8. Probabilistic, Causal, and Uncertainty-Aware AI

Introduction and Context

Artificial intelligence systems rarely operate with complete information. Sensors are noisy, databases contain missing values, labels may be ambiguous, and observations may support several competing explanations. In many applications, the central question is not simply whether a proposition is true, but how strongly the available evidence supports it, what would happen under an intervention, and how a decision should be made when outcomes remain uncertain.

Classical machine learning often expresses uncertainty indirectly through prediction errors or confidence scores. Advanced AI requires a more disciplined treatment. Probability theory provides a language for representing incomplete knowledge; probabilistic graphical models encode dependencies among variables; Bayesian inference updates beliefs in response to evidence; causal models distinguish association from intervention; and decision theory connects uncertain predictions with actions, utilities, and risks.

This chapter develops these ideas as mutually related but distinct perspectives. A probabilistic model describes uncertainty about variables. A causal model describes how a system generates outcomes and how interventions alter it. A decision model asks which action is preferable given beliefs about the world and the consequences of error. The distinctions are essential in medicine, finance, robotics, public policy, scientific discovery, and trustworthy machine learning.

Learning Objectives

By the end of this chapter, the reader should be able to:

1. Explain conditional probability, Bayes’ theorem, likelihood, prior and posterior distributions.

2. Construct and interpret Bayesian networks and Markov random fields.

3. Compare exact and approximate inference methods, including sampling and variational approaches.

4. Describe Gaussian processes and their use in uncertainty-aware prediction.

5. Distinguish observational association, intervention, and counterfactual reasoning.

6. Use structural causal models and causal graphs to analyze confounding, mediation, and selection bias.

7. Explain major approaches to causal discovery and their assumptions.

8. Assess aleatoric and epistemic uncertainty and evaluate calibration.

9. Formulate decisions under incomplete information using expected utility and value of information.

10. Critically evaluate technical, ethical, governance, and deployment issues in probabilistic and causal AI.

Probability as a Language for Incomplete Knowledge

Let a random variable represent an uncertain quantity. A discrete variable has a probability mass function, while a continuous variable is described by a density. Probability is not necessarily a physical property of an object; in Bayesian reasoning it can represent a rational degree of belief conditional on current information.

The fundamental identity is the product rule:

\[

P(A,B)=P(A\mid B)P(B).

\]

The marginal probability is obtained by summing or integrating over unobserved variables. Bayes’ theorem follows directly:

\[

P(H\mid E)=\frac{P(E\mid H)P(H)}{P(E)}.

\]

Here, \(H\) is a hypothesis and \(E\) is evidence. The prior \(P(H)\) expresses belief before observing the evidence, the likelihood \(P(E\mid H)\) describes how compatible the evidence is with the hypothesis, and the posterior \(P(H\mid E)\) is the updated belief. The denominator normalizes the competing hypotheses.

A common error is to confuse \(P(E\mid H)\) with \(P(H\mid E)\). For example, a diagnostic test may be highly sensitive among diseased patients, yet the probability that a person with a positive result is diseased can be modest when the disease is rare. Base-rate information is therefore not an optional refinement; it is mathematically central.

Independence simplifies probability models. Variables X and Y are independent if \(P(X,Y)=P(X)P(Y)\). Conditional independence means that X and Y become independent after conditioning on Z: \(P(X,Y\mid Z)=P(X\mid Z)P(Y\mid Z)\). Conditional independence is especially valuable because it permits complex joint distributions to be represented compactly.

Two interpretations of uncertainty are useful. Aleatoric uncertainty arises from inherent variability or randomness, such as fluctuations in demand. Epistemic uncertainty arises from limited knowledge, such as insufficient training data in a region of the input space. Aleatoric uncertainty may remain even with more data; epistemic uncertainty can often be reduced through additional observations, better experiments, or improved models.

Graphical Models: Representing Structure and Dependence

A probabilistic graphical model combines graph theory with probability distributions. Nodes represent random variables, and edges express relationships or dependencies. The graph does not automatically prove causation. Its meaning depends on the model class and assumptions attached to it.

Bayesian Networks

A Bayesian network is a directed acyclic graph (DAG) together with a conditional distribution for each node given its parents. If the graph contains variables A, B, and C, and A and B point to C, the joint distribution factorizes as:

\[

P(A,B,C)=P(A)P(B)P(C\mid A,B).

