Statistics in Toxicology II

Testing

Kirsten Schorning
Department of Statistics,
TU Dortmund University

Winter semester 2024/25

 1 Introduction

Introduction

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 1

 1 Introduction 1.1 Overview

1.1 Lecturer
JProf. Dr. Kirsten Schorning (lecture)
Mathematics Building, Room 725
Email: schorning@statistik.tu-dortmund.de

Study of Mathematics with minor in Computer linguistics in

Bochum;
JProfessor for Mathematical Statistics, Department of Statistics, TU
Dortmund University; with research areas: Statistical methods for
toxicology, optimal design

Leonie Schuermeyer (exercises)

Mathematics Building, Room 916
Email: schuermeyer@statistik.tu-dortmund.de

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 2

 1 Introduction 1.1 Overview

1.1 Organization
Course: Statistics in toxicology (testing)
Lecture dates:
Monday, 10:15 - 11:45 in C/HS 1
Exercise dates:
Wednesday, 8.30-10.00 in M/E 27
Moodle: Statistics in Toxicology II, WiSe 24/25
Website:
https://moodle.tu-dortmund.de/course/view.php?id=48552

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 3

 1 Introduction 1.1 Overview

1.1 Organization
Admission to exam
For admission to the written exam, 50% of the points of the graded
exercise sheet 4 must be achieved. The submission of sheet 4 is
mandatory, and it will be corrected in detail.
The exercises will consist of theoretical and software exercises.
Further information can be found on the Moodle page.
Admission to exam via BOSS system, deadline TBA via Moodle.
Written exam date
One hour exam
Room and data: TBA in Moodle
Written retry exam date
One hour exam
Room and data: TBA in Moodle

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 4

 1 Introduction 1.1 Overview

1.1 Organization
Modules
BS 14, BS 15 ”Spezialgebiete”, ”Quantitative Methoden”
MS 6, MS 7 ”Spezialgebiete”
MD E2-12 ”Applications - Elective Courses”
Prerequisites
For BS: Statistik I - III
For MD: MD Req4 Probability, MD Req5 Inference, MD Req6
Linear models
No deep knowledge in genetics and biology, but at least a strong
interest

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 5

 1 Introduction 1.1 Overview

1.1 Content
The main topic of the lecture is the analysis of data from toxicology,
especially dose-response measurements. Specific topics are:
Proof of hazard using simultaneous comparisons with a negative
control
Multiple testing
Tests for normally distributed endpoints and for proportions
Trend tests
Analysis of long-term effects in cancer studies
Survival analysis, tests for survival endpoint
Analysis of effects in mutagenicity assays
Mixture distributions
EM algorithm
Multiple testing for genomic data
Family-wise error rate
False discovery rate
Clinical trials phase I studies

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 6

 1 Introduction 1.1 Overview

1.1 Literature
Main book for this course:
Hothorn, L.A.: Statistics in Toxicology Using R, Chapman and
Hall/CRC, 2016 (SiTUR)

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 7

 1 Introduction 1.1 Overview

1.1 Motivation

There are three kinds of lies - lies, damned lies, and

statistics (Leonard Henry Courtney, 1832-1918)
To guess is cheap, to guess wrong is expensive (Chinese
proverb)

Statistics is (also) fun

Statistics is (also) intuition
Statistics is (also) surprise

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 8

 1 Introduction 1.1 Overview

1.1 Toxicology
Definitions
a science that deals with poisons and their effect and with the
problems involved (such as clinical, industrial, or legal problems)
(Merriam-Webster, 22.11.2023)
Toxicology is a scientific discipline, overlapping with biology,
chemistry, pharmacology, and medicine, that involves the study of
the adverse effects of chemical substances on living organisms and
the practice of diagnosing and treating exposures to toxins and
toxicants. The relationship between dose and its effects on the
exposed organism is of high significance in toxicology. Factors that
influence chemical toxicity include the dosage, duration of exposure
(whether it is acute or chronic), route of exposure, species, age, sex,
and environment. Toxicologists are experts on poisons and
poisoning. There is a movement for evidence-based toxicology as
part of the larger movement towards evidence-based practices.
(Wikipedia, 05.10.2024)

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 9

 1 Introduction 1.1 Overview

1.1 Toxicology
Definitions
Toxicology is a field of science that helps us understand the harmful
effects that chemicals, substances, or situations can have on people,
animals, and the environment. Some refer to toxicology as the
”Science of Safety” because, as a field, it has evolved from a science
focused on studying poisons and adverse effects of chemical
exposures to a science devoted to studying safety.
(NIH, https://www.niehs.nih.gov/health/topics/science/toxicology,
05.10.2024)

The contribution (or general role) of statistics respectively data

analysis is typically neglected, even in extended definitions.

