catalytic_glm_binomial

Yitong Wu ywu039@e.ntu.edu.sg

Introduction

This vignette provides an overview of how to use the functions in the catalytic package that focuses on GLM Logistic Regression. The other catalytic vignettes go into other model-estimating functions.

The goal of the catalytic package is to build framework for catalytic prior distributions. Stabilizing high-dimensional working models by shrinking them towards simplified models. This is achieved by supplementing observed data with weighted synthetic data generated from a predictive distribution under the simpler model.

For more information, see (Huang et al. 2020).

The two steps of using catalytic package for GLM Logistic Regression are

  1. Initialization: The cat_glm_initialization function constructs a cat_init object based on the formula provided by the user to generate synthetic data. The resulting cat_init object is tailored to facilitate further analysis, and is integral for subsequent modeling steps in the catalytic package.

  2. Choose Method(s): Users have the flexibility to choose from five main functions within the catalytic package: cat_glm, cat_glm_tune, cat_glm_bayes_joint_gibbs, cat_glm_bayes, and cat_glm_bayes_joint. Each function serves a specific purpose in modeling with catalytic priors and offers distinct capabilities tailored to different modeling scenarios for GLM Logistic Regression. This approach enables users to seamlessly incorporate synthetic data with varying weights from different method into GLM Logistic Regression analyses, providing flexibility and control over the modeling process.

Data Preparation

Creating a high-dimensional dataset with a low data size. This step involves increasing the number of features (dimensions) while keeping the number of observations (data size) relatively small. This is useful for testing the performance of catalytic models in high-dimensional settings.

The swim dataset is loaded and split into training (train_data) and test (test_data) datasets.

library(catalytic)

set.seed(1)

# Function for calculating the mean of logarithmic error between true response and estimated response
get_mean_logarithmic_error <- function(Y, est_Y) {
  Y <- pmax(0.0001, pmin(0.9999, Y))
  est_Y <- pmax(0.0001, pmin(0.9999, est_Y))
  return(mean(Y * log(Y / est_Y) + (1 - Y) * log((1 - Y) / (1 - est_Y))))
}

# Load swim dataset for binomial analysis
data("swim")
swim_data <- cbind(swim$x, swim$y)

# Seperate observation data into train and test data
n <- 5 * ncol(swim$x + 1) # Size for training data
train_idx <- sample(1:nrow(swim_data), n)
train_data <- swim_data[train_idx, ]
test_data <- swim_data[-train_idx, ]

dim(train_data)

In this section, we explore the foundational steps of fitting a logistic regression model (GLM) using the stats::glm function with the binomial family.

# Fit a logistic regression model (GLM)
glm_model <- stats::glm(
  formula = empyr1 ~ .,
  family = binomial,
  data = train_data
)

predicted_y <- predict(
  glm_model,
  newdata = test_data,
  type = "response"
)

cat(
  "MLE GLM Binomial Model - Logarithmic Error:",
  get_mean_logarithmic_error(
    Y = test_data$empyr1,
    est_Y = predicted_y
  )
)

Let us check the ROC curve of the predicted_y from glm_model versus the test_data$empyr1, this can be a great way to visually assess the accuracy and performance of the model.

roc_curve <- pROC::roc(test_data$empyr1, predicted_y)
plot(roc_curve, main = "ROC Curve (MLE)", col = "blue", lwd = 2)

Usage of catalytic

Step 1: Initialization

To initialize data for GLM Logistic Regression using the catalytic package, the cat_glm_initialization function is employed. This function facilitates the setup by preparing synthetic data tailored for modeling purposes.

Here’s a breakdown of the parameters used:

cat_init <- cat_glm_initialization(
  formula = empyr1 ~ 1,
  family = binomial, # Default: gaussian
  data = train_data,
  syn_size = 50, # Default: 4 times the number of predictors
  x_degree = NULL, # Default: NULL, which means degrees of all predictors are 1
  resample_only = FALSE, # Default: FALSE
  na_replace = mean # Default: stats::na.omit
)

cat_init

Here shows how users can simplify the input for cat_glm_initialization. User do not have to specify syn_size and other parameters, as they have default values, which mentioned above.

cat_init object contains a list of attributes, which is typically generated from above function cat_glm_initialization. These attributes provide comprehensive information for below modeling tasks or for user check.

