Frequently asked questions

How do I create a modified distribution?

Each modifier is a plain function taking a distribution and returning a distribution. The printed result below shows each wrapper displaying the distribution it wraps:

using ModifiedDistributions, Distributions

[affine(LogNormal(1.5, 0.5); scale = 2.0, shift = 1.0),
    weight(Normal(2.0, 1.0), 10.0),
    thin(Gamma(2.0, 1.0), 0.3),
    cumulative(Gamma(2.0, 1.0)),
    modify(Weibull(1.5, 2.0), 0.5)]

5-element Vector{ModifiedDistributions.AbstractModifiedDistribution{Distributions.Univariate}}:
 Affine(Distributions.LogNormal{Float64}(μ=1.5, σ=0.5))
 Weighted(Distributions.Normal{Float64}(μ=2.0, σ=1.0))
 Transformed(Distributions.Gamma{Float64}(α=2.0, θ=1.0))
 Transformed(Distributions.Gamma{Float64}(α=2.0, θ=1.0))
 Modified(Distributions.Weibull{Float64}(α=1.5, θ=2.0), 0.5; link=LogLink)

What are the three weight scenarios for weight?

1. A fixed constructor weight: weight(d, 10.0) multiplies logpdf by 10.

2. An observation-time weight: weight(d) stores missing and expects joint observations (value = x, weight = w).

3. Vectorised weights: weight(d, [3, 1, 4]) builds a Product of weighted components for vector observations.

Constructor and observation weights combine by multiplication when both are present.

In every form the result is still a real Distributions.jl distribution, unlike an ad hoc n * logpdf(d, x) term or Turing.jl's @addlogprob!. Sampling delegates to the base while only the likelihood contribution is scaled, so a Turing.jl model (or any PPL built on Distributions.jl) that uses a weighted distribution stays a complete generative model — prior simulation and posterior-predictive draws keep working.

Why does a zero or missing weight give -Inf rather than NaN?

0 * logpdf is NaN when logpdf is -Inf, which poisons a sampler. weight short-circuits zero and missing weights to -Inf directly, keeping the log-density well defined (and keeping automatic differentiation happy).

Why doesn't thin change logpdf?

thin and cumulative are forward-series transforms: they carry an operation for a count series a downstream layer (for example a convolution engine) produces. The distribution itself is unchanged — logpdf, rand, cdf and everything else delegate to the inner distribution. If you want thinning that changes the density, that is a different operation and lives with the convolution layer that owns the series.

What is the difference between get_dist and get_dist_recursive?

get_dist removes one layer of wrapping; get_dist_recursive keeps unwrapping until it reaches a distribution with no more layers:

nested = weight(affine(Normal(0, 1); scale = 2.0), 3.0)
get_dist(nested)           # the Affine wrapper

Affine(Distributions.Normal{Float64}(μ=0.0, σ=1.0))

get_dist_recursive(nested) # the Normal

Distributions.Normal{Float64}(μ=0.0, σ=1.0)

Can I combine modifiers in any order?

Yes. Each modifier is a thin wrapper around whatever you pass it, so weight(affine(d; scale = 2), 10) and affine(weight(d, 10); scale = 2) both work. Order matters for meaning, not validity: modify the distribution first, then weight the resulting likelihood term, unless you have a reason to do otherwise.

Why does modify reject my discrete distribution or negative additive effect?

The hazard-modification maths implemented here is the closed-form continuous path. The discrete (interval-censored) path lives upstream in CensoredDistributions.jl, where the interval types live. Negative additive effects need hazard clamping and numeric integration of the cumulative hazard (see CensoredDistributions#670) — not yet ported; use the log link for hazard reductions (modify(d, -0.5) with the default link scales the hazard by exp(-0.5)).

Does this work with composed distribution chains?

Yes. Loading ComposedDistributions.jl activates a package extension that lets the modifier verbs apply to a Sequential chain. A chain observes one scalar quantity — its convolved total — so a modifier on the chain modifies that observed scalar: the chain collapses to its convolved total first, then the modifier wraps the resulting univariate distribution. A Parallel has several independent endpoints and no single observed scalar, so the verbs are not defined for it. See the Modifiers across composed chains tutorial.

How do I cite ModifiedDistributions?

See the citation section of the README.