How to Read: The central role of the propensity score in observational studies for causal effects [Reading Guide]

This guide will help you navigate "The central role of the propensity score in observational studies for causal effects" by Rosenbaum and Rubin (1983), a seminal paper published in Biometrika. Given its age, this paper is foundational and highly theoretical, introducing a concept that has become indispensable in causal inference.

First pass (5–10 min)

The goal of this initial pass is to grasp the paper's core contribution and context.

• What to skim:

  • Title: "The central role of the propensity score in observational studies for causal effects." This immediately tells you the paper's focus: addressing causal inference in non-randomized settings using propensity scores. The word "central" hints at its foundational importance.
  • Introduction (Section 1, pp. 19-20): Read this section carefully. It lays out the problem—bias in observational studies compared to randomized experiments—and introduces the concept of a "balancing score" and, specifically, the "propensity score" as a solution. Look for the definition of "strong ignorability" as a key assumption.
  • Discussion (Section 9, pp. 28-29): Read the concluding remarks. This section summarizes the paper's contributions, discusses practical implications, and hints at future research directions.
  • Figures and Tables: Quickly glance at the figures and tables, especially Figure 1 (p. 20) and Figure 2 (p. 22), to get a visual sense of what a propensity score distribution looks like. Notice the presence of mathematical notation and proofs, indicating a rigorous statistical treatment.

• What to notice:

  • Contributions: The paper introduces and formalizes the propensity score as a method to balance covariates between treated and control groups in observational studies, thereby reducing bias when estimating causal effects.
  • Problem: How to draw valid causal inferences from observational data, where treatment assignment is not random, leading to confounding.
  • Setting: Observational studies, where researchers cannot control treatment assignment. The paper aims to make these studies mimic randomized experiments as much as possible.
  • Key Concept: The propensity score, defined as the conditional probability of receiving treatment given observed covariates.

Second pass (up to 1 hour)

Using SpatialRead in the Second Pass

• Use the Screenshot tool to capture figures, tables, diagrams, and formulae from the PDF and use AI for simple explanations.
• Ask: "Explain Figure X at a high level", "Compare Table 2 vs 3", or "What does the y-axis represent here?"
• Download SpatialRead →

Now, delve deeper into the paper's theoretical underpinnings and practical applications.

• What sections/figures to focus on:

  • Balancing Scores (Section 2, pp. 20-21): Understand the general definition of a "balancing score" and its properties. This sets the stage for the propensity score.
  • The Propensity Score (Section 3, pp. 21-22): This is the heart of the paper. Focus on the formal definition of the propensity score and, crucially, Theorem 1 and Theorem 2. These theorems establish the propensity score's fundamental property: conditioning on the propensity score balances all observed covariates.
  • Estimation of the Propensity Score (Section 4, pp. 22): This section provides practical guidance on how to estimate the propensity score, mentioning logistic regression as a primary method.
  • Methods for Using the Propensity Score (Sections 5, 6, 7, pp. 22-25):
    • Section 5: Matching on the Propensity Score: Understand how matching treated and control units on their estimated propensity scores works.
    • Section 6: Stratification on the Propensity Score: Learn about dividing units into strata based on propensity scores.
    • Section 7: Covariance Adjustment on the Propensity Score: See how the propensity score can be used in regression models.
  • Examples (Section 8, pp. 25-28): Review the two examples to see the methods in action.
    • Lung Cancer Example: Examine Table 1 (p. 22) for initial covariate imbalance. Then, compare it with Table 2 (p. 23) after matching and Table 3 (p. 24) after stratification to observe how the propensity score improves covariate balance. Table 4 (p. 25) shows the resulting estimated treatment effects.
    • Right Heart Catheterization Example: Similarly, review Table 5 (p. 26) for initial imbalance and Table 6 (p. 27) for improved balance after matching. Table 7 (p. 28) presents the estimated effects.

• Key terms/questions to note:

  • What is a balancing score?
  • What is the formal definition of the propensity score, $e(X)$?
  • What does "strong ignorability" mean, and why is it crucial?
  • How does the propensity score achieve covariate balance? (Refer to Theorem 1 and 2).
  • What are the different ways to use the propensity score (matching, stratification, adjustment)?
  • How do the examples demonstrate the utility of the propensity score in practice?

Third pass (1–5 hours)

Using SpatialRead in the Third Pass

• Highlight complex keywords, dense paragraphs, equations, or proofs and ask for a plain-English explanation that preserves notation.
• Use AI to explain the concepts in a way that is easy to understand.
• Ask for alternative phrasings or analogies to confirm understanding without changing technical meaning.
• Connect concepts with each other using AI to build a comprehensive understanding of the paper.
• Download SpatialRead →

This pass involves a deep, critical reading to fully understand the paper's nuances, assumptions, and implications.

• Deep reading guidance:

  • Mathematical Proofs: Go through the proofs in Sections 2 and 3 carefully. Understand the logical steps that lead to Theorems 1 and 2. This is essential for a complete grasp of why the propensity score works.
  • Assumptions: Critically evaluate the "strong ignorability" assumption. Understand its components: unconfoundedness (no unmeasured confounders) and common support (overlap in propensity scores between treated and control groups). Recognize that this assumption is untestable with observed data and is a major limitation of all observational causal inference.
  • Methods Comparison: Compare and contrast the different methods of using the propensity score (matching, stratification, covariance adjustment). Consider their strengths, weaknesses, and when one might be preferred over another.
  • Limitations and Extensions (Section 9): Re-read the Discussion section. What limitations do the authors acknowledge? What future research avenues do they suggest? This will help you contextualize the paper within the broader field.
  • Practical Implementation Details: While the paper is theoretical, pay attention to the practical advice in Section 4 (estimation) and the descriptions of matching/stratification. How would you implement these steps given a dataset?

• Replication pointers if applicable:

  • While this paper doesn't present a full empirical study for direct replication, the examples in Section 8 provide sufficient detail to understand the process.
  • To "replicate" the ideas, you would:
    1. Identify a dataset: Find an observational dataset with a binary treatment and multiple covariates.
    2. Estimate the propensity score: Use a method like logistic regression, as suggested in Section 4, to predict treatment assignment based on covariates.
    3. Check balance: Before and after using the propensity score (e.g., via matching or stratification), compare covariate means or distributions between treated and control groups, similar to how Tables 1-3 and 5-6 are structured.
    4. Estimate treatment effect: Apply one of the methods (matching, stratification, or adjustment) to estimate the causal effect, mirroring the results in Tables 4 and 7.
  • Focus on understanding how the propensity score transforms the data to achieve balance, which is the core contribution demonstrated in the examples.