Here’s a breakdown of all attributes except the input parameters:

For more details, please check ?cat_glm_initialization.

names(cat_init)

And of course, user can extract items mentioned above from cat_glm_initialization object.

# The number of observations (rows) in the original dataset (`obs_data`)
cat_init$obs_size

# The information detailing the process of resampling synthetic data
cat_init$syn_x_resample_inform

Step 2.1: Choose Method(s) - Estimation with fixed tau

The cat_glm function fits a Generalized Linear Model (GLM) with a catalytic prior on the regression coefficients. It utilizes information from the cat_init object generated during the initialization step, which includes both observed and synthetic data, plus other relevant information.

The GLM model is then fitted using the specified formula, family, and a single tau(synthetic data down-weight factor). The resulting cat_glm object encapsulates the fitted model, including estimated coefficients and family information, facilitating further analysis.

Here’s a breakdown of the parameters used:

cat_glm_model <- cat_glm(
  formula = empyr1 ~ female + agege35, # Same as `~ female + agege35`
  cat_init = cat_init,
  tau = 10 # Default: number of predictors / 4
)

cat_glm_model

Here shows how users can simplify the input for cat_glm. User do not have to specify tau, as tau has default value , which mentioned above.

Let’s check the prediction error.

cat_glm_predicted_y <- predict(
  cat_glm_model,
  newdata = test_data,
  type = "response"
)

cat(
  "Catalytic `cat_glm` - Logarithmic Error:",
  get_mean_logarithmic_error(
    Y = test_data$empyr1,
    est_Y = cat_glm_predicted_y
  )
)

Let us check the ROC curve of the cat_glm_predicted_y from cat_glm_model versus the test_data$empyr1, this can be a great way to visually assess the accuracy and performance of the model.

roc_curve <- pROC::roc(test_data$empyr1, cat_glm_predicted_y)
plot(roc_curve, main = "ROC Curve (cat_glm)", col = "blue", lwd = 2)

Both cat_glm_model and cat_glm_model objects are outputs from the cat_glm function, providing a list of attributes for further analysis or user inspection.

Here’s a breakdown of all attributes except the input parameters:

For more details, please check ?cat_glm.

names(cat_glm_model)

User can extract items mentioned above from cat_glm object.

# The formula used for modeling
cat_glm_model$formula

# The fitted GLM model object obtained from `stats::glm` with `tau`
cat_glm_model$coefficients

Step 2.2: Choose Method(s) - Estimation with Selective tau

The cat_glm_tune function fits a GLM with a catalytic prior on the regression coefficients and provides options for optimizing model performance over a range of tau values(tau_seq).

These methods empower users to fit and optimize GLM models with catalytic priors, leveraging both observed and synthetic data to enhance model performance and robustness in various statistical analyses.

Cross-validation (risk_estimate_method = “cross_validation”)

This method computes the partial likelihood across a specified range of tau values (tau_seq). It iterates through each tau value, evaluating its performance based on cross-validation folds (cv_fold_num) to select the optimal tau that minimizes the discrepancy error.

Here’s a breakdown of the parameters used:

  • formula and cat_init are same as above.

  • risk_estimate_method: Method for risk estimation, chosen from “parametric_bootstrap”, “cross_validation”, “mallows_estimate”, “steinian_estimate”. In this example, “cross_validation” is used.

  • discrepancy_method: Method for discrepancy calculation, chosen from “square_error”, “classification_error”, “logistic_deviance”. In this example, “logistic_deviance” is used because the family is binomial.

  • tau_seq: Vector of positive numeric values for down-weighting synthetic data. Defaults to a sequence around one fourth of the number of predictors.

  • cross_validation_fold_num: Number of folds for cross-validation. Defaults to 5.

cat_glm_tune_cv <- cat_glm_tune(
  formula = empyr1 ~ ., # Same as `~ .`
  cat_init = cat_init,
  risk_estimate_method = "cross_validation", # Default auto-select based on the data size
  discrepancy_method = "logistic_deviance", # Default auto-select based on family
  tau_seq = seq(0.1, 5.1, 0.5), # Default is a numeric sequence around the number of predictors / 4. Do not recommand to including 0 for cross validation.
  cross_validation_fold_num = 3 # Default: 5
)

cat_glm_tune_cv

User can plot the tau_seq (x) against discrepancy error (y) using the plot() function. This plot will show the lowest discrepancy error at the optimal tau value.

plot(cat_glm_tune_cv)

Bootstrap (risk_estimate_method = “parametric_bootstrap”)

This method estimates tau using bootstrap resampling, refining the model through iterative sampling to enhance robustness and accuracy.