\]

More generally,

\[

P(X_1,\ldots,X_n)=\prod_{i=1}^{n}P(X_i\mid Pa(X_i)),

\]

where \(Pa(X_i)\) denotes the parents of variable \(X_i\).

The graph conveys conditional independence through d-separation. A chain A → B → C implies that A and C are generally associated, but may become conditionally independent given B. A fork A ← B → C also produces association that may disappear after conditioning on B. A collider A → B ← C behaves differently: A and C may be independent before conditioning on B, but conditioning on the collider or its descendant can create dependence. This phenomenon is known as explaining away and is a major source of selection bias.

A Bayesian network can support diagnostic inference, predictive inference, and belief updating. Its construction may rely on expert knowledge, historical data, or both. Parameter learning estimates conditional probability tables for discrete variables or conditional densities for continuous variables. Structure learning seeks the graph itself, using constraint-based, score-based, or hybrid procedures.

Markov Random Fields

A Markov random field, or undirected graphical model, represents symmetric relationships using an undirected graph. Its distribution is often written in terms of clique potentials:

\[

P(X)=\frac{1}{Z}\prod_{c\in C}\psi_c(X_c),

\]

where \(\psi_c\) is a non-negative potential over clique c and Z is the partition function. Unlike a Bayesian network, an undirected model does not assign directional parent-child semantics. Markov random fields are useful when interactions are naturally symmetric, as in image segmentation, spatial modeling, and energy-based learning.

Conditional random fields are discriminative undirected models that represent the conditional distribution of labels given observations. They have been used in sequence labeling, natural language processing, and structured prediction. Their strength lies in modeling dependencies among output variables without requiring a generative model for the observations.

Latent-Variable Models

Latent variables are unobserved quantities introduced to explain observed regularities. Mixture models assume that observations arise from one of several hidden components. Hidden Markov models add temporal structure: a hidden state generates an observation and transitions probabilistically to the next hidden state. State-space models generalize this idea to continuous variables and nonlinear dynamics.

Latent-variable modeling creates an interpretive obligation. A hidden component may be mathematically useful without corresponding to a real-world entity. Researchers should avoid treating statistically identified factors as substantive causes unless independent evidence supports that interpretation.

Inference: From Evidence to Belief

Inference calculates a distribution over unknown variables given observed variables. Exact inference in a Bayesian network can be performed through enumeration, variable elimination, or message passing when the graph structure is manageable. The computational difficulty is often governed by graph width, such as treewidth, rather than simply the number of variables.

Belief propagation is exact on trees and certain tree-like graphs. On graphs with loops, loopy belief propagation may still perform well, but its convergence and accuracy are not guaranteed. Junction-tree algorithms convert a graph into a tree of cliques, making exact inference possible at the cost of potentially large clique sizes.

For large or continuous models, approximate inference is usually necessary. Markov chain Monte Carlo (MCMC) constructs a Markov chain whose stationary distribution is the target posterior. Metropolis–Hastings accepts or rejects proposals according to an acceptance ratio, while Gibbs sampling updates one variable at a time from its conditional distribution. Important diagnostics include burn-in, autocorrelation, effective sample size, and convergence across chains.

Sequential Monte Carlo, or particle filtering, represents a changing posterior with weighted samples. It is particularly useful for online inference in dynamic systems. Particle degeneracy can occur when most particles acquire negligible weight; resampling and proposal design are therefore important.

Variational inference approximates a difficult posterior with a tractable family of distributions. It commonly minimizes a divergence by optimizing an evidence lower bound (ELBO). Variational methods are often faster and more scalable than MCMC, but their approximation can be overconfident when the chosen family cannot represent posterior dependence or multimodality.

Neural probabilistic models, including Bayesian neural networks, deep ensembles, normalizing flows, and diffusion-based models, extend uncertainty modeling to high-dimensional data. They do not eliminate the need for calibration and model criticism. A flexible model may fit observations well while assigning unreliable probabilities outside the training distribution.

Gaussian Processes and Function-Space Uncertainty

A Gaussian process (GP) is a distribution over functions. It is specified by a mean function m(x) and a covariance kernel k(x,x′). For any finite collection of inputs, the corresponding function values follow a multivariate Gaussian distribution:

\[

f(x)\sim GP(m(x),k(x,x')).

