Modern Research & Data Science

Philosophy, Paradigms, and AI-Augmented Reproducibility

Workshop Introduction

9–10 September 2026

Workshop Overview

Modern research lies at the intersection of classical philosophy, rigorous statistical frameworks, and cutting-edge artificial intelligence. This workshop bridges these domains to build quality-conscious, reproducible workflows.

Part I: Philosophy of Knowledge
From ontology and epistemology to logical inference (induction, deduction, reduction).

Part II: Data Analysis Paradigms
Frequentist, Bayesian, and Artificial Intelligence paradigms.

Part III: Domain Focus
Methodological applications in Clinical/Pharmaceutical Research and Econometrics.

Part IV: The Modern AI-Augmented Stack
Literate programming in R & Quarto, accelerated by Generative AI.

Part I — Research Philosophy

From Ontology to Epistemology, and the Structures of Inquiry

Ontology & Epistemology

Philosophical foundations

Every study design choice is a downstream consequence of a philosophical position.

Ontology: What is reality?

  • Realism / Objectivism: There is a single, objective reality independent of our perception.
  • Relativism / Subjectivism: Reality is constructed socially, meaning multiple truths exist.

Epistemology: How do we know it?

  • Empiricism: Knowledge is derived from sensory, observational evidence.
  • Rationalism: Knowledge is derived from logical reasoning and intellect.

In clinical and quantitative econometrics, we search for objective, reproducible truths (realist ontology, empirical epistemology).

Positivism vs. Constructivism

Research Traditions

How do these ontological and epistemological choices shape our research traditions?

Feature Positivist Tradition Constructivist Tradition
Ontological Stance Single, objective, discoverable reality. Multiple, constructed, contextual realities.
Goal of Inquiry Formulate general laws, establish causal links. Understand subjective meanings and experiences.
Role of Researcher Detached, objective observer. Engaged, co-creator of meaning.
Typical Data Quantitative, structured measurements. Qualitative, unstructured interviews or text.

Logical Inference: Induction to Deduction

Generating and Testing Knowledge

Science moves in cycles between building theories and testing them.

Induction: Theory-Building

  • Starts with observations.
  • Detects patterns and regularities.
  • Formulates tentative hypotheses.
  • Generates general theories.

\[\text{Observations} \rightarrow \text{Pattern} \rightarrow \text{Theory}\]

Deduction: Theory-Testing

  • Starts with a theory.
  • Derives testable hypotheses.
  • Collects observations.
  • Confirms or rejects the hypothesis.

\[\text{Theory} \rightarrow \text{Hypothesis} \rightarrow \text{Validation}\]

Reductionism & Systems Thinking

How We Analyze Systems

Scientific inquiry has historically relied on breaking systems down, but modern science requires integration.

Reductionism

  • The belief that complex phenomena can be entirely explained by analyzing their simplest physical components.
  • Example: Explaining human disease solely at the level of cellular biology or molecular genetics.
R
“Divide each difficulty into as many parts as is feasible.” — René Descartes

Systems Thinking

  • The belief that the whole is greater than the sum of its parts.
  • Example: Recognizing that health outcomes are shaped by a complex web of biological, social, and economic factors.
  • Essential for understanding emergent properties in pharmacology and econometrics.

Part II — Three Paradigms of Data Analysis

Frequentist, Bayesian, and Artificial Intelligence

1. The Frequentist Paradigm

Paradigms of Data Analysis

The historical bedrock of clinical trials and classical econometrics.

  • Definition of Probability: The limit of relative frequencies in an infinite number of identical, hypothetical trials.
  • Core Characteristic: Parameters of interest (\(\theta\)) are considered fixed, unknown constants. Data (\(D\)) are random variables.
  • Key Tools: Null Hypothesis Significance Testing (NHST), \(p\)-values, confidence intervals.

\[\text{Hypothesis: } H_0: \theta = 0 \quad \text{vs.} \quad H_1: \theta \neq 0\]

A \(p\)-value represents \(P(D \mid H_0)\) — the probability of observing data at least as extreme as ours, assuming the null hypothesis is true. It is not the probability that the hypothesis is true.

2. The Bayesian Paradigm

Paradigms of Data Analysis

A coherent framework for updating beliefs as new evidence accumulates.

  • Definition of Probability: A measure of belief or certainty, conditional on current information.
  • Core Characteristic: Parameters (\(\theta\)) are random variables with probability distributions. Data (\(D\)) are fixed, observed constants.
  • Key Tool: Bayes’ Theorem.

\[P(\theta \mid D) = \frac{P(D \mid \theta) P(\theta)}{P(D)}\]

  • \(P(\theta \mid D)\): Posterior (belief after seeing data)
  • \(P(D \mid \theta)\): Likelihood (data explanation)
  • \(P(\theta)\): Prior (belief before seeing data)
  • \(P(D)\): Marginal Likelihood (normalizing constant)

3. The Artificial Intelligence Paradigm

Paradigms of Data Analysis

A shift from explaining parameters to predicting outcomes using high-dimensional data.