Here’s a breakdown of other parameters used:

  • tau_0: Initial tau value used for discrepancy calculation in risk estimation. Defaults to one fourth of the number of predictors for binomial and 1 for gaussian.

  • bootstrap_iteration_times: Number of bootstrap iterations for “parametric_bootstrap” risk estimation. Defaults to 100.

For the breakdown of other input parameters, please check section Cross Validation

cat_glm_tune_boots <- cat_glm_tune(
  formula = ~., # Same as `empyr1 ~ .`
  cat_init = cat_init,
  risk_estimate_method = "parametric_bootstrap", # Default auto-select based on the data size
  discrepancy_method = "logistic_deviance", # Default auto-select based on family
  tau_0 = 2, # Default: number of predictors / 4
  parametric_bootstrap_iteration_times = 10, # Default: 100
)

cat_glm_tune_boots

Steinian Estimate (risk_estimate_method = “steinian_estimate”)

This method computes the partial likelihood using a steinian estimate approach, optimizing the model based on observed and synthetic data.

For the breakdown of the input parameters, please check section Cross Validation and Bootstrap)

cat_glm_tune_stein <- cat_glm_tune(
  formula = ~., # Same as `empyr1 ~ .`
  cat_init = cat_init,
  risk_estimate_method = "steinian_estimate", # Default auto-select based on the data size
  discrepancy_method = "logistic_deviance", # default  auto select based on family
)

cat_glm_tune_stein

Recommendations for Choosing risk_estimate_method and discrepancy_method

Choosing the appropriate risk_estimate_method and discrepancy_method depends on the data size, model complexity, and the specific requirements of user’s analysis.

  1. Choosing risk_estimate_method
  • Small to Medium Data Size (observation data size <= 200)
    • Default: “parametric_bootstrap”
    • For smaller datasets, parametric bootstrap is generally chosen. It efficiently estimates the model’s performance using resampling techniques.
  • Large Data Size (observation data size > 200):
    • Default: “cross_validation”
    • For larger datasets, cross-validation is preferred. It provides a robust estimate of model performance by splitting the data into multiple folds and averaging the results.
  1. Setting discrepancy_method
  • GLM Family is binomial (family == “binomial”)
    • For binomial GLM families, “logistic_deviance” is the most appropriate option. It calculates the discrepancy based on the logistic regression. Besides “logistic_deviance”, user can also choose “classification_error” for binary outcomes.
  • GLM Family is not binomial (family != “binomial”):
    • Default: “square_error”
    • For other GLM families (e.g. gaussian), “square_error” is the only option. It measures the discrepancy using squared differences, appropriate for continuous outcomes. Please check catalytic_glm_gaussian for more details.

Automatic Parameter Selection

Of course, user don’t need to worry about specifying these parameters explicitly, and they just need to simply provide the cat_init object and the formula. then cat_glm_tune will automatically select risk_estimate_method and discrepancy_method based on the dataset size and GLM family type.

In this example, it is risk_estimate_method = "parametric_bootstrap" and discrepancy_method = "logistic_deviance".

For the breakdown of the input parameters, please check section Cross Validation and Bootstrap

cat_glm_tune_auto <- cat_glm_tune(
  formula = ~.,
  cat_init = cat_init
)

cat_glm_tune_auto

Let’s check the prediction error.

cat_glm_tune_predicted_y <- predict(
  cat_glm_tune_auto,
  newdata = test_data,
  type = "response"
)

cat(
  "Catalytic `cat_glm_tune` - Logarithmic Error:",
  get_mean_logarithmic_error(
    Y = test_data$empyr1,
    est_Y = cat_glm_tune_predicted_y
  )
)

Let us check the ROC curve of the cat_glm_tune_predicted_y from cat_glm_tune_auto versus the test_data$empyr1, this can be a great way to visually assess the accuracy and performance of the model.

roc_curve <- pROC::roc(test_data$empyr1, cat_glm_tune_predicted_y)
plot(roc_curve, main = "ROC Curve (cat_glm_tune)", col = "blue", lwd = 2)

All above objects in this section including cat_glm_tune_auto objects are outputs from the cat_glm_tune function, providing a list of attributes for further analysis or user inspection.

