Intended for healthcare professionals

CCBYNC Open access

Maternal and infant genetic variants, maternal periconceptional use of selective serotonin reuptake inhibitors, and risk of congenital heart defects in offspring: population based study

BMJ 2017; 356 doi: (Published 06 March 2017) Cite this as: BMJ 2017;356:j832

SSRI-gene interaction: less than meets the eye. Re: Maternal and infant genetic variants, maternal periconceptional use of selective serotonin reuptake inhibitors, and risk of congenital heart defects in offspring: population based study

The National Birth Defects Prevention Study data-base used by Dr. Nembhard, et al. in their recent paper provided an opportunity for interesting insights into any susceptibility conferred by genetic polymorphisms on the possible teratogenicity of SSRIs [1]. Unfortunately, excessive modeling and over-interpretation of the results have obfuscated rather than shed light on the question. Supplemental and sensitivity analyses would add substantial value to this paper.

There is a high chance of false positives in the study with 872 polymorphisms studied in two genotypes (maternal and infant) and in two phenotypes. Among the 3488 associations studied, the relative risks of the 20 reported associations for 1 copy of the risk allele are unremarkable, mostly ranging from 1.5 to 3.0. None would be statistically significant after Bonferroni adjustment (which corrects for associations studied, not SNPs; here 1 X 10e-5); the (preferred) False Discovery Rate adjustment is not provided but few of the reported associations would likely maintain statistical significance. From all the examined associations, it is unclear if the 20 reported are the only ones with a BFDPS <0.8; or if they have been selected for reporting, and on what criteria.

Readers will notice that the relative risk for 2 copies of the risk allele, in every analysis, are exactly the square of the copy-1 risk for both the SSRI exposed and unexposed, something which would never happen in nature. This is a function of an assumption in the log additive model and is not calculated from the data. Because this risk is forced by the choice of statistical model, it is entirely predictable and does not merit publication, a footnote would suffice. What would be of interest is the unadjusted and adjusted relative risk of copy-2 versus the referent 0-copy condition. This estimate is not predictable but one can surmise that all the published copy-2 relative risks would be much lower and some might be less than the copy-1 risk; such is the play of chance. Could the authors provide such a table as a supplement to their paper?

Significance testing used a Bayesian false discovery rate which assumed that the risk of missing a true association was four times “more costly” than falsely reporting an association. While this may be appropriate for a genetic screening study, it is opposite to the usual α=0.05 and β=0.2 typically used in association studies and where false positives are actually considered more costly than false negatives. This notwithstanding, the results are discussed as if they were testing hypotheses (rather than screening) and with no circumspection that the copy-2 result is simply the copy-1 result squared. One example is the caveat-free statement: “Risks ranged from 6.81 to 9.00 for the homozygous alleles on the respective single nucleotide polymorphisms” which is presented with no acknowledgment that this result is owed entirely to the assumptions in the statistical modelling. A frequentist analysis of the FDR would also have been of interest in a supplemental table.

The underlying rationale for selecting the genes to study is largely based on indirect evidence of unrelated conditions (ovarian cancer) and very general concepts of disease etiology (“supports protein synthesis,” “detoxification of xenobiotics”). The infant and maternal SNPs are not replicated and there is no replication with the few SNPs or even genes that have been previously examined for these heart phenotypes. There is even no replication from a study in the same population that used one of the identical phenotypes, the same polymorphisms and the same genotyping platform [2].

Finally, it should be noted that neither depression nor any clinical indication for SSRI use was modelled and so confounding by indication cannot be excluded as an explanation of the findings. Several large population-based non-genetic studies have reported that indication bias explains their observed associations with congenital malformations [3-5].

Without appropriate and multiple replications, it must be concluded that the play of chance is the most parsimonious explanation for the reported associations in this paper. The problems in the strategy used to analyze the data were anticipated by Muller, et al. [6]: “We caution against over-interpretation of results based on highly structured probability models and often arbitrary choices of utility functions. Data analysis for Bayesian Multiple Comparisons high-throughput gene expression data is particularly prone to problems arising from data pre-processing. Often it is more important to understand the pre-processing of the raw data, and correct it if necessary, than to spend effort on sophisticated modeling.”


1. Nembhard WN, Tang X, Hu Z, et al. Maternal and infant variants, maternal periconceptional use of selective serotonin reuptake inhibitors, and risk of congenital heart defects in offspring: population based study. BMJ 2017; 356: j832.
2. Li M, Cleves MA, Mallick H, et al. A genetic association study detects haplotypes associated with obstructive heart defects. Hum Genet 2014; 133: 1127-1138.
3. Huybrechts KF, Palmsten K, Avorn J, et al. Antidepressant use in pregnancy and the risk of cardiac defects. N Engl J Med 2014; 370: 2397-407.
4. Jimenez-Solem E, Andersen JT, Petersen M, et al. Exposure to selective serotonin reuptake inhibitors and the risk of congenital malformations: a nationwide cohort study. BMJ Open 2012; 2: e001148 doi:10.
5. Furu K, Kieler H, Haglund B, et al. Selective serotonin reuptake inhibitors and venlafaxine in early pregnancy and risk of birth defects: population based cohort study and sibling design. BMJ 2015; 350: h1798.
6. Muller P, Parmigiani G, Rice K. FDR and Bayesian multiple comparison rules. Proc Valencia/ISBA 8th World Meeting on Bayesian Statistics, Benidorm, Spain 2006.

Michael B. Bracken
Susan Dwight Bliss Professor Emeritus
Senior Research Scientist
Yale University
Professor Bracken is a consultant to manufacturers of SSRIs.

Competing interests: Professor Bracken is a consultant to manufacturers of SSRIs.

20 April 2017
Michael B. Bracken
Susan Dwight Bliss Professor Emeritus, Senior Research Scientist
Yale University
Yale Center for Perinatal, Pediatric and Environmental Epidemiology, One Church St., 6th Floor, New Haven, CT 06510