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 10

 1 Introduction 1.1 Overview

1.1 Statistics for toxicology

Statistics for toxicology: For statisticians in its basic form, maybe

standard, but for toxicologists, often a great challenge →
communication is key!
Central question still open: significant test of a criterion for hazard
or even more important is non-significance a criterion for safety?
Difference between statistical significance and biological relevance
Hypothesis testing or modeling?
Arguments pro testing (SiTUR book)
Guidelines often require hypothesis tests
Modelling not possible for only 2-3 different doses
Modelling is only correct when certain assumptions are fulfilled
Hypotheses tests often preferred in scientific publications
Arguments pro modeling
If enough doses and replicates are available and functional form can
be justified, then estimates are more precise (also at non-tested
doses)
Flexible frameworks exist (such as MCPMod)
Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 11
 1 Introduction 1.1 Overview

1.1 Statistics for toxicology

Research project at TU Dortmund (first phase 04/2021-09/2025):
DFG Research Training Group (RTG) 2624 ”Biostatistical methods for
high-dimensional data in toxicology”
https://grk2624.statistik.tu-dortmund.de/en/

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 12

 1 Introduction 1.2 Principles

1.2 Evaluation of short-term repeated toxicity

studies
General principles
Compound is often administered between 4 weeks and 3 months,
but other exposition times are possible
Type of endpoints (with different scales of measurements):
continuous (e.g., hemoglobin, a protein in blood that carries oxygen)
rates (proportions of histopathological findings)
ordered categorical data (e.g., graded histopathological findings)
Typical design: Negative control C and doses D1 , . . . , Dk , often for
k = 3 doses (concentrations)
Small sample size per dose is a common feature in toxicology
For the simple comparison dose-control (for different doses) the
following tests are ”standard” and will be discussed in this course:
Unadjusted two-sample test
Dunnett-type tests without order restriction
Williams-type tests with order restriction
Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 13
 1 Introduction 1.2 Principles

1.2 Data visualization: Barplots and boxplots

Common presentations and visualizations

Tables with measures for groups (doses): Means, standard
deviations, standard error, sample sizes
Barplots (barcharts) (with standard error, sometimes standard
deviation)
Disadvantages of barplots
Assumption of normal distribution not necessarily fulfilled
Individual data points not shown, also no extreme values
Alternatives:
Plots including all individual data points
Boxplots (median, quartile, extreme values): not yet used much in
toxicology, but provide additional information regarding variability,
symmetry, and potential outliers
Letters of significance (compact letter display, generated e.g. with
insert-and-absorb algorithm, see second next slide)

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 14

 1 Introduction 1.2 Principles

1.2 Data visualization: Barplots and boxplots

Measures of variability
Given are data points x1 , .r
. . , xn with mean x̄
n
1
P
Standard deviation: sd = n−1
(xi − x̄ )2 , measures variability of
i=1
individual data points
Standard error: √sdn , measures variability of x̄
MAD: median absolute deviation from median, more robust
alternative for sd
Boxplots
Many different ways to construct boxplots
Box indicates quartiles, with median (second quartile) as line within
the box
Whiskers extended to smallest value within 1.5 IQR of the lower
quartile and to largest value within 1.5 IQR of the upper quartile
More extreme points typically plotted as single points
Attention: For very small sample size (in toxicology, e.g., often 3
values) overinterpretation of five measures, for very large sample size
overinterpretation of extreme values (not necessarily outliers!)

