Data Upload & Configuration

Try Demo Data

Load a built-in demo dataset to explore the dashboard's features.

Data must be in long format — one row per observation/measurement.

Subset Filter Exclude mice, treatments, or cages from all downstream analyses.

Edit the lists above, then click Apply Filtering . The Dataset Summary and Data Preview below update to reflect what's currently applied. Filters apply to every analysis until you clear them.

Settings preset (column mapping + methods)

Save your column mappings, time / volume / formula settings, study design and subset filter to a small JSON file. Load it after uploading a dataset with the same column names to skip these steps.

Study report

Every analysis you have run in this session, with its settings, tables, plots and diagnostics, and the data check, in one HTML file to print or keep. Each tab's Download Results has its own report too.

Download the study report (HTML)

Tumor Growth Analysis

Configure analysis parameters and run statistical modeling. Ensure data is uploaded first from the Data Upload tab.


Overview

The Tumor Growth tab provides three statistical modeling approaches selectable from the Statistical Model dropdown:

  • Linear Mixed-Effects Model (LME4) — fits a mixed-effects model to repeated tumor volume measurements. Accounts for inter-animal variability and unequal observation times. Recommended for most longitudinal datasets.
  • Area Under Curve (AUC) — compares each arm’s total tumor burden as the area under its fitted growth curve, over the window in which every arm still has enough animals on study.
  • Bayesian LMM (brms) — the LME4 model fitted by MCMC, returning full posterior distributions instead of point estimates and p-values.

R14.3: this tab does not report Tumor Growth Inhibition. TGI is an endpoint summary — a ratio of two group means on one day — while every model here estimates a growth trajectory , and collapsing a trajectory to a single-day ratio discards most of what was measured. TGI is computed where it belongs to the question being asked: the Toxicity tab (as the efficacy half of the therapeutic window) and the Drug Synergy tab (as the fractional effect the Bliss model acts on). Both expose an endpoint-estimand selector, because a ratio of endpoint means is exactly where early removals bias the answer.

Which endpoint? Animals removed before study end make “the final timepoint” ambiguous, and taking it literally computes TGI among survivors only — which flatters the control arm, because the fastest-growing controls are exactly the ones that left. The Toxicity and Synergy tabs expose an Endpoint estimand selector for this; here TGI is derived from the fitted model, which uses every observation each animal contributed rather than only those present at the end.

LME4 Mode — Analyses & Outputs

Model

A linear mixed-effects model is fitted to transformed tumor volumes:

log(Volume) ~ Treatment × Day + (Day | Animal)

The Treatment × Day interaction captures whether treatment groups differ in their growth trajectories over time, not just at a single endpoint.

Transformation
Log (default) stabilises variance and linearises exponential growth. Square root suits moderate heteroscedasticity. Use 'None' only for already-linear data.
Random Effects
'Random slopes' (default) gives each animal its own baseline volume and its own growth rate. If that model cannot be fitted, the analysis falls back to uncorrelated slopes and then to an intercept only, and the Summary says so. 'Intercept only' assumes every animal in an arm grows at the same rate, which makes the treatment × time test find differences that are not there. 'None' drops the animal-level term entirely and is appropriate only when each animal contributes a single observation.
Cage Effect
Auto (default) classifies the housing design and applies the matching treatment, because the correct handling depends on the design rather than on preference:
  • Crossed — a cage holds more than one treatment. Cage is separable from treatment and enters as a fixed effect.
  • Nested, replicated — one treatment per cage, several cages per arm. Cage nests within treatment as a random effect; ignoring it treats cagemates as independent and understates standard errors.
  • Nested, partly replicated — some arms have several cages and others one. Cage stays in the model as a random effect, estimated from the replicated arms, and the Summary names the single-cage arms.
  • Nested, confounded — one cage per treatment. Cage and treatment are the same variable and no model can separate them. The result reports the effect as inseparable rather than silently crediting it to treatment.
The remaining options force a specific handling. Include if not collinear (legacy) is the pre-v0.12 behaviour, kept so earlier analyses can be reproduced.
Fixed and random effects
Fixed effects
What the study is about, estimated once for the whole study and reported as results: the treatment, time, their interaction (whether the arms change at different rates), and any covariate. The comparisons and their p-values or credible intervals come from these.
Random effects
How each animal departs from its arm's average: a random intercept for its own starting level, a random slope for its own rate. The model estimates how much animals vary (an SD) rather than reporting each one. A cage random effect does the same for cages, whose animals resemble each other.
Why both
Repeated measurements of one animal are not independent. Treating them as independent overstates the evidence and gives p-values that are too small; random effects account for it, so the treatment comparison uses the right amount of information. When animals differ in their rates, leaving out the random slope finds treatment differences that are not there.

ANOVA Tab

F-tests with Satterthwaite degrees of freedom, reported as NumDF / DenDF / F value / Pr(>F) , in this order: whether the arms' growth rates differ (treatment × time, the primary analysis target); whether the arms differ at the mean study day; and whether there is growth on average. A Tests column states each question.

The treatment row compares the arms at the mean study day, not at day 0, so it does not change when time is counted from implantation instead of randomisation. With few animals per arm, the randomisation test below makes fewer assumptions.


Comparisons Tab

Pairwise group comparisons at the mean observation day using estimated marginal means (EMMs). Two independent settings control this, and conflating them is the usual mistake:

Post-hoc Comparisons
Which comparisons are made — vs Reference Group (each treatment against the reference only) or All Comparisons (every pair). This defines the family.
Correction
How that family is corrected: Dunnett (exact, multivariate-t), Tukey HSD, Bonferroni, Holm, Benjamini-Hochberg FDR, or None. Dunnett is the default and is the right partner for a vs-reference family — it accounts for the correlation between comparisons that share a control arm, which Bonferroni ignores and therefore over-corrects.

The correction is applied over exactly the set of comparisons requested, and the table header states which correction was actually used rather than asserting one. Comparisons are marginalised at the mean study day; for differences in trajectory across time, see the Treatment Effects tab.


Randomisation Test

An optional exact test of the Treatment × Day interaction that avoids both the denominator-degrees-of-freedom approximation and the assumption of normally distributed random effects. Treatment labels are permuted and the model refitted; the p-value is the proportion of permutations giving a statistic at least as extreme as the observed one.

You must state the unit that was randomised. If animals were assigned to arms individually, permute animals. If whole cages were assigned, permute cages — permuting animals then breaks the exchangeability the test relies on and yields a p-value that is too small. This cannot be inferred from the data; it is a fact about how the experiment was run.

The panel reports the number of exchangeable units, how many permutations converged, whether the permutation set was enumerated exhaustively, and the smallest p-value the design can attain. With few units that floor can exceed 0.05, in which case no amount of signal can produce a significant result — and the asymptotic p-value beside it should not be trusted.


Treatment Effects Tab

Estimated marginal means (EMMs) of transformed volume per treatment group, at the mean observation day: the day the treatment row of the ANOVA compares the arms at. How the arms diverge over time is in the growth-rate test (ANOVA) and the Growth Rates tab.


Growth Rates Tab

Per-animal exponential growth rates estimated by log-linear regression of each animal’s volume trajectory independently. Reported as growth rate (log-scale slope), fold change, and % change per day with 95 % CIs and R². These are descriptive; for group-level inference use the ANOVA and Comparisons tabs.


Diagnostics Tab

Four residual plots, each answering a different question about whether the model's assumptions hold — which is what makes the p-values above mean anything:

Q–Q plot
Are residuals normally distributed? Systematic curvature at the tails suggests the transformation choice should be revisited.
Residuals vs fitted
Is the mean structure right? Curvature means log-linear growth does not hold over the whole study — read growth-rate comparisons with care.
Scale–location
Is the variance constant on the modelling scale? A rising trend means heteroscedasticity survived the transformation.
Random-effects Q–Q
Are the per-animal effects themselves normal? This is a separate assumption from residual normality and the one most often overlooked. If it fails, prefer the randomisation test, which assumes neither.

Influence measures (Cook's distance, DFBETAS) are also available: they identify individual animals whose removal would materially change the conclusion. A result resting on one mouse is worth knowing about before publication rather than after.

AUC Mode — Analyses & Outputs

Each arm's AUC is the area under its fitted growth curve, the geometric-mean curve of a mixed model of log volume with a smooth (natural spline) time course per arm and per-animal random slopes. It is integrated over one window common to every arm: from the first study day to the last day on which every arm is evaluable (at least 50 % of its enrolled animals, and at least 3, still on study). AUC is in volume × days.

AUC by arm
Each arm's AUC with a 95 % interval, from draws of the model.
Comparisons
The ratio of two arms' AUCs and their difference, with 95 % intervals and a Wald p-value from the same draws, corrected over the family selected in the sidebar. A ratio below 1 means less tumor burden than the comparator.
ANOVA
A Wald test that every arm has the same AUC (on the log scale).
Per-animal AUC
Each animal's own trapezoidal AUC, for description only.
Animals that die or are removed early still count. Every measurement an animal had goes into the model, which estimates its growth rate from them and keeps it in the arm's curve as if the animal had continued. So the AUC is not pulled down when the largest tumours leave the study, as an average of the animals still on study would be. In simulations with 10 animals per arm and about a third of the treated arm lost, removal at the volume limit moved the AUC by −0.2 % (the average of the survivors: −13 %). Deaths that the earlier measurements only partly foreshadow left a small underestimate, −3 % to −5 % (survivors: −18 % to −21 %). When many animals die for reasons their growth does not show, read the Survival analysis alongside the AUC.
Why not each animal's own AUC? An animal euthanised on day 14 has a smaller trapezoidal AUC than one followed to day 28 because it was measured for less time, not because its tumour was smaller. Controls leave first, so comparing per-animal AUCs made an effective drug look worse than control in about half of simulated studies. The per-animal values are shown under ‘AUC Individual Data’ for description, not compared.

Bayesian Mode — Method & Outputs

Method

The Bayesian LMM fits the same formula as LME4 using full Bayesian inference via brms / Stan. Instead of point estimates and p-values, every parameter has a posterior distribution : a probability distribution over plausible values given the data and priors.

log(Volume) ~ Treatment × Day + (Day | Animal)

Results are summarised as posterior means with 95 % credible intervals. A 95 % CrI has a direct probability interpretation: there is a 95 % posterior probability that the true parameter lies within the interval — unlike a frequentist confidence interval, which is a statement about the procedure rather than the parameter.

R14.7: two different interval types appear under the same Lower_95_CrI / Upper_95_CrI headings, and it is worth knowing which you are reading. The Summary tab reports raw model parameters straight from brms: an equal-tailed interval, cutting 2.5 % from each tail. The Treatment Effects and Comparisons tabs come from emmeans, which returns a genuine highest-density interval — the shortest interval containing 95 % of the posterior. The two coincide when a posterior is symmetric and diverge when it is skewed, which is exactly the case for variance components and for the back-transformed time ratios in the Bayesian survival model.


