Specific SSRIs and Birth Defects: New Data
Specific SSRIs and Birth Defects: New Data
For this analysis we used data from the NBDPS, a population based case-control study of birth defects. The study's methods have been described previously. Briefly, cases of birth defects were identified through birth defects surveillance systems in the US states of Arkansas, California, Georgia, Iowa, Massachusetts, New Jersey, New York, North Carolina, Texas, and Utah. Cases could be live born, stillborn, or induced abortions with one of over 30 major birth defects. The NBDPS excluded cases with known chromosomal or monogenic disorders. Unmatched liveborn controls from the same geographical region and time period were selected from birth certificates or birth hospital records. More cases than controls were included overall because the study was designed for the assessment of individual defects, where controls, which were the same for all case groups, outnumbered even the largest case group. Mothers were asked to participate in a telephone interview in English or Spanish between six weeks and two years after the estimated date of delivery.
For this analysis we included cases and controls if they were born on or after 1 October 1997 and had an estimated date of delivery on or before 31 December 2009. Overall participation was 67.4% for cases and 64.8% for controls. A previous NBDPS analysis of the association between SSRI use during pregnancy and birth defects, included in Table 1 as Alwan and colleagues, included data from 1997 to 2002. To avoid double counting, these NBDPS data were excluded from the meta-analyses used to calculate prior odds ratios in the current study, but we used significant findings from the previous analysis along with those of the other studies listed in Table 1 to determine which birth defect-SSRI combinations to assess. Because of the strong association between diabetes and birth defects, we excluded case and control mothers who reported pregestational diabetes (type 1 or type 2). We also excluded mothers who reported the use of any of the following known teratogenic treatments: misoprostol, methotrexate, mycophenolate mofetil, thalidomide, or isotretinoin.
During the interview for the NBDPS, no specific question addressed depression. Mothers were asked if they had any illnesses other than the ones already discussed (for example, hypertension or diabetes) and whether they took any medication for the illness. Women could report depression here, which would then be followed by a question about any medications taken for the illness. There were also specific medication related questions: "between three months before conception and [the baby's] date of birth, did you take any of the following medications? Prozac? Paxil? Zoloft? Celexa?" There was no specific question for Lexapro. For this analysis, we considered women exposed if they reported taking citalopram (Celexa), escitalopram (Lexapro), fluoxetine (Prozac), paroxetine (Paxil), or sertraline (Zoloft) at least once in the period from one month before conception through the third month of pregnancy. Women who reported taking more than one type of SSRI were only included in the multiple SSRI category. We considered women as unexposed if they did not take any antidepressants in the period from three months before to the end of the pregnancy and did not report any depression, anxiety, bipolar disorder, or obsessive compulsive disorder. Women who did not answer all medication related questions, used SSRIs in a period other than the period of interest, or took antidepressants other than SSRIs (for example, bupropion and venlafaxine) were excluded.
The NBDPS includes over 30 categories of major birth defects. After ascertainment in population based surveillance systems, the diagnostic information was reviewed by clinical geneticists at each site to establish eligibility. Designated individual clinical geneticists reviewed all cases with a particular defect to ensure consistency across sites. Because the primary intent of this analysis was to assess previously reported associations with SSRIs and to determine if those associations were supported by NBDPS data, we included only outcomes with at least one previous report in the peer reviewed literature suggesting a possible association with SSRIs: neural tube defects (international classification of diseases, ninth revision (ICD-9): 740–742.0), anencephaly (ICD-9: 740), all septal defects (ICD-9: 745), ventricular septal defects (ICD-9: 745.4), right ventricular outflow tract obstructions (ICD-9: 746.0–746.1), cleft palate (ICD-9: 749.0), cleft lip with or without cleft palate (ICD-9: 749.2–749.4), esophageal atresia (ICD-9: 750.3), anal atresia (ICD-9: 751.23–751.24), hypospadias (ICD-9: 752.6), any limb reduction defect (ICD-9: 755.2), craniosynostosis (ICD-9: 756.0), gastroschisis (ICD-9: 756.71), and omphalocele (ICD-9: 756.70). Some defects reported in other studies (for example, cystic kidney) could not be evaluated in this analysis because they were not included in NBDPS.