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 15

 1 Introduction 1.2 Principles

1.2 Data visualization: Letters of significance

Idea of letters of significance

Show statistically significant differences between variables
For all variables with the same letter, the difference between the
means is not statistically significant
More precisely: If two treatments (doses) do not share any letter,
then (and only then) they are statistically different
Algorithm: Insert-and-absorb algorithm (Piepho, 2004)
Given is a set of k (k − 1)/2 significance statements (often in the
form of p-values) corresponding to all pairwise comparisons, where k
is the number of treatments, adjusted for multiplicity where necessary
For every significant pair, perform a series of steps as described on
the next slide
Advantages and disadvantages
Often understandable presentation of groups of similar treatments
Worst-case runtime is exponential → for very large numbers of
treatments (which is rare) not always suitable
Exact p-value is unknown, unless included (e.g. in a table)

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 16

 1 Introduction 1.2 Principles

1.2 Data visualization: Compact letter display

Insert-and-absorb algorithm (Piepho, 2004,

https://doi.org/10.1198/1061860043515), see also exercises

1. Generate a column connecting all treatments.

2. For each signicant comparison do:
For each column currently in the display do:
If the column connects the two signicantly different treatments then do:
- Duplicate the column.
- In the rst of the two columns delete the letter corresponding to the one
treatment. If possible, absorb the column into another column.
- In the second of the two columns delete the letter corresponding to the other
treatment. If possible, absorb the column into another column.
End.
End.
End.

Fig. 1 : Insert-and-absorb algorithm

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 17

 1 Introduction 1.2 Principles

1.2 Presentation
Presentation of data using a table

Fig. 2 : Example of presentation in table, with group means and

corresponding standard errors, DMSO (dimethyl sulfoxide) is a solvent
that is often used as control, tBOOH (tert-Butyl hydroperoxide) is an
organic compound whose effects are of interest

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 18

 1 Introduction 1.2 Principles

1.2 Presentation
Presentation of data using a barplot

Fig. 3 : Example of presentation with barplot, with group means,

corresponding standard errors (only upwards), and letters of significance
(equal letters indicate non-significant difference between groups)
Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 19
 1 Introduction 1.2 Principles

1.2 Presentation

Presentation of data with individual points

Fig. 4 : Example of presentation with individual data points, with geometric

means (logarithmic scale!)

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 20

 1 Introduction 1.2 Principles

1.2 Presentations

Presentation of data with boxplots

Fig. 5 : Example of presentation with boxplots, with significant group

differences indicated by lines with stars

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 21

 1 Introduction 1.2 Principles

1.2 Presentation
Comparison of two graphical presentations (next slide)
Data represent litter-specific pup weights (of newborn rats)
Litter: a number of young animals born to an animal at one time
Barplot visualizes only means and standard deviations
Boxplot shows individual data (dots, jittered), parametric and
non-parametric summary measures (mean, standard deviation (both
next to boxplot), median, interquartile range), and per-litter
structure (including group-specific number of litters and number of
animals)

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 22

 1 Introduction 1.2 Principles

1.2 Presentation

Fig. 6 : Direct comparison of two graphical presentations, with detailed boxplots

(left) and barplots plus standard deviation (right), for the same dataset

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 23

 1 Introduction 1.2 Principles

1.2 Presentation of test results

Presentation styles
Rejection/non-rejection of H0 , letters indicate non-distinguishable
treatment groups
Rejection of H0 for three α-levels (0.05, 0.01, 0.001), visualized by
stars (∗, ∗∗, ∗∗∗)
p-value (p)
confidence interval
Stars are still common in toxicology, also p-values are frequently used
p-value has philosophical disadvantage
A small value does not prove H1 , but only demonstrates that H0
does not fit well to the data
In other words: a p-value is the probability of observing a sample
statistic that is at least as extreme as the sample statistic when one
assumes that the null hypothesis is true
A p-value refers to P(data|H0 ), not to P(H0 |data) !