Here’s a breakdown of all attributes except the input parameters:

  • function_name: The name of the function (cat_glm_tune).

  • tau: Selected optimal tau value from tau_seq that minimizes discrepancy error.

  • model: The fitted GLM model object obtained from stats::glm, with the selected optimal tau (tau).

  • coefficients: The estimated coefficients from the fitted GLM model (model).

  • risk_estimate_list: Collected risk estimates across different tau values.

For more details, please check ?cat_glm_tune.

names(cat_glm_tune_auto)

User can extract items mentioned above from cat_glm_tune object.

# The method used for risk estimation
cat_glm_tune_auto$risk_estimate_method

# Selected optimal tau value from `tau_seq` that minimizes discrepancy error
cat_glm_tune_auto$tau

Step 2.3: Choose Method(s) - Bayesian Posterior Sampling with Fixed Tau

Now, we will explore advanced Bayesian modeling techniques tailored for GLM Binomial using the catalytic package. Bayesian inference offers a powerful framework to estimate model parameters and quantify uncertainty by integrating prior knowledge with observed data.

Below functions enable Bayesian inference for GLM Logistic Regression Model with catalytic priors. This function utilizes Markov chain Monte Carlo (MCMC) methods, implemented using the rstan package, to sample from the posterior distribution of model parameters. Users can specify various options such as the number of MCMC chains (chains), iterations (iter), warmup steps (warmup), and the MCMC algorithm (algorithm). User could also apply other attributes to rstan::sampling, like refresh and control.

In this section, we explore Bayesian approaches using the cat_glm_bayes function from the catalytic package. This function can fit a Bayesian Generalized Linear Model (GLM) using a fixed tau value. The MCMC sampling process will generate posterior distributions for the coefficients based on the specified tau.

Here’s a breakdown of the parameters used:

For more details, please refer to ?cat_glm_bayes.

cat_glm_bayes_model <- cat_glm_bayes(
  formula = empyr1 ~ ., # Same as `~ .`
  cat_init = cat_init,
  tau = 50, # Default: number of predictors / 4
  chains = 1, # Default: 4
  iter = 100, # Default: 2000
  warmup = 50, # Default: 1000
  algorithm = "NUTS" # Default: NUTS
)

cat_glm_bayes_model

Here shows how users can simplify the input for cat_glm_bayes. User do not have to specify tau and other attributes, as tau and other attributes have default value, which mentioned above. Here we assign lower value to chains, iter and warmup for quicker processing time.

User can also get the trace plot of the rstan model by using traceplot() directly into the output from cat_glm_bayes.

traceplot(cat_glm_bayes_model)

Plus, user can use this catlaytic::traceplot just like the rstan::traceplot, user can add parameters used in rstan::traceplot, like include and inc_warmup.

traceplot(cat_glm_bayes_model, inc_warmup = TRUE)

Let’s check the prediction error.

cat_glm_bayes_predicted_y <- predict(
  cat_glm_bayes_model,
  newdata = test_data,
  type = "response"
)

cat(
  "MLE GLM Binomial Model - Logarithmic Error:",
  get_mean_logarithmic_error(
    Y = test_data$empyr1,
    est_Y = cat_glm_bayes_predicted_y
  )
)

Let us check the ROC curve of the cat_glm_bayes_predicted_y from cat_glm_bayes_model versus the test_data$empyr1, this can be a great way to visually assess the accuracy and performance of the model.

roc_curve <- pROC::roc(test_data$empyr1, cat_glm_bayes_predicted_y)
plot(roc_curve, main = "ROC Curve (cat_glm_bayes)", col = "blue", lwd = 2)

Both cat_glm_bayes_model and cat_glm_bayes_model objects are outputs from the cat_glm_bayes function, providing a list of attributes for further analysis or user inspection.

Here’s a breakdown of all attributes except the input parameters:

For more details, please refer to ?cat_glm_bayes.

names(cat_glm_bayes_model)

User can extract items mentioned above from cat_glm_bayes object.