Prior Strength

Priors express what you believe about parameter values before seeing the data. They regularise estimates and prevent the model from chasing noise in small samples. All presets use Normal priors on fixed effects and Exponential priors on variance components; only the widths differ.

Skeptical
Default. N(0, 0.25) on fixed effects; Exponential(2) on SDs. Encodes the reasonable belief that treatment effects are small — effects > ±0.5 log-units per day are considered very unlikely a priori. Recommended for exploratory preclinical studies where large effects have not been previously established.
Weakly informative
N(0, 1) on fixed effects; Exponential(1) on SDs. Broader than Skeptical — appropriate when prior work suggests the treatment category can produce moderate-to-large effects.
Informative
N(0, 0.5) on fixed effects; Exponential(2) on SDs. Between Skeptical and Weakly informative. Useful when effect sizes in the 0.1–0.5 log-unit range are plausible based on prior data.
Diffuse
N(0, 2.5) on fixed effects; Exponential(0.5) on SDs. Near-flat priors; use with caution on small datasets as estimates can be driven entirely by noisy data.
Manual
Specify the prior family and parameters for each model class independently. Fixed effects (b): Normal or Student-t. Intercept: same. Random-effect SD and residual sigma: Exponential or half-Normal.

Result Tabs
  • Summary — posterior median and 95 % CrI for every model parameter.
  • Treatment Effects — group-level estimated marginal means (posterior) at the mean study day.
  • Comparisons — posterior contrasts of each group vs the reference.
  • Posterior Summary — every model parameter's posterior estimate, 95 % credible interval, R-hat and effective sample sizes. It replaces the ANOVA tab for a Bayesian fit.
  • Diagnostics: posterior predictive check — overlay of observed data density vs draws from the posterior predictive distribution. Systematic mismatch indicates model mis-specification. A coverage plot reports whether the stated credible intervals actually contain the observed values at the claimed rate.
  • Diagnostics (Bayesian view) — Rhat and effective sample size (ESS) per parameter, with trace and rank plots. Rhat > 1.01 indicates the chains have not mixed; low ESS means the posterior summaries are themselves noisy. Rank plots detect the mixing failures that trace plots hide once chains overlap.
  • Influence — Pareto-k values from PSIS leave-one-out. Values above the threshold (0.7 at most; lower with fewer posterior draws, e.g. 0.667 at 1,000) mean the importance sampling is unreliable for that observation, which usually flags an animal the model fits badly rather than a coding error.
  • Growth Rates — per-animal log-linear regression rates (same as LME4, not MCMC-based).
Computation time: the first fit of a model compiles its Stan code (about two minutes on mouseexperiment.com, less on a fast computer); a later fit whose model has the same structure (same number of arms, random effects and prior preset) reuses the compiled model and starts sampling at once; on mouseexperiment.com compiled models are kept across restarts and updates of the server. Fits run one at a time on a background worker, so the app stays responsive while a model runs; the progress window shows how many fits are ahead of yours.

Necrotic Tumors — Handling Options

When a Necrotic Flag column is mapped in Data Upload, a handling selector appears in the sidebar. Necrotic tumors have a non-viable core that inflates caliper-based volume estimates without contributing to proliferating tissue.

Exclude (recommended)
Removes flagged observations from the model input. Conservative and publication-defensible. The Necrosis tab reports how many observations were flagged and which groups were affected, regardless of handling mode.
Covariate
Adds a 0/1 .necrotic_flag fixed effect to the LMM formula. Estimates the average volume bias attributable to necrosis. Useful when necrosis is approximately evenly distributed across groups and you want to retain all timepoints.
No adjustment
All observations are used as-is (current default when no necrotic column is mapped). The Necrosis summary table is still shown for documentation purposes.
The Necrosis tab in the results panel is inserted automatically when a necrotic column is mapped and an analysis has been run.

Shared Parameters

Reference Group
The study's control group, chosen under Study Design on Data Upload. With a dose column mapped, arms are labelled 'Treatment Dose' and the control maps to its own dose label.
Post-hoc Comparisons
Which family of comparisons to make — vs Reference Group (default) or All Comparisons. Paired with the separate Correction setting.
Correction
How the chosen family is corrected: Dunnett (exact multivariate-t, default), Tukey HSD, Bonferroni, Holm, Benjamini-Hochberg FDR, or None.
Extrapolation Points
Affects the growth-trajectory plots only (Plots tab, under Show extrapolated curves/points). Number of recent timepoints used to estimate a slope when projecting the curve for animals that left the study early; projected segments are drawn dashed. Set to 0 to disable. This no longer feeds any statistical model — treating an imputed point as measured data made standard errors ignore the imputation.
Time Column
Numeric study day, or calculated from a date column mapped in Data Upload (days elapsed from first observation).

Loading...
Download ANOVA
Posterior summary: each model parameter's posterior distribution, summarised (coefficients on the log tumour volume scale) .
  • Parameter / Term — the model coefficient, and what it measures in the study's own terms.
  • Estimate — the posterior mean: the parameter's most representative value given the data and the priors.
  • Est.Error — the posterior standard deviation: how uncertain that estimate is.
  • Lower_95_CrI / Upper_95_CrI — the 95 % credible interval: given the model and the priors, the parameter lies in it with 95 % probability. An interval that excludes 0 is a clear direction for that effect.
  • Rhat — whether the chains agree. Near 1.00 means they have converged on the same distribution; above 1.01 they have not, and the row should not be reported until a longer run brings it down.
  • Bulk_ESS — the effective number of independent draws behind the estimate. Above about 400 the estimate is stable; below, increase the draws.
  • Tail_ESS — the same for the ends of the interval, which need more draws than the middle. Above about 400 the interval's limits are stable.
Loading...
Per-animal exponential growth rates estimated from each subject’s individual tumor volume trajectory using linear regression on log₁ₚ(volume) vs time. Requires ≥3 time points per animal. All rate, fold change, and percent change values are on a per-day basis.
  • Treatment — treatment group.
  • ID — individual animal identifier.
  • Cage — cage identifier.
  • Growth Rate — slope of the log₁ₚ(volume)–time regression (units: log-volume per day). Represents the instantaneous exponential growth constant: a value of 0.10 means ~10% volume increase per day on a continuous compounding basis.
  • GR 95% CI Lower/Upper Bound — 95% confidence interval for the log-scale growth rate.
  • Fold Change — exp(Growth Rate): multiplicative volume change per day. A value of 1.10 means the tumor is 10% larger each day; values below 1.0 indicate regression.
  • FC 95% CI Lower/Upper Bound — exp(GR CI bounds): back-transformed confidence interval for the daily fold change.
  • Percent Change — (Fold Change − 1) × 100: daily volume change expressed as a percentage (e.g. +10.5%/day or −3.2%/day).
  • PC 95% CI Lower/Upper Bound — (exp(GR CI) − 1) × 100: back-transformed CI bounds for the percent change.
  • R² — coefficient of determination from the per-animal log-linear regression. Values close to 1.0 indicate a good exponential fit; lower values suggest irregular growth or too few time points.
These are descriptive per-animal values used to visualise within-group spread and for sample size estimation. For group-level inference, use the Treatment Effects tab.
Loading...
Download Growth Rates

Per-mouse drill-down Click rows in the table above to see those mice's trajectories; click again to remove one.
Loading...
Growth delay (T − C), from the data this run analysed, without the model: the Kaplan-Meier median time for an arm's tumours to reach the volume below, minus the reference arm's. Each animal's time is interpolated on the log scale between measurements; an animal that never reached the volume is censored at its last measurement. "Not reached" means fewer than half the arm's animals reached it.
  • Delay 95% interval — a bootstrap over animals within each arm.
  • Specific growth delay — T − C divided by the reference arm's volume doubling time: the delay in doublings, comparable across tumour models that grow at different rates.
  • p (log-rank) — each arm's times against the reference arm's, corrected over the arms by the method shown. It treats animals as independent, so it overstates the evidence when whole cages were randomised.

Each animal
Descriptive statistics for tumor volume by treatment group and timepoint, on the volume scale in the declared units, whatever transform the model used.
  • Treatment — treatment group.
  • Day — study day (measurement timepoint).
  • N — number of animals with a volume measured on that day.
  • Mean — arithmetic mean tumor volume for the group at that timepoint.
  • SD — standard deviation of volume across animals in the group.
  • SEM — standard error of the mean (SD / √N).
  • Geometric mean (log transform only) — the mean on the log scale the model uses, back-transformed; zeros count at the floor stated in the Summary. Tumour volumes are right-skewed, so it is usually below the arithmetic mean.
Useful for inspecting raw descriptive trends across time and confirming data entry. Statistical comparisons should use the ANOVA, Comparisons, or AUC Summary tabs as appropriate.
Loading...
Download Data Summary

Back-transforms log-scale model output to original units. Growth rates and contrasts become daily fold changes (exp(value)); forest plot uses a log10 axis centred on 1. Inference is still performed on the log scale.



Tumor Growth Curves

Loading...

Growth Rates Plot

Loading...

Area Under the Curve Plot

Loading...

Waterfall Plot

Loading...

Effect Size Forest Plot

Loading...

Posterior Distributions

Loading...
Model-fit diagnostics. Frequentist fits: residual diagnostics. Bayesian fits: MCMC convergence + posterior diagnostics. The view switches based on the selected model type.

Q-Q Plot of Residuals

How to read this plot
The residuals are standardised (centred and divided by their SD), so for a well-fitting model the points fall along the red dashed line. S-shape: residuals have heavier tails than normal — consider a robust likelihood or check for outliers. Bow: skewed residuals — check the volume transform (log vs sqrt vs none). A few outliers at the ends: common; not a problem unless they're clinically obvious errors.
Loading...

Residuals vs Fitted

How to read this plot
Points should scatter randomly around the red zero line. Curved blue trend: the model misses non-linearity — growth is not log-linear over the whole study; try a different volume transform. Fan shape (wider on the right): heteroscedasticity — variance grows with the mean. Switch to log transform or a model with variance structure.
Loading...

Scale-Location

How to read this plot
Plots sqrt(|residuals|) vs fitted to magnify variance trends. Flat blue line: constant variance — LMM assumption satisfied. Upward slope: heteroscedasticity — see Residuals vs Fitted notes above.
Loading...

Random-effects Q-Q

How to read this plot
Per-animal random intercepts (and slopes, if used) should be normally distributed. The values are standardised, so points along the red line = assumption met. A component with no spread (estimated variance zero) is left out. Heavy tails / clear S-shape: the LMM may be giving a few animals too much / too little shrinkage. Rare and usually safe to ignore at N ≤ 30.
Loading...

