Patient selection
A retrospective cohort analysis was conducted using a dataset collected from Lee Women’s Hospital in Taichung, Taiwan. This study was approved by the Institutional Review Board on Biomedical Science Research at Academia Sinica. The dataset we obtained consisted of 9,090 fresh cycles from 1998 to 2014. This is a protocol based upon the first fresh transfer after oocyte retrieval and cryopreservation is just considered between the covariates in order to describe better patients characteristics. The information recorded in this database for each cycle included patients’ age at cycle start, the date of oocyte retrieval, cycle status, cycle outcome, AMH, follicle stimulating hormone (FSH), estradiol (E2), along with other 64 variables. In order to explore the causal effects of embryo transfer day on odds of live birth and other associated outcomes, only patients with transfer day of Day 3 and Day 5 and without missing or incomplete outcome information were included. To minimize the potential bias caused by numerical outliers, cycles with outliers in certain variables, such as age, weight, height, total number of oocytes retrieved, total sperm count, sperm motility, normal sperm morphology, AMH, FSH, period day of retrieval, number of embryos transferred and peak E2 were excluded. Cycles with PGD/PGS (preimplantation genetic diagnosis/preimplantation genetic screening) cases were also excluded. Therefore, the final dataset contained 4,127 autologous IVF cycles, with known embryo transfer day, and confirmed information about live births, number of gestational sacs, and number of fetal heartbeats at 4 weeks of gestation. Details about the inclusion and exclusion criteria of the study were documented as a flowchart in Fig. 1. Controlled ovarian stimulation was achieved by a downregulation protocol in which patients were administered leuprolide acetate (Lupron, Takeda Chemical Industries, Ltd., Osaka, Japan) during the midluteal phase. Subsequently, patients received recombinant follicle stimulating hormone (Gonal-F; Serono, Bari, Italy) on day 3 for ovarian stimulation. The retrieved oocytes were cultured in Quinn’s Advantage Fertilization Medium (Sage Bio- Pharma, Inc., Trumbull, CT, USA) with 15% serum protein substitute (SPS, Sage BioPharma, Inc) at 37∘C with the gas mixture of 5% oxygen, 5% carbon dioxide and 90% of nitrogen. Following conventional insemination or intracytoplasmic sperm injection (ICSI), a fertilization medium (SAGE Biopharma, USA) with 15% serum protein substitute (SPS; SAGE Biopharma, USA) was used to further culture the embryos. 70 to 72 hours after insemination or ICSI, all cleaved embryos were group cultured in microdrops of a blastocyst medium (Sage BioPharma, Inc.) with 15% SPS. Further details about the clinical aspects of the IVF protocols or embryology for the study were documented in a previous paper [16].
Statistical methods
Comparison of baseline characteristics and handling missing data
To observe the discrepancies of marginal characteristics of the clinical characteristics between patients undergoing Day 3 embryo transfer and Day 5 embryo transfer, a series of statistical tests were performed. We applied Welch’s t-tests for continuous variables, chi-square tests for binary variables, and Poisson regressions for count variables. The comparisons of baseline characteristics were conducted both before and after matching to evaluate the performance of our constructed propensity score. Because data were largely missing in peak E2, we utilized Generalized Association Plots (GAP) [17] to retrieve the important variables highly associated with peak E2. A single imputation for peak E2 was then performed using the number of embryos frozen, number of oocytes retrieved, number of follicles, and embryo transfer day. In order to avoid the confounding effect of prescribed dosage given by multiple physicians, we converted Gonal-F (follitropin alfa), Menotropin and recombinant luteinizing hormone (rLH) into dichotomous values, and combined Menotropin and rLH into a single variable (whether or not the patient was injected with either Menotropin or rLH throughout the cycle).
Propensity score matching
To estimate the causal effect of embryos transfer day on different clinical outcomes, we implemented the method of propensity score matching to mimic a randomized controlled trial. The propensity score of receiving Day 5 transfer was constructed by a logistic regression where the probability of receiving Day 5 embryo transfer was modeled through a logit link function. The formula and deviation of the propensity score are detailed in Supple-ment. Embryo transfer day was regressed on a series of observed baseline covariates to derive a propensity score for each IVF cycle. The baseline covariates selected in the propensity score model were based on the literature as well as their clinical implications [18]. The variables (Xji,j=1,...,J) contributed to the propensity score included: age (linear and quadratic terms), body mass index (BMI), AMH, tubal factor (divided into six categories: undecided, removed, patent, sticking, ligation and obstructed for both left and right side of Fallopian tubes), sperm motility, sperm count, normal sperm morphology, azoospermia, polycystic ovarian syndrome (PCOS), Gonal-F, menotropin, rLH, peak estradiol, day of trigger, number of follicles, number of oocytes retrieved, number of embryos transferred, number of embryos frozen, number of mature oocytes and ICSI vs. IVF. Furthermore, to adjust for the secular trend regarding changes or modifications of the protocols made by clinical professionals, the year of cycles performed was also included as an additional matching factor, i.e., we only paired cycles that the embryos were implanted in the same years. Matched sets of Day 3 and Day 5 transfer subjects were thus formed: for each cycle whose embryo transfer was Day 5, we selected another cycle with the nearest propensity score from those with an embryo transfer Day 3. The method of greedy matching was used to produce balanced matched samples without replacement [19]. Those without matched cycles from the other transfer protocol were excluded from the subsequent analyses.
Post-matching analyses
Using the matched dataset, the effect of embryo transfer day on odds of live birth was then estimated through logistic regression, adjusting for propensity score and the remaining covariates that appeared to be statistically different between Day 3 and Day 5 transfers after propensity score matching. On the other hand, the effect of embryo transfer day on the number of gestational sacs and the number of fetal heartbeats was then estimated by Poisson regression where the number of embryos transferred was treated as an offset. Similarly, the Poisson regression adjusted for propensity score and the remaining significant covariates to account for residual confounding.
Stratification by age and AMH
Additional analyses were conducted stratifying cycles by age and AMH. We stratified cycles based on the age categories recommended by SART (age <35, 35 ≤age<37, 37 ≤age≤38 years) and the quantiles of AMH (AMH =0, 0 ≤AMH≤1.39, 1.39 <AMH≤3.95, AMH >3.95 ng/ml). The method of propensity score matching, logistic regressions and Poisson regressions were applied repeatedly to each subgroup to identify the causal effects of embryo transfer day on associated IVF outcomes.