# The number of Markov chains used during MCMC sampling in `rstan`.
cat_glm_bayes_model$chain

# The mean estimated coefficients from the Bayesian GLM model
cat_glm_bayes_model$coefficients

Step 2.4: Choose Method(s) - Bayesian Posterior Sampling with Adaptive Tau

In this section, we delve into Bayesian methodologies employing the cat_glm_bayes_joint function within the catalytic package. Unlike its non-adaptive counterpart (cat_glm_bayes), this method employs a joint tau prior approach where tau is treated as a parameter within the MCMC sampling process, improving the robustness and accuracy of parameter estimation in Bayesian Binomial modeling.

In this section, we explore Bayesian approaches using the cat_glm_bayes_joint function from the catalytic package. These functions are similar to their non-adaptive (non-joint) version (cat_glm_bayes), but corporate tau into the MCMC sampling process.

Here’s a breakdown of the parameters used:

For more details, please refer to ?cat_glm_bayes_joint.

cat_glm_bayes_joint_model <- cat_glm_bayes_joint(
  formula = empyr1 ~ ., # Same as `~ .`
  cat_init = cat_init,
  chains = 1, # Default: 4
  iter = 100, # Default: 2000
  warmup = 50, # Default: 1000
  algorithm = "NUTS", # Default: NUTS
  tau_alpha = 2, # Default: 2
  tau_gamma = 1 # Default: 1
)

cat_glm_bayes_joint_model

Here shows how users can simplify the input for cat_glm_bayes_joint. User do not have to specify tau_alpha and other attributes, as these attributes have default value, which mentioned above. Here we assign lower value to chains, iter and warmup for quicker processing time.

User can also get the traceplot of the rstan model by using traceplot() directly into the output from cat_glm_bayes_joint.

traceplot(cat_glm_bayes_joint_model)

Like the traceplot shown in the cat_glm_bayes function , user can add parameters used in rstan::traceplot, like include and inc_warmup.

traceplot(cat_glm_bayes_joint_model, inc_warmup = TRUE)

Let’s check the prediction error.

cat_glm_bayes_joint_predicted_y <- predict(
  cat_glm_bayes_joint_model,
  newdata = test_data,
  type = "response"
)

cat(
  "Catlytic `cat_glm_bayes_joint` - Logarithmic Error:",
  get_mean_logarithmic_error(
    Y = test_data$empyr1,
    est_Y = cat_glm_bayes_joint_predicted_y
  )
)

Let us check the ROC curve of the cat_glm_bayes_joint_predicted_y from cat_glm_bayes_joint_model versus the test_data$empyr1, this can be a great way to visually assess the accuracy and performance of the model.

roc_curve <- pROC::roc(test_data$empyr1, cat_glm_bayes_joint_predicted_y)
plot(roc_curve, main = "ROC Curve (cat_glm_bayes_joint)", col = "blue", lwd = 2)

Both cat_glm_bayes_joint_model and cat_glm_bayes_joint_model objects are outputs from the cat_glm_bayes_joint function, providing a list of attributes for further analysis or user inspection.

Here’s a breakdown of all attributes except the input parameters:

For more details, please refer to ?cat_glm_bayes_joint.

names(cat_glm_bayes_joint_model)

User can extract items mentioned above from cat_glm_bayes_joint object.

# The estimated tau parameter from the MCMC sampling `rstan::sampling`,
cat_glm_bayes_joint_model$tau

# The mean estimated coefficients from the Bayesian GLM model
cat_glm_bayes_joint_model$coefficients

Step 2.5: Choose Method(s) - Special Approaches for Binomial Models

Traditional methods like rstan often face challenges in accurately fitting Binomial models, such as computational inefficiencies and convergence issues. To overcome these limitations, we would like to introduce below four specialized approaches designed to improve parameter estimation accuracy, enhance convergence rates, and better handle the complexities inherent in Binomial modeling.

Gibbs Sampling

Besides the above two approaches, cat_glm_bayes_joint_gibbs function utilizes Gibbs sampling. enhancing performance in Bayesian modeling tasks.

This approach is particularly advantageous for handling complex data distributions and intricate variable inter-dependency. Gibbs sampling sequentially samples from conditional distributions, efficiently exploring joint posterior distributions and promoting faster convergence rates compared to direct rstan approaches.

