TERATOLOGY 4 4 2 15-223 ( 1991)
Relationship Between Fetal Weight and Malformation in Developmental Toxicity Studies LOUISE M. RYAN,' PAUL J. CATALANO,' CAROLE A. KIMMEL,' AND GARY L. KIMMEL' 'Department of Biostutistics, Harvard School of Public Health, Boston, Massachusetts 02115; and 2Reproductiue and Developmental Toxicology Brunch, RD689. Human Health Assessment Groua. Office of Health and Environmental Assessment, US.Environmental Protection Agency, Washington, DC 20460
ABSTRACT Exposure to developmental toxicants may cause fetal malformations, increase prenatal death rates and reduce fetal weight a t term. However, there has been little formal study of the relationship among these effects. Certainly, no statistical methods are currently available to jointly analyze these effects of exposure. As a preliminary step in developing such methods, simple exploratory analyses were conducted using a series of ten studies conducted for the National Toxicology Program. Because fetal weight and malformation status were both reported for all live fetuses, the data permitted a n exploration of the correlation between these two outcomes. The data show a clear pattern wherein malformed fetuses tended to be lighter at term than nonmalformed fetuses. While these patterns cannot be used to draw inferences regarding the biological relationship between fetal weight and malformation, they do suggest the potential value in developing statistical models for the joint effect of exposure on fetal weight and malformations. Developmental toxicity studies in laboratory animals play a n important role in the regulation of chemicals, drugs or other substances with potential danger during development. However, quantitative risk assessment for developmental toxicants is still relatively new and many statistical issues remain unresolved. Recent years have seen a n increasing awareness of the importance of including outcomes such as prenatal death and fetal weight in addition to malformations when assessing the risk of developmental toxicity. Several investigators recently discussed statistical models that view prenatal death and malformation a s representing different but related outcomes of the developmental process (Rai and Van Ryzin, '85; Chesson, '86; Chen e t al., '91; Ryan, '91). The relationship between fetal weight and other developmental outcomes has long been of interest (Brent and Jensh, '67; Szabo and Brent, '74; Szabo and Brent, '75; Aliverti et al., '79; Sucheston et al., '86). However, there has been relatively little systematic study of these relationships, and no formal statistical methods are currently 0 1991 WILEY-LISS,
INC.
available to examine the joint effects of exposure on resorption, malformation and weight. The development of appropriate models is complicated by several features of the observable data. For example, such models would have to account for the natural ordering of these outcome variables; weight and malformation status can be observed only among fetuses that survive the risk of prenatal death (Fig. 1). In addition, such analyses would have to allow for possible correlations between fetal weight and malformation, since both these outcomes are observed in each surviving fetus. As a preliminary step in developing suitable models, a study was conducted to explore the relationship among developmental outcomes in ten developmental toxicity studies conducted for the National Toxicology Program. The major goal was to explore
Received October 17, 1990; accepted March 5, 1991 Address reprint requests to Dr. Louise Ryan, Division of Biostatistics, Dana Farber Cancer Institute, 44 Binney Street, Boston, MA 02115.
216
L.M. RYAN ET AL. Implantation
Viable Fetus
P\
Malformation Status
Weight
Fig. 1. Framework for relationships among outcomes of the developmental process.
the relationship between fetal weight and malformation status. First, preliminary analyses were conducted to determine whether exposure affected the incidence of prenatal death, malformation, and fetal weight changes in a similar fashion. This issue was addressed by simply comparing the patterns of dose response with respect to each of the three outcomes. Next, the relationship between malformation status and fetal weight was explored further, since these two outcomes were recorded for all viable fetuses. All 10 studies revealed a similar tendency for malformed fetuses to have a lower weight at term than nonmalformed fetuses. These findings suggest that fetal weight and malformation are correlated, providing motivation for the development of multivariate statistical models that allow for the joint effect of exposure on these and other related outcomes. The Discussion section lists some statistical and technical issues that arise in the development of such models, and suggests avenues for future research.
