Merge pull request #3266 from aleksgorica/newrhat

Improved rhat diagnostic

src/stan/analyze/mcmc/compute_potential_scale_reduction.hpp

@@ -8,6 +8,7 @@
 #include <boost/accumulators/statistics/stats.hpp>
 #include <boost/accumulators/statistics/mean.hpp>
 #include <boost/accumulators/statistics/variance.hpp>
+#include <boost/math/distributions/normal.hpp>
 #include <algorithm>
 #include <cmath>
 #include <vector>
@@ -16,6 +17,156 @@
 namespace stan {
 namespace analyze {
 
+/**
+ * Computes normalized average ranks for draws. Transforming them to normal
+ * scores using inverse normal transformation and a fractional offset. Based on
+ * paper https://arxiv.org/abs/1903.08008
+ * @param chains stores chains in columns
+ * @return normal scores for average ranks of draws
+ */
+inline Eigen::MatrixXd rank_transform(const Eigen::MatrixXd& chains) {
+  const Eigen::Index rows = chains.rows();
+  const Eigen::Index cols = chains.cols();
+  const Eigen::Index size = rows * cols;
+
+  std::vector<std::pair<double, int>> value_with_index(size);
+
+  for (Eigen::Index i = 0; i < size; ++i) {
+    value_with_index[i] = {chains(i), i};
+  }
+
+  std::sort(value_with_index.begin(), value_with_index.end());
+
+  Eigen::MatrixXd rank_matrix = Eigen::MatrixXd::Zero(rows, cols);
+
+  // Assigning average ranks
+  for (Eigen::Index i = 0; i < size; ++i) {
+    // Handle ties by averaging ranks
+    Eigen::Index j = i + 1;
+    double sum_ranks = j;
+    Eigen::Index count = 1;
+
+    while (j < size && value_with_index[j].first == value_with_index[i].first) {
+      sum_ranks += j + 1;  // Rank starts from 1
+      ++j;
+      ++count;
+    }
+    double avg_rank = sum_ranks / count;
+    boost::math::normal_distribution<double> dist;
+    for (std::size_t k = i; k < j; ++k) {
+      double p = (avg_rank - 3.0 / 8.0) / (size - 2.0 * 3.0 / 8.0 + 1.0);
+      const Eigen::Index index = value_with_index[k].second;
+      rank_matrix(index) = boost::math::quantile(dist, p);
+    }
+    i = j - 1;  // Skip over tied elements
+  }
+  return rank_matrix;
+}
+
+/**
+ * Computes square root of marginal posterior variance of the estimand by the
+ * weigted average of within-chain variance W and between-chain variance B.
+ *
+ * @param chains stores chains in columns
+ * @return square root of ((N-1)/N)W + B/N
+ */
+inline double rhat(const Eigen::MatrixXd& chains) {
+  const Eigen::Index num_chains = chains.cols();
+  const Eigen::Index num_draws = chains.rows();
+
+  Eigen::RowVectorXd within_chain_means = chains.colwise().mean();
+  double across_chain_mean = within_chain_means.mean();
+  double between_variance
+      = num_draws
+        * (within_chain_means.array() - across_chain_mean).square().sum()
+        / (num_chains - 1);
+  double within_variance =
+      // Divide each row by chains and get sum of squares for each chain
+      // (getting a vector back)
+      ((chains.rowwise() - within_chain_means)
+           .array()
+           .square()
+           .colwise()
+           // divide each sum of square by num_draws, sum the sum of squares,
+           // and divide by num chains
+           .sum()
+       / (num_draws - 1.0))
+          .sum()
+      / num_chains;