Within the cat_glm_bayes_joint_gibbs function, Hamiltonian Monte Carlo (HMC) is employed for more efficient exploration of the parameter space, as described by Radford M. Neal (2010).

The Gibbs algorithm can be used to sample from the posterior distribution under the joint prior \(\pi_{\alpha, \gamma}(\tau, \beta)\). The Gibbs algorithm iterative samples one component from the conditional distribution holding the other component fixed. Given coefficients (\(\beta\)), an update of τ can be sampled from the Gamma distribution.

\[ \pi_{\alpha, \gamma}(\tau | \beta, \mathbf{Y}, \mathbf{X}) \propto\Gamma_{\alpha, \gamma}(\tau) \exp \left( \frac{\tau}{M} \sum_{i=1}^{M} \log(f(Y_i^* | \beta^\top \mathbf{X}_i^*)) \right) \]

\[ = \tau^{c-1} \exp \left( -\tau \left( \kappa + \frac{1}{\gamma} - \frac{1}{M} \sum_{i=1}^{M} \log(f(Y_i^* | \beta^\top \mathbf{X}_i^*)) \right) \right) \]

Given tau (\(\tau\)), an update of coefficients (\(\beta\)) should be drawn from

\[ \pi(\beta | \tau, \mathbf{Y}, \mathbf{X}) \propto f(Y^*|X^*, \beta)^{\tau/M}f(Y|X, beta) \]

It can be sampled by various methods such as the Metropolis-Hasting algorithm and the Hamiltonian Monte Carlo (HMC). We recommend to use HMC with random step size and with adaptive variances for the momentum variables. Before running HMC within Gibbs, the adaptive variances are set at the diagonal entries of the inverse Hessian matrix of the negative log density at the posterior mode. The initial point of such a MCMC step should be the most recent sample of coefficients (\(\beta\)).

For more information, see (Huang et al. 2020) and (Neal 2012).

Here’s a breakdown of the parameters used:

  • formula, and cat_init are same as above.

  • iter: Total number of Gibbs sampling iterations. Defaults to 2000.

  • warmup: Number of warmup iterations in Gibbs sampling. Defaults to 1000.

  • coefs_iter: Number of iterations in the HMC sampling for coefficients. Defaults to 5.

  • tau_0: Initial value for tau (down-weighting factor for synthetic data). Defaults to one forth of the number of predictors.

  • tau_alpha: Shape parameter for the gamma distribution used in tau sampling. Defaults to 2.

  • tau_gamma: Rate parameter for the gamma distribution used in tau sampling. Defaults to 1.

  • refresh: Logical flag indicating whether to print progress in Gibbs sampling. Defaults to TRUE.

  • ...(ellipsis): Denotes additional arguments that can be passed directly to the underlying stats::glm, for generating init_coefficients if user not defined, such as control, offset, or other model-specific settings.

For more details, please refer to ?cat_glm_bayes_joint_gibbs. Here we assign lower value to chains, iter and warmup for quicker processing time.

cat_glm_bayes_joint_gibbs_model <- cat_glm_bayes_joint_gibbs(
  formula = empyr1 ~ ., # Same as `~ .`
  cat_init = cat_init,
  iter = 100, # Default: 1000
  warmup = 50, # Default: 500
  coefs_iter = 2, # Default: 5
  tau_0 = 1, # Default: number of predictors / 4
  tau_alpha = 2, # Default: 2
  tau_gamma = 1, # Default: 1
  refresh = TRUE # Default: TRUE
)

cat_glm_bayes_joint_gibbs_model

Here shows how users can simplify the input for cat_glm_bayes_joint_gibbs. User do not have to specify tau_0 and other attributes, as these attributes have default value, which mentioned above. Here we assign lower value to iter and warmup for quicker processing time.