\]

The kernel encodes assumptions about smoothness, periodicity, scale, and similarity. Common kernels include the squared-exponential, Matérn, and periodic kernels. Given noisy observations, GP regression produces a posterior mean and variance at each test point. The posterior variance generally increases in regions distant from observed data, making epistemic uncertainty explicit.

Gaussian processes are attractive for scientific and engineering applications because they work effectively with relatively small datasets and provide interpretable uncertainty. Their classical computational cost, often cubic in the number of observations, limits direct use at very large scale. Sparse inducing-point methods, structured kernels, and approximate linear algebra reduce this burden.

A GP’s uncertainty is only as meaningful as its kernel, likelihood, and data assumptions. Heteroscedastic noise, nonstationarity, measurement error, and distribution shift may require specialized models. Predictive intervals should be evaluated empirically rather than accepted merely because they arise from a Gaussian calculation.

Causal Graphs and Structural Reasoning

Prediction and causation answer different questions. A predictive model may estimate that people who receive treatment T have outcome Y, but this association alone does not identify what would happen if a particular person were treated. Causal analysis requires assumptions about the data-generating process and the meaning of interventions.

A structural causal model (SCM) represents variables using structural equations, for example:

\[

Y=f_Y(X,U_Y),

\]

where X may be a treatment, U represents exogenous influences, and f describes how the outcome is generated. A causal graph summarizes these relationships. The do-operator, written \(do(X=x)\), denotes an intervention that sets X to x and removes the usual mechanisms determining X.

The observational quantity \(P(Y\mid X=x)\) is generally different from the interventional quantity \(P(Y\mid do(X=x))\). If a backdoor path connects X and Y through a confounder Z, adjustment for an appropriate set of variables can identify the intervention effect:

\[

P(Y\mid do(X=x))=\sum_z P(Y\mid X=x,Z=z)P(Z=z),

\]

provided the adjustment set satisfies the relevant graphical conditions.

A confounder is a common cause of treatment and outcome. Adjusting for confounders can reduce bias. A mediator lies on the causal pathway from treatment to outcome; adjusting for it may remove part of the total effect and is inappropriate when the total effect is the target. A collider is a common effect of two variables. Conditioning on it can create a noncausal association and introduce bias.

Counterfactual reasoning concerns alternative outcomes for the same unit under different possible actions. The counterfactual statement “What would this patient’s outcome have been without treatment?” involves a potential outcome that is not simultaneously observed with the treated outcome. SCMs provide a formal procedure: abduction uses observed evidence to infer latent factors, action modifies the structural equations, and prediction computes the counterfactual outcome. Counterfactuals are therefore more demanding than ordinary prediction and depend strongly on structural assumptions.

Causal effects can be summarized in several ways: average treatment effect, conditional treatment effect, individual treatment effect, mediation effect, and policy value. Each estimand requires clarity about population, treatment, outcome, timing, and intervention. A technically sophisticated analysis can still be invalid if its estimand is poorly defined.

Causal Discovery

Causal discovery attempts to infer graph structure from data. Constraint-based methods, such as those based on conditional independence tests, search for graphs compatible with observed independencies. Score-based methods optimize a criterion balancing fit and complexity. Functional causal model approaches exploit assumptions about functional form or noise asymmetry.

In many settings, observational data identify only a Markov equivalence class: several DAGs imply the same conditional independencies. The data may determine some edge orientations but not all. Latent confounding, selection bias, measurement error, feedback, and finite samples make discovery harder.

Interventions can improve identifiability because they break ordinary dependencies and reveal how variables respond to manipulation. Randomized experiments remain a powerful source of causal evidence, although they may be expensive, ethically constrained, or limited in external validity. Modern causal discovery increasingly combines observational data, experiments, temporal information, domain knowledge, and invariant prediction across environments.

Researchers must resist interpreting a discovered graph as unquestionable truth. Causal discovery is model selection under assumptions, not an assumption-free recovery of reality. Sensitivity analysis, falsification tests, replication, and expert scrutiny are essential.

Quantifying and Evaluating Uncertainty

Uncertainty quantification concerns how a model represents, propagates, and communicates uncertainty. A point prediction alone is inadequate when decisions carry asymmetric costs. Predictive intervals, posterior distributions, confidence sets, and calibrated probabilities provide richer information.

