Statistical Analysis of Rodent Body Weight Data is Robust to Departures from Normality in Historical National Toxicology Program Studies Dated 1980-2013

Abstract

Parametric statistical tests used to assess body weight changes in rodent experiments assume a normal distribution, and the actual distribution of the rodent body weights is often assumed to be approximately normal. In order for statistical tests to be deemed appropriate without routinely confirming the normal distribution for rodent body weight data, the tests must be powerful enough to detect meaningful changes even when a population deviates from a normal distribution. Here, we present a novel analysis to assess the normality of rodent body weight data for control animals in 1,386 National Toxicology Program (NTP) studies and determined how robust a set of procedures are to detect departures from normality. The distributions of terminal body weight measurements from 90 day and chronic NTP studies were evaluated for normality using graphical and statistical testing methods. The percent of studies with terminal body weights that were not normally distributed in normality tests was typically higher in 90-day studies for Fischer 344/N (F344/N) rats and B6C3F1/N (B6C3F1) mice than Harlan Sprague-Dawley (HSD) rats across all routes of administration evaluated (feed, drinking water, gavage or inhalation). Through simulation studies, the t-test indicated adequate power to detect a difference in body weights in male B6C3F1 mice and F344/N rats in 90-day studies, even under a skew normal distribution. According to these results, common parametric tests display enough power to accurately detect body weight differences from populations not following a normal distribution, confirming the general notion that the study designs are appropriately powered. In addition to providing adequate power, the False Positive Rate (FPR) was controlled around 5% in all simulations. These results suggest that parametric tests are robust enough to give reliable results of body weight analysis in NTP studies where this is an important endpoint. Therefore, parametric testing approaches are appropriate to detect body weight changes in NTP studies when body weight distributions do not deviate too far from normality. Future steps will look at the distributions of non-terminal body weights in chronic studies, organ weights, and other species and strains of rodents.

Figure 1.. F344/N Rat Body Weight Distributions.

Boxplots depicting distribution of terminal F344/N rat body weights in NTP 90-day studies according to diet consumed and route of administration of the test article. Represented in the figure by the box is the median, first quartile, third quartile, and the vertical bars represent the 1.5*IQR. The IQR is calculated by finding the difference between the 25^th percentile and the 75^th percentile of the data (Diez et al., 2019). The type of diet and dose route is denoted on the x-axis by DF= Dosed Feed, DW = Dosed Water, G = Gavage, I = Inhalation. F07 refers to females (F) given the NIH-07 diet, F00 refers to females given the NTP-2000 diet. M07 refers to males (M) given the NIH-07 diet, M00 refers to males given the NTP-2000 diet. Body weight (in grams) is on the y-axis. The number below the diet and dose route indicates how many data points are included in each boxplot. A. Body weight distribution of F344/N rats. Median female F344/N body weights are 191.7-197.5 grams while median male F344/N body weights are 325.0-354.3 grams. B. Body weight distribution of mean body weights of F344/N rats per study subgroup. The median of the mean female and male body weights are approximately the same as in A, but there are fewer data points exceeding the 1.5 IQR limits and the spread of the data is generally smaller.

Figure 2.. B6C3F1 Mouse Body Weight Distributions.

Boxplots depicting the distribution of B6C3F1 mouse body weights in NTP 90-day studies according to diet consumed and route of administration of the test article. Represented in the figure by the box is the median, first quartile, third quartile, and the vertical bars represent the 1.5*IQR. The IQR is calculated by finding the difference between the 25^th percentile and the 75^th percentile of the data (Diez et al., 2019). The type of diet and dose route is denoted on the x-axis by DF= Dosed Feed, DW = Dosed Water, G = Gavage, I = Inhalation. F07 refers to females (F) given the NIH-07 diet, F00 refers to females given the NTP-2000 diet. Body weight (in grams) is on the y-axis. M07 refers to males (M) given the NIH-07 diet, M00 refers to males given the NTP-2000 diet. The number below the diet and dose route indicates how many data points are included in each boxplot. A. Body weight distribution of B6C3F1 mice. Female B6C3F1 mice median body weights range from 26.5-31.3 grams and male B6C3F1 median body weights range from 32.6-40.1 grams. B. Body weight distribution of mean body weights of B6C3F1 mice per study subgroup. The median of the mean and female body weights are approximately the same as in A, but there are fewer data points exceeding the 1.5 IQR limits and the spread is generally smaller.

Figure 3.. Graphical Comparison of Normally Distributed vs Non-normally Distributed Samples.

A, B. Distribution of body weights from 90-day female B6C3F1 mice exposed to 1,1,1-Trichloroethane (71-55-6) via dosed feed on the NTP diet. The SW test for normality (Royston, 1982) produced a p-value of 0.419 and therefore, the null hypothesis of a normal distribution is not rejected. C, D. Distribution of body weights from 90-day female B6C3F1 mice exposed to 2-Hydroxy-4-methoxybenzophenone (131-57-7) via dosed feed on the NIH diet. The point in blue is a point that lies beyond the 1.5 interquartile range (Diez et al., 2019). The SW test for normality produced a p-value of 0.004 and therefore the null hypothesis of a normal distribution is rejected.

Figure 4.. Distribution of the Coefficient of Variation of Each Rodent Strain.

Boxplots depicting the distribution of each strain’s coefficient of variance (CVs) for terminal body weight in NTP 90-day studies according to route of administration and diet. Represented in the figure by the box is the median, first quartile, third quartile, and the vertical bars represent the 1.5*IQR. On the x-axis, the dose route is denoted by DF= Dosed Feed, DW = Dosed Water, G = Gavage, I = Inhalation. F07 refers to females given the NIH-07 diet, F00 refers to females given the NTP-2000 diet. M07 refers to males given the NIH-07 diet, M00 refers to males given the NTP-2000 diet. The y-axis shows the range of the proportion representing the coefficient of variation. The number below the diet and dose route indicates how many data points are included in each boxplot. A. Distribution of CVs for F344/N rats. The Interquartile Range (IQR) ranged from 0.01-0.12 grams. B. Distribution of CVs for B6C3F1 mice. The IQR ranged from 0.01-0.15 grams. The IQR is calculated by finding the difference between the 25^th percentile and the 75^th percentile of the data (Diez et al., 2019).