  • Core Focus: Generalization and predictive accuracy (\(y \approx \hat{f}(X)\)) rather than isolating specific, low-dimensional parameter estimates (\(\beta\)).
  • Mechanism: Algorithmic pattern recognition. The model architecture learns representations from vast, complex datasets without requiring rigid, pre-specified functional forms.
  • Generative AI: Learning the underlying probability distribution of data (\(P(X)\) or \(P(X \mid Y)\)) to generate new, realistic samples (e.g., text, molecules, synthetic patient cohorts).

The Trade-off: While frequentist and Bayesian models favor interpretability and causality, the AI paradigm favors raw predictive capacity and handling high-dimensional complexity.

Comparing the Three Paradigms

Synthesis

Dimension Frequentist Bayesian Artificial Intelligence
Goal Isolate effects; test hypotheses Update belief distributions Maximize predictive accuracy
Parameter Status Fixed, unknown constant Random variable Implicitly learned representations
Key Output Point estimates, \(p\)-values, CIs Posterior distributions, Credible intervals Predictions, generated text/data
Assumptions Strong distributional assumptions Requires specifying a prior Weak structural assumptions
Best Used For Regulatory clinical trials Sequential decision-making High-dimensional prediction

Part III — Domain-Specific Applications

Clinical, Pharmaceutical, and Econometric Research

Clinical & Pharmaceutical Research

Domain Focus

In healthcare and drug development, reproducibility and safety are paramount.

Methodological Needs

  • Gold Standard: Randomized Controlled Trials (RCTs) to eliminate selection bias.
  • Safety & Efficacy: Survival Analysis (Cox Proportional Hazards), pharmacokinetic (PK/PD) modeling.
  • Regulatory Rigor: Standardized protocols, pre-registration of trials, strict statistical bounds.

Role of Generative AI

  • Synthetic Control Arms: Modeling placebo groups based on historical trial data to reduce trial costs.
  • Drug Discovery: Using generative chemistry (diffusion models) to design novel molecules targeting specific receptors.
  • Drafting Protocols: Generating drafts of regulatory filings, consent forms, and ethics applications.

Econometric Analysis

Domain Focus

Econometrics focuses on identifying causal relationships from observational data where experiments are impossible.

Methodological Needs

  • Causal Inference: Techniques to bypass confounding and establish causality:
    • Difference-in-Differences (DiD)
    • Instrumental Variables (IV)
    • Regression Discontinuity (RDD)
    • Propensity Score Matching
  • Policy Evaluation: Isolating the net impact of macro policies or social interventions.

Role of Generative AI

  • Synthetic Controls: Automating the creation of synthetic control groups using predictive modeling.
  • Unstructured Data Extraction: Parsing millions of policy documents, financial filings, or news transcripts into structured variables for econometric models.
  • Advanced Controls: Integrating machine learning (Double/Debiased Machine Learning) to isolate causal effects in high-dimensional settings.

Part IV — Quality, Reproducibility, & Generative AI

Elevating Research Quality in the Era of AI

The Reproducibility Crisis

Quality-Conscious Research

Too many published findings cannot be replicated by independent researchers.

Common Pitfalls

  • \(p\)-Hacking: Searching through multiple variables or sub-groups until a statistically significant result emerges.
  • HARKing: Hypothesizing After Results are Known.
  • Data Dredging: Running regression after regression without a pre-registered analysis plan.
  • Manual Errors: Copypasting numbers from statistical software into word documents.

The Modern Data Science Solution

  • Literate Programming: Combining code, math, outputs, and text in a single document (Quarto).
  • Version Control: Tracking code history and collaboration with Git.
  • Environment Management: Recording package versions (e.g., renv) to ensure code runs the same way on any machine, today or in ten years.

Quarto: The Reproducible Pipeline

Quality-Conscious Research

Quarto is a multi-language, next-generation version of R Markdown.

  • No Copy-Pasting: If the data changes, simply re-run the file; tables, figures, and inline values update automatically.
  • Interoperability: Write in R, Python, Julia, or SQL inside the same document.

Leveraging Generative AI Responsibly

Generative AI in Research

Generative AI is a powerful assistant, not a replacement for domain expertise.

What GenAI Does Exceptionally Well

  • Accelerating Code: Writing boilerplate R, RStudio functions, or complex ggplot visualisations.
  • Literature Syntheses: Drafting initial summaries of complex papers (always check the source).
  • Document Formatting: Formatting raw text into markdown tables or slide structures.
  • Debugging: Explaining error logs and suggesting corrections.

The Critical Risks

  • Hallucinations: Inventing citations, data points, or R functions that do not exist.
  • Logical Gaps: Generating code that runs but calculates the wrong statistical metric.
  • Compliance Issues: Uploading sensitive, non-anonymized patient data to external servers.

The AI-Augmented Workflow

The Way Forward

Quality-conscious researchers use Generative AI as an iterative partner under strict human control.

  • Rule of Thumb: Never publish a line of code or a citation you have not personally verified.
  • Goal: Use AI to handle the cognitive friction of syntax, freeing your mind to focus on study design, methodology, and interpretation.