Methodologic issues and statistical approaches to the analysis of multiple cycle data from couples undergoing in vitro fertilization (IVF)

Abstract

Multiple cycle IVF data are routinely used to identify important predictors of success or failure. However, current statistical approaches do not always make full use of the data due to its complexity. Our objective was to evaluate and identify valid statistical methods for the analysis of multiple cycle IVF data. A prospective cohort of 2687 couples undergoing 5816 cycles of ART from 1994–2003. Six statistical methods – including multiple versions of logistic and proportional hazards regression – were applied to identify predictors of livebirth. All models included female age and BMI, male age, study site, study period, prior livebirth, primary infertility diagnosis, gonadotropin dose, GnRH-a regimen, day 3 E2 level, number of oocytes retrieved, ICSI, and number of embryos transferred (when restricted to cycles with a transfer). Analyses restricted to first cycle only tended to overestimate while last cycle only analyses underestimated relations; both exhibited wider confidence intervals and decreased statistical power to detect associations. Among the four approaches using data from all cycles, effect estimates were smallest for the continuous-time Cox proportional hazards model – most likely due to bias associated with the standard approach to adjusting for ties. Of the six methods employed, only discrete-time survival analysis used data from all cycles and appropriately addressed the issues of within-couple dependence of cycles, cessation of treatment upon success, censoring (i.e., discontinuation of treatment among cycle failures), and heavily tied event times (i.e., a sizable proportion of couples experience success on the same cycle attempt). Across the statistical methods, variation in effect estimates of predictors of success was between 10 and 32%. In addition, care must be taken in interpretation of patient characteristics and treatment variables when including mid-cycle outcomes (e.g., number of embryos created) in the multivariable model, as control for intermediates along the causal pathway may result in biased estimates. Discrete-time survival analysis appears to be the most appropriate method for the analysis of multiple IVF cycles and is easily implemented using standard software. Inclusion of intermediate treatment outcomes in multivariable models must be considered critically when evaluating the relation between successful IVF and variables that chronologically precede these intermediate outcomes.