The source and why it matters
The paper under review is Richter (2024), "Challenging current scientific practice: how a shift in research methodology could reduce animal use", published as a Comment in Lab Animal. Its author works in behavioural biology at the University of Muenster. It is a short methodological commentary, not an empirical study, and it should be read and cited as such.
The commentary starts from the 3Rs. Replacement, reduction and refinement were set out by Russell and Burch (1959) and are now written into law: Directive 2010/63/EU requires member states to ensure that the principles of replacement, reduction and refinement are applied in every procedure (European Parliament & Council of the European Union, 2010). Richter argues that progress has been uneven and that reduction in particular has not been implemented as well as it could be, because the way most experiments are planned, conducted and analysed has changed little.
Her diagnosis focuses on sample size. If too few animals are used, a study cannot detect a real effect and the animals have been used to little purpose. If too many are used, reduction is not being honoured. Richter suggests that both errors are common, and proposes a change of method rather than a change of equipment or facilities.
What Richter argues
The problem with fixing sample size in advance
In conventional practice, sample size is calculated before the experiment from an assumed effect size and variability, often taken from limited earlier work. Richter notes that when these assumptions are wrong, the real power of the study differs from the planned power. She cites evidence that much animal research is underpowered, a problem documented in detail for neuroscience by Button et al. (2013). She also argues that a fixed design gives no formal way to adjust group sizes once the study has started.
Bayesian updating and historical controls
Richter's first building block is Bayesian updating: prior information is combined with new data to give a posterior estimate. She draws on a simulation study by Bonapersona et al. (2021), who showed that using data from control groups of earlier studies as a prior could halve the minimum sample size needed to reach 80% power, or increase power for the same number of animals. Richter is careful to note the condition: the approach depends on suitable historical data being available.
Mini-experiments
The second building block is the mini-experiment design, in which one experiment is split into several smaller, identically designed replicates run with independent cohorts at different times. Richter's group had previously reported that this design improved the reproducibility of results in a single laboratory (von Kortzfleisch et al., 2020). Because mini-experiments are separated in time, they create natural points for interim analysis.
The combined proposal
Richter proposes combining the two. Data accumulate mini-experiment by mini-experiment, a Bayesian analysis is updated after each, and the study stops when enough information has been gained. She recommends that interim analyses be carried out by someone independent of the experimenters, likens the approach to adaptive designs in clinical trials (Berry, 2006), and acknowledges two costs: experiments may take longer, and researchers will need training in Bayesian methods.
Placing the proposal among the design options
Richter contrasts her proposal with fixed frequentist designs. The literature offers more options than that contrast suggests, and a doctoral researcher should know them all before choosing one.
| Design | When sample size is set | Interim looks | How errors are controlled | Main risk |
|---|---|---|---|---|
| Fixed frequentist | Before the study | None | Alpha and power fixed in advance | Wrong assumptions give an under- or overpowered study (Button et al., 2013) |
| Group sequential (frequentist) | Maximum set in advance | Pre-planned | Adjusted stopping boundaries | Requires discipline to follow the plan (Neumann et al., 2017) |
| Bayesian with historical priors | Before the study, but smaller | Optional | Depends on the prior | Bias if past controls differ from current ones (Bonapersona et al., 2021; Walley et al., 2016) |
| Sequential Bayes factor | Open, with a maximum | After each batch | Error rates estimated by simulation | Misleading evidence if thresholds are poorly chosen (Schönbrodt et al., 2017) |
| Richter's combined design | Emerges during the study | After each mini-experiment | Not specified in the commentary | Operating characteristics untested in animal studies |
Strengths of the commentary
- It treats reduction as a design problem. Linking an ethical principle to concrete statistical choices gives researchers something they can act on, rather than an exhortation to "use fewer animals".
- It builds on evidence. The two building blocks are supported by a published simulation study (Bonapersona et al., 2021) and a published experimental study (von Kortzfleisch et al., 2020), not only by argument.
- It is realistic about costs. Richter names longer timelines and the need for training, which many advocacy pieces omit.
- It protects against bias. Recommending that interim analyses be done by someone independent of the experimenters is sound practice and mirrors the separation used in clinical trials.
Limitations and methodological questions
The contrast with frequentist methods is too sharp
Frequentist statistics do not forbid interim analyses. Group sequential designs, standard in clinical trials, set a maximum sample size and pre-planned interim looks with stopping boundaries that keep the overall false-positive rate at the chosen level. Neumann et al. (2017) examined these designs for preclinical research and concluded that they can make animal studies more efficient. A fair comparison would set Richter's proposal against group sequential designs, not only against fixed ones.
Bayesian does not mean free to stop at will
Richter reports that Bayesian methods are less vulnerable to false-positive inflation from repeated interim inspection. That claim needs qualifying. de Heide and Grünwald (2021) show that optional stopping can be a problem for Bayesian analyses that use default or pragmatic priors, which is how most applied analyses are done. Schönbrodt et al. (2017) recommend that sequential Bayes factor designs be planned with simulations of how often they yield misleading evidence and how many observations they need. In clinical research, the FDA's January 2026 draft guidance on Bayesian methods asks sponsors to justify success criteria and priors and to evaluate operating characteristics (U.S. Food and Drug Administration, 2026). Animal studies deserve the same discipline.