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 24

 1 Introduction 1.2 Principles

1.2 Presentation of test results

Confidence intervals have the advantage that they contain

information about
rejection/non-rejection of H0 , by inclusion/non-inclusion of the value
of H0 (e.g., 0 for difference and 1 for ratio)
interpretation of biological relevance, by the distance of the
confidence limits to this value of H0 , in terms of the measured unit
(difference) or percentage change (ratio)
the directional decision (whether increasing or decreasing effects
occur)
Confidence limits are also recommended in guidelines for clinical
trials (ICH E9), for toxicological studies no such recommendations
exist yet (SiTUR)
For confidence intervals an adequate effect size is required
(difference, ratio, odds ratio)

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 25

 1 Introduction 1.2 Principles

1.2 Effect sizes

Let πi be the proportion of events in group i, typically estimated by
π̂i = Ynii , where ni is the size of group i and Yi the number of events in
group i.
The risk difference between group i and control group 0 is defined as

πi − π0

The relative risk between group i and control group 0 is defined as

πi
π0
The odds ratio between group i and control group 0 is defined as

πi /(1 − πi )
π0 /(1 − π0 )

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 26

 1 Introduction 1.2 Principles

1.2 Presentation of test results

Example data set
Data set with outcome (serum) triglyceride (Blutfette)
High value is risk factor for cardiovascular diseases
Sodium dichromate dihydrate given to F344 rats (inbreeding strain
of rats, dates back to the year 1920), often used in cancer research
and toxicology research
One control group, five (experimental) doses (62.5, 125, 250, 500,
1000) in a 13-week-study
Figure on next slide:
Left: Boxplots and letters
Right: Boxplots, individual values (with jitter), Dunnett-type
confidence intervals (see chapter 2) and stars (for difference between
corresponding dose and control)

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 27

 1 Introduction 1.2 Principles

1.2 Presentation of test results

Fig. 7 : Example: Visualization of triglyceride measurements in rats, left:

boxplots with letter-type representation of significances, right: boxplots with
Dunnett-type stars (for simultaneous comparison against control)

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 28

 1 Introduction 1.2 Principles

1.2 Proof of hazard or proof of safety?

Main question: Is a compound (new drug, chemical) in general

harmless, harmless up to a specified dose, or harmful?
Typical problem when testing with p-values:
For large sample sizes, potentially even small (biologically
non-relevant) effects are statistically significant
For small sample sizes, potentially even large (biologically) relevant
effects are not statistically significant

Important insight: Absence of evidence is not evidence of absence!

This means: If the result is non-significance, one cannot draw the
conclusion that there is no effect!

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 29

 1 Introduction 1.2 Principles

1.2 Sample size

Minimal required sample size for a toxicological assay is specified in

many guidelines (e.g., 3 or 5 per dose)
A minimal required sample size helps to increase the power of a test
(e.g., for the proof of hazard), but depends on the variance of the
outcome values, the test level α, and assumptions like normality, for
the proof of safety also on a tolerance threshold ϑ
Note: Properties of tests often do not hold for very small sample
size (e.g., asymptotic distribution is normal distribution)
Problematic interpretation of p-value: Direct function of sample
size, thus be careful with the interpretation in unplanned studies
→ Simulation study: Analysis of data sets with the same values for mean
and variance, but different sample size 5, 10, 15 (in R: the function
ermvnorm simulates normal distributed data with exact values for
parameters)

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 30

 1 Introduction 1.2 Principles

1.2 Two-step test decisions

In the case of potential variance heterogeneity (different variances

for different doses) often first a pretest on variance heterogeneity is
performed and then (depending on the result) a corresponding test
for the desired effect(s) is used
Problems with this procedure
Level α might not be kept
Nonparametric tests might be unsuitable in case of variance
heterogeneity
Conditional tests are often problematic

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 31

 1 Introduction 1.2 Principles

1.2 Importance of control groups

Negative control group is typically required in toxicology!

