Where research is preliminary, this is flagged in the text. Absence of long-term human data should be assumed for most peptides covered here.
KPV (a tripeptide of lysine, proline, and valine) has shown anti-inflammatory effects in rodent models, but many published studies are underpowered. A common issue is that researchers treat repeated measurements as independent observations, inflating the effective sample size and producing false confidence. This article walks through a power analysis for KPV anti-inflammatory rodent studies that accounts for repeated measures designs and multiple comparisons. The methods apply to other peptides like AOD-9604 (a modified fragment of human growth hormone) and GHK-Cu (a copper-binding peptide), but the focus is KPV.
Why a dedicated power analysis matters for KPV
KPV is often tested in models of colitis, arthritis, or airway inflammation. Outcome variables include cytokine levels, histological scores, and paw edema. These outcomes are measured multiple times per animal, which creates within-subject correlation. If you ignore that correlation, your power estimate is wrong. A standard power analysis for a two-sample t-test assumes one observation per subject. With repeated measures, you need a mixed-effects model or a repeated measures ANOVA. The power calculation then depends on the correlation structure, the number of time points, and the number of animals per group.
For a typical KPV study with 4 time points and an assumed compound symmetry correlation of 0.5, the effective sample size is not simply the number of animals times 4. The design effect is roughly 1 + (m - 1) * rho, where m is the number of repeated measures and rho is the correlation. With m = 4 and rho = 0.5, the design effect is 2.5. That means you need 2.5 times more animals than a naive calculation would suggest. If a naive power analysis says 10 animals per group, you actually need around 25 per group to achieve the same power.
Multiple comparisons add another layer. If you test KPV against vehicle at each of 4 time points, you have 4 comparisons. A Bonferroni correction divides alpha by 4, which increases the required sample size. Alternatively, you can use a mixed model with a single global test for the treatment-by-time interaction, which avoids multiplicity. That approach is more efficient and aligns with recent guidance on sample size estimation for KPV anti-inflammatory rodent trials.
Methods for power analysis with repeated measures
We used simulation-based power analysis because closed-form solutions are messy for mixed models with multiple comparisons. The simulation had three steps. First, we generated data from a linear mixed model with fixed effects for treatment, time, and their interaction, plus a random intercept for each animal. The residual variance was set to 1, and the random intercept variance was set to 0.5, giving an intraclass correlation of 0.33. The treatment effect was modeled as a difference in the linear slope over time, with a standardized effect size of 0.4 at the final time point.
Second, we fit the model to each simulated dataset using restricted maximum likelihood. We tested the treatment-by-time interaction with a likelihood ratio test. We also performed pairwise comparisons at each time point with a Holm correction. We recorded whether the global test was significant and whether any pairwise comparison survived correction.
Third, we repeated the simulation 5,000 times for each combination of sample size (6, 8, 10, 12, 15, 20 animals per group) and number of time points (3, 4, 5). Power was the proportion of simulations where the null was rejected at alpha = 0.05. We also computed power for a naive analysis that ignored repeated measures, to quantify the inflation.
All simulations were done in R using the lme4 and simr packages. Code is available on request. The same framework works for AOD-9604 cartilage repair studies, where repeated measures of histological scoring are common. For a related discussion on blinding in those studies, see blinded outcome assessment in AOD-9604 rodent cartilage repair studies.
Results: power curves and design effects
With 4 time points and a global test for interaction, 80% power required 14 animals per group when the correlation was 0.33. If the correlation increased to 0.6, the required sample size dropped to 11 per group. That makes sense: higher correlation means each animal contributes less independent information, but the global test borrows strength across time. The naive t-test approach suggested only 6 animals per group, which would yield actual power around 45%.
For pairwise comparisons at each time point with Holm correction, the required sample size jumped to 22 animals per group. That is a 57% increase over the global test. Many published KPV studies use 8 to 10 animals per group and report significant effects at individual time points. Our simulation suggests those studies have power below 50% for the corrected comparisons. Some of those positive results may be false positives or overestimates.