Historical data are the weak link
Borrowing from past controls works when those controls resemble the current ones. Animal results vary with batch, season, housing, supplier and laboratory (Voelkl et al., 2020), so control data from even a few years earlier may not be exchangeable with today's. Walley et al. (2016) show the benefits of Bayesian borrowing in regularly repeated, stable studies, which is precisely the setting in which exchangeability is most plausible. Many doctoral projects are one-off studies with no such history.
The commentary leaves the design unspecified
The commentary does not say how to choose the prior, what stopping rule to apply, what maximum number of mini-experiments is acceptable, or how to account for batch in the analysis. These are not small details: each changes the error rates and the number of animals used. Mini-experiments introduce a time or batch factor that must be included in the statistical model, otherwise variation between batches is mistaken for noise or for a treatment effect.
Regulation and ethics committees
Under Directive 2010/63/EU, project evaluation includes an assessment of whether the 3Rs, including reduction, have been applied (European Parliament & Council of the European Union, 2010). An open-ended design is easier to authorise when the application states a maximum number of animals, pre-specified stopping rules and who will perform interim analyses. ARRIVE 2.0 asks authors to explain how the sample size was decided (Percie du Sert et al., 2020); an adaptive design must therefore be documented from the outset so that it can be reported honestly.
The wider context in 2026
Richter's argument sits within a broader move away from unexamined animal use. In April 2025 the FDA published a roadmap aiming to make animal studies the exception rather than the norm for preclinical safety testing, starting with monoclonal antibodies (U.S. Food and Drug Administration, 2025). In November 2025 the UK government published a strategy to support the development, validation and uptake of alternative methods, which includes an aim to offer 3Rs training to PhD students and early career scientists (Department for Science, Innovation and Technology et al., 2025). These initiatives emphasise replacement, but reduction remains essential wherever animals are still needed, and better design is the most direct route to it.
There is also a reproducibility argument. Splitting a study into mini-experiments deliberately samples variation across time, which can make a result more likely to hold in another laboratory (von Kortzfleisch et al., 2020; Voelkl et al., 2020). This benefit does not depend on Bayesian analysis at all, and it may be the most immediately usable part of Richter's proposal.
Worked example: planning an adaptive mini-experiment study
The scenario below is hypothetical and purely illustrative; it shows the decisions to make, not results. Suppose a doctoral researcher plans to compare a treatment with a vehicle on a behavioural outcome in mice.
- Fix the maximum. Decide the largest number of mini-experiments (for example three) and the group size in each, so the maximum number of animals is known for the project application.
- Define the smallest meaningful effect on biological grounds, before any data are seen.
- Choose and justify the prior. If historical controls are used, document their source, age and conditions and plan a check for conflict between prior and new data (Bonapersona et al., 2021). Otherwise use a weakly informative prior and say so.
- Write the stopping rules for efficacy and futility, stated as posterior probabilities or Bayes factor thresholds.
- Simulate the design under no effect and under the smallest meaningful effect, and report the false-positive rate, power and expected number of animals (Schönbrodt et al., 2017). Compare with a group sequential design (Neumann et al., 2017).
- Assign an independent analyst to run interim analyses, so experimenters remain blind to accumulating results.
- Include mini-experiment as a factor in the final model and report it.
If the simulations show that the adaptive design saves animals while keeping error rates acceptable, the case for it is strong and can be put to the ethics committee. If they do not, a fixed or group sequential design may be the better ethical choice. Either way, the simulation itself is a valuable thesis chapter.
What to report
| Item | What to state |
|---|---|
| Design type | Fixed, group sequential or Bayesian sequential, and why it was chosen |
| Maximum sample size | Per group and in total, as authorised |
| Interim analyses | Number, timing and who performed them |
| Stopping rules | Thresholds for efficacy and futility, fixed in advance |
| Prior | Distribution, source of any historical data, and sensitivity analysis |
| Operating characteristics | Simulated error rates, power and expected animal numbers |
| Batch structure | How mini-experiments were scheduled and modelled |
| Actual numbers used | Animals used and why the study stopped (Percie du Sert et al., 2020) |
Common mistakes and what this means for your PhD
- Citing Richter (2024) as evidence that Bayesian designs reduce animal use, when it is a proposal that still needs testing.
- Looking at the data after each batch without pre-specified rules and calling the design "Bayesian".
- Borrowing historical controls from different strains, ages, suppliers or laboratories without checking exchangeability.
- Ignoring batch or mini-experiment as a factor in the analysis.
- Failing to state the maximum number of animals in the project application.
For a doctoral researcher, the commentary is best read as a well-defined research question: under what conditions does a Bayesian mini-experiment design use fewer animals than a fixed or group sequential design, without inflating error rates? That question can be answered by simulation, by re-analysis of existing data sets, or prospectively in a pilot programme, and any of these would make a substantial contribution.
Key takeaways
- Richter (2024) is a methodological commentary proposing Bayesian updating combined with mini-experiments to make reduction a design principle.
- Historical controls can cut sample sizes in simulation, but only when past and present controls are comparable (Bonapersona et al., 2021; Voelkl et al., 2020).
- Frequentist group sequential designs also allow interim stopping and should be the comparison (Neumann et al., 2017).
- Bayesian sequential designs need simulated operating characteristics, as regulators now expect in clinical work (Schönbrodt et al., 2017; U.S. Food and Drug Administration, 2026).
- Document the maximum sample size, stopping rules and priors in advance so the study can be authorised and reported (European Parliament & Council of the European Union, 2010; Percie du Sert et al., 2020).
