I build statistical models that turn incomplete, noisy data into estimates and predictions with honest uncertainty. Typical questions: how a drug's concentration evolves across a patient population and how much individuals differ; which readings in a sensor network are implausible and why; how regions compare once small sample sizes are taken into account. Results are delivered as reproducible analyses, reports and interactive dashboards, as open-source software.
Modelling showcases
Population PK/PD: warfarin
How do drug concentrations and a biomarkerA measured quantity that follows the effect of a drug, such as a blood count. A safety marker is a biomarker that shows harm.Glossary → respond to a dose, how much do patients differ, and which daily dose keeps most patients in range? A population modelA model of a whole group of patients: the typical patient, and how much individual patients differ from that.Glossary → of the classic warfarin data: plasma concentrations and clotting activity in 32 people after one dose.
Plasma warfarin after one dose of 1.5 mg/kg: measurements (dots) and each subject's fitted curve (medianThe middle value: half of the plausible values lie below it and half above.Glossary →, 50% and 95% intervalsThe range that contains the unknown value with the stated probability, according to the model. The plots show a 50% interval (narrow band) and a 95% interval (wide band).Glossary →).
Clotting activity falls with a delay of about two days and recovers as warfarin is cleared; the turnover modelA model of a quantity that the body produces and removes all the time. A drug that slows the production makes the quantity fall with a delay, and it recovers when the drug is gone.Glossary → captures both.
Simulated from the fitted population: share of new patients whose clotting activity is in an illustrative target range after 14 days of a daily dose (highest at 10 mg), and the share pushed below it (medianThe middle value: half of the plausible values lie below it and half above.Glossary →, 50% and 95% intervalsThe range that contains the unknown value with the stated probability, according to the model. The plots show a 50% interval (narrow band) and a 95% interval (wide band).Glossary → over the posteriorWhat a Bayesian model says about an unknown quantity after it has seen the data: the plausible values and how plausible each is.Glossary →).
Data: O'Reilly & Aggeler (1963, 1968), as distributed in the nlmixr2data R package.
health
pharmacometrics
multilevel
dosing
Regional trends across Germany's districts
How fast is the share of young children receiving integration assistance changing in each district? Six yearly values give each district a noisyThe chance variation in data: measurement error and everything else a model does not explain.Glossary → trend; a Fay–Herriot-type small-area modelA multilevel model for regions that each have few observations. The Fay–Herriot model is the classic form.Glossary →poolsEstimating each unit partly from its own data and partly from all units together. A unit with little data leans more on the others, which steadies its estimate.Glossary → each district towards its state as far as its data warrant.
Estimated change per year, 2020–2025 (posteriorWhat a Bayesian model says about an unknown quantity after it has seen the data: the plausible values and how plausible each is.Glossary →medianThe middle value: half of the plausible values lie below it and half above.Glossary →); hover a district for its raw trend and 95% intervalThe range that contains the unknown value with the stated probability, according to the model. The plots show a 50% interval (narrow band) and a 95% interval (wide band).Glossary →. The middle 95% of raw district trends spans −18% to +29% per year; pooledEstimating each unit partly from its own data and partly from all units together. A unit with little data leans more on the others, which steadies its estimate.Glossary →, −10% to +20%. Colours stop at ±20% per year; 11 districts lie beyond. Published counts are rounded to multiples of 5.
How fast is Germany warming, and how sure can we be? A smooth trend fitted to annual means since 1881 and continued ten years ahead with its uncertainty.
Dots: Germany's annual mean temperature, 1881–2025. Line and bands: the estimated trend (medianThe middle value: half of the plausible values lie below it and half above.Glossary →, 50% and 95% intervalsThe range that contains the unknown value with the stated probability, according to the model. The plots show a 50% interval (narrow band) and a 95% interval (wide band).Glossary →), about 2.6 °C warmer than in 1881, continued to 2035. The continuation is a statistical extrapolation of the trend, not a climate projection.
Has every part of Germany warmed alike? Each state's smoothed temperature trend against its own 1961–1990 level, as an animated map from 1881 to today.
Each state's smoothed temperature trend (posteriorWhat a Bayesian model says about an unknown quantity after it has seen the data: the plausible values and how plausible each is.Glossary →medianThe middle value: half of the plausible values lie below it and half above.Glossary →), relative to its own 1961–1990 level; press play to step through the years from 1881. By 2025 every state had warmed by 1.8 to 2.0 K. Berlin, Hamburg and Bremen share DWD's series with their neighbours.
