Ecochard, R. (1997). Random Effect Models in the Statistical Analysis of Human Fecundability Data: Application to artificial insemination with sperm from donor. Open Research Online (The Open University). https://doi.org/10.21954/ou.ro.0000f5e2
Ecochard R. Random Effect Models in the Statistical Analysis of Human Fecundability Data: Application to artificial insemination with sperm from donor. Open Research Online (The Open University); 1997. doi:10.21954/ou.ro.0000f5e2
Ecochard, Rene. Random Effect Models in the Statistical Analysis of Human Fecundability Data: Application to artificial insemination with sperm from donor. Open Research Online (The Open University), 1997.
Sperm donors vary widely in fertility even after semen measures
Fertility varied widely between sperm donors, even after semen measures were included, a 1997 doctoral thesis found. The thesis analyzed 12,100 donor-insemination cycles in 1,901 French women. In first attempts, the conception rate per cycle was about 13 out of 100 at the first cycle and about 6 out of 100 at the twelfth.
Key Findings
In first attempts, 909 of 1,901 women (47.8%) conceived; the rate per cycle was 0.1261 at the first cycle and 0.0573 (20 of 349) at the twelfth.
Between-donor variance fell from 0.382 to 0.222 when sperm count, motility and quality after thaw were included, so these measures appear to explain over one third of donor differences.
Semen measures related more strongly to conception as donor averages than as single-donation deviations from them: for quality after thaw, the log-odds were 0.366 against 0.099.
In the crossed model, odds ratios for conception were 1.56 per Insler score unit, 0.58 for insemination more than two days before estimated ovulation, and 0.83 for insemination after it.
In the crossed model, Penalized Quasi Likelihood gave variances about 50% lower between women and 30% lower between donors than Gibbs sampling, which estimated 0.993 and 0.335.
Interpretation
This 1997 doctoral thesis in biostatistics fits random-effect models to observational records from one French CECOS center, 1985 to 1994. The women were considered normally fertile. Their husbands had no sperm or only a few barely mobile sperm. Donors were matched to women with no systematic rule beyond blood-group and appearance checks, and physicians could favor donors with past successes. The author concludes that an apparent rise in success with repeated use of a donor's sperm was due to selection bias. The author describes the fall in conception rates across cycles as a selection effect, because the most fertile women conceive earlier.
RRM Context
Restorative reproductive medicine is couple-centered and uses cycle charting. Here, conception was related to the Insler score, a rating that reflects partly mucus quality and partly the opening and appearance of the cervix. It was also related to insemination timing against estimated ovulation. The models separate the woman's part from the man's. Each varied between individuals.
Our editorial summary of this paper, not the article's abstract.
Abstract
The main aim of this dissertation is to explore methodological approaches to correlated binary data and to assess their suitability for the analysis of data on human fertility. The dataset concerns a study of Artificial Insemination by Donor (AID). AID represents an unusual research opportunity to study both male and female fecundability simultaneously. In each attempt to conceive, artificial insemination is carried out in consecutive ovulatory cycles until conception or change of treatment. The probability of conception may differ between women, so that the data are discrete time survival data with censoring and between-subject heterogeneity. There is also potential heterogeneity between donors. Non-systematic allocation of the donor to recipient ensures that the same woman receives semen from several donors, This added heterogeneity as well as other cycle dependent covariates have to be taken into account. The analysis must also take account of covariates, most of them time-varying. Our dataset have a crossed hierarchical structure due to the presence of both, female and male factors. The rather complicated "design" calls for unit specific regression models. These models are presented as well as their lack of tractability except in some rather specific cases. The motivation for choosing Gaussian random effects in unit specific regression models is discussed. We demonstrate the use of an approximate inference method (Penalized Quasi Likelihood). This method is shown to be a useful and practical way of carrying out preliminary data analysis. Finally a Bayesian procedure (Gibbs sampling) provides validation and more accurate results despite the intensive computation it needs. The main substantive finding of the analysis is the unexpectedly pronounced heterogeneity of donor fecundability, even after inclusion of conventional measures of sperm quality into the model. These measures were shown to be predictive at the donor level but not at the level of individual donation.
01502ca60 