+
+  return sqrt((between_variance / within_variance + num_draws - 1) / num_draws);
+}
+
+/**
+ * Computes the potential scale reduction (Rhat) using rank based diagnostic for
+ * the specified parameter across all kept samples. Based on paper
+ * https://arxiv.org/abs/1903.08008
+ *
+ * Current implementation assumes draws are stored in contiguous
+ * blocks of memory.  Chains are trimmed from the back to match the
+ * length of the shortest chain.
+ *
+ * @param chain_begin stores pointers to arrays of chains
+ * @param sizes stores sizes of chains
+ * @return potential scale reduction for the specified parameter
+ */
+inline std::pair<double, double> compute_potential_scale_reduction_rank(
+    const std::vector<const double*>& chain_begins,
+    const std::vector<size_t>& chain_sizes) {
+  std::vector<const double*> nonzero_chain_begins;
+  std::vector<std::size_t> nonzero_chain_sizes;
+  nonzero_chain_begins.reserve(chain_begins.size());
+  nonzero_chain_sizes.reserve(chain_sizes.size());
+  for (size_t i = 0; i < chain_sizes.size(); ++i) {
+    if (chain_sizes[i]) {
+      nonzero_chain_begins.push_back(chain_begins[i]);
+      nonzero_chain_sizes.push_back(chain_sizes[i]);
+    }
+  }
+  if (!nonzero_chain_sizes.size()) {
+    return {std::numeric_limits<double>::quiet_NaN(),
+            std::numeric_limits<double>::quiet_NaN()};
+  }
+  std::size_t num_nonzero_chains = nonzero_chain_sizes.size();
+  std::size_t min_num_draws = nonzero_chain_sizes[0];
+  for (std::size_t chain = 1; chain < num_nonzero_chains; ++chain) {
+    min_num_draws = std::min(min_num_draws, nonzero_chain_sizes[chain]);
+  }
+
+  // check if chains are constant; all equal to first draw's value
+  bool are_all_const = false;
+  Eigen::VectorXd init_draw = Eigen::VectorXd::Zero(num_nonzero_chains);
+  Eigen::MatrixXd draws_matrix(min_num_draws, num_nonzero_chains);
+
+  for (std::size_t chain = 0; chain < num_nonzero_chains; chain++) {
+    Eigen::Map<const Eigen::Matrix<double, Eigen::Dynamic, 1>> draws(
+        nonzero_chain_begins[chain], nonzero_chain_sizes[chain]);
+
+    for (std::size_t n = 0; n < min_num_draws; n++) {
+      if (!std::isfinite(draws(n))) {
+        return {std::numeric_limits<double>::quiet_NaN(),
+                std::numeric_limits<double>::quiet_NaN()};
+      }
+      draws_matrix(n, chain) = draws(n);
+    }
+
+    init_draw(chain) = draws(0);
+    are_all_const |= !draws.isApproxToConstant(draws(0));
+  }
+  // If all chains are constant then return NaN
+  if (are_all_const && init_draw.isApproxToConstant(init_draw(0))) {
+    return {std::numeric_limits<double>::quiet_NaN(),
+            std::numeric_limits<double>::quiet_NaN()};
+  }
+
+  double rhat_bulk = rhat(rank_transform(draws_matrix));
+  double rhat_tail = rhat(rank_transform(
+      (draws_matrix.array() - math::quantile(draws_matrix.reshaped(), 0.5))
+          .abs()));
+
+  return std::make_pair(rhat_bulk, rhat_tail);
+}
+
 /**
  * Computes the potential scale reduction (Rhat) for the specified
  * parameter across all kept samples.
@@ -101,6 +252,29 @@ inline double compute_potential_scale_reduction(
   return sqrt((var_between / var_within + num_draws - 1) / num_draws);
 }
 