Plausibility checks in a radiation monitoring network
Which readings of Germany's ambient gamma dose rateThe level of gamma radiation in the open air, from natural and artificial sources, measured continuously by a network of probes.Glossary → network need a second look? A Bayesian modelA statistical model that describes everything unknown with probabilities and updates them with the data. Its results are ranges of plausible values, not single numbers.Glossary → of each probe (baselineThe level of a quantity before anything acts on it: a patient's value before the dose, or a probe's usual reading without rain.Glossary →, daily cycle, rain washoutRain carries naturally occurring radioactive particles from the air to the ground, which raises the measured dose rate for a few hours.Glossary →, rare unexplained readings) separates rain peaks from readings it cannot explain.
Kehl ST Auenheim, 16 October 2026 to 23 October 2026: hourly readings (dots), the model's expected level (medianThe middle value: half of the plausible values lie below it and half above.Glossary →, 50% and 95% intervalsThe range that contains the unknown value with the stated probability, according to the model. The plots show a 50% interval (narrow band) and a 95% interval (wide band).Glossary →, blue) and the level without rain washoutRain carries naturally occurring radioactive particles from the air to the ground, which raises the measured dose rate for a few hours.Glossary → (dashed); dotted red: a fixed threshold (median plus 3 robust standard deviationsA measure of spread. For bell-shaped data about two thirds of the values lie within one standard deviation of the mean. Abbreviated SD.Glossary →). The reading is above the threshold in 15 hours, 14 of them during rain, which the model explains; it flags 1 hour it cannot explain (red). The last 33 hours came after the model's fit and are assessed against it. Below: the radar rain class at the probe.
Data: Bundesamt für Strahlenschutz (BfS), ODL-Info open data, hourly dose rate and radar precipitation, 16–23 October 2026; model results as of 23 October 2026.
environment
monitoring
multilevel
anomalies
Fewer false pre-alarms
A fixed threshold raises a pre-alarmAn automatic notice that a probe's reading is above its threshold and needs a second look.Glossary → whenever rain lifts the reading. Across the whole network, the model explains rain peaks and keeps only the pre-alarms it cannot explain.
Probes with at least one pre-alarmAn automatic notice that a probe's reading is above its threshold and needs a second look.Glossary → under a fixed threshold rule in the week to 23 October 2026 (1,562 of 1,564). The model keeps pre-alarms at 594 of them (red) and explains the others by rain. Of 21,872 hours above the threshold it leaves 1,603 (7%) unexplained; 33 of 90 pre-alarm episodes remain.
Did a programme change the outcome, and when? Municipalities adopt it in different years; an event-study modelA comparison of units that adopt a programme with units that have not adopted it, year by year relative to the year of adoption. Estimates near zero before adoption support the comparison.Glossary → compares adopters with non-adopters year by year.
Estimated programme effect by year relative to adoption (medianThe middle value: half of the plausible values lie below it and half above.Glossary →, 50% and 95% intervalsThe range that contains the unknown value with the stated probability, according to the model. The plots show a 50% interval (narrow band) and a 95% interval (wide band).Glossary →); diamonds: the effects the data were simulated with. Before adoption the estimates stay near zero, which checks the comparison: 30 adopting and 30 non-adopting municipalities.
Synthetic data: 60 simulated municipalities over 12 years.
A smooth curve with no fixed formula, and noiseThe chance variation in data: measurement error and everything else a model does not explain.Glossary → whose size changes over time: a heteroscedasticWith noise whose size is not constant, for example larger at some times than at others.Glossary →Gaussian processA model for an unknown smooth curve. It assumes how similar the curve is at nearby points, not a formula for its shape. Abbreviated GP.Glossary → for the classic motorcycle crash data, approximated with 20 Hilbert-space basis functionsA set of fixed wave-shaped curves whose weighted sum approximates a Gaussian process at a fraction of the computing cost. Each curve has one basis weight.Glossary →.
Head acceleration after the impact (dots) and the estimated mean (medianThe middle value: half of the plausible values lie below it and half above.Glossary →, 50% and 95% intervalsThe range that contains the unknown value with the stated probability, according to the model. The plots show a 50% interval (narrow band) and a 95% interval (wide band).Glossary →).