Influence diagnostics (case-deletion)

Cook's distance measures the overall effect of leaving out each animal, one row per animal (arm, ID, cage). Values above the threshold (4 / number of animals) flag animals that move the fit. Inspect them individually — they are not automatically wrong, but they are driving the result. DFBETAS breaks the influence down per fixed effect: each column is one coefficient, each row one animal. Values | > 1 | are conventionally flagged.
Cook's distance
Loading...
DFBETAS
Loading...
MCMC convergence diagnostics for all fixed-effect parameters. All chains should have converged (R̂ ≈ 1.00) and sufficient effective samples before interpreting results.
  • Parameter — model parameter name.
  • Rhat — potential scale-reduction factor; values > 1.01 indicate non-convergence.
  • Bulk_ESS — bulk effective sample size; aim for > 400 per chain.
  • Tail_ESS — tail effective sample size; important for credible interval accuracy.
  • Converged — TRUE when R̂ ≤ 1.01.
If Rhat > 1.01, try increasing iterations or chains before reporting results.
What these diagnostics mean
Divergent transitions
The sampler met a region of the posterior it could not follow accurately. Even a few can bias the estimates. Run more warmup iterations, choose a stronger prior preset, or simplify the random effects, and refit.
Max-treedepth hits
The sampler reached its step limit before finishing a trajectory. This is about efficiency, not validity: the draws are still correct, but the fit was slow and the effective sample sizes may be low. Many hits often point to a weakly identified parameter, such as a random-effect SD the data say little about.
E-BFMI
How well the sampler moved between the posterior's typical values and its tails. Below 0.3 it explored the tails poorly, and variance parameters and the ends of credible intervals may be unreliable. The remedies are those for divergences.
Bayes R²
The share of the variation in the modelled values that the model explains, with its credible interval. It includes the random effects, so differences between animals count as explained. It describes fit; it is not a test of the treatment.
PPC coverage
How often the observed values fall inside the model's 50 %, 80 % and 95 % predictive intervals. A well-calibrated model gives about 50 %, 80 % and 95 %. Much lower means its intervals are too narrow (overconfident); much higher, too wide.
PSIS-LOO (elpd_loo)
Leave-one-out cross-validation, estimated without refitting: how well the model predicts each observation when that observation is left out. elpd_loo means little on its own; it compares models fitted to the same data, where higher is better and a difference smaller than about twice its SE is not meaningful.
Pareto-k
For each observation, how much leaving it out would change the posterior. Above the threshold (about 0.7) the leave-one-out estimate for that observation is unreliable, and the observation is influential: usually a measurement the model fits badly, such as a data-entry error or an animal whose growth departs from its arm's. Check those rows in the table below against the raw data. The estimates are not invalidated, but they lean on those observations.
Loading...

Influential observations (Pareto-k > 0.7)
Measurements that move the fit too much for the leave-one-out estimate, with their animal. Check them against the raw data; 'What these diagnostics mean' above explains Pareto-k.
Loading...

Posterior predictive check
How to read this plot
The model simulates new datasets from the posterior and overlays their densities (y_rep, 50 draws) on the observed, transformed volumes (y). The observed density inside the simulated envelope: the model reproduces the data. Outside it: consider a different transformation or random-effects specification.
Loading...
Download Plot

Pareto-k distribution (all observations)
How to read this plot
Histogram of LOO Pareto-k values across all observations. All bars left of 0.5 (orange dotted): LOO is reliable. Some bars between 0.5 and 0.7: moderate influence; estimates usable but inspect. Bars right of 0.7 (red dashed): those observations are too influential for PSIS-LOO to estimate reliably — see the table above for which.
Loading...
Posterior predictive coverage
How to read this plot
Each bar = empirical coverage of the corresponding nominal posterior-predictive interval. Bars near red dashed line: well-calibrated. Bars below: under-coverage — model intervals are too narrow. Bars above: over-coverage — too wide; model is being conservative.
Loading...

MCMC trace plots
How to read this plot
Each line is one chain's draws of a parameter, after warmup. Converged chains overlap and look like stationary noise around one level (a ‘fuzzy caterpillar’). A drift, a trend, or a chain at a different level: the chains have not converged and the estimates are unreliable; run more warmup iterations or simplify the model.
Loading...

MCMC rank plots
How to read these plots
Modern convergence diagnostic. Each chain's draws are rank-normalised; chains that have converged look like uniform draws (flat histograms) and resemble each other. Distinctly non-uniform shapes or chain-specific patterns: the chain hasn't mixed and the posterior estimate is unreliable.
Loading...

Treatment-parameter posterior densities
How to read this plot
Posterior density (filled area) for each treatment coefficient. Look for: (a) modes shifted from the prior (the data informed the parameter), (b) credible-interval bars that don't include zero (significant posterior direction), (c) heavy tails / multiple modes (potential identifiability issue).
Loading...
Bayesian residuals vs day
How to read this plot
Posterior-mean residuals plotted against study day, faceted by treatment, with a loess smoother. Flat horizontal line at zero per group = the linear-time assumption is satisfied. Systematic curvature: the log-linear time assumption is violated; read growth-rate comparisons with care.
Loading...

Choose what to download. It all goes into one ZIP file, with a README.txt that lists every file, the settings of the run and the package versions.

Download selected (ZIP)

Survival Analysis (Time-to-Endpoint)


Note on ‘survival’ in mouse tumour experiments: Most IACUC protocols require euthanasia once tumour volume exceeds a set limit (e.g. 2000 mm³). The recorded endpoint is therefore time until tumour volume threshold , not true death. This is mathematically identical to survival analysis, but results should be labelled as ‘Time-to-Endpoint’ or ‘Time-to-Tumour-Volume-Threshold’ rather than ‘survival’ in publications.
Occasionally a mouse dies before reaching the volume limit. Competing-risks models (e.g. Fine–Gray) are the rigorous solution, but they require large per-group sample sizes rarely achievable in mouse experiments. The recommended pragmatic approach is to treat any death as an event (‘endpoint reached’), because the animal did not survive the study regardless of cause. Map a removal-reason column on Data Upload and the sidebar offers exactly that. Without one, with an Event/Censor column events are exactly the rows coded 1, so code such deaths 1 to follow this approach; when events are derived from a volume threshold, only a tumour reaching the limit is an event and other removals are censored. The sidebar says which applies.

Overview

The Survival tab fits Kaplan–Meier (KM) survival curves and a Cox proportional hazards model to time-to-endpoint data. In mouse tumor experiments the 'endpoint' is typically the day a tumor exceeds the IACUC volume limit, not true animal death.

Three outputs are produced:

  • Kaplan–Meier curves , with the log-rank test across all arms (not adjusted for multiple comparisons).
  • Cox proportional hazards model reporting hazard ratios (HR) per group relative to the reference.
  • Risk table showing animals at risk, events, and KM estimates at each event time.

Kaplan–Meier Estimator

The KM estimator computes a non-parametric step-function survival probability S(t) = P(T > t) from event times. Each downward step corresponds to one or more events on that day.

Log-rank test
Tests whether the survival curves of two or more groups are the same. The plot shows the test across all arms at once, unadjusted. The comparisons of each arm with the reference are in Model Results: Cox (or Firth) p-values, adjusted for the number of comparisons; when the Cox model cannot be fitted, they come from pairwise log-rank tests.
Median survival
Day by which 50 % of animals reached the endpoint. 'Not reached' if fewer than half had events before study end.
Censoring
Animals that never reach the endpoint contribute their last observation day as a censored time. This correctly uses partial information rather than excluding them.

When an arm has no events, its hazard ratio is estimated with Firth's correction, which stays finite. If that fit fails, the comparisons come from pairwise log-rank tests.

Cox Proportional Hazards Model

The Cox model estimates the hazard ratio (HR) for each treatment group relative to the reference:

h(t) = h₀(t) × exp(β₁ × Treatment)
HR < 1
Lower hazard than reference — animals reach the endpoint more slowly (better outcome).
HR = 1
Identical hazard to reference.
HR > 1
Higher hazard than reference — animals reach the endpoint faster (worse outcome).

Firth's bias-reduced estimation is applied automatically when any group has zero events. Standard Cox partial likelihood produces infinite coefficient estimates in this case; Firth's penalisation regularises them.

The proportional hazards assumption requires that the HR is constant over time. Crossing KM curves indicate a time-varying HR, which violates this assumption — interpret Cox results with caution in that case.

Endpoint Configuration

Mapped censor column
When an Event/Censor column is mapped in Data Upload (0 = censored, 1 = event), it is used directly.
No censor column
Without a mapped Event/Censor column, choose 'Derive from tumour volume threshold' as the event source. An animal's disappearance is not treated as an event by itself.
Volume threshold
Optionally generate an event indicator by flagging animals whose tumor volume exceeds a specified threshold at any timepoint.
Natural death
With a removal-reason column mapped on Data Upload, the sidebar offers to count deaths from other causes as events (recommended): an animal whose reason is marked 'Death or moribund' under Study Design is an event at its last measurement, with either event source. Without one, or with the box unticked: Event/Censor column, an animal that died of other causes is an event only if its row is coded 1; volume threshold, it is censored at its last measurement, which assumes its later risk was like that of the animals still on study; if such deaths are treatment-related, this favours the treatment.
Event time
The first day an animal's row is flagged 1 (or its tumour reaches the limit). An animal whose every row is flagged has an animal-level status, and its last day is used. Animals are identified by treatment, ID and cage, so an ear tag reused in another arm is a different animal.

Time-to-Endpoint summary: Per-group event counts, total subjects, event rate, and median time-to-endpoint (the day by which 50 % of animals reached the endpoint — typically tumour volume exceeding the IACUC limit). ‘Not reached’ means fewer than half the group experienced an event by study end.
Loading...

                      
Accelerated failure time model. Each arm's time to the endpoint compared with the reference's, as a time ratio : 1.8 means the arm's animals take 1.8 times as long to reach the endpoint, 0.5 half as long. It does not assume proportional hazards; it assumes a distribution for the times (Diagnostics compares them).
  • Time_Ratio — with its 95 % confidence interval (Lower / Upper 95% CI).
  • P_Value / P_Adjusted — Wald test that the time ratio is 1, then corrected over the arms compared with the reference.
  • Median_Predicted — the arm's median time to the endpoint from the model.
  • Note — an arm in which no animal reached the endpoint has no finite time ratio: every time is censored.
Cox proportional hazards model. Hazard Ratios (HR) are computed relative to the reference group. Firth’s bias-reduced method is applied automatically when a group has zero events.
  • Group — treatment group name.
  • HR — Hazard Ratio: the instantaneous risk of the event in this group relative to the reference. HR < 1 means lower hazard (better survival); HR > 1 means higher hazard.
  • Lower / Upper 95% CI — 95 % confidence interval for the HR, not adjusted for multiple comparisons. It can exclude 1.0 while the adjusted P_Value is above 0.05; judge significance by P_Value.
  • P_Value — p-value for the comparison with the reference, adjusted for the number of comparisons (P_Adjust_Method; unadjusted in P_Value_Unadjusted).
  • P_Method — shown when whole cages were randomised (Study Design on Data Upload): the p-value comes from a cage-level permutation log-rank, and the hazard ratio has no interval.
  • Events — number of animals that reached the endpoint (tumour volume exceeded the IACUC threshold, or died before reaching it if natural death is treated as an event).
  • Total — total number of animals in the group.
  • Event_Rate — Events ÷ Total. Values close to 1.0 mean most animals reached the endpoint.
  • Median_Survival — day by which 50 % of the group reached the endpoint. ‘Not reached’ if fewer than half had events by study end.
  • Note — flags the reference group row.
Loading...