Calibration asks whether predicted probabilities correspond to empirical frequencies. If a classifier assigns probability 0.8 to many cases, approximately 80 percent of those cases should belong to the predicted class under the relevant population and conditions. Reliability diagrams, proper scoring rules, the Brier score, and logarithmic score are useful tools. Sharpness is also important: intervals that are extremely wide may be calibrated but practically uninformative.

Confidence intervals and credible intervals have different interpretations. A frequentist confidence procedure concerns the long-run behavior of an estimation method, whereas a Bayesian credible interval assigns posterior probability to a parameter conditional on the model and prior. Neither interpretation should be casually transferred to the other.

Uncertainty can be propagated through a pipeline. Measurement uncertainty affects estimated parameters; parameter uncertainty affects predictions; predictive variability affects decisions. Monte Carlo propagation, Bayesian posterior sampling, bootstrap procedures, and interval arithmetic are possible approaches. In deployed systems, monitoring should detect changes in data distributions, calibration, missingness patterns, and error rates.

Out-of-distribution detection is closely related to epistemic uncertainty. A model may be confidently wrong when unfamiliar inputs resemble familiar examples in superficial ways. Ensembles, density models, representation-based detectors, conformal prediction, and abstention policies can help, but no method guarantees reliable uncertainty under arbitrary distribution shift.

Conformal prediction provides prediction sets or intervals with finite-sample coverage under exchangeability assumptions. Its appeal lies in relatively weak distributional assumptions; its limitation is that coverage can fail when the deployment environment differs substantially from the calibration environment. Group-conditional coverage may also involve trade-offs with efficiency and sample size.

Decisions Under Incomplete Information

Inference becomes operationally meaningful when linked to decisions. Let a be an action, θ a state of the world, and U(a,θ) a utility. Under a posterior distribution P(θ|D), the Bayes action is:

\[

a^*=\arg\max_a E[U(a,\theta)\mid D].

\]

This formulation shows why the most probable outcome is not always the best basis for action. A low-probability event with severe consequences may justify precaution, while a highly probable event may not warrant intervention if the benefit is small.

Decision thresholds should reflect utilities, costs, constraints, and fairness requirements rather than arbitrary probability cutoffs. A medical triage system, for example, may accept more false positives when missing a dangerous condition is costly. Thresholds can also vary across operational contexts, but changes should be documented and evaluated for disparate impact.

The value of information measures whether obtaining additional evidence is worth its cost. Expected value of perfect information compares decisions with complete knowledge against decisions under current uncertainty. Expected value of sample information evaluates a particular proposed test or data collection activity. These concepts support active learning, experiment design, sensor placement, and resource allocation.

Robust decision-making is appropriate when the probability model itself is disputed. Distributionally robust optimization considers a set of plausible distributions and seeks an action that performs acceptably across them. Robust Bayesian analysis examines how conclusions change across reasonable priors or likelihood specifications. Such methods are particularly important in high-stakes environments where model misspecification is consequential.

Applications Across Domains

In healthcare, probabilistic models support diagnosis, risk stratification, treatment-effect estimation, and clinical decision support. Causal methods help distinguish prognostic variables from treatment-effect modifiers. However, clinical deployment requires temporal validation, calibration across hospitals, privacy protection, clinician oversight, and careful analysis of missing-not-at-random data.

In finance, uncertainty-aware models are used for credit risk, fraud detection, portfolio optimization, and stress testing. Financial data are nonstationary and vulnerable to feedback effects: predictions influence behavior, which changes future data. Causal analysis can help assess the effect of lending policies, but confounding and selective observation remain serious concerns.

In robotics and autonomous systems, Bayesian filtering combines noisy sensor observations with dynamic models. A robot must maintain beliefs about location, objects, and intentions while choosing actions under uncertainty. Particle filters, Kalman filters, Gaussian processes, and partially observable Markov decision processes are relevant tools.

In climate and environmental science, probabilistic ensembles express uncertainty in forecasts and parameters. Causal analysis can evaluate interventions such as emissions policies, while hierarchical models combine observations from multiple locations and scales. Communication of uncertainty is as important as numerical accuracy because decisions affect communities with unequal exposure and resources.

In cybersecurity, probabilistic graphical models can combine weak signals from logs, endpoints, and network activity. Causal reasoning can support analysis of attack pathways and intervention points. Yet adversarial adaptation violates assumptions of stable data generation, requiring continuous monitoring and robust design.

Illustrative Case Study: A Hospital Readmission Decision System

Background

Consider a fictional regional hospital developing a system to identify patients at high risk of readmission within thirty days. The proposed system will support discharge planning, not automatically deny discharge or allocate treatment.

