Example: Logistic Regression with Random Effects

We will use the Seeds for demonstration. This example concerns the proportion of seeds that germinated on each of 21 plates. Here, we transform the data into a NamedTuple:

data = (
    r = [10, 23, 23, 26, 17, 5, 53, 55, 32, 46, 10, 8, 10, 8, 23, 0, 3, 22, 15, 32, 3],
    n = [39, 62, 81, 51, 39, 6, 74, 72, 51, 79, 13, 16, 30, 28, 45, 4, 12, 41, 30, 51, 7],
    x1 = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
    x2 = [0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1],
    N = 21,
)

where r[i] is the number of germinated seeds and n[i] is the total number of the seeds on the $i$-th plate. Let $p_i$ be the probability of germination on the $i$-th plate. Then, the model is defined by:

\[\begin{aligned} b_i &\sim \text{Normal}(0, \tau) \\ \text{logit}(p_i) &= \alpha_0 + \alpha_1 x_{1 i} + \alpha_2 x_{2i} + \alpha_{12} x_{1i} x_{2i} + b_{i} \\ r_i &\sim \text{Binomial}(p_i, n_i) \end{aligned}\]

where $x_{1i}$ and $x_{2i}$ are the seed type and root extract of the $i$-th plate. The original BUGS program for the model is:

model
{
    for( i in 1 : N ) {
        r[i] ~ dbin(p[i],n[i])
        b[i] ~ dnorm(0.0,tau)
        logit(p[i]) <- alpha0 + alpha1 * x1[i] + alpha2 * x2[i] +
        alpha12 * x1[i] * x2[i] + b[i]
    }
    alpha0 ~ dnorm(0.0, 1.0E-6)
    alpha1 ~ dnorm(0.0, 1.0E-6)
    alpha2 ~ dnorm(0.0, 1.0E-6)
    alpha12 ~ dnorm(0.0, 1.0E-6)
    tau ~ dgamma(0.001, 0.001)
    sigma <- 1 / sqrt(tau)
}

Modeling Language

Writing a Model in BUGS

Language References:

Implementations in C++ and R:

Language Syntax:

Writing a Model in Julia

We provide a macro which allows users to write down model definitions using Julia:

seeds = @bugs begin
    for i in 1:N
        r[i] ~ dbin(p[i], n[i])
        b[i] ~ dnorm(0.0, tau)
        p[i] = logistic(alpha0 + alpha1 * x1[i] + alpha2 * x2[i] + alpha12 * x1[i] * x2[i] + b[i])
    end
    alpha0 ~ dnorm(0.0, 1.0E-6)
    alpha1 ~ dnorm(0.0, 1.0E-6)
    alpha2 ~ dnorm(0.0, 1.0E-6)
    alpha12 ~ dnorm(0.0, 1.0E-6)
    tau ~ dgamma(0.001, 0.001)
    sigma = 1 / sqrt(tau)
end

BUGS syntax carries over almost one-to-one to Julia, with minor exceptions. Modifications required are minor: curly braces are replaced with begin ... end blocks, and for loops do not require parentheses. In addition, Julia uses f(x) = ... as a shorthand for function definition, so BUGS' link function syntax is disallowed. Instead, user can call the inverse function of the link functions on the RHS expressions.

Support for Legacy BUGS Programs

The @bugs macro also works with original (R-like) BUGS syntax:

seeds = @bugs("""
model{
    for( i in 1 : N ) {
        r[i] ~ dbin(p[i],n[i])
        b[i] ~ dnorm(0.0,tau)
        logit(p[i]) <- alpha0 + alpha1 * x1[i] + alpha2 * x2[i] +
        alpha12 * x1[i] * x2[i] + b[i]
    }
    alpha0 ~ dnorm(0.0,1.0E-6)
    alpha1 ~ dnorm(0.0,1.0E-6)
    alpha2 ~ dnorm(0.0,1.0E-6)
    alpha12 ~ dnorm(0.0,1.0E-6)
    tau ~ dgamma(0.001,0.001)
    sigma <- 1 / sqrt(tau)
}
""", true, true)

By default, @bugs will translate R-style variable names like a.b.c to a_b_c, user can pass false as the second argument to disable this. User can also pass true as the third argument if model { } enclosure is not present in the BUGS program. We still encourage users to write new programs using the Julia-native syntax, because of better debuggability and perks like syntax highlighting.

