Statistical Inference Unit 14 ReviewDecision Theory in Statistical Inference

Key Concepts in Decision Theory

• Decision theory provides a framework for making optimal decisions under uncertainty
• Involves specifying a set of possible actions, states of nature, and consequences
• Consequences are determined by the action taken and the true state of nature
• Incorporates prior knowledge or beliefs about the states of nature (prior probabilities)
• Aims to minimize expected loss or maximize expected utility
  • Expected loss is the average loss incurred over all possible states of nature
  • Expected utility is the average utility gained over all possible states of nature
• Requires defining a loss function or utility function to quantify the consequences of actions
• Distinguishes between two main approaches: Bayesian and frequentist decision theory

Statistical Decision Problems

• Arise when making decisions based on statistical data or inference
• Involve choosing an action from a set of possible actions based on observed data
• The true state of nature is unknown but can be described probabilistically
• Goal is to make the best decision given the available information and uncertainty
• Examples include:
  • Hypothesis testing (deciding whether to reject or fail to reject a null hypothesis)
  • Parameter estimation (choosing an estimator for an unknown parameter)
  • Classification (assigning an object to one of several categories based on its features)
• Requires specifying the following components:
  • Parameter space: the set of possible states of nature or true parameter values
  • Action space: the set of possible actions or decisions that can be taken
  • Loss function: a function that quantifies the loss or cost associated with each action-state pair

Loss Functions and Risk

• A loss function $L(\theta, a)$ quantifies the loss incurred when taking action $a$ if the true state of nature is $\theta$
• The choice of loss function depends on the specific problem and the consequences of different actions
• Common loss functions include:
  • Squared error loss: $L(\theta, a) = (\theta - a)^2$
  • Absolute error loss: $L(\theta, a) = |\theta - a|$
  • 0-1 loss: $L(\theta, a) = \begin{cases} 0 & \text{if } a = \theta \\ 1 & \text{if } a \neq \theta \end{cases}$
• The risk function $R(\theta, \delta)$ is the expected loss of a decision rule $\delta$ under the true state of nature $\theta$
  • $R(\theta, \delta) = \mathbb{E}_\theta[L(\theta, \delta(X))]$, where $X$ is the observed data
• A decision rule $\delta$ is a function that maps the observed data $X$ to an action $a$
• The goal is to find a decision rule that minimizes the risk function over all possible states of nature

Bayesian Decision Theory

• Incorporates prior knowledge or beliefs about the states of nature through a prior probability distribution $\pi(\theta)$
• Updates the prior distribution using the observed data $X$ to obtain a posterior distribution $\pi(\theta|X)$ via Bayes' theorem
• The Bayes risk of a decision rule $\delta$ is the expected loss averaged over both the data distribution and the prior distribution
  • $r(\pi, \delta) = \mathbb{E}_\pi[\mathbb{E}_\theta[L(\theta, \delta(X))]]$
• The Bayes decision rule $\delta^*$ minimizes the Bayes risk among all possible decision rules
• Allows for the incorporation of subjective prior information and provides a principled way to update beliefs based on data
• Useful when prior information is available and can lead to better decisions by leveraging this information

Frequentist Decision Theory

• Focuses on the long-run performance of decision rules under repeated sampling
• Does not incorporate prior distributions and relies solely on the observed data
• Aims to find decision rules that perform well on average across all possible states of nature
• Minimax principle: choose the decision rule that minimizes the maximum risk over all states of nature
  • $\delta^* = \arg\min_\delta \max_\theta R(\theta, \delta)$
• Admissibility: a decision rule is admissible if no other rule has smaller or equal risk for all states of nature and strictly smaller risk for at least one state
• Unbiasedness: a decision rule is unbiased if its risk function satisfies certain symmetry properties
• Frequentist decision theory provides a framework for evaluating and comparing decision rules based on their long-run performance

Minimax and Admissible Decision Rules

• Minimax decision rules aim to minimize the maximum risk over all possible states of nature
• Useful when the goal is to protect against the worst-case scenario
• The minimax risk is the smallest possible maximum risk that can be attained by any decision rule
  • $R^* = \min_\delta \max_\theta R(\theta, \delta)$
• A decision rule $\delta^*$ is minimax if it achieves the minimax risk, i.e., $\max_\theta R(\theta, \delta^*) = R^*$
• Admissible decision rules are those for which no other rule has smaller or equal risk for all states of nature and strictly smaller risk for at least one state
• Admissible rules are Pareto optimal: cannot be improved upon without increasing the risk for some state of nature
• Minimax rules are always admissible, but not all admissible rules are minimax
• Admissible rules form a subset of all possible decision rules and are of interest because they cannot be universally improved upon

Applications in Statistical Inference

• Hypothesis testing: deciding whether to reject or fail to reject a null hypothesis based on observed data
  • Loss functions can be defined to penalize Type I and Type II errors differently
  • Minimax and Bayes decision rules can be derived for various testing problems
• Parameter estimation: choosing an estimator for an unknown parameter based on observed data
  • Loss functions such as squared error or absolute error can be used to quantify the accuracy of estimators
  • Minimax and Bayes estimators can be derived to minimize the maximum or average risk
• Classification: assigning an object to one of several categories based on its features
  • Loss functions can be defined to penalize different types of misclassification errors
  • Bayes and minimax classifiers can be derived to minimize the expected or worst-case misclassification risk
• Model selection: choosing the best model from a set of candidate models based on observed data
  • Loss functions can be defined to balance model fit and complexity (e.g., AIC, BIC)
  • Bayes and frequentist model selection criteria can be derived using decision-theoretic principles

Advanced Topics and Current Research

• Robust decision theory: making decisions that are insensitive to deviations from assumed models or distributions
  • Minimax regret: minimizing the maximum regret (difference between the loss of the chosen action and the best possible action) over a set of possible models
  • Robust Bayes: incorporating uncertainty in the prior distribution and finding decision rules that perform well over a range of priors
• Sequential decision theory: making a series of decisions over time, where each decision may depend on previous observations and actions
  • Dynamic programming: breaking down a sequential decision problem into smaller subproblems and solving them recursively
  • Multi-armed bandits: balancing exploration and exploitation when making decisions with uncertain rewards
• Causal decision theory: making decisions based on causal relationships between variables, rather than just statistical associations
  • Causal graphs: representing the causal structure of a problem using directed acyclic graphs
  • Interventions: evaluating the effects of actions by considering their impact on the causal system
• Algorithmic decision theory: studying the computational complexity and tractability of decision-making algorithms
  • Approximation algorithms: finding decision rules that are provably close to optimal while being computationally efficient
  • Online learning: making decisions and updating beliefs in real-time as new data becomes available