Situation and Problem

An initial predictive model uses age, previous admissions, diagnosis codes, medication history, and neighborhood-level variables. It achieves strong discrimination on historical data. However, its probability estimates are poorly calibrated across demographic groups, and administrators propose using the score to assign home-care resources.

A further complication is that previous readmission is influenced by access to primary care. Access is also associated with income, transportation, and the likelihood of receiving follow-up services. Thus, historical utilization reflects both health needs and unequal opportunities to obtain care.

Analysis

The team first defines the estimand. Prediction of readmission is not the same as estimating the causal effect of offering a follow-up intervention. A patient may have high predicted risk but derive little benefit from the proposed service, while another patient with moderate risk may benefit substantially.

A causal graph is constructed with health status, access to care, prior utilization, discharge support, intervention, and readmission. Prior utilization is treated cautiously because it may be a proxy for both illness and access. The team examines whether variables are confounders, mediators, proxies, or consequences of earlier institutional decisions.

The model is recalibrated using temporally separated data. Reliability diagrams and Brier scores are reported by clinically relevant subgroups. Missingness is analyzed rather than silently imputed, because missing follow-up records may reflect lack of access rather than absence of need. The organization also evaluates distribution shift across hospitals.

Approach and Solution

The hospital separates three tasks: risk prediction, estimation of intervention benefit, and resource allocation. A calibrated probabilistic model estimates baseline risk. A causal or quasi-experimental analysis estimates heterogeneous benefit from follow-up care where assumptions are defensible. A decision layer then combines predicted benefit, cost, capacity, and ethical constraints.

The system provides an uncertainty interval and an abstention flag when a patient differs substantially from the training population. Clinicians receive the principal contributing factors, data-quality warnings, and a clear statement that the score is decision support rather than a causal explanation. Allocation rules are reviewed by a clinical governance committee and audited for disparate error rates.

Outcome

The hospital does not adopt a single high-risk threshold. Instead, it prioritizes patients using estimated expected benefit and available capacity, with manual review for uncertain cases. Monitoring shows improved calibration and identifies a subgroup for which the model is unreliable because outpatient records are incomplete. Additional data collection and a prospective evaluation are commissioned before wider deployment.

Learning Points

The case illustrates that prediction, causation, and decision-making are distinct. It also demonstrates why calibration, missing-data analysis, uncertainty communication, subgroup evaluation, and governance must accompany technical performance. A model can be accurate on average yet unsuitable for a specific intervention or population.

Challenges, Risks, and Common Errors

A frequent mistake is treating correlation as causation. Predictive importance, attention weights, or feature attribution do not by themselves establish an intervention effect. Another error is conditioning on post-treatment variables when estimating a total effect, or conditioning on colliders because they appear predictive.

Data leakage can produce deceptively strong performance when information unavailable at decision time is included during training. Selection bias occurs when the observed sample is generated by a process related to both exposure and outcome. Label bias arises when recorded outcomes reflect institutional decisions rather than the underlying construct of interest.

Probabilistic systems may be miscalibrated because of class imbalance, dataset shift, overfitting, or inappropriate likelihood assumptions. Bayesian methods are not automatically objective: priors encode assumptions, and posterior conclusions can be sensitive to them. Approximate inference may introduce additional bias or false certainty.

Causal discovery can fail when important variables are unmeasured, feedback is present, time ordering is unclear, or conditional independence tests have low power. Counterfactual claims are especially vulnerable to unverifiable assumptions. Researchers should state assumptions explicitly and perform sensitivity analyses rather than presenting causal conclusions as purely data-derived.

Ethics, Governance, and Professional Practice

Uncertainty must be communicated in a way that supports informed human judgment. Suppressing uncertainty encourages automation bias; presenting technical distributions without interpretation may create confusion. High-stakes systems need documented intended use, known limitations, escalation procedures, audit trails, and mechanisms for contesting decisions.

Fairness evaluation should consider not only group-level predictive metrics but also unequal access to interventions, differences in error consequences, and the possibility that protected attributes are proxies for structural disadvantage. Removing a sensitive variable does not necessarily remove its influence.

Privacy is relevant to both probabilistic and causal modeling. Linkage of records can increase identifiability, while causal analysis may encourage collection of sensitive variables. Data minimization, secure computation, access controls, retention limits, and institutional review are part of responsible practice.