                      
Loading...
Numbers at risk table: one row per group per event time point, summarising what happened to the cohort at that moment.
  • Group — treatment group name.
  • Time — the day on which one or more events occurred.
  • N_Risk — number of animals still alive and uncensored at the start of this time interval.
  • N_Events — number of animals that experienced the endpoint at exactly this time.
  • N_Censored — number of animals whose last observation was at this time without reaching the endpoint (e.g. still within tumour volume limit at study end, or removed from study).
  • Survival — Kaplan–Meier estimated survival probability at this time point (1.0 = 100 % surviving).
N_Risk decreases over time as events and censoring accumulate; a rapid drop indicates many events in a short window.
Loading...

Group colours are mapped automatically from the analysis groups.


Annotation



Kaplan-Meier Survival Curves

Loading...

Forest Plot

Loading...

Posterior Distributions

Loading...
Model-fit diagnostics. Cox PH: proportional-hazards assumption check (Schoenfeld). AFT: the distributions compared, and the fit against Kaplan-Meier. Bayesian (brms): MCMC convergence + posterior diagnostics. The view switches based on the selected Statistical Model.
Cox proportional-hazards assumption check via Schoenfeld residuals (cox.zph). The Cox model assumes the hazard ratio is constant over time. A small global p-value indicates the assumption is violated and the reported HRs / CIs may be misleading (a treatment whose effect accelerates or attenuates over time gives inconsistent HRs depending on when you look).
  • chisq — Schoenfeld-residual test statistic per covariate, plus a GLOBAL row.
  • df — degrees of freedom for the test.
  • p — p-value. < 0.05 in the GLOBAL row means the PH assumption is rejected at the 5 % level.
Only available for the standard Cox path. NA for Firth (coxphf) and log-rank fallbacks.
Loading...

Schoenfeld residuals plot

How to read this plot. Each panel shows the smoothed Schoenfeld residuals for one covariate against transformed survival time. Flat horizontal smoother around 0 (within the shaded CI band) = PH assumption met for that term. Monotonic trend = the hazard ratio drifts over time (treatment effect strengthens or weakens). Sinusoidal / non-monotonic shape = a more complex time-varying effect — usually means a Cox model with a time-varying coefficient (or a parametric AFT) fits better. The dashed line at zero is the null.
Loading...

The distributions compared

Each distribution's fit to the data, by AIC (lower is better; a difference under about 2 is not a real difference). 'Used' marks the one the results come from.

Fit against Kaplan-Meier

How to read this plot
Each arm's Kaplan-Meier curve (steps) with the AFT model's fitted curve (dashed). The dashed curves follow the steps: the distribution fits. They part systematically: try another distribution, or the Cox model.
Loading...
MCMC convergence diagnostics for all fixed-effect parameters. All chains should have converged (R̂ ≈ 1.00) and have sufficient effective samples before interpreting results.
  • Rhat — potential scale-reduction factor; values > 1.01 indicate non-convergence.
  • Bulk_ESS — bulk effective sample size; aim for > 400 per chain.
  • Tail_ESS — tail ESS; important for credible interval accuracy.
  • Converged — TRUE when R̂ ≤ 1.01.
If Rhat > 1.01, try increasing iterations or chains before reporting results.
What these diagnostics mean
Divergent transitions
The sampler met a region of the posterior it could not follow accurately. Even a few can bias the estimates. Run more warmup iterations, choose a stronger prior preset, or simplify the random effects, and refit.
Max-treedepth hits
The sampler reached its step limit before finishing a trajectory. This is about efficiency, not validity: the draws are still correct, but the fit was slow and the effective sample sizes may be low. Many hits often point to a weakly identified parameter, such as a random-effect SD the data say little about.
E-BFMI
How well the sampler moved between the posterior's typical values and its tails. Below 0.3 it explored the tails poorly, and variance parameters and the ends of credible intervals may be unreliable. The remedies are those for divergences.
Bayes R²
The share of the variation in the modelled values that the model explains, with its credible interval. It includes the random effects, so differences between animals count as explained. It describes fit; it is not a test of the treatment.
PPC coverage
How often the observed values fall inside the model's 50 %, 80 % and 95 % predictive intervals. A well-calibrated model gives about 50 %, 80 % and 95 %. Much lower means its intervals are too narrow (overconfident); much higher, too wide.
PSIS-LOO (elpd_loo)
Leave-one-out cross-validation, estimated without refitting: how well the model predicts each observation when that observation is left out. elpd_loo means little on its own; it compares models fitted to the same data, where higher is better and a difference smaller than about twice its SE is not meaningful.
Pareto-k
For each observation, how much leaving it out would change the posterior. Above the threshold (about 0.7) the leave-one-out estimate for that observation is unreliable, and the observation is influential: usually a measurement the model fits badly, such as a data-entry error or an animal whose growth departs from its arm's. Check those rows in the table below against the raw data. The estimates are not invalidated, but they lean on those observations.
Loading...

Influential observations (Pareto-k > 0.7)
Measurements that move the fit too much for the leave-one-out estimate, with their animal. Check them against the raw data; 'What these diagnostics mean' above explains Pareto-k.
Loading...

Posterior predictive check
How to read this plot
The model simulates new datasets from the posterior and overlays their densities (y_rep) on the observed survival times (y). The observed density inside the simulated envelope: the model reproduces the data. Outside it: consider the other likelihood family.
Loading...

MCMC trace plots
How to read this plot
Each line is one chain's draws of a parameter, after warmup. Converged chains overlap and look like stationary noise around one level (a ‘fuzzy caterpillar’). A drift, a trend, or a chain at a different level: the chains have not converged and the estimates are unreliable; run more warmup iterations or simplify the model.
Loading...

Choose what to download. It all goes into one ZIP file, with a README.txt that lists every file, the settings of the run and the package versions.

Download selected (ZIP)

Toxicity Analysis

Analyze body weight toxicity, or combine with tumor efficacy to assess the therapeutic window. Requires a Weight/Mass column configured in the Data Upload tab.


Overview

The Toxicity tab provides two complementary analysis modes selectable from the Statistical Model dropdown:

  • Body Weight Analysis — models longitudinal weight trajectories using a linear mixed-effects model, and times each animal to a weight-loss threshold (Kaplan-Meier).
  • Therapeutic Window — sets each arm's tumor growth inhibition beside its worst body-weight loss, each with an interval, and flags whether the weight loss is tolerated. Requires a tumor volume column.

Body weight is used as the toxicity surrogate throughout. When 'Subtract estimated tumor weight' is enabled, the gross weight is corrected by subtracting the tumour mass, volume × density (g/cm³), with the volume in the units declared under Study Design on Data Upload (mm³ is divided by 1000), giving a net host weight.

Body Weight Analysis — Analyses & Metrics

Mixed-Effects Model

Fits a linear mixed-effects model to all repeated weight measurements:

Weight ~ Treatment × Day + [covariates] + (Day | Mouse)

Random slopes allow each mouse's weight trend to deviate from the group average. Fixed effects test whether trajectories differ by treatment group over time.

The model is fitted by restricted maximum likelihood (REML), the usual choice for reporting a model's estimates and variance components.

Fixed and random effects
Fixed effects
What the study is about, estimated once for the whole study and reported as results: the treatment, time, their interaction (whether the arms change at different rates), and any covariate. The comparisons and their p-values or credible intervals come from these.
Random effects
How each mouse departs from its arm's average: a random intercept for its own starting level, a random slope for its own rate. The model estimates how much mouses vary (an SD) rather than reporting each one. A cage random effect does the same for cages, whose animals resemble each other.
Why both
Repeated measurements of one mouse are not independent. Treating them as independent overstates the evidence and gives p-values that are too small; random effects account for it, so the treatment comparison uses the right amount of information. When mouses differ in their rates, leaving out the random slope finds treatment differences that are not there.

Covariates (optional):

  • Tumor Volume — controls for tumor-driven weight changes independent of drug toxicity. Offered only when 'Subtract estimated tumor weight' is off: with tumour mass subtracted, adjusting for volume as well would count tumour burden twice.
  • Sex — adjusts for sex-specific baseline weight differences.
  • Initial Mass — adjusts for pre-treatment weight variation across animals.

Weight Loss KM Curve

Treats each mouse as an event-time observation where the event is crossing the weight-loss threshold (default 20%). Mice that never reach the threshold are censored at their last observation. A Kaplan-Meier curve is fitted per group; group differences are tested with the log-rank test. When an arm has no events, Firth's penalised Cox model is used for the hazard ratios.

Competing risks. A mouse euthanised for tumour burden can no longer experience a weight-loss event, so treating it as an ordinary censoring overstates the weight-loss risk — censoring assumes the animal remained at risk, and it did not. Which animals were removed, and why, is known only from a removal-reason column (mapped on Data Upload). With one, a removal for weight loss or body condition counts as a weight-loss event, a planned end (end of study, scheduled sacrifice, data cut) as censoring, and any other removal as a competing risk, with an Aalen-Johansen cumulative incidence reported. Without one, an animal whose record ends early is censored at its last day, which assumes its later risk was like that of the animals still on study; the panel states which assumption applied. The censoring summary breaks down, per arm, how each animal left the study.

Weight Loss Threshold: The % body weight loss that constitutes a toxicity event. The default (20%) is a common IACUC endpoint; adjust to match study-specific welfare criteria.

Therapeutic Window — Analyses & Metrics

Efficacy

Efficacy is tumor growth inhibition at the endpoint day, relative to the reference group: TGI = (1 − V_treated / V_control) × 100 . Values > 0 indicate tumor suppression. Each arm's endpoint volume comes from the endpoint estimand described below; it is the same TGI as on the Drug Synergy and Dose Response tabs.

Evaluable days

An arm is evaluable on a day when at least 50% of its enrolled animals, and at least 3 animals, are still on study (measured on that day or later). The endpoint day is the last day on which every arm is evaluable; later days are not used, because an arm's mean there would rest on a minority of its animals or be extrapolated past them. The results name the day and list the days and arms the rule excluded.


Tolerability

Each animal's worst weight loss is its largest drop below its own first weighing, in percent, over its whole record (weight gain counts as no loss). The arm's worst weight loss is the mean over its animals, with a 95% interval from resampling animals. Over threshold counts the animals whose own worst loss reached the threshold.