Historical controls might be useful, but often unclear how similar
they are to simultaneous control experiments (e.g., due to batch
effects)
Positive controls are not common, but can also be useful
to demonstrate that an assay is sensitive (changes in outcome
variable can actually be achieved)
to demonstrate the relevance of a change versus negative control
(using a non-inferiority test w.r.t. positive control)

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 32

 1 Introduction 1.2 Principles

1.2 Importance of control groups

Example: Micronucleus data
6 treatments: negative control (Vehicle, C–), Hydro30, Hydro50,
Hydro75, Hydro100, positive control (Cycle25)
Dunnett-type comparison: all 4 doses of Hydro once tested versus
C– and once versus C+
Description of the figure on next slide:
upper p-value for comparison against C– (vehicle)
lower p-value for comparison against C+ (non-inferiority against C+)
result, for example for Hydro 100: significant increase compared to
C– (p = 0.001), but no significant inferiority against C+ (p = 0.639)
p-value for Cycle25 corresponds to t-test for C– versus C+ (here
p < 0.01)

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 33

 1 Introduction 1.2 Principles

1.2 Importance of control groups

upper p-value
comparison
against vehicle C–
lower p-value
comparison
against C+
p-value for
Cycle25
corresponds to
t-test for C–
versus C+

Fig. 8 : Example with comparisons against negative and

positive control

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 34

 1 Introduction 1.2 Principles

1.2 Statistical significance and biological relevance

Statistical significance versus biological relevance:

Difference between the two terms is important
Example (simulated data, five doses D1-D5, negative control NC)
Biological relevance, when estimated effect e is above an a priori
defined relevance threshold c (here c = 2.1):
D1-NC: Statistically not significant, biological relevance unclear
D2-NC: Statistically significant (0 < e < c), biologically not relevant
D3-NC: Statistically significant (0 < e), not significantly below
threshold c, biological relevance unclear
D4-NC: Statistically significant, probably biologically relevant
D5-NC: Statistically significant (c < e), biologically relevant

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 35

 1 Introduction 1.2 Principles

1.2 Statistical significance and biological relevance

Fig. 9 : Confidence intervals for an artificial example with five 5 doses

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 36

 1 Introduction 1.2 Principles

1.2 Proof of hazard using two-sample comparisons

Proof of hazard is the typical (most frequent) test situation:

Null hypothesis H0 corresponds to equality of effects and is rejected
or not rejected
For normal distributed continuous endpoint: often t-test
(respectively Welch test) is used, sometimes also Wilcoxon test
Example:
F344 rats, treated 13 weeks with sodium dichromate dihydrate,
6 different doses
18 endpoints, including BUN (blood urea nitrigen), Creat (serum
creatinin), ALB (albumin), SerumGlucose (serum glucose),
CreatKinase (creatine kinase), ALT (anilin aminotransferase)
Confidence intervals of Welch test: scale-specific
Alternatives
Confidence intervals for ratio-to-control comparisons
Confidence intervals for log-normal data
Nonparametric confidence intervals for ratio-to-control comparisons

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 37

 1 Introduction 1.2 Principles

1.2 Proof of hazard using two-sample comparisons

Dataset clin: 13-week study with sodium dichromate dihydrate

administered to rats, 6 doses, 10 animals per dose, and 18 biological
endpoints for each animal
In R: mclin: Data set joined with command melt
R code:
str(mclin)

## 'data.frame': 1080 obs. of 3 variables:

## $ dose : Factor w/ 6 levels "0","62.5","125",..: 1 1 1 1 1 1 1 1 1 1 ...
## $ variable: Factor w/ 18 levels "BUN","Creat",..: 1 1 1 1 1 1 1 1 1 1 ...
## $ value : num 17.3 15 15.7 16.5 16.7 15.2 16.2 16.6 17.9 16.7 ...

levels(mclin$dose); levels(mclin$variable)

## [1] "0" "62.5" "125" "250" "500" "1000"

## [1] "BUN" "Creat" "TP " "ALB "
## [5] "SerumGlucose" "CreatKinase" "ALT" "SDH"
## [9] "ALP" "Nuc" "TBA" "Cholesterol"
## [13] "Triglyceride" "Chloride" "Sodium" "Calc"
## [17] "PHOS" "Potassium"