User can also use traceplot() to check the coefficients from cat_glm_bayes_joint_gibbs.

traceplot(cat_glm_bayes_joint_gibbs_model)

And user can select the columns they would like to check.

traceplot(cat_glm_bayes_joint_gibbs_model, pars = c("female", "hsdip", "numchild"))

Let’s check the prediction error.

cat_glm_bayes_joint_gibbs_predicted_y <- predict(
  cat_glm_bayes_joint_gibbs_model,
  newdata = test_data,
  type = "response"
)

cat(
  "Catalytic `cat_glm_bayes_joint_gibbs` - Logarithmic Error:",
  get_mean_logarithmic_error(
    Y = test_data$empyr1,
    est_Y = cat_glm_bayes_joint_gibbs_predicted_y
  )
)

Let us check the ROC curve of the cat_glm_bayes_joint_gibbs_predicted_y from cat_glm_bayes_joint_gibbs versus the test_data$empyr1, this can be a great way to visually assess the accuracy and performance of the model.

roc_curve <- pROC::roc(test_data$empyr1, cat_glm_bayes_joint_gibbs_predicted_y)
plot(roc_curve, main = "ROC Curve (cat_glm_bayes_joint_gibbs)", col = "blue", lwd = 2)

Both cat_glm_bayes_joint_gibbs_model and cat_glm_bayes_joint_gibbs_model objects are outputs from the cat_glm_bayes_joint_gibbs function, providing a list of attributes for further analysis or user inspection.

Here’s a breakdown of all attributes except the input parameters:

  • function_name: The name of the function (cat_glm_bayes_joint_gibbs).

  • sys_time: The date and time when sampling ended.

  • gibbs_iteration_log: A matrix containing the Gibbs sampling results for both tau and coefficients.

  • inform_df: The summary statistics including mean, standard error of mean, standard deviation, quantiles, and effective sample size for coefficients and tau.

  • tau: The mean of estimated tau from sampled value.

  • coefficients: The mean of estimated coefficients from sampled value.

For more details, please refer to ?cat_glm_bayes_joint_gibbs.

names(cat_glm_bayes_joint_gibbs_model)

User can extract items mentioned above from cat_glm_bayes_joint_gibbs object.

# The Date and time when sampling ended.
cat_glm_bayes_joint_gibbs_model$sys_time

# The summary statistics
cat_glm_bayes_joint_gibbs_model$inform_df

Setting Higher Tau Lower Bound

Since the inefficiency in sampling is caused by small values of tau, a straightforward solution is to modify the joint prior by truncating tau to prevent it from approaching 0. The cat_glm_bayes_joint function incorporates binomial_tau_lower within the rstan model to enhance its performance. Adjusting this lower bound transforms the parameter space, a critical step in leveraging Bayesian inference to manage complex data distributions effectively.

binomial_tau_lower (default is 0.05) sets the minimum allowable value for the tau parameter in the Bayesian model, ensuring tau is greater than or equal to number_of_predictors * binomial_tau_lower, which defines a suitable range for model calculations. This adjustment improves the coherence of the posterior density and enhances sampling outcomes. Additionally, the density of the second derivative is not too small, indicating improved performance.

For breakdown of input parameters and output attributes, please check section Bayesian Posterior Sampling with Adaptive Tau. Here we assign lower value to chains, iter and warmup for quicker processing time.

cat_glm_bayes_joint_tau_lower <- cat_glm_bayes_joint(
  formula = ~., # Same as `empyr1 ~ .`
  cat_init = cat_init,
  binomial_tau_lower = 0.5, # Default: 0.05
  chains = 1, # Default: 4
  iter = 100, # Default: 2000
  warmup = 50 # Default: 1000
)

cat_glm_bayes_joint_tau_lower

Let’s check the prediction error.

cat_glm_bayes_joint_tau_lower_predicted_y <- predict(
  cat_glm_bayes_joint_tau_lower,
  newdata = test_data,
  type = "response"
)

cat(
  "Catalytic `cat_glm_bayes_joint_tau_lower` - Logarithmic Error:",
  get_mean_logarithmic_error(
    Y = test_data$empyr1,
    est_Y = cat_glm_bayes_joint_tau_lower_predicted_y
  )
)

Let us check the ROC curve of the cat_glm_bayes_joint_tau_lower_predicted_y from cat_glm_bayes_joint_tau_lower versus the test_data$empyr1, this can be a great way to visually assess the accuracy and performance of the model.

roc_curve <- pROC::roc(test_data$empyr1, cat_glm_bayes_joint_tau_lower_predicted_y)
plot(roc_curve, main = "ROC Curve (MLE)", col = "blue", lwd = 2)

Incorporating theta (1/tau)

For GLM Binomial models, the joint distribution of tau and coefficients displays a banana-shape contour, which makes it challenging for the no-U-turn sampler to effectively transit over the parameter space. To address this, the cat_glm_bayes_joint function introduces theta = 1/tau into the rstan model.

