Portrait of Nikolas Siccha

Nikolas Siccha

Bayesian statistical modelling and data analysis

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

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 noisy trend; a Fay–Herriot-type small-area model pools each district towards its state as far as its data warrant.

Estimated change per year, 2020–2025 (posterior median); hover a district for its raw trend and 95% interval. The middle 95% of raw district trends spans −18% to +29% per year; pooled, −10% to +20%. Colours stop at ±20% per year; 11 districts lie beyond. Published counts are rounded to multiples of 5.

Data: © Statistische Ämter des Bundes und der Länder, Deutschland, 2026 (Regionalstatistik 22161-01-01-4, 12411-04-02-4; reference dates 2020–2025), dl-de/by-2-0. Kreis boundaries: © GeoBasis-DE / BKG (2026), dl-de/by-2-0.

  • government
  • regional
  • geospatial
  • multilevel
  • trends

Temperature trend and outlook

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 (median, 50% and 95% intervals), about 2.6 °C warmer than in 1881, continued to 2035. The continuation is a statistical extrapolation of the trend, not a climate projection.

Data: Deutscher Wetterdienst (DWD), annual mean air temperature for Germany, modified (CC BY 4.0).

  • environment
  • climate
  • forecasting

Warming across the states

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 (posterior median), 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.

Data: Deutscher Wetterdienst (DWD), annual mean air temperature by state, modified (CC BY 4.0). State boundaries: © GeoBasis-DE / BKG (2026), dl-de/by-2-0.

  • environment
  • climate
  • geospatial

Fewer false pre-alarms

A fixed threshold raises a pre-alarm 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-alarm 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.

Data: Bundesamt für Strahlenschutz (BfS), ODL-Info open data, 16–23 October 2026; model results as of 23 October 2026. State boundaries: © GeoBasis-DE / BKG (2026), dl-de/by-2-0.

  • environment
  • monitoring
  • geospatial
  • anomalies

Programme evaluation

Did a programme change the outcome, and when? Municipalities adopt it in different years; an event-study model compares adopters with non-adopters year by year.

Estimated programme effect by year relative to adoption (median, 50% and 95% intervals); 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.

  • government
  • evaluation
  • causal

Gaussian processes and their approximations

Originally published on the Generable blog ↗︎

A smooth curve with no fixed formula, and noise whose size changes over time: a heteroscedastic Gaussian process for the classic motorcycle crash data, approximated with 20 Hilbert-space basis functions.

Head acceleration after the impact (dots) and the estimated mean (median, 50% and 95% intervals).
The noise SD changes over time too: near zero before the impact, largest around 30 ms. A second Gaussian process models its logarithm.

Data: motorcycle crash test, Silverman (1985), as distributed in the MASS R package (GPL-2 | GPL-3).

  • statistics
  • nonparametric
  • gaussian-processes

Measurements below the limit of quantification

Originally published on the Generable blog ↗︎

What to do with lab values reported as "below the LLOQ"? Dropping them, or replacing them by half the limit, biases a dose-response analysis; treating them as censored recovers the truth. A simulation study with three ways to fit the same data.

One subject at the highest dose: true level (line), measurements (circles; squares are below the LLOQ of 0.1, drawn at the LLOQ) and each method's estimate (median, 50% and 95% intervals).
Calibration over all 495 samples after the dose: a method whose curve follows the diagonal places the true level where its posterior says it should be. The central 95% intervals 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 baseline per dose group (median, 50% and 95% intervals); the dashed line is the truth. Samples below the LLOQ: 5%, 26%, 50%, 73%, 78% by dose.

Simulated data (55 subjects in five dose groups, ten samples each).

  • health
  • pharmacometrics
  • censoring
  • simulation

Divergences in Gaussian-process posteriors

Originally published on the Generable blog ↗︎

Why does the sampler report divergent transitions for the motorcycle model, and does it help to learn, for each basis weight, how far to centre it? The same model sampled three ways, with 40 chains at three target acceptance rates.

The centeredness (0 non-centered, 1 centered) of each basis weight: median, 50% and 95% intervals over the 120 chains that adapted it in their own warm-up (40 at each target acceptance rate), and the value learned once from one chain of 10,000 draws, 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 noise SD.
Effective draws of the worst-mixing GP hyperparameter (the amplitudes and length scales) per 1000 gradient evaluations, chain by chain: median, 50% and 95% intervals over each method's chains of 1000 draws. Medians at a target acceptance rate of 0.8: adapted in each chain's warm-up 2.4, non-centered 1.7, 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

Open-source software

ReactiveKernels.jl

Describe a computation as a graph, say what is known and what is wanted, and get a specialised kernel that computes only what is needed.

Repository ↗︎Documentation ↗︎

  • computation

NLMEEstimation.jl

Estimators for nonlinear mixed-effects models that work with any model providing a small set of functions.

Repository ↗︎Documentation ↗︎

  • estimation
  • pharmacometrics
No items match the current filter.

CV

since 2025Independent consultant: Bayesian pharmacometric modellingGenerable Inc.
2022 – 2026Postdoctoral research in probabilistic modellingAalto University; University of Cambridge
2021PhD in MathematicsRWTH Aachen University
2017BSc and MSc in Computational Engineering ScienceRWTH Aachen University