A bayesian approach requires specification of prior distributions for each of the variables in the model used to estimate the potential association between risk of birth defect and use of SSRIs. These prior distributions are probabilistic summaries of beliefs about the true values of the unknown variables before assessment of new data. The bayesian approach allowed us to incorporate existing information on the association between each SSRI and the birth defect outcome of interest. Prior distributions were developed based on literature review. A systematic review identified six studies published before 2010 that had available specific information on SSRI-birth defect combinations (Table 1).
We created categories of birth defects based on those reports, which corresponded as closely as possible to the NBDPS birth defect categories (Table 1). The approach used to summarize the information presented in these publications for each of the SSRI-birth defect combinations of interest depended on the number of available studies; to avoid duplication of cases in the current analysis, we did not include the results from the earlier NBDPS analysis. If one published assessment was only available, then the prior distribution for the log of the odds ratio relating the birth defect and use of SSRIs was assumed to be normal, with a mean given by the log odds ratio estimate reported in the study and variance defined using the corresponding reported confidence interval. If two or more studies were identified, we used bayesian meta-analysis methods to summarize the results. The goal of the meta-analysis was to produce an estimate of the log odds ratio relating the specific birth defect and SSRI across studies, taking into account the study specific estimates and their associated sampling errors. The assumed meta-analysis model included a term corresponding to the true underlying log odds ratio relating SSRI use and risk of birth defects and a collection of random study level effects. All available information was included in developing the meta-analysis based prior estimates, including results indicating no association between risk of birth defects and SSRI use from other published studies. If no information other than the previous NBDPS analysis was identified, then we assumed the log odds ratio relating birth defects and SSRI use to have a non-informative prior distribution, defined using a normal distribution with mean zero and a variance of 1000. Using this non-informative prior places virtually the entire weight in developing the bayesian estimates on the information contained in the full NBDPS data.
An alternative approach to this analysis would be to develop estimates of the association between SSRI consumption and risk of birth defects using frequentist methods only focused on the 1997–2009 NBDPS data. These results could be summarized and then included as an additional point in a larger meta-analysis of available information. We chose the bayesian approach for two primary reasons. Firstly, we viewed the collection of information summarized by the meta-analysis as the state of current knowledge concerning potential association between SSRIs and risk of birth defects and the NBDPS data as new information available to update that knowledge. This view is consistent with the bayesian updating paradigm as applied in this analysis. In addition, we believe that only utilizing summary values (for example, estimated odds ratios and their standard errors) from the NBDPS data would be an unnecessary sacrifice of information as opposed to utilizing the individual level data informed by the meta-analysis priors.
We used a bayesian approach to develop estimates for a logistic regression model relating the log odds of a specific defect and the mother's use of SSRIs (see supplementary appendix 1). In addition to a term reflecting the log odds ratio relating birth defects and use of SSRIs, the model also included confounders selected a priori and obtained through the maternal interview: maternal race/ethnicity (non-Hispanic white versus other), maternal education (0–12 years versus >12 years), obesity (body mass index <30 versus ≥30), and smoking (any smoking versus no smoking from one month before to the end of the first trimester). Although prior probabilities for the variable relating birth defect risk and use of SSRIs were developed, we assumed non-informative priors for the odds ratios associated with maternal race/ethnicity, education, obesity, and smoking in the logistic regression model. Posterior estimates for the model variables were developed using Markov Chain Monte Carlo methods. BUGS was used for the bayesian analyses.
The primary results presented here were derived using an analysis based only on NBDPS participants who reported values for all the variables used in the logistic regression model. We also conducted sensitivity analyses focused on assessing the impact of not including participants with missing information, including consideration of a potential association between being missing and the unknown outcome. This analysis utilized bayesian imputation for missing data both under an assumption that the missing information was missing at random and under plausible assumptions on mechanisms for informative missingness (see supplementary appendix 1).