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 38

 1 Introduction 1.2 Principles

1.2 Proof of hazard using two-sample comparisons

In this example many comparisons are of interest, and simultaneous

confidence intervals are suitable (required)
Rat example: Four types of confidence intervals are calculated
Welch t-test confidence intervals based on original data (classical,
symmetric in additive sense)
Ratio-to-control confidence intervals, suitable e.g. for skewed
distributions, related to logarithmic transformation (symmetric in
multiplicative sense)
Ratio-to-control confidence intervals (Fieller) for log-normal
endpoints (alternative to transformation)
Non-parametric ratio-to-control confidence intervals
(Hodges-Lehmann, based on ranks and test of median differences, as
in Wilcoxon test)
Result: In general similar results in all 4 cases, but different types of
symmetry and also different lengths of confidence intervals,
especially close to difference 0

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 39

 1 Introduction 1.2 Principles

1.2 Confidence intervals

BUN : 62.5 − 0
BUN : 125 − 0
BUN : 250 − 0
BUN : 500 − 0
BUN : 1000 − 0
Creat : 62.5 − 0
Creat : 125 − 0
Creat : 250 − 0
Creat : 500 − 0
Creat : 1000 − 0
TP : 62.5 − 0
TP : 125 − 0
TP : 250 − 0
TP : 500 − 0
TP : 1000 − 0
ALB : 62.5 − 0
ALB : 125 − 0
ALB : 250 − 0
ALB : 500 − 0
ALB : 1000 − 0
SerumGlucose : 62.5 − 0
SerumGlucose : 125 − 0
SerumGlucose : 250 − 0
SerumGlucose : 500 − 0
SerumGlucose : 1000 − 0
CreatKinase : 62.5 − 0
CreatKinase : 125 − 0
CreatKinase : 250 − 0
CreatKinase : 500 − 0
CreatKinase : 1000 − 0
ALT : 62.5 − 0
ALT : 125 − 0
ALT : 250 − 0
ALT : 500 − 0
ALT : 1000 − 0
SDH : 62.5 − 0
SDH : 125 − 0
SDH : 250 − 0
SDH : 500 − 0
SDH : 1000 − 0
ALP : 62.5 − 0
ALP : 125 − 0
ALP : 250 − 0
ALP : 500 − 0
ALP : 1000 − 0
Nuc : 62.5 − 0
Nuc : 125 − 0
Nuc : 250 − 0
Nuc : 500 − 0
Nuc : 1000 − 0
TBA : 62.5 − 0
TBA : 125 − 0
TBA : 250 − 0
TBA : 500 − 0
TBA : 1000 − 0
Cholesterol : 62.5 − 0
Cholesterol : 125 − 0
Cholesterol : 250 − 0
Cholesterol : 500 − 0
Cholesterol : 1000 − 0
Triglyceride : 62.5 − 0
Triglyceride : 125 − 0
Triglyceride : 250 − 0
Triglyceride : 500 − 0
Triglyceride : 1000 − 0
Chloride : 62.5 − 0
Chloride : 125 − 0
Chloride : 250 − 0
Chloride : 500 − 0
Chloride : 1000 − 0
Sodium : 62.5 − 0
Sodium : 125 − 0
Sodium : 250 − 0
Sodium : 500 − 0
Sodium : 1000 − 0
Calc : 62.5 − 0
Calc : 125 − 0
Calc : 250 − 0
Calc : 500 − 0
Calc : 1000 − 0
PHOS : 62.5 − 0
PHOS : 125 − 0
PHOS : 250 − 0
PHOS : 500 − 0
PHOS : 1000 − 0
Potassium : 62.5 − 0
Potassium : 125 − 0
Potassium : 250 − 0
Potassium : 500 − 0
Potassium : 1000 − 0