Basic Workflow

Compilation

The model definition returned by @bugs is callable: call it with the data (a NamedTuple) to construct a BUGSModel, which implements the LogDensityProblems.jl interface.

model = seeds(data)
BUGSModel (parameters are in transformed (unconstrained) space, with dimension 26):

  Model parameters:
    alpha2
    b[21], b[20], b[19], b[18], b[17], b[16], b[15], b[14], b[13], b[12], b[11], b[10], b[9], b[8], b[7], b[6], b[5], b[4], b[3], b[2], b[1]
    tau
    alpha12
    alpha1
    alpha0

  Variable sizes and types:
    b: size = (21,), type = Vector{Float64}
    p: size = (21,), type = Vector{Float64}
    n: size = (21,), type = Vector{Int64}
    alpha2: type = Float64
    sigma: type = Float64
    alpha0: type = Float64
    alpha12: type = Float64
    N: type = Int64
    tau: type = Float64
    alpha1: type = Float64
    r: size = (21,), type = Vector{Int64}
    x1: size = (21,), type = Vector{Int64}
    x2: size = (21,), type = Vector{Int64}

Parameter values will be sampled from the prior distributions in the original space.

`compile` is now mostly internal

Calling the model definition is the supported way to construct a BUGSModel. It wraps compile, which is kept for backward compatibility (compile(seeds, data) does the same thing) but is now mostly an internal function — you shouldn't need to call it directly.

We can provide initial parameter values after construction with initialize!:

initializations = (alpha = 1, beta = 1)
initialize!(model, initializations)
BUGSModel (parameters are in transformed (unconstrained) space, with dimension 26):

  Model parameters:
    alpha2
    b[21], b[20], b[19], b[18], b[17], b[16], b[15], b[14], b[13], b[12], b[11], b[10], b[9], b[8], b[7], b[6], b[5], b[4], b[3], b[2], b[1]
    tau
    alpha12
    alpha1
    alpha0

  Variable sizes and types:
    b: size = (21,), type = Vector{Float64}
    p: size = (21,), type = Vector{Float64}
    n: size = (21,), type = Vector{Int64}
    alpha2: type = Float64
    sigma: type = Float64
    alpha0: type = Float64
    alpha12: type = Float64
    N: type = Int64
    tau: type = Float64
    alpha1: type = Float64
    r: size = (21,), type = Vector{Int64}
    x1: size = (21,), type = Vector{Int64}
    x2: size = (21,), type = Vector{Int64}

initialize! also accepts a flat vector. In this case, the vector should have the same length as the number of parameters, but values can be in transformed space:

initialize!(model, rand(26))
BUGSModel (parameters are in transformed (unconstrained) space, with dimension 26):

  Model parameters:
    alpha2
    b[21], b[20], b[19], b[18], b[17], b[16], b[15], b[14], b[13], b[12], b[11], b[10], b[9], b[8], b[7], b[6], b[5], b[4], b[3], b[2], b[1]
    tau
    alpha12
    alpha1
    alpha0

  Variable sizes and types:
    b: size = (21,), type = Vector{Float64}
    p: size = (21,), type = Vector{Float64}
    n: size = (21,), type = Vector{Int64}
    alpha2: type = Float64
    sigma: type = Float64
    alpha0: type = Float64
    alpha12: type = Float64
    N: type = Int64
    tau: type = Float64
    alpha1: type = Float64
    r: size = (21,), type = Vector{Int64}
    x1: size = (21,), type = Vector{Int64}
    x2: size = (21,), type = Vector{Int64}

Inference

For gradient-based inference, construct your model with an AD backend using the adtype keyword (see Automatic Differentiation for details). We use AdvancedHMC.jl:

# Construct with gradient support
model = seeds(data; adtype=AutoMooncake(; config=nothing))

n_samples, n_adapts = 2000, 1000

D = LogDensityProblems.dimension(model); initial_θ = rand(D)

samples_and_stats = AbstractMCMC.sample(
                        model,
                        NUTS(0.8),
                        n_samples;
                        chain_type = VNChain,
                        n_adapts = n_adapts,
                        init_params = initial_θ,
                        discard_initial = n_adapts,
                        progress = false
                    )
summarystats(samples_and_stats)
╭─FlexiSummary (9 statistics) ─────────────────────────────────────────────────╮
   iter    collapsed                                                          
   chain   collapsed                                                          
 ↓ stat  = [mean, std, mcse, ess_bulk, ess_tail, rhat, q5, q50, q95]          
                                                                              