The noiseThe chance variation in data: measurement error and everything else a model does not explain.Glossary →SDA measure of spread. For bell-shaped data about two thirds of the values lie within one standard deviation of the mean. Abbreviated SD.Glossary → changes over time too: near zero before the impact, largest around 30 ms. A second Gaussian processA model for an unknown smooth curve. It assumes how similar the curve is at nearby points, not a formula for its shape. Abbreviated GP.Glossary → models its logarithm.
Data: motorcycle crash test, Silverman (1985), as distributed in the MASS R package (GPL-2 | GPL-3).
What to do with lab values reported as "below the LLOQLower limit of quantification: the smallest concentration a laboratory method measures reliably. Lower values are reported only as "below the LLOQ".Glossary →"? Dropping them, or replacing them by half the limit, biases a dose-response analysis; treating them as censoredKnown only to lie below or above a limit. A censored analysis uses exactly that information instead of dropping the value or inventing one.Glossary → recovers the truth. A simulation studyAn experiment on invented data whose truth is known, to see how well a method recovers it.Glossary → with three ways to fit the same data.
One subject at the highest dose: true level (line), measurements (circles; squares are below the LLOQLower limit of quantification: the smallest concentration a laboratory method measures reliably. Lower values are reported only as "below the LLOQ".Glossary → of 0.1, drawn at the LLOQ) and each method's estimate (medianThe middle value: half of the plausible values lie below it and half above.Glossary →, 50% and 95% intervalsThe range that contains the unknown value with the stated probability, according to the model. The plots show a 50% interval (narrow band) and a 95% interval (wide band).Glossary →).
Calibration over all 495 samples after the dose: a method whose curve follows the diagonal places the true level where its posteriorWhat a Bayesian model says about an unknown quantity after it has seen the data: the plausible values and how plausible each is.Glossary → says it should be. The central 95% intervalsThe range that contains the unknown value with the stated probability, according to the model. The plots show a 50% interval (narrow band) and a 95% interval (wide band).Glossary → contain the true level for 77% (M1), 94% (M3), 60% (M5) of the samples.
Error of each method's estimate of the average largest drop below baselineThe level of a quantity before anything acts on it: a patient's value before the dose, or a probe's usual reading without rain.Glossary → per dose group (medianThe middle value: half of the plausible values lie below it and half above.Glossary →, 50% and 95% intervalsThe range that contains the unknown value with the stated probability, according to the model. The plots show a 50% interval (narrow band) and a 95% interval (wide band).Glossary →); the dashed line is the truth. Samples below the LLOQLower limit of quantification: the smallest concentration a laboratory method measures reliably. Lower values are reported only as "below the LLOQ".Glossary →: 5%, 26%, 50%, 73%, 78% by dose.
Simulated data (55 subjects in five dose groups, ten samples each).
Why does the samplerThe algorithm that produces the posterior draws of a Bayesian model.Glossary → report divergent transitionsA move of the sampler that failed numerically. It warns that the draws may miss a part of the posterior.Glossary → for the motorcycle model, and does it help to learn, for each basis weightA set of fixed wave-shaped curves whose weighted sum approximates a Gaussian process at a fraction of the computing cost. Each curve has one basis weight.Glossary →, how far to centre it? The same model sampled three ways, with 40 chainsOne independent run of the sampler. Several chains that agree are evidence that the draws describe the posterior.Glossary → at three target acceptance ratesA setting of the sampler: a higher value makes it take smaller, more careful steps, at the price of more computation.Glossary →.
The centerednessTwo equivalent ways of writing a multilevel model. Which one the sampler handles well depends on how much the data say about each unit; partly centered forms lie in between.Glossary → (0 non-centered, 1 centered) of each basis weightA set of fixed wave-shaped curves whose weighted sum approximates a Gaussian process at a fraction of the computing cost. Each curve has one basis weight.Glossary →: medianThe middle value: half of the plausible values lie below it and half above.Glossary →, 50% and 95% intervalsThe range that contains the unknown value with the stated probability, according to the model. The plots show a 50% interval (narrow band) and a 95% interval (wide band).Glossary → over the 120 chainsOne independent run of the sampler. Several chains that agree are evidence that the draws describe the posterior.Glossary → that adapted it in their own warm-upThe first part of a chain, in which the sampler tunes itself to the model. Its draws are not used.Glossary → (40 at each target acceptance rateA setting of the sampler: a higher value makes it take smaller, more careful steps, at the price of more computation.Glossary →), and the value learned once from one chain of 10,000 drawsOne plausible set of values for a model's unknowns, drawn from the posterior. A few thousand draws together describe the posterior.Glossary →, the post's protocol (dots). Low-frequency weights, which the data pin down, tend to be centered; high-frequency ones, which the data barely inform, stay non-centered, most clearly for the noiseThe chance variation in data: measurement error and everything else a model does not explain.Glossary →SDA measure of spread. For bell-shaped data about two thirds of the values lie within one standard deviation of the mean. Abbreviated SD.Glossary →.
