drmTMB fits distributional regression models for one or
two responses. A distributional model can describe the expected response
and other features of its distribution, such as residual variation, tail
shape, zero probability, or residual correlation. This page helps you
find the public guide that matches your question.
If this is your first drmTMB model, begin with Distributional regression with drmTMB. It
introduces the formula interface, fits a small Gaussian location–scale
model, and gives a short path through checking, interpretation, and
reporting.
Choose a guide by task
| Your task | Start here | What you will learn |
|---|---|---|
| Decide whether a model is available | What can I fit today? | The current one-response, two-response, and structural-dependence scope. |
| Choose a response distribution | Choosing response families | How the measurement process determines the family and the meaning of each distributional parameter. |
| Decide which kind of variation to model | Which scale are you modelling? | The difference between residual sigma, group-level
standard deviations, known sampling variance, likelihood weights, and
residual rho12. |
| Write a formula | Formula grammar | The supported mu, sigma, shape,
zero-probability, hurdle, and two-response formula forms. |
| Check and interpret a fitted model | Checking and using fitted models |
check_drm(), summaries, intervals, predictions, fitted
distributions, residuals, and simulation. |
| Decide whether an estimate is ready to report | Can I fit and report this model? | The exact fitted scope, inference limits, and reporting cautions for specialised models. |
| Find a function quickly | Function map and cheat sheet | The main fitting, checking, extraction, prediction, and reporting functions. |
Choose by response type
| Response or scientific question | Family route | Worked guide |
|---|---|---|
| Continuous response with changing mean or residual variation | Gaussian; Student-t for heavy tails; Skew-normal for asymmetry | When variance carries signal, Part 1, Part 2, and Robust continuous responses |
| Positive continuous response | Lognormal, Gamma, or Tweedie, depending on the mean–variance relationship and support | Choosing response families |
| Counts, extra variation, extra zeros, or a hurdle process | Poisson, negative-binomial 2, zero-inflated, zero-truncated, or hurdle models | Count abundance and extra zeros |
| Continuous proportions, exact boundary values, or successes out of trials | Beta, Zero-one beta, binomial, or beta-binomial | Proportions and success rates |
| Ordered categories | Cumulative-logit model | Choosing response families |
| Two continuous outcomes with a direct joint distribution | Bivariate Gaussian, lognormal, or Student-t | Changing residual coupling with rho12 and Bivariate non-Gaussian models |
| Two outcomes of different kinds | Staged latent-normal association for reviewed family pairs | Association between mixed outcome pairs |
Choose the family from how the response was generated, not from which
model is most familiar. The family guide states the response support and
the scale on which mu, sigma, nu,
zi, hu, zoi, and coi
are interpreted.
Choose a dependence guide
Ordinary random effects describe exchangeable groups. Structured effects add known relationships among levels, such as shared ancestry, coordinates, or a supplied relatedness matrix.
| Dependence in the data | Public guide |
|---|---|
| Repeated observations or grouped individuals | Distributional regression with drmTMB and Formula grammar |
| Shared phylogenetic history | Phylogenetic structured effects |
| Spatial coordinates | Coordinate-spatial structured effects |
| Pedigree or animal-model relatedness | Animal models and relatedness |
| A supplied covariance or precision matrix for latent effects | Known-matrix relmat models |
| Known sampling variance or covariance | Meta-analysis with known sampling variances |
| An overview comparing these choices | Structural dependence overview |
Known sampling covariance and latent relatedness answer different
questions. Use meta_V(V = V) when V describes
known observation-level sampling error; use relmat(K = K)
or relmat(Q = Q) when the matrix describes related latent
effects.
Keep correlation layers separate
For two responses, rho12 is residual correlation within
an observation after the fitted mean and scale models.
corpairs() reports correlations among group-level random
effects. A staged mixed-outcome association estimates a latent-normal
association after fitting and fixing the marginal models. These
quantities are not interchangeable; choose the guide whose correlation
scale matches the scientific question.
If your exact model is not shown
Do not infer support from a nearby formula. Check What can I fit today?, then read Can I fit and report this model?
for the exact family, distributional parameter, dependence structure,
and interval scope. After fitting, run check_drm(fit)
before interpreting estimates and use profile_targets(fit)
before requesting profile intervals. If the requested combination is not
currently supported, fit the closest simpler model described on those
pages rather than relying on unlisted syntax.