Tolerated
the interval lies below the tolerability threshold
Not tolerated
the interval lies at or above it
Unclear
the interval contains it

The flag describes the arm's average animal, so a Tolerated arm can still have an animal over the threshold; read the two columns together. Weighings between calliper days count: for the tumour-mass correction, each animal's volume is interpolated between its calliper readings.


Reading the two axes

The Therapeutic Window table and plot set each arm's TGI (efficacy) beside its worst weight loss (tolerability). They are read separately: the table is not a ranking. A ratio of the two was reported before version 26.09.005; it was withdrawn because an arm that loses no weight needed an arbitrary floor, and one number cannot show both effects.

Intervals
Under the model estimand, TGI intervals come from draws of the fitted model; under the observed estimands, from resampling animals. Weight-loss intervals always come from resampling animals.
Interval draws
Number of draws or resamples behind those intervals. Set to 0 to skip, which returns point estimates with no interval.
Endpoint estimand
Model (default) estimates group means at the endpoint from the fitted model, using every observation each animal contributed. Observed uses only animals still present at the endpoint. The two differ whenever animals are removed early, and the difference is not small: because the fastest-growing controls are the ones that leave, the observed estimand flatters the control arm and understates TGI. Per-arm attrition is reported beneath the table so the size of the issue is visible.

Shared Parameters

Reference Group
The study's control group, chosen under Study Design on Data Upload; it is the denominator for TGI and the reference for every comparison.
Subtract Tumor Weight
Removes the estimated tumor mass (volume × density, in the declared volume units) from gross body weight before all calculations, so that body weight reflects host tissue rather than tumor burden. It needs the volume units declared under Study Design on Data Upload.
Tumor Density
Used only when tumor weight adjustment is enabled. Default 1.0 g/cm³ (water density). Typical soft-tissue tumor density is 0.9–1.1 g/cm³.
Baseline Day
The study day used to compute % change from baseline. Defaults to the earliest available day.

Analysis summary. Body Weight Analysis: mixed-effects model output, fixed effects, and each arm's rate of weight change. Therapeutic Window: TGI and worst weight loss per arm, with the tolerability flag.
Loading...

                      

Fixed Effects

Loading...
Download Fixed Effects CSV

Rate of Weight Change by Treatment (g/day)

Each arm's estimated rate, with its interval. Whether an arm differs from the reference is tested in the next table, with p-values corrected for the number of comparisons.
Loading...
Download Rates CSV

Difference in Rate vs the Reference (g/day)

Loading...
Download Comparisons CSV
Time-to-threshold event data. Each row is one mouse. Event = 1 when the weight-loss threshold was crossed; Event = 0 (censored) when the mouse never reached the threshold. This table feeds the Kaplan-Meier plot in the Plots tab.

Time-to-Threshold Event Data

Loading...
Download Events CSV



Body Weight Trajectory

Loading...

Weight Loss Kaplan-Meier Curve

Loading...

Therapeutic Window

Loading...
Body-weight model diagnostics. Residual, random-effects and influence diagnostics for the body-weight mixed model.

Q-Q Plot of Residuals

How to read this plot
The residuals are standardised, so for a well-fitting model the points fall along the red reference line. S-shape: heavy-tailed residuals — outliers or a transform issue. Bow: skewed residuals — usually fixed by changing the weight transform.
Loading...

Residuals vs Fitted

How to read this plot
Random scatter around zero = LMM assumptions met. Curved blue line: weight does not change linearly with time, so the Day slope averages over a curved trajectory — read it alongside the trajectory plot. Fan shape: heteroscedasticity — consider log- transforming or weighting the response.
Loading...

Scale-Location

How to read this plot
sqrt(|residuals|) vs fitted — magnifies variance trends. Flat blue line = constant variance. Upward slope = heteroscedasticity.
Loading...

Random-effects Q-Q

How to read this plot
Per-animal random intercepts should be normally distributed. Heavy tails are rare; safe to ignore at typical preclinical N.
Loading...

Influence diagnostics (case-deletion)

Cook's distance measures the overall effect of leaving out each animal (one row per animal). Above the threshold (4 / number of animals) = an animal that moves the fit. DFBETAS shows per-coefficient influence; values | > 1 | are flagged.
Cook's distance
Loading...
DFBETAS
Loading...
Model diagnostics are only available in Body Weight Analysis mode. Switch the Statistical Model in the sidebar to access them.

Choose what to download. It all goes into one ZIP file, with a README.txt that lists every file, the settings of the run and the package versions.

Download selected (ZIP)

Dose-Response Analysis

Analyze tumor volume responses across dose levels using linear regression, ANOVA, and non-linear dose-response models. Upload and configure your data (including a Dose column) in the Data Upload tab first, then run Tumor Growth Analysis before proceeding here.


Overview

The Dose-Response tab models how tumor volume (or TGI) varies across dose levels. Three complementary analyses are performed simultaneously:

  • Linear regression — tests whether there is a significant linear trend in tumor volume with increasing dose.
  • One-way ANOVA — tests whether any dose group differs from the others, followed by Tukey HSD post-hoc comparisons if significant.
  • Non-linear dose-response model (4PL) — fits a four-parameter log-logistic (Hill) model to quantify ED50 and curve steepness.

One agent at a time. A dose-response curve describes one agent at several doses, against the study's control. Choose the agent's arms under Dose series in the sidebar; the dashboard suggests the largest set of single-agent arms that share a name and differ in dose. Arms that share a dose, or other agents and combinations, cannot join the same curve, and the analysis stops rather than pool them. When one treatment name is used at every dose, the rows with dose 0 are the control.

A Dose column must be mapped in Data Upload.

Evaluable days

An arm is evaluable on a day when at least 50% of its enrolled animals, and at least 3 animals, are still on study (measured on that day or later).

The analysis uses one day on which every dose group is evaluable: by default the last such day, with the animals measured on it. The day selector lists only evaluable days. Before, the default was each animal's own last measurement, so controls removed at the volume limit contributed capped, earlier volumes and the EC50 was biased. The results name the day and list the days and dose groups the rule excluded.

TGI by dose is read at the same day from the model that the Drug Synergy and Toxicity tabs use, so the three tabs agree.

Linear Regression

Fits Volume ~ Dose by ordinary least squares. The slope quantifies the change in mean tumor volume per unit increase in dose.

Slope
Change in volume per dose unit. A negative slope indicates dose-dependent tumor suppression.
R²
Proportion of variance in volume explained by dose. Low R² with a significant slope can occur when individual variability is high relative to group differences.
p-value
Significance of the slope. A non-significant p-value does not exclude a non-linear relationship — inspect the 4PL model output and plots as well.

One-Way ANOVA & Tukey HSD

Treats dose as a categorical factor and tests whether group means differ. A significant F-test (p < 0.05) triggers Tukey HSD post-hoc comparisons to identify which specific dose pairs differ.

F-statistic
Ratio of between-group to within-group variance. Larger values indicate greater separation between dose groups relative to within-group noise.
Tukey HSD
Family-wise error rate corrected pairwise differences. Values in 'p adj' below 0.05 indicate a significant difference at the 5 % level after correction for all comparisons.

ANOVA assumes equal variances across groups (homoscedasticity). When this is violated, interpret alongside the linear regression, which is more robust to mild heteroscedasticity.

Non-Linear Model (4-Parameter Log-Logistic)

A four-parameter log-logistic (4PL / Hill) model is fitted via the drc package:

Volume = c + (d − c) / (1 + (Dose / ED50)^b)
b
Hill coefficient: steepness of the dose-response curve. Larger absolute values indicate a sharper transition.
c
Lower asymptote (response at saturating inhibitory dose).
d
Upper asymptote (response in the absence of drug).
ED50
Dose producing 50 % of the maximal effect. The primary pharmacodynamic summary statistic.

The EC50 reported is the dose giving half of the fitted response range, with a 95 % interval computed on log dose, so it stays positive. The lower asymptote cannot go below zero, and the five-parameter (asymmetric) curve is considered only with at least six dose levels. The summary flags an EC50 outside the tested doses, and a curve with as many parameters as dose levels, which fits the dose means exactly and so says nothing about its own shape; read either as descriptive.

Which curve: the linear model and the log-logistic curves (Emax, 3PL, and 4PL with at least 5 doses) are fitted to the same data and compared by AIC, which is valid across them: each is the full Gaussian log-likelihood of the same observations. The selected curve is the simplest within 2 AIC of the best (Model Results, and the default trend line in Plots). Until 0.58.0 this note said they could not be compared; they can.

Trend Tests

A dose-response study usually asks whether response is monotone in dose — not merely whether the groups differ somewhere. Three tests address that directly:

Jonckheere–Terpstra
The test that actually respects dose ordering. It is sensitive to a consistent monotone trend and, unlike ANOVA, gains power from the ordering rather than discarding it. This is the primary trend result. It is two-sided; the summary also says whether volumes fall or rise with dose. (It was one-sided in the direction the data suggested, which doubled its false-positive rate.)
Linear trend
p-value for a linear term in dose. Significant means a straight-line component is present; it says nothing about whether the relationship is only linear.
Quadratic trend
Detects curvature — a plateau or a reversal at high dose. A significant quadratic term alongside a significant linear one usually means the response saturates, which matters when extrapolating to untested doses.

Dose-response summary: Results of linear regression (slope, R², p-value) and one-way ANOVA across dose levels. A significant linear slope indicates a systematic increase or decrease in tumor volume with dose. A significant ANOVA F-test indicates that at least one dose group differs from the others.
Loading...

                      
Per-dose summary statistics of tumor volume at the selected time point, one row per dose level.
  • Dose — dose level (concentration or quantity as provided in your data).
  • n — number of animals at this dose.
  • mean_volume — arithmetic mean tumor volume across all animals at this dose.
  • sd_volume — standard deviation of tumor volume; reflects within-dose variability.
  • sem_volume — standard error of the mean (SD ÷ √n); indicates precision of the mean estimate.
  • ci95_lower / ci95_upper — lower and upper bounds of the 95 % confidence interval for the mean (using the t-distribution). Non-overlapping intervals between adjacent doses suggest a significant dose step.
Loading...

TGI by dose

At the analysis day, from the endpoint model (the TGI the Drug Synergy and Toxicity tabs report), with 95% intervals from draws of the model.

Full statistical test results. The text output includes linear regression (slope, intercept, p-value, R²), one-way ANOVA table, and polynomial trend p-values. Non-linear model AIC/BIC are shown when the drc package fit converges. When ANOVA p < 0.05 the Tukey HSD post-hoc table is shown below:
  • Comparison — the two dose levels being compared (‘DoseA - DoseB’).
  • diff — estimated mean difference in tumor volume (Dose A minus Dose B).
  • lwr / upr — lower and upper bounds of the Tukey-adjusted 95 % confidence interval for the difference.
  • p adj — Tukey family-wise error rate adjusted p-value. Values below 0.05 indicate a significant difference between that dose pair.
Loading...

                      
Loading...