In the rstan program used in this example, theta serves as the inverse of tau, allowing the model to adjust its parameterization dynamically based on the the characteristics of data.

This approach enhances flexibility in model fitting, especially when dealing with datasets exhibiting diverse and complex patterns across different dimensions.

For breakdown of input parameters and output attributes, please check section Bayesian Posterior Sampling with Adaptive Tau. Here we assign lower value to chains, iter and warmup for quicker processing time.

cat_glm_bayes_joint_theta <- cat_glm_bayes_joint(
  formula = ~., # Same as `empyr1 ~ .`
  cat_init = cat_init,
  binomial_joint_theta = TRUE, # Default: FALSE
  chains = 1, # Default: 4
  iter = 100, # Default: 2000
  warmup = 50 # Default: 1000
)

cat_glm_bayes_joint_theta

Let’s check the prediction error.

cat_glm_bayes_joint_theta_predicted_y <- predict(
  cat_glm_bayes_joint_theta,
  newdata = test_data,
  type = "response"
)

cat(
  "Catalytic `cat_glm_bayes_joint_theta` - Logarithmic Error:",
  get_mean_logarithmic_error(
    Y = test_data$empyr1,
    est_Y = cat_glm_bayes_joint_theta_predicted_y
  )
)

Let us check the ROC curve of the cat_glm_bayes_joint_theta_predicted_y from cat_glm_bayes_joint_theta versus the test_data$empyr1, this can be a great way to visually assess the accuracy and performance of the model.

roc_curve <- pROC::roc(test_data$empyr1, cat_glm_bayes_joint_theta_predicted_y)
plot(roc_curve, main = "ROC Curve (cat_glm_bayes_joint_theta)", col = "blue", lwd = 2)

Incorporating theta (1/tau) with Adaptive Alpha

Additionally, in scenarios where adjusting the spread of data (controlled by alpha) significantly impacts model performance, the cat_glm_bayes_joint function extends thetheta = 1/tau approach by introducing a adaptive/joint alpha parameter. This enhancement in the rstan model allows simultaneous adjustment of both theta and alpha parameters, aiming for better convergence and more accurate parameter estimates.

For breakdown of input parameters and output attributes, please check section Bayesian Posterior Sampling with Adaptive Tau. Here we assign lower value to chains, iter and warmup for quicker processing time.

cat_glm_bayes_joint_alpha <- cat_glm_bayes_joint(
  formula = ~., # Same as `empyr1 ~ .`
  cat_init = cat_init,
  binomial_joint_theta = TRUE, # Default: FALSE
  binomial_joint_alpha = TRUE, # Default: FALSE
  chains = 1, # Default: 4
  iter = 100, # Default: 2000
  warmup = 50 # Default: 1000
)

cat_glm_bayes_joint_alpha

Let’s check the prediction error.

cat_glm_bayes_joint_alpha_predicted_y <- predict(
  cat_glm_bayes_joint_alpha,
  newdata = test_data,
  type = "response"
)

cat(
  "Catalytic `cat_glm_bayes_joint_alpha` - Logarithmic Error:",
  get_mean_logarithmic_error(
    Y = test_data$empyr1,
    est_Y = cat_glm_bayes_joint_alpha_predicted_y
  )
)

Let us check the ROC curve of the cat_glm_bayes_joint_alpha_predicted_y from cat_glm_bayes_joint_alpha versus the test_data$empyr1, this can be a great way to visually assess the accuracy and performance of the model.

roc_curve <- pROC::roc(test_data$empyr1, cat_glm_bayes_joint_alpha_predicted_y)
plot(roc_curve, main = "ROC Curve (cat_glm_bayes_joint_alpha)", col = "blue", lwd = 2)

References

Huang, Dongming, Nathan Stein, Donald B. Rubin, and S. C. Kou. 2020. “Catalytic Prior Distributions with Application to Generalized Linear Models.” ResearchGate. https://www.researchgate.net/publication/341417038_Catalytic_prior_distributions_with_application_to_generalized_linear_models.
Neal, Radford M. 2012. “MCMC Using Hamiltonian Dynamics.” https://arxiv.org/pdf/1206.1901.