 Parameters (27) ── VarName                                                   
  Float64  tau, alpha12, alpha2, alpha1, alpha0, b[21], b[20], b[19], b[18],  
           b[17], b[16], b[15], b[14], b[13], b[12], b[11], b[10], b[9],      
           b[8], b[7], b[6], b[5], b[4], b[3], b[2], b[1], sigma              
                                                                              
 Extras (13)                                                                  
  Float64  lp, n_steps, is_accept, acceptance_rate, log_density,              
           hamiltonian_energy, hamiltonian_energy_error,                      
           max_hamiltonian_energy_error, tree_depth, numerical_error,         
           step_size, nom_step_size, is_adapt                                 
                                                                              
 Summary                                                                      
     param      mean       std     mcse   ess_bulk   ess_tail    rhat      
       tau  136.4383  388.9133  67.8417    42.7743    40.8389  1.0055      
   alpha12   -0.8217    0.4331   0.0170   640.2339   847.5808  1.0036      
    alpha2    1.3572    0.2695   0.0107   640.8132   926.3066  1.0018      
    alpha1    0.0832    0.3177   0.0135   561.3976   640.7380  1.0007      
    alpha0   -0.5525    0.1958   0.0080   593.0728   828.8067  1.0016      
     b[21]   -0.0422    0.2842   0.0069  1808.0148   884.6912  0.9998      
     b[20]    0.2188    0.2678   0.0198   192.9872  1025.6187  1.0002      
     b[19]   -0.0091    0.2457   0.0048  2842.2560  1417.3466  1.0044      
     b[18]    0.0400    0.2378   0.0062  1590.6001   942.1665  1.0030      
     b[17]   -0.2121    0.3243   0.0175   323.4617   834.3455  1.0013      
     b[16]   -0.1405    0.3232   0.0142   696.0796   436.3219  1.0014      
     b[15]    0.2305    0.2770   0.0209   175.9375   492.7498  1.0007      
     b[14]   -0.1334    0.2666   0.0113   697.0344   758.1380  1.0089      
     b[13]   -0.0638    0.2547   0.0066  1574.0950   831.5875  1.0136      
     b[12]    0.1198    0.2757   0.0133   608.1638   498.9922  1.0005      
     b[11]    0.0774    0.2765   0.0082  1275.8069   913.0056  1.0001      
     b[10]   -0.2502    0.2454   0.0219   124.7840  1151.6049  1.0038      
      b[9]   -0.1324    0.2347   0.0120   492.5755   977.4851  1.0017      
      b[8]    0.1700    0.2327   0.0145   285.5426  1115.8088  1.0024      
      b[7]    0.0536    0.2118   0.0058  1426.2754  1077.6915  1.0025      
      b[6]    0.0738    0.2965   0.0079  1921.4423   883.4450  1.0003      
      b[5]    0.1085    0.2430   0.0117   460.6033   702.8310  1.0061      
      b[4]    0.2616    0.2691   0.0253   114.5194   756.5463  1.0073      
      b[3]   -0.1907    0.2306   0.0145   272.7104   936.3670  1.0009      
      b[2]    0.0057    0.2186   0.0066  1101.6833   929.4083  1.0016      
      b[1]   -0.1848    0.2636   0.0140   384.3690   963.1904  1.0017      
     sigma    0.2772    0.1559   0.0230    42.7743    40.8389  1.0055      
╰──────────────────────────────────────────────────────────────────────────────╯

This is consistent with the result in the OpenBUGS seeds example.

Here chain_type = VNChain collects the samples into a FlexiChains.FlexiChain keyed by variable name. MCMCChains is also still supported: load it and pass chain_type = MCMCChains.Chains instead (or convert an existing chain with MCMCChains.Chains(samples_and_stats)).

Next Steps

More Examples

We have transcribed all the examples from the first volume of the BUGS Examples (original and transcribed). All programs and data are included, and can be compiled using the steps described in the tutorial above.

More worked Julia scripts are available in the examples directory, including SIR, Gaussian process, and Bayesian neural network examples using Mooncake-backed gradient sampling.