Professional governance should distinguish model risk from ordinary software defects. Model cards, data sheets, causal diagrams, uncertainty reports, version control, reproducible pipelines, and post-deployment monitoring support accountability. Human oversight should be substantive rather than merely ceremonial.

Current Developments and Best Practices

Recent research increasingly combines deep representation learning with probabilistic inference, causal structure, and foundation models. Probabilistic programming systems make complex generative models easier to specify and infer. Causal representation learning seeks latent features that remain stable under interventions or environmental changes, although identifiability remains an active research problem.

Diffusion models and score-based methods provide powerful generative modeling tools, but their uncertainty is not automatically calibrated for decision support. Bayesian deep learning, ensemble methods, conformal prediction, and distribution-shift detection are being integrated into safety-oriented pipelines. Digital twins and scientific machine learning use mechanistic models together with data-driven components, raising questions about uncertainty propagation and causal validity.

For doctoral research and professional deployment, the following practices are particularly important:

• Define the estimand, decision, population, and time horizon before selecting a model.
• Draw a causal diagram when intervention or policy claims are involved.
• Separate training, calibration, temporal validation, and prospective evaluation.
• Report uncertainty, calibration, subgroup performance, and sensitivity to assumptions.
• Distinguish aleatoric from epistemic uncertainty and identify what new data could reduce uncertainty.
• Treat approximate inference as an engineering and statistical approximation requiring diagnostics.
• Use abstention, human review, or conservative policies when uncertainty is consequential.
• Monitor deployed systems for drift, feedback, changing base rates, and unequal impact.
• Document assumptions, provenance, limitations, and governance responsibilities.

Chapter Summary

Probabilistic AI represents incomplete knowledge through distributions, conditional dependence, and Bayesian updating. Graphical models make complex distributions more interpretable and computationally manageable, while exact and approximate inference provide mechanisms for reasoning from evidence. Gaussian processes offer a principled function-space approach to prediction with uncertainty.

Causal AI goes beyond association by modeling data-generating mechanisms, interventions, and counterfactuals. Structural causal models and directed graphs clarify confounding, mediation, colliders, adjustment, and identifiability. Causal discovery can provide useful hypotheses but remains dependent on assumptions, data quality, and experimental validation.

Uncertainty quantification requires attention to calibration, distribution shift, predictive intervals, and the distinction between aleatoric and epistemic uncertainty. Decision theory connects beliefs to actions through utilities, costs, value of information, and robustness. Across healthcare, finance, robotics, environmental science, and cybersecurity, reliable AI depends not only on predictive performance but also on transparent assumptions, causal validity, ethical governance, and continuous monitoring.

Key Terms

Bayesian network: A directed acyclic graph whose joint distribution factorizes according to conditional distributions given parent variables.

Bayes’ theorem: A rule for updating the probability of a hypothesis using prior belief and observed evidence.

Conditional independence: Independence between variables after conditioning on one or more other variables.

D-separation: A graphical criterion for determining conditional independencies in a directed acyclic graph.

Markov random field: An undirected graphical model that represents dependencies through clique potentials.

Latent variable: An unobserved variable introduced to explain patterns among observed variables.

MCMC: A family of sampling methods that approximate a target distribution using a Markov chain.

Variational inference: An approximation method that optimizes a tractable distribution to approximate an intractable posterior.

Gaussian process: A probability distribution over functions defined by a mean function and covariance kernel.

Aleatoric uncertainty: Uncertainty arising from inherent variability or randomness in the process being modeled.

Epistemic uncertainty: Uncertainty caused by limited knowledge, data, or model inadequacy.

Structural causal model: A model using structural equations to represent how variables are generated and how interventions operate.

Intervention: An external operation represented by do(X=x) that sets a variable independently of its usual causes.

Confounder: A variable that causally influences both an exposure and an outcome.

Mediator: A variable that transmits part of the causal effect from an exposure to an outcome.

Collider: A variable caused by two other variables; conditioning on it can create spurious association.

Counterfactual: An outcome that would have occurred under an alternative condition or intervention.

Calibration: Agreement between predicted probabilities and observed event frequencies.

Expected utility: The probability-weighted utility of an action across possible states of the world.

Value of information: The improvement in expected decision quality obtained by acquiring additional information.