Which dose-response curve

Each candidate curve fitted to the same data, compared by AIC (lower is better).
  • Linear — volume linear in dose.
  • Emax — a hyperbolic curve: baseline, maximum effect and ED50.
  • 3PL — a log-logistic curve falling to 0: slope, upper limit and ED50.
  • 4PL — the same with its own lower limit; fitted with at least 5 doses, since with fewer it has nearly as many parameters as doses.
  • Selected — the simplest curve within 2 AIC of the best: a smaller difference is not evidence for the extra parameters. It is the default trend line in Plots.

Appearance

Options

Title, Labels & Subtitle

Plot Size / Export

Tumor Volume vs Dose

Loading...

Tumor Growth Inhibition vs Dose

Loading...
Model-fit diagnostics. Hill / Emax lack-of-fit and residuals.
Only available when the drc Hill / Emax fit converged; in pure-linear mode this section will be empty.

Lack-of-fit test

How to read this table
drc::modelFit compares the Hill fit against a one-mean- per-dose smoother via an F-test. Large p (≥ 0.05): the Hill model is as good as the smoother — functional form is adequate. Small p (< 0.05): the smoother fits significantly better — the Hill equation may be the wrong shape; inspect the residuals plot for clues.
Loading...

Residuals vs Dose

How to read this plot
Points should scatter around the red zero line with no trend. Curved blue smoother: the Hill function misses curvature at some dose range — try a different model. Fan shape (variance grows with dose): heteroscedasticity; consider log-transforming the volume axis or weighting.
Loading...

Choose what to download. It all goes into one ZIP file, with a README.txt that lists every file, the settings of the run and the package versions.

Download selected (ZIP)

Drug Synergy Analysis

Test whether a drug combination shows synergistic, additive, or antagonistic effects using the Bliss independence model. Ensure tumor growth data is uploaded and configured in the Data Upload tab.


Overview

The Drug Synergy tab assesses whether a drug combination produces greater tumor suppression than predicted from the individual drugs acting independently. The reference model is Bliss independence : the drugs act independently on the fraction of tumour each one leaves, so the expected combined effect is FE_A + FE_B − FE_A × FE_B (details below).

Four treatment groups must be assigned in the sidebar: Control, Drug A only, Drug B only, and the Combination (A + B). Analysis can be run at a single timepoint or tracked over the full study duration.

Bliss Independence

Under Bliss independence the drugs act independently on the fraction of tumour that survives each one. With FE = TGI / 100 for each single agent, the expected combined fractional effect is:

Expected FE = FE_A + FE_B − FE_A × FE_B
Expected TGI = 100 × Expected FE
Bliss Difference
FE_combo − Expected FE, shown in percentage points. Positive = synergy; negative = antagonism; near zero = independent action.
Ceiling effect: when both single agents are strong, the Bliss expectation is already close to 100 % (50 % and 50 % give 75 %; 80 % and 80 % give 96 %), which leaves little room for the combination to exceed it. Read small Bliss differences cautiously when single-agent TGIs are high.

Uncertainty — how far to trust the verdict

The synergy verdict comes from the 95% interval for the Bliss excess: Synergy when the interval lies above zero, Antagonism when it lies below, and Additive (no departure from Bliss detected) when it contains zero. With Interval draws set to 0 there is no interval; the verdict then compares the point estimate with symmetric bands of ±0.1 and says it has no interval. Earlier versions labelled any positive excess “Synergy” but negative ones down to −0.1 “Additivity”, so a Bliss-additive combination was called synergistic in about half of simulated studies.

Interval draws
The 95 % intervals describe the reported estimates. Under the model estimand they come from draws of the fitted model's coefficients; under the per-animal estimands, from resampling animals within each arm, control included. Either way the uncertainty of the control mean, the denominator of every TGI, is carried through. Set to 0 to skip; the verdict is then a point estimate with no stated uncertainty.
Interval callout
Beneath the verdict, the 95 % interval for the Bliss excess that decided it, and whether it excludes zero.
Endpoint estimand
Model (default) estimates endpoint group means from a mixed model of log volume with a smooth time course for each arm and a growth rate for each animal, using every observation each animal contributed. Observed uses only animals still present at the endpoint. Because the fastest-growing controls are the ones removed early, the observed estimand shrinks the control mean and inflates every fractional effect derived from it — which propagates directly into both Bliss and the Bliss expectation. Per-arm attrition is shown so the exposure is visible.

Evaluable days

An arm is evaluable on a day when at least 50% of its enrolled animals, and at least 3 animals, are still on study (measured on that day or later).

Synergy is reported only on days when all four arms are evaluable. By default the analysis uses the last such day; the day selector lists only evaluable days. Later days are not analysed, because an arm's mean there would rest on a minority of its animals or be extrapolated past them — the control arm, removed first at the volume limit, is usually the one that runs out. Each result names the day used and lists the days and arms the rule excluded.

Over-Time Analysis

With ‘All evaluable days’ selected, the Bliss score is computed on every evaluable day, from one fit of the model, and plotted as a synergy trend. This shows whether synergy emerges, fades, or reverses over the course of treatment.

Per-day p-values (Combination vs Drug A and vs Drug B, on the same estimand as the TGIs) are included in the Synergy Metrics table. They are unadjusted for multiple testing across days — interpret overall patterns rather than individual p-values at isolated days.


Synergy summary: Overall assessment from the Bliss independence model. Reports whether the drug combination is synergistic, additive, or antagonistic at the evaluated timepoint, and lists the key quantitative scores.
Loading...

                      
Quantitative synergy scores. When ‘Analyze over time’ is disabled, the table has two columns ( Metric / Value ) summarising scores at the selected timepoint:
  • Bliss Difference — observed combination effect minus the Bliss-independent expected effect (TGI combo − TGI Bliss). Positive = synergy; negative = antagonism.
  • Interpretation — the verdict, from the 95% interval for the Bliss excess: Synergy (above zero), Antagonism (below zero) or Additive (contains zero). Without an interval, symmetric bands of ±0.1 around zero, and the label says there is no interval.
With ‘All evaluable days’, each row is an evaluable day with columns:
  • Time_Point — study day.
  • TGI_Drug_A / TGI_Drug_B / TGI_Combo — % tumor growth inhibition for Drug A, Drug B, and the combination relative to control.
  • Bliss_Expected_TGI — predicted TGI under the Bliss independence model.
  • Bliss_Difference — TGI_Combo minus the expected TGI; positive values indicate synergy.
  • *_Lower / *_Upper — 95% interval of each TGI and of the Bliss difference.
  • P_Value_vs_Drug_A / P_Value_vs_Drug_B — p-value comparing the combination group to each mono-therapy on the same estimand (a Wald test of the volume ratio under the model estimand); values below 0.05 indicate a statistically significant difference.
  • Synergy_Assessment — the verdict on each day, from that day's interval for the Bliss excess.
  • Bliss_Applies — FALSE where a single agent did not inhibit growth that day. Bliss independence is defined only for inhibitory agents, so the Bliss columns are blank on those days and they cannot be the peak.
Loading...




Synergy Plot

Loading...

Synergy Plot

Loading...

Tumor Growth Inhibition

Loading...
Checks on the estimate that is reported. Under the model estimand: the endpoint model's residuals by arm, and each arm's fitted curve over its data up to the evaluation day. Under the per-animal estimands: the per-animal volumes used, by arm.
Loading...
Loading...

Choose what to download. It all goes into one ZIP file, with a README.txt that lists every file, the settings of the run and the package versions.

Download selected (ZIP)

Power Analysis

Estimate statistical power and sample sizes using a priori methods. Configure parameters in the sidebar and click Run Analysis.


Effect Size




Tumor Growth Parameters


Variance Components (log scale)


Study Design


Simulation


Alpha Levels


Target Power



Overview

The Power Analysis tab supports prospective (a priori) sample-size planning before an experiment is run. Both methods estimate the probability of detecting a true effect of a specified magnitude at a chosen significance level.

  • A priori — Analytic — closed-form power for each treated-vs-control comparison (two-sample t-test), with the alpha split across the k − 1 comparisons a k-group study makes. Fast; no simulation required.
  • A priori — LMM Simulation — Monte Carlo simulation of longitudinal tumor-growth data. Fits the same mixed-effects model used in the Tumor Growth tab and extracts power from the Treatment × Day interaction. More realistic for repeated-measures designs.

Post-hoc power analysis (computed from observed data after the experiment) is not provided. Post-hoc power is a monotone function of the p-value and adds no information beyond the test result itself (Hoenig & Heisey, 2001).

Analytic Mode — Method & Parameters

Method

Power is computed for one treated-vs-control comparison: a two-sample t-test (non-central t-distribution) with Cohen's d = Δ / SD_pooled. That is how these studies are analysed, as k − 1 comparisons against the control arm.

Two groups: one comparison at the chosen alpha.

Three or more groups: each comparison is tested at alpha / (number of comparisons) under the default Bonferroni correction, so each treated arm that differs from control by d is detected with the reported power. Dunnett, the Tumor Growth default, is slightly less conservative, so the N shown is a small overestimate for it.

Attrition: Required_N is the number that must reach the endpoint; with an expected attrition above 0 %, Enroll_N is the number to start with.


Effect Size Input
Cohen's d
Enter one or more d values directly (comma-separated). Conventional benchmarks: small ≈ 0.2, medium ≈ 0.5, large ≈ 0.8.
Mean difference + SD
Enter the expected mean difference (Δ) and pooled within-group SD. Cohen's d is computed internally. 'Seed SD from loaded data' fills in the pooled SD of your uploaded data at one endpoint day, from each animal's last volume up to that day. Choose the day your study's comparison is planned at; Δ should be a difference at that day.

Solve For
Find N
Returns the required N per group to achieve each target power × alpha combination. Results shown in the Scenario Table.
Find Power
Returns achieved power at specified N values. Useful when sample size is constrained by practical considerations.

SD Sensitivity Tab

Shows how the required N changes when the assumed SD is perturbed by ±20 % and ±40 %. Only available in Mean Difference + Pooled SD mode. Use this to gauge sensitivity of your sample size estimate to uncertainty in the expected variance.

LMM Simulation Mode — Method & Parameters

Method

Synthetic tumor-growth datasets are generated on the log scale:

log(Volume) = log(baseline) + b₀ + (rate + b₁) × t + ε

where b₀ ~ N(0, random intercept SD), b₁ ~ N(0, random slope SD), and ε ~ N(0, residual SD). Volumes are exponentiated back to the original scale.

For each simulated dataset, log(Volume) ~ Treatment × Day + (Day | ID) is fitted as in the Tumor Growth analysis (REML; uncorrelated slopes, then a random intercept, when a fit fails). Each treated arm's growth rate is compared with the control's by a Satterthwaite t-test at α/(k−1). Power = proportion of simulations in which a given arm is significant; the Scenario Table also gives the chance that any arm, or every arm, is significant.