Effective drawsHow many independent draws a chain's draws are worth. Successive draws of a chain resemble each other, so the number is smaller than the number of draws.Glossary → of the worst-mixingHow freely a chain moves through the plausible values. A chain that mixes badly stays in one region for long and needs many more draws.Glossary →GPA model for an unknown smooth curve. It assumes how similar the curve is at nearby points, not a formula for its shape. Abbreviated GP.Glossary →hyperparameterA number that sets the general behaviour of a model component and not a single value of it. For a Gaussian process: the amplitude and the length scale.Glossary → (the amplitudesHow far a Gaussian-process curve typically moves away from its mean.Glossary → and length scalesHow far apart two points have to be before the curve can differ noticeably between them.Glossary →) per 1000 gradient evaluationsThe slope of the posterior at one point. Samplers evaluate it at every step, so the number of gradient evaluations measures the computing cost.Glossary →, chainOne independent run of the sampler. Several chains that agree are evidence that the draws describe the posterior.Glossary → by chain: medianThe middle value: half of the plausible values lie below it and half above.Glossary →, 50% and 95% intervalsThe range that contains the unknown value with the stated probability, according to the model. The plots show a 50% interval (narrow band) and a 95% interval (wide band).Glossary → over each method's chains of 1000 drawsOne plausible set of values for a model's unknowns, drawn from the posterior. A few thousand draws together describe the posterior.Glossary →. Medians at a target acceptance rateA setting of the sampler: a higher value makes it take smaller, more careful steps, at the price of more computation.Glossary → of 0.8: adapted in each chain's warm-upThe first part of a chain, in which the sampler tunes itself to the model. Its draws are not used.Glossary → 2.4, non-centeredTwo equivalent ways of writing a multilevel model. Which one the sampler handles well depends on how much the data say about each unit; partly centered forms lie in between.Glossary → 1.6, learned once 1.2. One chain that adapted its centering in its own warm-up, at a target acceptance rate of 0.9, settled at a length scale of the mean curve about 2 times that of the other chains and stayed there. It is left out here: mixing within such a chain says nothing about the region it misses.
Data: motorcycle crash test, Silverman (1985), as distributed in the MASS R package (GPL-2 | GPL-3).
statistics
sampling
gaussian-processes
Dose finding with a PK/PD model
A phase 1 cancer trialThe first trial of a new drug in patients. In oncology it searches for the highest dose patients tolerate, in a few dozen patients.Glossary → searches for the highest tolerable doseThe highest dose at which the share of patients with a dose-limiting toxicity stays below an agreed limit; 33% in the study shown here. Abbreviated MTD.Glossary →, usually by counting dose-limiting toxicitiesA side effect severe enough to limit the dose a patient can be given. Abbreviated DLT.Glossary →. Does the search do better when it models what the trial measures anyway: drug concentrations and a safety markerA measured quantity that follows the effect of a drug, such as a blood count. A safety marker is a biomarker that shows harm.Glossary →? A published simulation studyAn experiment on invented data whose truth is known, to see how well a method recovers it.Glossary →, rebuilt on simulated patients and extended with a population PK/PDPharmacokinetics and pharmacodynamics. Pharmacokinetics is what the body does to the drug (its concentration over time); pharmacodynamics is what the drug does to the body (its effects). A PK/PD model describes both.Glossary → model that is refitted after every cohortA small group of patients treated at the same dose, here three, before the dose of the next group is decided.Glossary →.