−100 0 100 200 300 400 500

Welch t−test confidence intervals

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 40

 1 Introduction 1.2 Principles

1.2 Confidence intervals

BUN : 62.5 / 0
BUN : 125 / 0
BUN : 250 / 0
BUN : 500 / 0
BUN : 1000 / 0
Creat : 62.5 / 0
Creat : 125 / 0
Creat : 250 / 0
Creat : 500 / 0
Creat : 1000 / 0
TP : 62.5 / 0
TP : 125 / 0
TP : 250 / 0
TP : 500 / 0
TP : 1000 / 0
ALB : 62.5 / 0
ALB : 125 / 0
ALB : 250 / 0
ALB : 500 / 0
ALB : 1000 / 0
SerumGlucose : 62.5 / 0
SerumGlucose : 125 / 0
SerumGlucose : 250 / 0
SerumGlucose : 500 / 0
SerumGlucose : 1000 / 0
CreatKinase : 62.5 / 0
CreatKinase : 125 / 0
CreatKinase : 250 / 0
CreatKinase : 500 / 0
CreatKinase : 1000 / 0
ALT : 62.5 / 0
ALT : 125 / 0
ALT : 250 / 0
ALT : 500 / 0
ALT : 1000 / 0
SDH : 62.5 / 0
SDH : 125 / 0
SDH : 250 / 0
SDH : 500 / 0
SDH : 1000 / 0
ALP : 62.5 / 0
ALP : 125 / 0
ALP : 250 / 0
ALP : 500 / 0
ALP : 1000 / 0
Nuc : 62.5 / 0
Nuc : 125 / 0
Nuc : 250 / 0
Nuc : 500 / 0
Nuc : 1000 / 0
TBA : 62.5 / 0
TBA : 125 / 0
TBA : 250 / 0
TBA : 500 / 0
TBA : 1000 / 0
Cholesterol : 62.5 / 0
Cholesterol : 125 / 0
Cholesterol : 250 / 0
Cholesterol : 500 / 0
Cholesterol : 1000 / 0
Triglyceride : 62.5 / 0
Triglyceride : 125 / 0
Triglyceride : 250 / 0
Triglyceride : 500 / 0
Triglyceride : 1000 / 0
Chloride : 62.5 / 0
Chloride : 125 / 0
Chloride : 250 / 0
Chloride : 500 / 0
Chloride : 1000 / 0
Sodium : 62.5 / 0
Sodium : 125 / 0
Sodium : 250 / 0
Sodium : 500 / 0
Sodium : 1000 / 0
Calc : 62.5 / 0
Calc : 125 / 0
Calc : 250 / 0
Calc : 500 / 0
Calc : 1000 / 0
PHOS : 62.5 / 0
PHOS : 125 / 0
PHOS : 250 / 0
PHOS : 500 / 0
PHOS : 1000 / 0
Potassium : 62.5 / 0
Potassium : 125 / 0
Potassium : 250 / 0
Potassium : 500 / 0
Potassium : 1000 / 0

0 2 4 6 8

Ratio−to−control confidence intervals

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 41

 1 Introduction 1.2 Principles

1.2 Confidence intervals

BUN : 62.5 / 0
BUN : 125 / 0
BUN : 250 / 0
BUN : 500 / 0
BUN : 1000 / 0
Creat : 62.5 / 0
Creat : 125 / 0
Creat : 250 / 0
Creat : 500 / 0
Creat : 1000 / 0
TP : 62.5 / 0
TP : 125 / 0
TP : 250 / 0
TP : 500 / 0
TP : 1000 / 0
ALB : 62.5 / 0
ALB : 125 / 0
ALB : 250 / 0
ALB : 500 / 0
ALB : 1000 / 0
SerumGlucose : 62.5 / 0
SerumGlucose : 125 / 0
SerumGlucose : 250 / 0
SerumGlucose : 500 / 0
SerumGlucose : 1000 / 0
CreatKinase : 62.5 / 0
CreatKinase : 125 / 0
CreatKinase : 250 / 0
CreatKinase : 500 / 0
CreatKinase : 1000 / 0
ALT : 62.5 / 0
ALT : 125 / 0
ALT : 250 / 0
ALT : 500 / 0
ALT : 1000 / 0
SDH : 62.5 / 0
SDH : 125 / 0
SDH : 250 / 0
SDH : 500 / 0
SDH : 1000 / 0
ALP : 62.5 / 0
ALP : 125 / 0
ALP : 250 / 0
ALP : 500 / 0
ALP : 1000 / 0
Nuc : 62.5 / 0
Nuc : 125 / 0
Nuc : 250 / 0
Nuc : 500 / 0
Nuc : 1000 / 0
TBA : 62.5 / 0
TBA : 125 / 0
TBA : 250 / 0
TBA : 500 / 0
TBA : 1000 / 0
Cholesterol : 62.5 / 0
Cholesterol : 125 / 0
Cholesterol : 250 / 0
Cholesterol : 500 / 0
Cholesterol : 1000 / 0
Triglyceride : 62.5 / 0
Triglyceride : 125 / 0
Triglyceride : 250 / 0
Triglyceride : 500 / 0
Triglyceride : 1000 / 0
Chloride : 62.5 / 0
Chloride : 125 / 0
Chloride : 250 / 0
Chloride : 500 / 0
Chloride : 1000 / 0
Sodium : 62.5 / 0
Sodium : 125 / 0
Sodium : 250 / 0
Sodium : 500 / 0
Sodium : 1000 / 0
Calc : 62.5 / 0
Calc : 125 / 0
Calc : 250 / 0
Calc : 500 / 0
Calc : 1000 / 0
PHOS : 62.5 / 0
PHOS : 125 / 0
PHOS : 250 / 0
PHOS : 500 / 0
PHOS : 1000 / 0
Potassium : 62.5 / 0
Potassium : 125 / 0
Potassium : 250 / 0
Potassium : 500 / 0
Potassium : 1000 / 0