rats (Tyl et al., ,881, ethylene glycol (EG) in mice and rats (Price et al., '85), triethylene glycol dimethyl ether (TGDM) in mice and rabbits (George et al., '87a,b), diethylene glycol dimethyl ether (DYME) in mice and rabbits (Price et al., '87a,b) and theophylline (THEO) in mice and rats (Lindstrom et al., '90). With the exception of DEHP, which involved exposure throughout gestation, all studies were conducted using a standard Segment I1 design. This design involves exposure of timed-pregnant animals to the test agent over the period of organogenesis, beginning just after implantation. The effect on in-utero development is evaluated just before parturition would normally occur. Each study included a vehicle control group and 3 or 4 dosed groups, each including 20-30 pregnant dams for mice and rats or 15-25 pregnant dams for rabbits. Measured outcomes included the number of implantation sites in each female and the incidence of resorptions andlor fetal deaths. Fetuses surviving to sacrifice were weighed and evaluated for the presence of a variety of different types of malformations. The standard segment I1 design does not permit an assessment of other important developmental outcomes, such as functional abnormalities.
Statistical methods This section describes the statistical tools used in the analysis, much of which is descriptive and relies heavily on graphical techniques. The graphs and plots were generated using the statistical package S (Becker et al, '88). Boxplots provide a useful graphical display of the distribution of individual fetal weight by dose group. S constructs a boxplot by drawing a line through the median of the data, and a box around the interquartile range (25th to 75th percentile) of the data. Lines then extend out METHODS from the box to the furthest point within 1.5 Data times the interquartile range. Points beThe data were collected from ten develop- yond these lines are designated as stars. mental toxicity studies conducted by the Re- Boxplots provide a useful view of any oversearch Triangle Institute under contract to all trends in the data as well as a quick the National Toxicology Program (NTP). visual assessment of symmetry and the These studies were used because they were identification of extreme observations. It is readily available in a computerized, stan- important to note that boxplots may overdard format, and each had been conducted lap, even when groups are significantly difin two species, The studies investigated the ferent. The reason for this is that the effects of exposure to five chemicals: di(2- strength of a statistical difference between ethyhexy1)phthalate (DEHP) in mice and two or more groups depends on the standard
FETAL WEIGHT AND MALFORMATION
errors of the means, not the standard deviations of the data. The height of the boxplot reflects the latter, and in order to assess statistical significance of a n observed difference, one must perform the appropriate statistical test. Formal tests for trend required specification of a statistical model that accounts for litter effects or the tendency for littermates to be more alike than are animals from different litters (Haseman and Kupper, '79). Quasi-likelihood methods for overdispersed binary data were applied to categorical outcomes such a s resorption or malformation (McCullagh and Nelder, '89). Specifically, the statistical package GLIM was used to fit a logistic model with a n overdispersion parameter to account for the litter effect. The dose effect was characterized by a linear model on the logit scale with dose measured in mgikgiday, as shown in Table 1. Tests with fetal weight as the outcome variable were based on normal linear models, with random effects to account for litter effects (Laird and Ware, '82). The models were fit using the Fortran program REML, based on the theory and algorithms described by Laird, Lange, and Stram ('87). To allow for study-to-study variation, the models allowed a different mean weight for each study and dose group. Nonparametric smoothing techniques (Cleveland, '79) were used to illustrate and compare the shapes of the dose response curves with respect to fetal weight among normal and malformed fetuses. While the objective of the analysis was not to impose specific parametric assumptions on the shape of the dose response curves, simply plotting means within each dose group yielded very noisy and uninterpretable plots. The application of nonparametric smoothing techniques offered a n acceptable compromise, yielding relatively smooth plots but still allowing the data to determine the basic shape of the plot. RESULTS
Table 1 summarizes the data from the 10 studies. The P values for the number of live and malformed fetuses are based on trend tests for correlated binary data. The P values for fetal weight are based on a model for normally distributed data. In general, the relationship between dose level and the percentages of live births, malformations and mean fetal weight suggests the presence of