Review Questions

Short Answer Questions

1. Distinguish a prior, likelihood, and posterior in Bayesian inference.

2. What does conditional independence contribute to the construction of a Bayesian network?

3. Explain the difference between aleatoric and epistemic uncertainty.

4. Why is conditioning on a collider potentially harmful in causal analysis?

5. What is the purpose of calibration in probabilistic prediction?

Long Answer Questions

1. Explain Bayesian networks, d-separation, and factorization, using a suitable example.

2. Compare MCMC, variational inference, and particle filtering in terms of assumptions, computational behavior, and diagnostic requirements.

3. Explain how structural causal models distinguish observational, interventional, and counterfactual quantities.

4. Discuss Gaussian processes, including kernels, posterior uncertainty, computational limitations, and practical applications.

5. Analyze the role of uncertainty quantification and decision theory in the deployment of high-stakes AI systems.

Analytical / Application Questions

1. A predictive model identifies customers likely to default, but the proposed intervention is a counseling program. Design an analysis that distinguishes default-risk prediction from estimation of counseling benefit. Identify likely confounders and possible selection biases.

2. Construct a causal graph for an educational intervention in which prior achievement affects both participation and final performance, while teacher recommendation is influenced by both achievement and classroom behavior. Explain which variables should and should not be adjusted for when estimating the total intervention effect.

3. A model is well calibrated in the development hospital but overconfident in a rural clinic with missing outpatient records. Propose a monitoring, recalibration, uncertainty, and governance strategy before deployment at scale.

22. Multiple Choice Questions with Answers

Question 1

A medical test has high sensitivity, but the disease is rare in the tested population. Which quantity is most directly affected by the low disease prevalence?

A. The prior probability of disease

B. The sensitivity conditional on disease

C. The definition of a collider

D. The graph’s treewidth

Correct Answer: A. The prior probability of disease

Question 2

In a Bayesian network containing A → B → C, which conditional independence is generally implied?

A. A and B are independent without conditioning

B. A and C are independent given B

C. B and C are independent given A

D. A, B, and C are mutually independent

Correct Answer: B. A and C are independent given B

Question 3

Why can conditioning on a collider create a spurious association?

A. It removes all causal paths between the variables

B. It changes a previously closed path into an open path

C. It guarantees that the variables become independent

D. It replaces observational probability with a counterfactual probability

Correct Answer: B. It changes a previously closed path into an open path

Question 4

Which statement best characterizes variational inference?

A. It always obtains exact posterior distributions

B. It approximates a posterior using an optimized tractable family

C. It can only be applied to discrete Bayesian networks

D. It requires a randomized experiment for every parameter

Correct Answer: B. It approximates a posterior using an optimized tractable family

Question 5

What is a principal advantage of a Gaussian process over a standard point-prediction model?

A. It removes the need to specify any assumptions about similarity

B. It provides a distribution over functions and predictive uncertainty

C. It guarantees accurate predictions outside the training distribution

D. It has constant computational cost regardless of sample size

Correct Answer: B. It provides a distribution over functions and predictive uncertainty

Question 6

Which expression represents an intervention rather than an observational conditional probability?

A. P(X)

B. P(Y | X=x)

C. P(Y | do(X=x))

D. P(X,Y)

Correct Answer: C. P(Y | do(X=x))

Question 7

A variable is a common cause of both treatment and outcome. In causal analysis, it is most appropriately described as a:

A. Mediator

B. Collider

C. Confounder

D. Posterior

Correct Answer: C. Confounder

Question 8

A model assigns probability 0.8 to a group of events. If approximately 80 percent of those events occur, the model demonstrates good:

A. Calibration

B. Treewidth

C. Identifiability

D. Exchangeability violation

Correct Answer: A. Calibration

Question 9

Which decision rule selects an action with the greatest probability-weighted utility?

A. Choose the most frequent feature value

B. Choose the action maximizing expected utility

C. Choose the action with the largest posterior variance

D. Choose the action with the highest training accuracy

Correct Answer: B. Choose the action maximizing expected utility

Question 10

In the hospital case, why was risk prediction separated from intervention-benefit estimation?

A. Risk prediction is always causal, whereas benefit estimation is not

B. A patient at high baseline risk may not receive the greatest benefit from the intervention

C. Intervention benefit cannot be evaluated using any data

D. Calibration is irrelevant to treatment decisions

Correct Answer: B. A patient at high baseline risk may not receive the greatest benefit from the intervention