0 2 4 6 8 10 12

Ratios−to−control for log−normal endpoints

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 42

 1 Introduction 1.2 Principles

1.2 Confidence intervals

BUN : 62.5 / 0
BUN : 125 / 0
BUN : 250 / 0
BUN : 500 / 0
BUN : 1000 / 0
TP : 62.5 / 0
TP : 125 / 0
TP : 250 / 0
TP : 500 / 0
TP : 1000 / 0
ALB : 62.5 / 0
ALB : 125 / 0
ALB : 250 / 0
ALB : 500 / 0
ALB : 1000 / 0
SerumGlucose : 62.5 / 0
SerumGlucose : 125 / 0
SerumGlucose : 250 / 0
SerumGlucose : 500 / 0
SerumGlucose : 1000 / 0
CreatKinase : 62.5 / 0
CreatKinase : 125 / 0
CreatKinase : 250 / 0
CreatKinase : 500 / 0
CreatKinase : 1000 / 0
ALT : 62.5 / 0
ALT : 125 / 0
ALT : 250 / 0
ALT : 500 / 0
ALT : 1000 / 0
SDH : 62.5 / 0
SDH : 125 / 0
SDH : 250 / 0
SDH : 500 / 0
SDH : 1000 / 0
ALP : 62.5 / 0
ALP : 125 / 0
ALP : 250 / 0
ALP : 500 / 0
ALP : 1000 / 0
Nuc : 62.5 / 0
Nuc : 125 / 0
Nuc : 250 / 0
Nuc : 500 / 0
Nuc : 1000 / 0
TBA : 62.5 / 0
TBA : 125 / 0
TBA : 250 / 0
TBA : 500 / 0
TBA : 1000 / 0
Cholesterol : 62.5 / 0
Cholesterol : 125 / 0
Cholesterol : 250 / 0
Cholesterol : 500 / 0
Cholesterol : 1000 / 0
Triglyceride : 62.5 / 0
Triglyceride : 125 / 0
Triglyceride : 250 / 0
Triglyceride : 500 / 0
Triglyceride : 1000 / 0
Chloride : 62.5 / 0
Chloride : 125 / 0
Chloride : 250 / 0
Chloride : 500 / 0
Chloride : 1000 / 0
Sodium : 62.5 / 0
Sodium : 125 / 0
Sodium : 250 / 0
Sodium : 500 / 0
Sodium : 1000 / 0
Calc : 62.5 / 0
Calc : 125 / 0
Calc : 250 / 0
Calc : 500 / 0
Calc : 1000 / 0
PHOS : 62.5 / 0
PHOS : 125 / 0
PHOS : 250 / 0
PHOS : 500 / 0
PHOS : 1000 / 0
Potassium : 62.5 / 0
Potassium : 125 / 0
Potassium : 250 / 0
Potassium : 500 / 0
Potassium : 1000 / 0

0 2 4 6 8 10

Nonparametric ratio−to−control comparisons

Kirsten Schorning: Statistics in Toxicology II Winter semester 2024/25 43