Methods
Study Population
For this analysis we used data from the NBDPS, a population based case-control study of birth defects. The study's methods have been described previously. Briefly, cases of birth defects were identified through birth defects surveillance systems in the US states of Arkansas, California, Georgia, Iowa, Massachusetts, New Jersey, New York, North Carolina, Texas, and Utah. Cases could be live born, stillborn, or induced abortions with one of over 30 major birth defects. The NBDPS excluded cases with known chromosomal or monogenic disorders. Unmatched liveborn controls from the same geographical region and time period were selected from birth certificates or birth hospital records. More cases than controls were included overall because the study was designed for the assessment of individual defects, where controls, which were the same for all case groups, outnumbered even the largest case group. Mothers were asked to participate in a telephone interview in English or Spanish between six weeks and two years after the estimated date of delivery.
For this analysis we included cases and controls if they were born on or after 1 October 1997 and had an estimated date of delivery on or before 31 December 2009. Overall participation was 67.4% for cases and 64.8% for controls. A previous NBDPS analysis of the association between SSRI use during pregnancy and birth defects, included in Table 1 as Alwan and colleagues, included data from 1997 to 2002. To avoid double counting, these NBDPS data were excluded from the meta-analyses used to calculate prior odds ratios in the current study, but we used significant findings from the previous analysis along with those of the other studies listed in Table 1 to determine which birth defect-SSRI combinations to assess. Because of the strong association between diabetes and birth defects, we excluded case and control mothers who reported pregestational diabetes (type 1 or type 2). We also excluded mothers who reported the use of any of the following known teratogenic treatments: misoprostol, methotrexate, mycophenolate mofetil, thalidomide, or isotretinoin.
SSRI use
During the interview for the NBDPS, no specific question addressed depression. Mothers were asked if they had any illnesses other than the ones already discussed (for example, hypertension or diabetes) and whether they took any medication for the illness. Women could report depression here, which would then be followed by a question about any medications taken for the illness. There were also specific medication related questions: "between three months before conception and [the baby's] date of birth, did you take any of the following medications? Prozac? Paxil? Zoloft? Celexa?" There was no specific question for Lexapro. For this analysis, we considered women exposed if they reported taking citalopram (Celexa), escitalopram (Lexapro), fluoxetine (Prozac), paroxetine (Paxil), or sertraline (Zoloft) at least once in the period from one month before conception through the third month of pregnancy. Women who reported taking more than one type of SSRI were only included in the multiple SSRI category. We considered women as unexposed if they did not take any antidepressants in the period from three months before to the end of the pregnancy and did not report any depression, anxiety, bipolar disorder, or obsessive compulsive disorder. Women who did not answer all medication related questions, used SSRIs in a period other than the period of interest, or took antidepressants other than SSRIs (for example, bupropion and venlafaxine) were excluded.
Birth Defects
The NBDPS includes over 30 categories of major birth defects. After ascertainment in population based surveillance systems, the diagnostic information was reviewed by clinical geneticists at each site to establish eligibility. Designated individual clinical geneticists reviewed all cases with a particular defect to ensure consistency across sites. Because the primary intent of this analysis was to assess previously reported associations with SSRIs and to determine if those associations were supported by NBDPS data, we included only outcomes with at least one previous report in the peer reviewed literature suggesting a possible association with SSRIs: neural tube defects (international classification of diseases, ninth revision (ICD-9): 740–742.0), anencephaly (ICD-9: 740), all septal defects (ICD-9: 745), ventricular septal defects (ICD-9: 745.4), right ventricular outflow tract obstructions (ICD-9: 746.0–746.1), cleft palate (ICD-9: 749.0), cleft lip with or without cleft palate (ICD-9: 749.2–749.4), esophageal atresia (ICD-9: 750.3), anal atresia (ICD-9: 751.23–751.24), hypospadias (ICD-9: 752.6), any limb reduction defect (ICD-9: 755.2), craniosynostosis (ICD-9: 756.0), gastroschisis (ICD-9: 756.71), and omphalocele (ICD-9: 756.70). Some defects reported in other studies (for example, cystic kidney) could not be evaluated in this analysis because they were not included in NBDPS.
