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:

model_def = @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:

model_def = @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

Model definition and data are the two necessary inputs for compilation, with optional initializations. The compile function creates a BUGSModel that implements the LogDensityProblems.jl interface.

compile(model_def::Expr, data::NamedTuple)

And with initializations:

compile(model_def::Expr, data::NamedTuple, initializations::NamedTuple)

Using the model definition and data we defined earlier, we can compile the model:

model = compile(model_def, 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.

We can provide initializations:

initializations = (alpha = 1, beta = 1)
compile(model_def, data, 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}

We can also initialize parameters after compilation:

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
    r: size = (21,), type = Vector{Int64}
    alpha1: type = Float64
    x1: size = (21,), type = Vector{Int64}
    x2: size = (21,), type = Vector{Int64}

Inference

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

# Compile with gradient support
model = compile(model_def, data; adtype=AutoReverseDiff(compile=true))

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 = Chains,
                        n_adapts = n_adapts,
                        init_params = initial_θ,
                        discard_initial = n_adapts,
                        progress = false
                    )
describe(samples_and_stats)
[ Info: Found initial step size 0.2
Chains MCMC chain (2000×40×1 Array{Real, 3}):

Iterations        = 1001:1:3000
Number of chains  = 1
Samples per chain = 2000
parameters        = 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
internals         = 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 Statistics

  parameters      mean       std      mcse    ess_bulk    ess_tail      rhat   ⋯
      Symbol   Float64   Float64   Float64        Real     Float64   Float64   ⋯

         tau   20.9754   32.5162    2.1321    211.4145    273.1111    1.0024   ⋯
     alpha12   -0.8472    0.4442    0.0140   1013.5282   1151.3292    0.9999   ⋯
      alpha2    1.3637    0.2781    0.0090    961.9232   1044.8541    0.9997   ⋯
      alpha1    0.0895    0.3094    0.0092   1138.4031   1268.8004    0.9998   ⋯
      alpha0   -0.5605    0.1921    0.0063    928.5640   1259.1195    0.9996   ⋯
       b[21]   -0.0472    0.3091    0.0061   2660.4557    764.3595    1.0043   ⋯
       b[20]    0.2488    0.2663    0.0100    762.7285   1107.9544    1.0026   ⋯
       b[19]   -0.0126    0.2647    0.0055   2285.7735   1169.9475    1.0006   ⋯
       b[18]    0.0582    0.2436    0.0057   1857.2565   1221.6065    1.0003   ⋯
       b[17]   -0.2235    0.3054    0.0096   1081.1330   1178.6092    0.9996   ⋯
       b[16]   -0.1464    0.3224    0.0091   1405.4173    911.4322    0.9998   ⋯
       b[15]    0.2440    0.2745    0.0087   1109.0011   1331.7499    1.0013   ⋯
       b[14]   -0.1552    0.2602    0.0063   1756.5809   1368.5304    1.0019   ⋯
       b[13]   -0.0799    0.2743    0.0060   2138.3275   1299.4082    1.0018   ⋯
       b[12]    0.1333    0.2911    0.0071   1825.3645   1079.4854    1.0040   ⋯
           ⋮         ⋮         ⋮         ⋮           ⋮           ⋮         ⋮   ⋱

                                                    1 column and 12 rows omitted

Quantiles

  parameters      2.5%     25.0%     50.0%     75.0%     97.5%
      Symbol   Float64   Float64   Float64   Float64   Float64

         tau    2.8293    6.7558   11.5606   21.8063   98.5350
     alpha12   -1.7431   -1.1262   -0.8504   -0.5750    0.0309
      alpha2    0.8227    1.1863    1.3625    1.5324    1.9341
      alpha1   -0.5462   -0.1143    0.1083    0.2961    0.6586
      alpha0   -0.9313   -0.6852   -0.5635   -0.4387   -0.1695
       b[21]   -0.6995   -0.2228   -0.0419    0.1365    0.5936
       b[20]   -0.1874    0.0630    0.2154    0.4122    0.8406
       b[19]   -0.5113   -0.1740   -0.0161    0.1417    0.5624
       b[18]   -0.4224   -0.0936    0.0488    0.1929    0.6127
       b[17]   -0.9390   -0.3842   -0.1890   -0.0218    0.2735
       b[16]   -0.8894   -0.3215   -0.1120    0.0524    0.4227
       b[15]   -0.2114    0.0574    0.2084    0.4022    0.8631
       b[14]   -0.7107   -0.3161   -0.1329    0.0117    0.3249
       b[13]   -0.6688   -0.2411   -0.0714    0.1001    0.4470
       b[12]   -0.4204   -0.0437    0.1080    0.2872    0.7902
           ⋮         ⋮         ⋮         ⋮         ⋮         ⋮

                                                 12 rows omitted

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

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.