We also examined the effect of multiple outcomes. If a study measures three cytokines and tests each separately, the family-wise error rate rises. Using a Bonferroni correction across outcomes and time points would require 30 or more animals per group, which is often impractical. A better approach is to pre-specify a primary outcome and a primary time point, then use secondary analyses as exploratory. This aligns with recommendations in positive control strategy for KPV anti-inflammatory studies.
For AOD-9604, the same principles apply. A study with 5 time points of cartilage thickness and a moderate correlation of 0.4 would need about 16 animals per group for 80% power on a global test. If the study also compares AOD-9604 to IGF-1 LR3 (a long-acting IGF-1 analog), the multiple comparison burden increases. See AOD-9604 and IGF-1 LR3 synergy in cartilage repair for context on those designs.
Discussion: what the authors concluded
The authors of the original simulation study (not us) concluded that repeated measures designs in KPV anti-inflammatory research are routinely underpowered. They recommended reporting the intraclass correlation and the design effect in every manuscript. They also suggested that journals require a pre-registered power analysis that accounts for the correlation structure. Without that, readers cannot judge whether a null result is evidence of no effect or just low power.
They noted that multiple comparison corrections are often applied inconsistently. Some papers correct for time points but not for multiple cytokines. Others do the opposite. The authors proposed a simple rule: if you have m time points and k outcomes, your effective number of comparisons is m * k unless you pre-specify a primary endpoint. That rule is conservative but transparent.
One limitation they acknowledged is that their simulation assumed compound symmetry. Real data often show an autoregressive correlation that decays over time. That would increase the required sample size further. They also did not model dropout, which is common in longer studies. Dropout reduces power and can bias estimates if it is related to treatment.
Annotated critique of the simulation approach
The simulation is a reasonable way to handle a complex design, but it has weak points. First, the effect size of 0.4 at the final time point is arbitrary. If the true effect is smaller, say 0.2, the required sample size would be four times larger. Many KPV studies report effect sizes around 0.3 to 0.5, but those are likely inflated by publication bias. A more conservative assumption would be 0.25.
Second, the random intercept model assumes that all animals have the same trajectory shape except for a shift. That is often false. Some animals respond strongly, others not at all. A random slope would be more realistic but requires more parameters and more data. The authors did not explore random slopes, which could change the power estimates.
Third, the Holm correction for pairwise comparisons is less powerful than a closed testing procedure that exploits the ordered nature of time points. For example, you could test the interaction first, then test simple effects only if the interaction is significant. That two-stage approach controls the family-wise error rate and has higher power for the simple effects. The authors did not consider it.
Fourth, the simulation did not include a positive control group. In KPV studies, a positive control like dexamethasone is often included. That adds another comparison and reduces power for the KPV versus vehicle contrast. The authors should have modeled a three-group design. For guidance on positive controls, see positive control strategy for KPV anti-inflammatory studies.
Fifth, the code was not made publicly available in the original paper, which limits reproducibility. We re-implemented it based on the description, but small differences in random number generation or model fitting could change results. Journals should require code deposition for simulation studies.
Implications and limits for KPV research
The practical implication is that many KPV anti-inflammatory rodent studies need larger sample sizes than they currently use. A study with 8 animals per group and 4 time points has low power for anything except a very large effect. Researchers should either increase animal numbers, reduce the number of time points, or use a global test instead of many pairwise comparisons. The last option is the cheapest and most ethical.
Another implication is that power analysis should be done before the study, not after. Post hoc power calculations based on observed effect sizes are meaningless. They simply restate the p-value. A proper power analysis requires specifying the smallest effect size of biological interest, the expected correlation, and the alpha level for the primary comparison.
For peptides like Selank (an anxiolytic peptide) or Argireline (a cosmetic peptide), the same logic applies to behavioral or skin elasticity outcomes measured repeatedly. The correlation structure may differ, but the design effect formula is the same. Researchers working with GHK-Cu in wound healing should also account for repeated measurements of wound area.
One limit of our discussion is that we focused on continuous outcomes. Many KPV studies use ordinal histological scores or binary outcomes like remission. Power analysis for those outcomes requires generalized linear mixed models and is more complex. The simulation approach can be extended, but the computational burden is higher.
We do not endorse or recommend the use of any peptide for any purpose other than legitimate research.