217
dose effects with respect to all three outcomes. This impression is supported for the most part by significant P values of
Relationship Between Fetal Weight and Malformation in Developmental Toxicity Studies LOUISE M. RYAN,' PAUL J. CATALANO,' CAROLE A. KIMMEL,' AND GARY L. KIMMEL' 'Department of Biostutistics, Harvard School of Public Health, Boston, Massachusetts 02115; and 2Reproductiue and Developmental Toxicology Brunch, RD689. Human Health Assessment Groua. Office of Health and Environmental Assessment, US.Environmental Protection Agency, Washington, DC 20460
ABSTRACT Exposure to developmental toxicants may cause fetal malformations, increase prenatal death rates and reduce fetal weight a t term. However, there has been little formal study of the relationship among these effects. Certainly, no statistical methods are currently available to jointly analyze these effects of exposure. As a preliminary step in developing such methods, simple exploratory analyses were conducted using a series of ten studies conducted for the National Toxicology Program. Because fetal weight and malformation status were both reported for all live fetuses, the data permitted a n exploration of the correlation between these two outcomes. The data show a clear pattern wherein malformed fetuses tended to be lighter at term than nonmalformed fetuses. While these patterns cannot be used to draw inferences regarding the biological relationship between fetal weight and malformation, they do suggest the potential value in developing statistical models for the joint effect of exposure on fetal weight and malformations. Developmental toxicity studies in laboratory animals play a n important role in the regulation of chemicals, drugs or other substances with potential danger during development. However, quantitative risk assessment for developmental toxicants is still relatively new and many statistical issues remain unresolved. Recent years have seen a n increasing awareness of the importance of including outcomes such as prenatal death and fetal weight in addition to malformations when assessing the risk of developmental toxicity. Several investigators recently discussed statistical models that view prenatal death and malformation a s representing different but related outcomes of the developmental process (Rai and Van Ryzin, '85; Chesson, '86; Chen e t al., '91; Ryan, '91). The relationship between fetal weight and other developmental outcomes has long been of interest (Brent and Jensh, '67; Szabo and Brent, '74; Szabo and Brent, '75; Aliverti et al., '79; Sucheston et al., '86). However, there has been relatively little systematic study of these relationships, and no formal statistical methods are currently 0 1991 WILEY-LISS,
INC.
available to examine the joint effects of exposure on resorption, malformation and weight. The development of appropriate models is complicated by several features of the observable data. For example, such models would have to account for the natural ordering of these outcome variables; weight and malformation status can be observed only among fetuses that survive the risk of prenatal death (Fig. 1). In addition, such analyses would have to allow for possible correlations between fetal weight and malformation, since both these outcomes are observed in each surviving fetus. As a preliminary step in developing suitable models, a study was conducted to explore the relationship among developmental outcomes in ten developmental toxicity studies conducted for the National Toxicology Program. The major goal was to explore
Received October 17, 1990; accepted March 5, 1991 Address reprint requests to Dr. Louise Ryan, Division of Biostatistics, Dana Farber Cancer Institute, 44 Binney Street, Boston, MA 02115.
216
L.M. RYAN ET AL. Implantation
Viable Fetus
P\
Malformation Status
Weight
Fig. 1. Framework for relationships among outcomes of the developmental process.
the relationship between fetal weight and malformation status. First, preliminary analyses were conducted to determine whether exposure affected the incidence of prenatal death, malformation, and fetal weight changes in a similar fashion. This issue was addressed by simply comparing the patterns of dose response with respect to each of the three outcomes. Next, the relationship between malformation status and fetal weight was explored further, since these two outcomes were recorded for all viable fetuses. All 10 studies revealed a similar tendency for malformed fetuses to have a lower weight at term than nonmalformed fetuses. These findings suggest that fetal weight and malformation are correlated, providing motivation for the development of multivariate statistical models that allow for the joint effect of exposure on these and other related outcomes. The Discussion section lists some statistical and technical issues that arise in the development of such models, and suggests avenues for future research.
