation of fertility treatment strategies. As practice patterns evolve, and in the absence of new randomized trials, target trial emulation using real world data will be needed to inform future guidelines for fertility treatments.

The use of fertility treatments is increasing, particularly in developed countries ( 1 – 9 ). Intrauterine insemination (IUI) with ovarian stimulation and assisted reproductive technology (ART), including in vitro fertilization (IVF) and intracytoplasmic sperm injection (ICSI), are commonly used treatments for infertility. Because each attempt of IUI is less costly and invasive than ART, IUI is often the first-line treatment in couples with unexplained or mild male infertility ( 10 ). However, guidelines are inconsistent: the UK National Institute for Health and Care Excellence (NICE) recommended ART as the first-line therapy for people with unexplained infertility, mild endometriosis, or mild male factor infertility ( 11 ), whereas the American Society of Reproductive Medicine (ASRM) recommended IUI before proceeding to ART for couples with unexplained infertility ( 10 ). These recommendations were based on randomized trials and observational studies that assessed pregnancy and live birth rates, but not maternal and neonatal complications such as preeclampsia and congenital anomalies ( 12 – 17 ). In the only trial (INeS) that included neonatal outcomes, 3 cycles of ART with single embryo transfer and 6 cycles of stimulated IUI resulted in similar probabilities of live birth, multiple births, and a composite outcome of healthy infants ( 12 ). However, the sample size of the INeS trial (602 couples and 333 live births) was insufficient to study specific maternal and neonatal outcomes ( 12 ). Observational studies with larger sample sizes were not designed to assess the comparative effectiveness of multiple IUI cycles versus ART ( 18 – 24 ). Therefore, we need improved observational approaches that extend the results from randomized trials to the study of maternal and neonatal complications, and that can be used for ongoing evaluation of a variety of IUI and ART dynamic strategies, given that there is no guarantee that additional randomized trials will be conducted for many contrasts of interest. Here, we used a large observational healthcare database to emulate a randomized trial—a target trial ( 25 – 27 )—of 1 cycle of ART versus 3 cycles of IUI. The outcomes of interest were pregnancy, live birth, pregnancy complications, and neonatal outcomes.

Supplemental Figure 1 . Study flow chart for a target trial of initiating intrauterine insemination or assisted reproductive technology in the IBM MarketScan Commercial Claims and Encounters Database, USA, January 2011-October 2015. Abbreviations: ART, assisted reproductive technology; IUI, intrauterine insemination (IUI), PGS/PGD, preimplantation genetic screening/diagnosis Supplemental Figure 2 . The estimated gestational age of end of pregnancy following IUI or ART.