+/**
+ * Computes the potential scale reduction (Rhat) using rank based diagnostic for
+ * the specified parameter across all kept samples. Based on paper
+ * https://arxiv.org/abs/1903.08008
+ *
+ * See more details in Stan reference manual section "Potential
+ * Scale Reduction". http://mc-stan.org/users/documentation
+ *
+ * Current implementation assumes draws are stored in contiguous
+ * blocks of memory. Chains are trimmed from the back to match the
+ * length of the shortest chain. Argument size will be broadcast to
+ * same length as draws.
+ *
+ * @param chain_begins stores pointers to arrays of chains
+ * @param size stores sizes of chains
+ * @return potential scale reduction for the specified parameter
+ */
+inline std::pair<double, double> compute_potential_scale_reduction_rank(
+    const std::vector<const double*>& chain_begins, size_t size) {
+  std::vector<size_t> sizes(chain_begins.size(), size);
+  return compute_potential_scale_reduction_rank(chain_begins, sizes);
+}
+
 /**
  * Computes the potential scale reduction (Rhat) for the specified
  * parameter across all kept samples.
@@ -124,6 +298,42 @@ inline double compute_potential_scale_reduction(
   return compute_potential_scale_reduction(draws, sizes);
 }
 
+/**
+ * Computes the potential scale reduction (Rhat) using rank based diagnostic for
+ * the specified parameter across all kept samples. Based on paper
+ * https://arxiv.org/abs/1903.08008
+ *
+ * When the number of total draws N is odd, the (N+1)/2th draw is ignored.
+ *
+ * See more details in Stan reference manual section "Potential
+ * Scale Reduction". http://mc-stan.org/users/documentation
+ *
+ * Current implementation assumes draws are stored in contiguous
+ * blocks of memory. Chains are trimmed from the back to match the
+ * length of the shortest chain.
+ *
+ * @param chain_begins stores pointers to arrays of chains
+ * @param chain_sizes stores sizes of chains
+ * @return potential scale reduction for the specified parameter
+ */
+inline std::pair<double, double> compute_split_potential_scale_reduction_rank(
+    const std::vector<const double*>& chain_begins,
+    const std::vector<size_t>& chain_sizes) {
+  size_t num_chains = chain_sizes.size();
+  size_t num_draws = chain_sizes[0];
+  for (size_t chain = 1; chain < num_chains; ++chain) {
+    num_draws = std::min(num_draws, chain_sizes[chain]);
+  }
+
+  std::vector<const double*> split_draws
+      = split_chains(chain_begins, chain_sizes);
+
+  size_t half = std::floor(num_draws / 2.0);
+  std::vector<size_t> half_sizes(2 * num_chains, half);
+
+  return compute_potential_scale_reduction_rank(split_draws, half_sizes);
+}
+
 /**
  * Computes the split potential scale reduction (Rhat) for the
  * specified parameter across all kept samples.  When the number of
@@ -156,6 +366,32 @@ inline double compute_split_potential_scale_reduction(
   return compute_potential_scale_reduction(split_draws, half_sizes);
 }
 
+/**
+ * Computes the potential scale reduction (Rhat) using rank based diagnostic for
+ * the specified parameter across all kept samples. Based on paper
+ * https://arxiv.org/abs/1903.08008
+ *
+ * When the number of total draws N is odd, the (N+1)/2th draw is ignored.
+ *
+ * See more details in Stan reference manual section "Potential
+ * Scale Reduction". http://mc-stan.org/users/documentation
+ *
+ * Current implementation assumes draws are stored in contiguous
+ * blocks of memory.  Chains are trimmed from the back to match the
+ * length of the shortest chain.  Argument size will be broadcast to
+ * same length as draws.
+ *
+ * @param chain_begins stores pointers to arrays of chains
+ * @param size stores sizes of chains
+ * @return potential scale reduction for the specified parameter
+ */
+inline std::pair<double, double> compute_split_potential_scale_reduction_rank(
+    const std::vector<const double*>& chain_begins, size_t size) {
+  size_t num_chains = chain_begins.size();
+  std::vector<size_t> sizes(num_chains, size);
+  return compute_split_potential_scale_reduction_rank(chain_begins, sizes);
+}
+
 /**
  * Computes the split potential scale reduction (Rhat) for the
  * specified parameter across all kept samples.  When the number of

src/stan/mcmc/chains.hpp

@@ -595,6 +595,22 @@ class chains {
     return split_effective_sample_size(index(name));
   }
 
+  std::pair<double, double> split_potential_scale_reduction_rank(
+      const int index) const {
+    int n_chains = num_chains();
+    std::vector<const double*> draws(n_chains);
+    std::vector<size_t> sizes(n_chains);
+    int n_kept_samples = 0;
+    for (int chain = 0; chain < n_chains; ++chain) {
+      n_kept_samples = num_kept_samples(chain);
+      draws[chain]
+          = samples_(chain).col(index).bottomRows(n_kept_samples).data();
+      sizes[chain] = n_kept_samples;
+    }
+
+    return analyze::compute_split_potential_scale_reduction_rank(draws, sizes);
+  }
+
   double split_potential_scale_reduction(const int index) const {
     int n_chains = num_chains();
     std::vector<const double*> draws(n_chains);
@@ -610,6 +626,11 @@ class chains {
     return analyze::compute_split_potential_scale_reduction(draws, sizes);
   }
 
+  std::pair<double, double> split_potential_scale_reduction_rank(
+      const std::string& name) const {
+    return split_potential_scale_reduction_rank(index(name));
+  }
+
   double split_potential_scale_reduction(const std::string& name) const {
     return split_potential_scale_reduction(index(name));
   }