What a phase 1 trialThe first trial of a new drug in patients. In oncology it searches for the highest dose patients tolerate, in a few dozen patients.Glossary → measures besides toxicities: a safety markerA measured quantity that follows the effect of a drug, such as a blood count. A safety marker is a biomarker that shows harm.Glossary → (think of a neutrophil countThe number of neutrophils in the blood, the white blood cells that fight infections. Many cancer drugs lower it, and a very low count is dangerous.Glossary →) of 10 simulated patients at 200 mg. A value below the dotted line is a dose-limiting toxicityA side effect severe enough to limit the dose a patient can be given. Abbreviated DLT.Glossary → (DLT), here in 3 patients. Usual designs keep only that yes or no.
A population PK/PDPharmacokinetics and pharmacodynamics. Pharmacokinetics is what the body does to the drug (its concentration over time); pharmacodynamics is what the drug does to the body (its effects). A PK/PD model describes both.Glossary → model, refitted after every cohortA small group of patients treated at the same dose, here three, before the dose of the next group is decided.Glossary → of three patients, turns concentrations and marker values into the DLTA side effect severe enough to limit the dose a patient can be given. Abbreviated DLT.Glossary → rate of the next patient at each dose (medianThe middle value: half of the plausible values lie below it and half above.Glossary →, 50% and 95% intervalsThe range that contains the unknown value with the stated probability, according to the model. The plots show a 50% interval (narrow band) and a 95% interval (wide band).Glossary → after the 24 patients of one simulated trial; red: the true rates).
Chance that a safe and effective drug ends with a positive phase 3 trialThe large trial that compares the drug with the standard treatment, usually in hundreds of patients. A positive phase 3 trial is the basis of an approval.Glossary →, by phase 1 designThe rule that decides the dose of the next cohort and the dose recommended at the end of a phase 1 trial.Glossary →, averaged over the 8 toxicity scenariosOne assumed truth of a simulation: the probability of a dose-limiting toxicity at each dose.Glossary → of Ruppert & Shoben (2018). Grey: designs that read only whether a toxicity occurred (exact). Blue: the PK/PDPharmacokinetics and pharmacodynamics. Pharmacokinetics is what the body does to the drug (its concentration over time); pharmacodynamics is what the drug does to the body (its effects). A PK/PD model describes both.Glossary → design (medianThe middle value: half of the plausible values lie below it and half above.Glossary →, 50% and 95% intervalsThe range that contains the unknown value with the stated probability, according to the model. The plots show a 50% interval (narrow band) and a 95% interval (wide band).Glossary → over 8,000 simulated trials). It recommends a too-toxic dose in 11.7 [11.0, 12.3]% of trials (BOINBayesian optimal interval design: a phase 1 design that raises, keeps or lowers the dose by comparing the toxicity rate observed at the current dose with two fixed boundaries.Glossary →: 22.6%). The gain needs a recommendation rule that uses the fit's uncertainty: with the paper's own rule the design reaches 50.5 [49.8, 51.2]%, no more than BOIN's 50.2%, and in 1 of the 8 scenarios it does worse than BOIN under both rules.
At a mountain probe the ambient gamma dose rateThe level of gamma radiation in the open air, from natural and artificial sources, measured continuously by a network of probes.Glossary → drops by up to a third in winter: snow absorbs the radiation that comes from the ground. Nobody measures the snow at the probe. A Bayesian modelA statistical model that describes everything unknown with probabilities and updates them with the data. Its results are ranges of plausible values, not single numbers.Glossary → follows it from precipitation and temperature, learns from ten winters of readings how strongly it shields, and predicts the next winter's readings from the weather alone.
Dots: daily mean dose rateThe level of gamma radiation in the open air, from natural and artificial sources, measured continuously by a network of probes.Glossary → at the probe Schauinsland (1,195 m) in a winter the model was not fitted to. Band: what the model predicts for each day from precipitation and temperature alone (medianThe middle value: half of the plausible values lie below it and half above.Glossary →, 50% and 95% intervalsThe range that contains the unknown value with the stated probability, according to the model. The plots show a 50% interval (narrow band) and a 95% interval (wide band).Glossary →). Dashed: the same prediction without the snow. Since 1 November 2025, 353 of 362 readings lie inside their 95% interval.
Nobody measures the snow at the probe. The model follows it day by day: precipitation falls as snow when it is cold, and melts with warmth and sun. Shown is the water held as snow in that winter (medianThe middle value: half of the plausible values lie below it and half above.Glossary →, 50% and 95% intervalsThe range that contains the unknown value with the stated probability, according to the model. The plots show a 50% interval (narrow band) and a 95% interval (wide band).Glossary →); at its most it took about 44 nSv/h from the reading.
