Common Methodologic Issues
No single approach can definitively establish the safety or risk of drugs, because of several underlying difficulties.
Most congenital malformations occur rarely, and many teratogens, even when known to be associated with an increased risk of a given malformation, do not affect the great majority of exposed fetuses. In fact, very few drugs increase the total malformation rate by a factor of more than two (isotretinoin and thalidomide are two such drugs). If, for example, the risk of major malformations in a given population is 3 percent, then at least 220 pregnancies with the specific exposure and a similar number of control pregnancies will be required to show a risk that is increased by a factor of 2.5, with a power of 80 percent.
Effect of Maternal Diseases
Apart from drug therapy, many medical conditions themselves increase fetal risks. For example, pregnant women with hypertension or cancer are more likely to have infants with intrauterine growth retardation, and pregnant women with epilepsy or diabetes mellitus are more likely to have infants with malformations. Therefore, any attempt to establish the role of fetal exposure to drugs must also address the contributing and confounding risk of the underlying maternal illness.
Recall Bias in Retrospective Studies
There is ample evidence of partial memory and bias in the way women recall the drugs they took during pregnancy. For example, women treated with a prescribed drug for a chronic illness tend to recall their treatment better than women who took an over-the-counter drug. Women who have given birth to malformed children may be more likely to remember the course of their pregnancies, in the effort to understand what went wrong, than women who have given birth to healthy children, thus giving rise to false positive associations. The initial suggestions that benzodiazepines, spermicides, and Bendectin, for example, were teratogenic were based on retrospective case–control studies subsequently refuted by other, larger studies (Table 2 - Common Drugs Initially Thought to Be Teratogenic but Subsequently Proved Safe).
With improved epidemiologic methods, the reliability of the case–control design has improved. For example, recruiting mothers of infants with a different major malformation as controls may eliminate or at least reduce the problem of differential maternal recall. In a recent study, this approach was used to document the effect of the mothers’ knowledge of the study hypothesis (that folic acid deficiency causes spina bifida) on the information they reported.
Nonrandomized Observational Studies
With prospective observational studies, the treatment decisions have not been made by the investigators collecting the data. As a result, the indications for treatment and concurrent exposures are not standardized. Therefore, in comparisons of treated and untreated pregnant women or pregnant women who received two different drugs, preexisting confounding factors are not randomly distributed between the two groups. For example, in comparing the outcome of pregnancy in women who received carbamazepine and women who received phenytoin, one must address the issue of whether the two groups of women had the same type and severity of seizure disorder.
Observational studies of neurobehavioral development require longer follow-up than observational studies of other abnormalities, and interpretation of the results is often complicated by numerous confounding factors. Maternal and paternal IQ, socioeconomic status, and educational levels all affect cognitive development in children. Any attempt to address the developmental effects of drugs without controlling for these factors is likely to be futile.
The information received by drug manufacturers is often a mix of prospective and retrospective case reports. The quality of the information about exposure is usually poor, and outcome data are sparse because of high rates of loss to follow-up. Most important, women and health professionals who contact manufacturers are likely to report adverse fetal outcomes, not uneventful ones. For example, the pivotal study that described retinoid embryopathy contained two parts: prospectively collected data from a study cohort, with a malformation rate of 36 percent, and data from voluntary retrospective reporting to the manufacturer, with a malformation rate of 80 percent.
A common concern regarding the use of meta-analysis is the inevitable combination of data from studies that are not equivalent in terms of quality and methods. In addition, there is the concern that negative studies (i.e., those that do not reject the null hypothesis) are less likely to be published than positive studies and that an overall positive association may therefore merely reflect unbalanced reporting.