Key Parameters
Control log-growth rate
Exponential growth constant for the control group on the log scale. A rate of 0.15 corresponds to ~16 % daily volume increase.
Treatment rate reduction
Reduction in growth rate for treated vs control, on the log scale.
Random intercept SD
Between-animal variability in baseline volume (log scale). Does not affect the power estimate — the test is on Treatment × Day, a contrast of growth rates, and the per-animal random intercept absorbs baseline spread exactly. Varying it from 0.05 to 1.50 leaves power unchanged. The same is true of Baseline SD, with which it is aliased.
Random slope SD
Between-animal variability in growth rate (log scale). This is the dominant driver of power — it is the noise the treatment effect must be detected against. In a worked 2-arm, n = 8 design, moving it from 0.01 to 0.15 dropped power from 1.00 to 0.15. Estimate it from your own historical control data before trusting a sample-size answer.
Residual SD
Within-animal measurement noise (log scale).
Euthanasia volume limit
Removes an animal's observations once its simulated volume crosses the limit, mimicking an IACUC removal. Leaving it blank assumes complete data and overstates power — about 6 points at 30 % attrition, 12 at 40 %. Realised attrition is reported with the results.
Simulations
Number of Monte Carlo draws. Use ≥ 1000 for publication-quality estimates; ≤ 500 for quick exploration. Power estimates are shown with ±1.96 × SE uncertainty bands in the Power Curves tab.

Interpretation

Alpha (α)
Significance threshold (Type I error rate). Conventional choices are 0.05 (5 %) and 0.01 (1 %). Lower alpha requires a larger N to achieve the same power.
Power (1−β)
Probability of detecting a true effect. 0.80 is a common minimum; 0.90 is recommended when false negatives are costly.
N per group
Sample size per treatment arm. Total study animals = N per group × number of groups.

Power estimates assume the specified effect size is exactly correct. If the true effect is smaller, actual power will be lower than estimated. Use the SD Sensitivity tab to understand robustness to variance assumptions.



Loading...
A priori power / sample size table: For each combination of effect size, alpha, and target power, shows the required N per group (Find N mode) or achieved power (Find Power / LMM mode).
  • Effect_Size — Cohen’s d (analytic) or the simulated scenario (LMM).
  • N_Groups — number of treatment groups in the design.
  • Alpha — significance threshold.
  • Required_N — N per group needed to reach the target power (Find N mode).
  • Achieved_Power / Power — power reached at the specified N (Find Power / LMM mode).
  • Alpha_Per_Comparison — the alpha each treated-vs-control comparison is tested at after the correction; under Dunnett it depends on N.
  • Hazard_Ratio, Control_Event_Prob, Treated_Event_Prob, Required_Events — survival endpoint only: the hazard ratio, the share of each arm expected to have the event, and the events one comparison needs (Schoenfeld).
Loading...
Download
Power vs N per group: prospective power curves for the specified effect size scenario(s). Each line represents one Cohen’s d / alpha combination. Dashed lines mark 0.80 and 0.90 power thresholds.
Loading...
Download Plot
Sensitivity to the assumed SD: shows how the required N changes when the pooled within-group SD used to convert a raw mean difference into Cohen’s d is perturbed by ±20 % and ±40 %. Only available when using the Mean difference + Pooled SD input mode.
  • SD_Change — percentage change applied to the assumed SD.
  • Assumed_SD — resulting SD value.
  • Cohens_d — Cohen’s d after adjusting for the new SD.
  • Required_N — sample size required at the reference alpha / target power.
Loading...
Download

Choose what to download. It all goes into one ZIP file, with a README.txt that lists every file, the settings of the run and the package versions.

Download selected (ZIP)

Results

Every analysis run in this session, newest first. Open one to show it again in its tab, with its tables, plots and downloads; select several to download them together, each in a folder of its own. Runs are kept until the session ends.


Download selected (ZIP)

Help & Documentation

Disclaimer

mouseExperiment is provided "as is" for research purposes only, without warranty of any kind, express or implied. The authors and contributors accept no liability for errors, inaccuracies, or any consequences — including but not limited to incorrect results, data loss, or decisions made on the basis of outputs from this software. All results must be independently validated before use in publications, reports, regulatory submissions, clinical decisions, or any other application.


Quick start

  1. On the Data tab, under Try Demo Data , choose All Features and click Load Demo Data .
  2. Below the column mapping, check Study Design : volume units mm³ and the control group Vehicle.
  3. Open Tumor Growth and click Run Analysis ; the results open on the Summary tab.
  4. Look through the other result tabs (Plots, Comparisons, Diagnostics), then save what you need from Download Results . Every other analysis tab works the same way.

Basic Workflow

  1. Data tab — upload a CSV file and map your columns (ID, time, volume, treatment). Optional columns: cage, dose, body weight, survival/censor indicator. The Data check below the summary then lists likely errors in the file (duplicates, negative volumes, large jumps, missed days) with their line numbers.
  2. Study Design (on the Data tab) — declare the tumour volume units (mm³ or cm³), the control group, and, when a cage column is mapped, whether cage labels are unique across the study or numbered within each arm (C1 and C2 in every arm are then different cages) and whether individual animals or whole cages were randomised. With a removal-reason column mapped, say what each reason means (death, weight loss, the volume limit, a planned end); Survival and Toxicity use it. Every analysis tab uses these settings; none has its own reference selector.
  3. Analysis tabs — each tab activates once the required columns are mapped:
    • Tumor Growth — always available after upload
    • Survival — requires a survival/censor column (or use the volume-threshold derivation)
    • Toxicity — requires a body weight column
    • Dose Response — requires a dose column
    • Drug Synergy — requires control, single-agent A, single-agent B, and combination groups
    • Power Analysis — always available; no data upload needed for a priori modes
  4. Select the statistical method in the sidebar, configure parameters, and click Run Analysis . The beside an option explains it: hover over it, or click or tap it.
  5. Explore results across sub-tabs (summary, plots, pairwise comparisons, model output). Each Diagnostics tab has a How to read this plot note under every plot.
  6. Download one table or plot from the button beside it, or choose any mix of tables, plots, diagnostics and data from a tab's Download Results tab: one ZIP, with plots in PNG, PDF or SVG and a README listing every file, the settings of the run and the package versions. Every tab also offers an R script that repeats the run with the mouseExperiment package, from the analysed data that comes with it; tick Include my original file and it starts from the file you uploaded, replaying your exclusions, the study days, the volumes and the arm labels. Tumor Growth also offers an HTML report.
  7. Long analyses (Bayesian fits, randomisation tests, the power simulation, the body-weight model) run on a background worker, so the app stays usable. You can close the browser and come back: a banner at the top lets you resume or cancel.
  8. To reuse the column mapping and settings with another file of the same layout, save them with Save Preset at the bottom of the Data tab and load them next time.

Data Format

Upload data in long format : one row per measurement per animal, as a CSV file (.csv) of up to 30 MB. Commas, semicolons or tabs can separate the columns (with a semicolon or tab, a decimal comma is read correctly), and the file can be UTF-8, Latin-1 or UTF-16.

Demo data

No data of your own? The Data tab ships four demo datasets:

  • Combo Treatment — four arms with survival data, for the Drug Synergy tab.
  • Dose Levels — graded doses with survival, for Dose Response.
  • Weight — four arms with body weight, for Toxicity.
  • All Features — the most complete, and the one that reaches every module: six arms (vehicle, a three-dose series, a second drug and a combination), length/width/height, body weight, dose, two cages per arm, necrosis flags, survival censoring and a removal-reason column (tumour volume limit, weight loss, found dead, end of study).

Column auto-detection

The dashboard detects columns by name. Recognised patterns (case-insensitive):

Role Recognised names Required?
Animal ID ID, Mouse_ID, Animal_ID, Subject, Tag, Ear Tag, Mouse Tag Yes
Time / Day Day, Time, Timepoint, Day_Number Yes ‡
Date Date, Measurement_Date Yes ‡
Tumor volume Volume, Tumor_Volume, TV Yes †
Tumor length Length, Tumor_Length Yes †
Tumor width Width, Tumor_Width Yes †
Tumor height mapped manually in Data Upload No
Treatment / Group Treatment, Group, Treatment_Group Yes
Cage Cage, Cage_ID No
Dose Dose, Dose_Level, Concentration, Conc No
Body weight Weight, Mass, Body_Weight, Body Mass, BW No
Sex Sex, Gender, M/F No
Necrotic flag Necrotic, Necrosis, Necr, Is_Necrotic, Necrotic_Flag No
Removal reason Removal_Reason, Reason, Cause_Of_Death, Fate No
Survival / Censor Survival_Censor, Censor, Event, Survival_Event No

† Provide either a pre-calculated Volume column, or Length + Width columns. When volume is calculated from calipers, six formulas are selectable in Data Upload: ellipsoid (default, π·L·W²/6 ), modified ellipsoid ( L·W²/2 ) — the common preclinical convention — plus 3-axis ellipsoid, cylinder, sphere and box. The 3-axis and box formulas use a Height column if one is mapped; the others approximate height by width. Volumes may be supplied in mm³ or cm³.

‡ Supply either a numeric Day column or a Date column; elapsed days are derived from the date and a reference start date.

Survival/Censor must be numeric: 1 = endpoint reached (e.g. tumour exceeded IACUC volume limit), 0 = censored (alive or study-end censored). If absent, use the volume-threshold derivation in the Survival tab sidebar. An animal's time to event is its last measurement day, or with the volume threshold, the first day its tumour reaches the limit; no separate time column is needed.


Key Concepts

Evaluable days

An arm is evaluable on a day when at least 50% of its enrolled animals, and at least 3 animals, are still on study (measured on that day or later).

Comparisons on one day, or over a window, use only days on which every arm compared is evaluable. The AUC's window ends at the last such day. Dose Response and the therapeutic window use the last evaluable day unless you choose another, and Drug Synergy offers every evaluable day or one of them. Later than that, an arm is represented by too few animals, and by the ones whose tumours grew slowest, for a fair comparison.

Glossary

Reference group
The control arm chosen under Study Design. Comparisons are made against it, or between all pairs of arms where a tab offers that.
Geometric mean
The average taken on the log scale and converted back: the typical tumour volume. A few very large tumours move it far less than they move the ordinary (arithmetic) mean. Model-based estimates of tumour volume are geometric means.
TGI (tumour growth inhibition)
(1 − treated volume / control volume) × 100 at the evaluation day, from each arm's mean volume: the geometric mean under the model-based estimand, the arithmetic mean under the observed ones. 100 % means no tumour left, 0 % the same as control, and a negative value a larger tumour than control.
Estimand
What a number is meant to estimate, decided before looking at the data. For example, each arm's mean volume on day 21 had every animal stayed on study (model-based), or the mean of the animals still present on day 21 (survivors). They differ when animals leave early, and the Toxicity and Drug Synergy sidebars let you choose.
Random effects (random slopes)
Each animal's own starting size and growth rate, around its arm's. They let a model use repeated measurements of the same animal correctly, instead of treating them as independent animals.
Adjusted and unadjusted p-values
Comparing several arms with the control makes several tests, and some p-values will be small by chance. An adjusted p-value (Dunnett, Tukey, Holm, Bonferroni or FDR) accounts for the whole family of comparisons shown, and adjusted intervals do the same for intervals. Report adjusted values when several arms are compared; an unadjusted value suits only a single planned comparison.
Confidence and credible intervals
A 95 % confidence interval (frequentist) comes from a procedure that covers the true value in 95 % of repeated studies. A 95 % credible interval (Bayesian) contains the value with 95 % probability, given the data, the model and its priors.
Hazard ratio and time ratio
A hazard ratio below 1 means the endpoint is reached at a lower rate than in the reference arm. A time ratio above 1 (Bayesian survival) means it takes longer to reach.
Bliss difference
The combination's observed effect minus the effect expected if the two drugs acted independently, in percentage points. Positive suggests synergy, negative antagonism; its 95 % interval decides the verdict.