The share of the ground's gamma radiation that passes a layer of snow, by the water it holds (medianThe middle value: half of the plausible values lie below it and half above.Glossary →, 50% and 95% intervalsThe range that contains the unknown value with the stated probability, according to the model. The plots show a 50% interval (narrow band) and a 95% interval (wide band).Glossary →). Half is absorbed by 92 [65, 134] mm of water, about 31 cm of settled snow. Grey: the range assumed before the fit, from published measurements. Ten winters of readings narrow it.
Data: Bundesamt für Strahlenschutz (BfS), ODL-Info, daily mean dose rates of the probe Schauinsland, October 2015 to November 2026. Datenbasis: Deutscher Wetterdienst, eigene Bearbeitung (stations Feldberg, Breitnau and Freiburg).
environment
monitoring
time-series
physics
Open-source software
BayesianRegressionModels.jl
Regression models written as brms-style formulasA widely used R package in which a regression model is written as a short formula.Glossary →, compiled to StanA widely used programming language and toolkit for Bayesian models.Glossary →.
Write and compose StanA widely used programming language and toolkit for Bayesian models.Glossary → models in JuliaA programming language for scientific computing. The open-source packages on this page are written in it.Glossary →: models compile to Stan and can be sampled or optimised directly.
Adaptive warmupThe first part of a chain, in which the sampler tunes itself to the model. Its draws are not used.Glossary → for Hamiltonian Monte CarloThe standard sampler for models with many unknowns. It moves through the plausible values by following the slope of the posterior.Glossary →, including automatic reparametrizationWriting the same model in other coordinates, so that the sampler can move through it more easily. The results do not change.Glossary → of posteriorsWhat a Bayesian model says about an unknown quantity after it has seen the data: the plausible values and how plausible each is.Glossary → that are hard to sample.
Estimators for nonlinear mixed-effects modelsAnother name for a multilevel model, usual in pharmacometrics. Nonlinear means that the model describes a curve, such as a drug concentration over time, not a straight line.Glossary → that work with any model providing a small set of functions.
Posterior drawsOne plausible set of values for a model's unknowns, drawn from the posterior. A few thousand draws together describe the posterior.Glossary → and model predictions as named, nested arrays, summarised over named axes (draws, chainsOne independent run of the sampler. Several chains that agree are evidence that the draws describe the posterior.Glossary →, subjects) without flattening them into long tables. The intervalsThe range that contains the unknown value with the stated probability, according to the model. The plots show a 50% interval (narrow band) and a 95% interval (wide band).Glossary → on this page are computed with it.
A fast, allocation-free implementation of the No-U-Turn samplerA variant of Hamiltonian Monte Carlo that chooses the length of each move by itself. Abbreviated NUTS.Glossary →.
since 2025Independent consultant: Bayesian pharmacometric modellingGenerable Inc.2022 – 2026Postdoctoral research in probabilistic modellingAalto University; University of Cambridge2021PhD in MathematicsRWTH Aachen University2017BSc and MSc in Computational Engineering ScienceRWTH Aachen University
Glossary
Statistics and Bayesian modelling
Bayesian model
A statistical model that describes everything unknown with probabilities and updates them with the data. Its results are ranges of plausible values, not single numbers.See also: Prior, Posterior
Prior
What a Bayesian model assumes about an unknown quantity before it sees the data.
Posterior
What a Bayesian model says about an unknown quantity after it has seen the data: the plausible values and how plausible each is.
Posterior draw
One plausible set of values for a model's unknowns, drawn from the posterior. A few thousand draws together describe the posterior.
Median
The middle value: half of the plausible values lie below it and half above.
Interval
The range that contains the unknown value with the stated probability, according to the model. The plots show a 50% interval (narrow band) and a 95% interval (wide band).
Student-t distribution
A bell-shaped distribution like the normal distribution, but one that makes values far from the centre more likely.
Noise
The chance variation in data: measurement error and everything else a model does not explain.
Standard deviation
A measure of spread. For bell-shaped data about two thirds of the values lie within one standard deviation of the mean. Abbreviated SD.
Standard error
The typical size of the chance error of an estimate.
Percentage points
The difference between two percentages: from 50% to 56% is 6 percentage points.
