Add rand safe version of Poisson and Negative binomial distributions (

#418)

* SafePoisson with safety for large means

* better selection for conversion to Int or BigInt

* add SafeNegativeBinomial

* add unit tests to doctests

* reformat

* Add type promotion so AD works with distribution constructor

* Add logpdf grad call unit tests for Safe discrete dists

* reformat

* change neg bin param to (r, p)

* Update utils.jl

* reformat

* change empirical var test to more principled approach

* add default nadapts rather than just 50% of target sampling

* Update NUTSampler.jl

* set dist check_args = false

* Set nadapts to Turing Default

* reformat

EpiAware/src/EpiAwareUtils/EpiAwareUtils.jl

@@ -8,13 +8,13 @@ using ..EpiAwareBase
 using DataFramesMeta: DataFrame, @rename!
 using DynamicPPL: Model, fix, condition, @submodel, @model
 using MCMCChains: Chains
-using Random: AbstractRNG
+using Random: AbstractRNG, randexp
 using Tables: rowtable
 
 using Distributions, DocStringExtensions, QuadGK, Statistics, Turing
 
 #Export Structures
-export HalfNormal, DirectSample
+export HalfNormal, DirectSample, SafePoisson, SafeNegativeBinomial
 
 #Export functions
 export scan, spread_draws, censored_pmf, get_param_array, prefix_submodel
@@ -32,5 +32,7 @@ include("turing-methods.jl")
 include("DirectSample.jl")
 include("post-inference.jl")
 include("get_param_array.jl")
+include("SafePoisson.jl")
+include("SafeNegativeBinomial.jl")
 
 end

EpiAware/src/EpiAwareUtils/SafeNegativeBinomial.jl

@@ -0,0 +1,137 @@
+@doc raw"
+Create a Negative binomial distribution with the specified mean that avoids `InExactError`
+when the mean is too large.
+
+
+# Parameterisation:
+We are using a mean and cluster factorization of the negative binomial distribution such
+that the variance to mean relationship is:
+
+```math
+\sigma^2  = \mu + \alpha^2 \mu^2
+```
+
+The reason for this parameterisation is that at sufficiently large mean values (i.e. `r > 1 / p`) `p` is approximately equal to the
+standard fluctuation of the distribution, e.g. if `p = 0.05` we expect typical fluctuations of samples from the negative binomial to be
+about 5% of the mean when the mean is notably larger than 20. Otherwise, we expect approximately Poisson noise. In our opinion, this
+parameterisation is useful for specifying the distribution in a way that is easier to reason on priors for `p`.
+
+# Arguments:
+
+- `r`: The number of successes, although this can be extended to a continous number.
+- `p`: Success rate.
+
+# Returns:
+
+- A `SafeNegativeBinomial` distribution with the specified mean.
+
+# Examples:
+
+```jldoctest SafeNegativeBinomial
+using EpiAware, Distributions
+
+bigμ = exp(48.0) #Large value of μ
+σ² = bigμ + 0.05 * bigμ^2 #Large variance
+
+# We can calculate the success rate from the mean to variance relationship
+p = bigμ / σ²
+r = bigμ * p / (1 - p)
+d = SafeNegativeBinomial(r, p)
+# output
+EpiAware.EpiAwareUtils.SafeNegativeBinomial{Float64}(r=20.0, p=2.85032816548187e-20)
+```
+
+```jldoctest SafeNegativeBinomial
+cdf(d, 100)
+# output
+0.0
+```
+
+```jldoctest SafeNegativeBinomial
+logpdf(d, 100)
+# output
+-850.1397180331871
+```
+
+```jldoctest SafeNegativeBinomial
+mean(d)
+# output
+7.016735912097631e20
+```
+
+```jldoctest SafeNegativeBinomial
+var(d)
+# output
+2.4617291430060293e40
+```
+"
+struct SafeNegativeBinomial{T <: Real} <: DiscreteUnivariateDistribution
+    r::T
+    p::T
+
+    function SafeNegativeBinomial{T}(r::T, p::T) where {T <: Real}
+        return new{T}(r, p)
+    end
+end
+
+#Outer constructors make AD work
+function SafeNegativeBinomial(r::T, p::T) where {T <: Real}
+    return SafeNegativeBinomial{T}(r, p)
+end
+
+SafeNegativeBinomial(r::Real, p::Real) = SafeNegativeBinomial(promote(r, p)...)
+
+# helper function
+_negbin(d::SafeNegativeBinomial) = NegativeBinomial(d.r, d.p; check_args = false)
+
+### Support
+
+Base.minimum(d::SafeNegativeBinomial) = 0
+Base.maximum(d::SafeNegativeBinomial) = Inf
+Distributions.insupport(d::SafeNegativeBinomial, x::Integer) = x >= 0
+
+#### Parameters
+
+Distributions.params(d::SafeNegativeBinomial) = _negbin(d) |> params
+Distributions.partype(::SafeNegativeBinomial{T}) where {T} = T
+
+Distributions.succprob(d::SafeNegativeBinomial) = _negbin(d).p
+Distributions.failprob(d::SafeNegativeBinomial{T}) where {T} = one(T) - _negbin(d).p
+
+#### Statistics
+
+Distributions.mean(d::SafeNegativeBinomial) = _negbin(d) |> mean
+Distributions.var(d::SafeNegativeBinomial) = _negbin(d) |> var
+Distributions.std(d::SafeNegativeBinomial) = _negbin(d) |> std
+Distributions.skewness(d::SafeNegativeBinomial) = _negbin(d) |> skewness
+Distributions.kurtosis(d::SafeNegativeBinomial) = _negbin(d) |> kurtosis
+Distributions.mode(d::SafeNegativeBinomial) = _negbin(d) |> mode
+function Distributions.kldivergence(p::SafeNegativeBinomial, q::SafeNegativeBinomial)
+    kldivergence(_negbin(p), _negbin(q))
+end
+
+#### Evaluation & Sampling
+
+Distributions.logpdf(d::SafeNegativeBinomial, k::Real) = logpdf(_negbin(d), k)
+
+Distributions.cdf(d::SafeNegativeBinomial, x::Real) = cdf(_negbin(d), x)
+Distributions.ccdf(d::SafeNegativeBinomial, x::Real) = ccdf(_negbin(d), x)
+Distributions.logcdf(d::SafeNegativeBinomial, x::Real) = logcdf(_negbin(d), x)
+Distributions.logccdf(d::SafeNegativeBinomial, x::Real) = logccdf(_negbin(d), x)
+Distributions.quantile(d::SafeNegativeBinomial, q::Real) = quantile(_negbin(d), q)
+Distributions.cquantile(d::SafeNegativeBinomial, q::Real) = cquantile(_negbin(d), q)
+Distributions.invlogcdf(d::SafeNegativeBinomial, lq::Real) = invlogcdf(_negbin(d), lq)
+Distributions.invlogccdf(d::SafeNegativeBinomial, lq::Real) = invlogccdf(_negbin(d), lq)
+
+## sampling
+function Base.rand(rng::AbstractRNG, d::SafeNegativeBinomial)
+    if isone(d.p)
+        return 0
+    else
+        return rand(rng, SafePoisson(rand(rng, Gamma(d.r, (1 - d.p) / d.p))))
+    end
+end
+
+Distributions.mgf(d::SafeNegativeBinomial, t::Real) = mgf(_negbin(d), t)
+Distributions.cgf(d::SafeNegativeBinomial, t) = cgf(_negbin(d), t)
+Distributions.cf(d::SafeNegativeBinomial, t::Real) = cf(_negbin(d), t)