Development of Priors
A bayesian approach requires specification of prior distributions for each of the variables in the model used to estimate the potential association between risk of birth defect and use of SSRIs. These prior distributions are probabilistic summaries of beliefs about the true values of the unknown variables before assessment of new data. The bayesian approach allowed us to incorporate existing information on the association between each SSRI and the birth defect outcome of interest. Prior distributions were developed based on literature review. A systematic review identified six studies published before 2010 that had available specific information on SSRI-birth defect combinations (Table 1).
We created categories of birth defects based on those reports, which corresponded as closely as possible to the NBDPS birth defect categories (Table 1). The approach used to summarize the information presented in these publications for each of the SSRI-birth defect combinations of interest depended on the number of available studies; to avoid duplication of cases in the current analysis, we did not include the results from the earlier NBDPS analysis. If one published assessment was only available, then the prior distribution for the log of the odds ratio relating the birth defect and use of SSRIs was assumed to be normal, with a mean given by the log odds ratio estimate reported in the study and variance defined using the corresponding reported confidence interval. If two or more studies were identified, we used bayesian meta-analysis methods to summarize the results. The goal of the meta-analysis was to produce an estimate of the log odds ratio relating the specific birth defect and SSRI across studies, taking into account the study specific estimates and their associated sampling errors. The assumed meta-analysis model included a term corresponding to the true underlying log odds ratio relating SSRI use and risk of birth defects and a collection of random study level effects. All available information was included in developing the meta-analysis based prior estimates, including results indicating no association between risk of birth defects and SSRI use from other published studies. If no information other than the previous NBDPS analysis was identified, then we assumed the log odds ratio relating birth defects and SSRI use to have a non-informative prior distribution, defined using a normal distribution with mean zero and a variance of 1000. Using this non-informative prior places virtually the entire weight in developing the bayesian estimates on the information contained in the full NBDPS data.
An alternative approach to this analysis would be to develop estimates of the association between SSRI consumption and risk of birth defects using frequentist methods only focused on the 1997–2009 NBDPS data. These results could be summarized and then included as an additional point in a larger meta-analysis of available information. We chose the bayesian approach for two primary reasons. Firstly, we viewed the collection of information summarized by the meta-analysis as the state of current knowledge concerning potential association between SSRIs and risk of birth defects and the NBDPS data as new information available to update that knowledge. This view is consistent with the bayesian updating paradigm as applied in this analysis. In addition, we believe that only utilizing summary values (for example, estimated odds ratios and their standard errors) from the NBDPS data would be an unnecessary sacrifice of information as opposed to utilizing the individual level data informed by the meta-analysis priors.
Bayesian Analysis
We used a bayesian approach to develop estimates for a logistic regression model relating the log odds of a specific defect and the mother's use of SSRIs (see supplementary appendix 1). In addition to a term reflecting the log odds ratio relating birth defects and use of SSRIs, the model also included confounders selected a priori and obtained through the maternal interview: maternal race/ethnicity (non-Hispanic white versus other), maternal education (0–12 years versus >12 years), obesity (body mass index <30 versus ≥30), and smoking (any smoking versus no smoking from one month before to the end of the first trimester). Although prior probabilities for the variable relating birth defect risk and use of SSRIs were developed, we assumed non-informative priors for the odds ratios associated with maternal race/ethnicity, education, obesity, and smoking in the logistic regression model. Posterior estimates for the model variables were developed using Markov Chain Monte Carlo methods. BUGS was used for the bayesian analyses.
The primary results presented here were derived using an analysis based only on NBDPS participants who reported values for all the variables used in the logistic regression model. We also conducted sensitivity analyses focused on assessing the impact of not including participants with missing information, including consideration of a potential association between being missing and the unknown outcome. This analysis utilized bayesian imputation for missing data both under an assumption that the missing information was missing at random and under plausible assumptions on mechanisms for informative missingness (see supplementary appendix 1).