Multilevel model
A model for many similar units, such as districts, patients or probes. It estimates each unit from its own data and from what the units have in common.See also: Pooling
Pooling
Estimating each unit partly from its own data and partly from all units together. A unit with little data leans more on the others, which steadies its estimate.
Mixed-effects model
Another name for a multilevel model, usual in pharmacometrics. Nonlinear means that the model describes a curve, such as a drug concentration over time, not a straight line.See also: Multilevel model
Small-area model
A multilevel model for regions that each have few observations. The Fay–Herriot model is the classic form.See also: Multilevel model, Pooling
Heteroscedastic
With noise whose size is not constant, for example larger at some times than at others.
Censored
Known only to lie below or above a limit. A censored analysis uses exactly that information instead of dropping the value or inventing one.
Logistic model
A model for a yes-or-no outcome. Here: the probability of a dose-limiting toxicity as a smooth S-shaped curve over the dose, with two unknown numbers.
Simulation study
An experiment on invented data whose truth is known, to see how well a method recovers it.
Monte Carlo estimate
A number computed from random draws. It carries a chance error that shrinks as the number of draws grows.
Sampling and computation
Sampler
The algorithm that produces the posterior draws of a Bayesian model.
Hamiltonian Monte Carlo
The standard sampler for models with many unknowns. It moves through the plausible values by following the slope of the posterior.
No-U-Turn sampler
A variant of Hamiltonian Monte Carlo that chooses the length of each move by itself. Abbreviated NUTS.
Chain
One independent run of the sampler. Several chains that agree are evidence that the draws describe the posterior.
Warm-up
The first part of a chain, in which the sampler tunes itself to the model. Its draws are not used.
Divergent transition
A move of the sampler that failed numerically. It warns that the draws may miss a part of the posterior.
R-hat
A check that the chains of a fit agree. Values near 1 are good; above 1.05 the fit should not be trusted without a closer look.
Effective draws
How many independent draws a chain's draws are worth. Successive draws of a chain resemble each other, so the number is smaller than the number of draws.
Mixing
How freely a chain moves through the plausible values. A chain that mixes badly stays in one region for long and needs many more draws.
Target acceptance rate
A setting of the sampler: a higher value makes it take smaller, more careful steps, at the price of more computation.
Centered and non-centered
Two equivalent ways of writing a multilevel model. Which one the sampler handles well depends on how much the data say about each unit; partly centered forms lie in between.
Reparametrization
Writing the same model in other coordinates, so that the sampler can move through it more easily. The results do not change.
Gradient
The slope of the posterior at one point. Samplers evaluate it at every step, so the number of gradient evaluations measures the computing cost.
Drugs and clinical trials
Pharmacometrics
The quantitative modelling of how drugs behave in patients: concentrations, effects and how both differ between people.
PK/PD
Pharmacokinetics and pharmacodynamics. Pharmacokinetics is what the body does to the drug (its concentration over time); pharmacodynamics is what the drug does to the body (its effects). A PK/PD model describes both.See also: Population model
Population model
A model of a whole group of patients: the typical patient, and how much individual patients differ from that.
One-compartment model
The simplest model of a drug in the body: one well-mixed volume that the drug enters and leaves.
Clearance
The volume of blood the body frees of the drug per unit of time. At the same dose, a patient with a low clearance has more drug in the body for longer.
Exposure
How much drug a patient's body sees. Here: the dose divided by the patient's clearance.
Turnover model
A model of a quantity that the body produces and removes all the time. A drug that slows the production makes the quantity fall with a delay, and it recovers when the drug is gone.
Baseline
The level of a quantity before anything acts on it: a patient's value before the dose, or a probe's usual reading without rain.
Biomarker
A measured quantity that follows the effect of a drug, such as a blood count. A safety marker is a biomarker that shows harm.
Neutrophil count
The number of neutrophils in the blood, the white blood cells that fight infections. Many cancer drugs lower it, and a very low count is dangerous.
LLOQ
Lower limit of quantification: the smallest concentration a laboratory method measures reliably. Lower values are reported only as "below the LLOQ".
Phase 1 trial
The first trial of a new drug in patients. In oncology it searches for the highest dose patients tolerate, in a few dozen patients.See also: Phase 2 trial, Phase 3 trial
Phase 2 trial
A trial in a few dozen patients that looks for first signs that the drug works, at the dose phase 1 recommended.
Phase 3 trial
The large trial that compares the drug with the standard treatment, usually in hundreds of patients. A positive phase 3 trial is the basis of an approval.