Analysis Modules

Tumor Growth

LME4 (linear mixed-effects)
Log-transforms tumor volume and fits a linear mixed model with a Treatment × Day interaction. Each animal has its own intercept and growth rate (random slopes), falling back to simpler structures only if the fit fails; the ANOVA reports F-tests with Satterthwaite degrees of freedom. Cage handling is chosen to match the housing design rather than applied uniformly — see the tab's Info panel. Pairwise contrasts via emmeans; the comparison family and its correction are two separate settings, defaulting to vs-reference with an exact Dunnett correction.
AUC (area under the curve)
The area under each arm's fitted growth curve over the window in which every arm is evaluable (at least 50 % of its enrolled animals, and at least 3, on study). Animals that die or are removed early count through all their measurements, so early deaths of large tumours do not pull the AUC down. Arms are compared by the ratio of their AUCs, with intervals and Wald p-values from the model and the selected correction. Each animal's own trapezoidal AUC is shown for description only: an animal removed early accumulates less area from being measured for less time.
Bayesian LMM (brms)
The LME4 model fitted via MCMC (Stan). Returns posterior distributions for treatment effects with credible intervals, plus convergence diagnostics (Rhat, ESS, trace and rank plots) and PSIS-LOO influence measures. Requires the brms package.
Randomisation test (optional, any frequentist mode)
Permutes treatment labels and refits, giving an exact p-value for the Treatment × Day interaction that assumes neither a denominator-df approximation nor normal random effects. The unit permuted, individual animals or whole cages, is the unit of randomisation declared under Study Design on the Data tab — the test is only valid if that matches how the experiment was actually run.
Growth Delay
A per-animal endpoint from the data the run analysed, without the model: T − C, the median time for an arm's tumours to reach a volume you set, minus the reference arm's, with a bootstrap interval, a log-rank test against the reference, and the specific growth delay (T − C in the reference arm's doubling times). A table gives each animal's time.

Survival Analysis

Kaplan-Meier
Non-parametric survival estimates with 95% confidence bands. Log-rank test for group differences.
Cox proportional hazards
Semi-parametric Cox PH model. Hazard ratios with 95% CI for each treatment vs. reference. When whole cages were randomised (Study Design), each arm is instead compared with the reference by a cage-level permutation log-rank, and the hazard ratios have no interval. Firth's bias-reduced estimation is applied automatically when any arm has zero events, where the standard partial likelihood would return infinite coefficients. The proportional-hazards assumption is tested and reported — if it fails, the hazard ratio is an average over a period in which the true ratio changed, and should not be quoted as a single number.
Accelerated failure time (AFT)
A parametric model of the time to the endpoint (Weibull, log-normal or log-logistic, the best by AIC unless you choose). Each arm is compared with the reference as a time ratio: 1.8 means its animals take 1.8 times as long to reach the endpoint. It does not assume proportional hazards, so use it when the Cox model's check rejects them. Diagnostics compares the distributions and draws the fit against Kaplan-Meier.
Bayesian AFT (brms)
Accelerated failure time model with a Weibull or log-normal distribution. Returns time ratios and derived hazard ratios with credible intervals.

Toxicity (Body Weight)

LME4 body weight model
Linear mixed model of body weight over time, with each animal's own intercept and slope. Each arm's rate of weight change (per day) is compared with the reference arm's; weight at the mean study day is reported as well. Tumour mass can be subtracted from body weight first (on by default).
Therapeutic window
Sets each arm's tumor growth inhibition beside its mean worst weight loss (each animal's largest drop below its own first weighing), each with a 95% interval, and flags the weight loss as Tolerated, Not tolerated or Unclear against a threshold you set (default 20%). The two axes are read separately; a single efficacy-toxicity ratio was reported before version 26.09.005 and was withdrawn.
Weight-loss time-to-event
Kaplan-Meier for time to crossing a weight-loss threshold. With a removal-reason column mapped on Data Upload, removals for weight loss count as events, planned ends as censoring, and other removals (such as tumour burden) as competing events, with an Aalen-Johansen cumulative incidence estimate. Without one, an animal whose record ends early is censored, and the tab says so. A per-arm breakdown of how animals left is reported.
Endpoint estimand
Shared by the therapeutic window and the Drug Synergy tab. Model-based (recommended) estimates each arm's geometric mean at the endpoint day from a mixed model of every observation, so animals removed earlier still count. Each animal's last observation uses every animal's final measurement, whatever day it was taken. Survivors uses only animals still present at the endpoint day; because the fastest-growing tumours leave first, it flatters the control arm and understates TGI, and is kept only to reproduce older figures.

Drug Synergy

Bliss Independence
The expected combined effect if the two drugs act independently: with FE = TGI / 100, Expected FE = FE_A + FE_B − FE_A × FE_B. The observed combination is compared with it at one day or across every evaluable day, and the difference is reported in percentage points. Positive = synergy; negative = antagonism.
Why there is no Combination Index
Loewe additivity, which a Combination Index measures, is dose-equivalence: it needs a dose-response curve for each agent, which single-dose designs do not provide. The usual single-dose stand-in is response additivity , which fails the sham-combination test: combining a drug with itself it predicts twice the effect, so an agent at 50 % inhibition should reach 100 %, and when it does not the method calls the drug antagonistic with itself. Measured across the effect grid with the combination set to exactly Bliss-additive, 42 % of cells were labelled antagonistic and none synergistic.
Interval on the verdict
The synergy verdict comes from the 95 % interval for the Bliss excess: Synergy when it lies above zero, Antagonism when it lies below, and Additive (no departure from Bliss detected) when it contains zero. Without intervals the label says so.

Dose-Response

One agent's dose series
A curve describes one agent at several doses against the study's control. Choose the agent's arms in the sidebar; other agents, combinations, and arms sharing a dose cannot join the same curve.
Hill / Emax
Fits a four-parameter log-logistic (4PL) dose-response curve via nonlinear least squares; with six or more doses a five-parameter (asymmetric) curve is also tried and kept when AIC prefers it. The lower asymptote cannot fall below zero. Reports the EC50 (the dose giving half the fitted response range) with a 95% interval on log dose, the Hill slope, and the fitted curve; an EC50 outside the tested doses is flagged.
Trend tests
Alongside the curve fit, the tab reports linear and quadratic trend p-values and a two-sided Jonckheere-Terpstra test. The last is the one that respects dose ordering: it asks whether response is monotone in dose, which is the question a dose-response study is usually posing, rather than merely whether the groups differ.

Power Analysis

Analytic (a priori)
Classical power for each treated-vs-control comparison (two-sample t-test), with the alpha split across the comparisons a multi-arm study makes (Dunnett or Bonferroni) and an optional attrition allowance, across a range of effect sizes and sample sizes.
Survival endpoint (a priori)
Under the analytic mode, choose "Hazard ratio" to plan the time to an event, such as the tumour volume limit, compared by the log-rank test. Schoenfeld's approximation gives the events one comparison needs from the hazard ratio, and the animals per arm from the share of controls expected to have the event by the end. It is asymptotic, so with few events treat the N as a minimum.
LMM simulation
Simulates repeated-measures data under a linear mixed model and estimates power empirically across a grid of sample sizes. An optional euthanasia volume limit models attrition; leaving it blank assumes every animal is measured at every timepoint, which overstates power by roughly 6 points at 30 % attrition. Note that baseline SD and random-intercept SD do not affect the answer — the test is on growth rates, which are orthogonal to per-animal baselines. Random-slope SD is the parameter that matters.

Results

Every run, kept
Each analysis you run is listed on the Results tab, newest first, until the session ends. Open one to show it again in its tab, with its tables, plots and Download Results (the sidebar keeps its current settings). Select several to download them together: one folder each, with what the tab's Download Results offers, the plots drawn with the tab's current settings.
No refit for the same fit
A Bayesian fit with the same data and settings as one already run in the session is shown again rather than refitted; the seed is fixed, so a refit would give the same results.

Troubleshooting

File won't upload
Save it as CSV (.csv), up to 30 MB. If a row is reported with the wrong number of fields, a value probably contains the separator, such as a comma inside a treatment name: quote that value or remove the comma.
Columns not auto-detected
Use the column mapping dropdowns in the Data tab to assign columns manually.
Analysis tab is greyed out
Upload and configure data first. Survival requires a censor column (or enable volume-threshold derivation). Toxicity requires a body weight column. Dose Response requires a dose column. Drug Synergy requires at least four treatment groups.
Analysis fails or produces no output
Verify that the selected columns contain valid, non-missing data. Tumor volumes must be positive. Time points must be numeric. At least two groups must be present for any comparison.
Bayesian model is slow to start
The first fit of a model compiles it, which takes about two minutes on mouseexperiment.com and less on a fast computer; later fits with the same structure reuse it and usually take under a minute. On mouseexperiment.com compiled models are kept across restarts and updates of the server, so each structure is compiled once; on your own computer, see ?clear_compiled_model_cache to keep them between R sessions. Fits run one at a time on a background worker, so other users and tabs stay responsive, and yours may wait for one ahead of it.
Plot download produces a blank file
Run the analysis first, then click the download button. Download Results draws every plot of the run with the current settings (colours, titles, sizes), including plot types you have not opened.

About

mouseExperiment

Version 0.59.1
Built 2026-10-30
Source github.com/sciOmics/mouseExperiment (MIT licence)
Run it yourself remotes::install_github("sciOmics/mouseExperiment") , then mouseExperiment::run_dashboard() . Your data then never leaves your computer.

Contact

Report bugs, request features and ask questions on GitHub Issues . Issues are public, so please don't attach unpublished study data: a few made-up rows, or a demo dataset, that show the problem are enough.

Usage analytics

mouseexperiment.com counts page views and which features are used, with self-hosted Umami analytics: no cookies, and nothing sent to a third party. Events carry only the analysis type, the method, counts such as the number of rows, and yes/no flags. They never include your data, file names, column names, treatment names or error messages. Uploaded data stays in your session on this server.