rats (Tyl et al., ,881, ethylene glycol (EG) in mice and rats (Price et al., '85), triethylene glycol dimethyl ether (TGDM) in mice and rabbits (George et al., '87a,b), diethylene glycol dimethyl ether (DYME) in mice and rabbits (Price et al., '87a,b) and theophylline (THEO) in mice and rats (Lindstrom et al., '90). With the exception of DEHP, which involved exposure throughout gestation, all studies were conducted using a standard Segment I1 design. This design involves exposure of timed-pregnant animals to the test agent over the period of organogenesis, beginning just after implantation. The effect on in-utero development is evaluated just before parturition would normally occur. Each study included a vehicle control group and 3 or 4 dosed groups, each including 20-30 pregnant dams for mice and rats or 15-25 pregnant dams for rabbits. Measured outcomes included the number of implantation sites in each female and the incidence of resorptions andlor fetal deaths. Fetuses surviving to sacrifice were weighed and evaluated for the presence of a variety of different types of malformations. The standard segment I1 design does not permit an assessment of other important developmental outcomes, such as functional abnormalities.
Statistical methods This section describes the statistical tools used in the analysis, much of which is descriptive and relies heavily on graphical techniques. The graphs and plots were generated using the statistical package S (Becker et al, '88). Boxplots provide a useful graphical display of the distribution of individual fetal weight by dose group. S constructs a boxplot by drawing a line through the median of the data, and a box around the interquartile range (25th to 75th percentile) of the data. Lines then extend out METHODS from the box to the furthest point within 1.5 Data times the interquartile range. Points beThe data were collected from ten develop- yond these lines are designated as stars. mental toxicity studies conducted by the Re- Boxplots provide a useful view of any oversearch Triangle Institute under contract to all trends in the data as well as a quick the National Toxicology Program (NTP). visual assessment of symmetry and the These studies were used because they were identification of extreme observations. It is readily available in a computerized, stan- important to note that boxplots may overdard format, and each had been conducted lap, even when groups are significantly difin two species, The studies investigated the ferent. The reason for this is that the effects of exposure to five chemicals: di(2- strength of a statistical difference between ethyhexy1)phthalate (DEHP) in mice and two or more groups depends on the standard
FETAL WEIGHT AND MALFORMATION
errors of the means, not the standard deviations of the data. The height of the boxplot reflects the latter, and in order to assess statistical significance of a n observed difference, one must perform the appropriate statistical test. Formal tests for trend required specification of a statistical model that accounts for litter effects or the tendency for littermates to be more alike than are animals from different litters (Haseman and Kupper, '79). Quasi-likelihood methods for overdispersed binary data were applied to categorical outcomes such a s resorption or malformation (McCullagh and Nelder, '89). Specifically, the statistical package GLIM was used to fit a logistic model with a n overdispersion parameter to account for the litter effect. The dose effect was characterized by a linear model on the logit scale with dose measured in mgikgiday, as shown in Table 1. Tests with fetal weight as the outcome variable were based on normal linear models, with random effects to account for litter effects (Laird and Ware, '82). The models were fit using the Fortran program REML, based on the theory and algorithms described by Laird, Lange, and Stram ('87). To allow for study-to-study variation, the models allowed a different mean weight for each study and dose group. Nonparametric smoothing techniques (Cleveland, '79) were used to illustrate and compare the shapes of the dose response curves with respect to fetal weight among normal and malformed fetuses. While the objective of the analysis was not to impose specific parametric assumptions on the shape of the dose response curves, simply plotting means within each dose group yielded very noisy and uninterpretable plots. The application of nonparametric smoothing techniques offered a n acceptable compromise, yielding relatively smooth plots but still allowing the data to determine the basic shape of the plot. RESULTS
Table 1 summarizes the data from the 10 studies. The P values for the number of live and malformed fetuses are based on trend tests for correlated binary data. The P values for fetal weight are based on a model for normally distributed data. In general, the relationship between dose level and the percentages of live births, malformations and mean fetal weight suggests the presence of
217
dose effects with respect to all three outcomes. This impression is supported for the most part by significant P values of