Dose-limiting toxicity
A side effect severe enough to limit the dose a patient can be given. Abbreviated DLT.
Maximum tolerated dose
The highest dose at which the share of patients with a dose-limiting toxicity stays below an agreed limit; 33% in the study shown here. Abbreviated MTD.
Cohort
A small group of patients treated at the same dose, here three, before the dose of the next group is decided.
Phase 1 design
The rule that decides the dose of the next cohort and the dose recommended at the end of a phase 1 trial.
3+3 design
The traditional phase 1 design. Three patients get a dose: after no toxicity the dose rises, after one toxicity three more patients get the same dose, and after two the dose counts as too high.
BOIN
Bayesian optimal interval design: a phase 1 design that raises, keeps or lowers the dose by comparing the toxicity rate observed at the current dose with two fixed boundaries.
CCD
Cumulative cohort design: a phase 1 design that raises, keeps or lowers the dose according to the toxicity rate observed so far at the current dose.
mTPI
Modified toxicity probability interval design: a phase 1 design that raises, keeps or lowers the dose according to which range of toxicity rates the data at the current dose support most: too low, on target or too high. mTPI-2 divides the ranges more finely.
CRM
Continual reassessment method: a phase 1 design that fits a simple dose-toxicity curve to the toxicities seen so far and gives the next cohort the dose whose estimated toxicity rate is closest to the target.
Toxicity profile
One assumed truth of a simulation: the probability of a dose-limiting toxicity at each dose.
Efficacy profile
One assumed truth of a simulation: how well the drug works at each dose.
Overall success
The probability that a drug development program succeeds: phase 1 recommends a dose that is not too toxic, and the phase 2 and phase 3 trials at that dose are both positive.
Operating characteristics
How a trial design behaves over many repetitions of the trial: how often it recommends each dose and how many patients it exposes to toxicity.
Monitoring and environment
Ambient gamma dose rate
The level of gamma radiation in the open air, from natural and artificial sources, measured continuously by a network of probes.
Rain washout
Rain carries naturally occurring radioactive particles from the air to the ground, which raises the measured dose rate for a few hours.See also: Radon decay products
Radon decay products
Short-lived radioactive atoms that form in the air when the natural gas radon decays, mainly lead-214 and bismuth-214. They emit gamma radiation and stick to dust.
Cosmic radiation
Radiation that reaches the ground from space. There is more of it at altitude, where less air lies above.
Snow water
The water a snow cover holds, given as the depth in mm of the layer it would form when melted (the snow water equivalent). Unlike the snow depth, it stays the same when the snow settles.
Attenuation
The weakening of radiation on its way through matter, such as snow or water.
Nanosievert per hour
The unit of the dose rate on this site, written nSv/h. Natural levels in Germany lie roughly between 50 and 200 nSv/h.
Pre-alarm
An automatic notice that a probe's reading is above its threshold and needs a second look.
Evaluation
Event study
A comparison of units that adopt a programme with units that have not adopted it, year by year relative to the year of adoption. Estimates near zero before adoption support the comparison.
Gaussian processes
Gaussian process
A model for an unknown smooth curve. It assumes how similar the curve is at nearby points, not a formula for its shape. Abbreviated GP.
Hilbert-space basis functions
A set of fixed wave-shaped curves whose weighted sum approximates a Gaussian process at a fraction of the computing cost. Each curve has one basis weight.
Hyperparameter
A number that sets the general behaviour of a model component and not a single value of it. For a Gaussian process: the amplitude and the length scale.
Amplitude
How far a Gaussian-process curve typically moves away from its mean.
Length scale
How far apart two points have to be before the curve can differ noticeably between them.
Software
Stan
A widely used programming language and toolkit for Bayesian models.
Julia
A programming language for scientific computing. The open-source packages on this page are written in it.
brms
A widely used R package in which a regression model is written as a short formula.
This website sets no cookies, uses no tracking or analytics, and loads no content from third parties. It is hosted on GitHub Pages; when you visit, GitHub Inc. processes technical access data such as your IP address to deliver the site and protect it from misuse (Art. 6(1)(f) GDPR); see GitHub's privacy statement. Controller: Nikolas Siccha (see legal notice). You have the rights of access, rectification, erasure, restriction and objection, and the right to lodge a complaint with a supervisory authority.