Two Ways to Think About Confounding

There are two ways to approach confounding assessment. Understanding both reveals why a systematic assessment matters. (This tool is for educational purposes.)

The Adjustment Approach

1
List potential confounders

Identify variables that might affect both exposure and outcome based on subject-matter knowledge.

2
Measure and adjust

Include confounders in regression models or use stratification to control for their effects.

3
Check model fit

Evaluate whether the model fits the data well and whether coefficients change when adding covariates.

Beyond Adjustment

+
Ask: What's still missing?

No matter how many confounders you measure, unmeasured factors may remain. Economists ask what those factors would need to look like to matter.

+
Quantify the threat

How strong would an unmeasured confounder need to be? How prevalent? If the answer is "implausibly strong," findings are more credible.

+
Seek identification

Rather than hoping confounding is small, find sources of variation in treatment that are independent of confounders.

What Is Sensitivity Analysis for Unmeasured Confounding?

Sensitivity analysis asks: "How would my conclusions change if there were an unmeasured confounder?" Rather than assuming your adjustment is complete, you explore scenarios where it isn't.

The goal is not to prove confounding is absent. You cannot prove a negative. The goal is to characterize how large and prevalent an unmeasured confounder would need to be to explain away your findings.

  • If a tiny unmeasured confounder could flip your conclusion, findings are fragile
  • If only an implausibly large confounder would matter, findings are more robust
  • This doesn't make observational findings causal, but it calibrates confidence

The key question economists add:

What would unmeasured confounding need to look like to matter? The checklist in the next panel provides a structured way to answer this.

The Confounding Assessment Checklist

This six-part checklist structures your assessment of confounding threats. Work through each section systematically. The goal: characterize both measured and unmeasured confounding, then evaluate what unmeasured confounding would need to look like to change your conclusions.

1 Identify the Treatment and Outcome

a
Define the treatment clearly

What exactly is the exposure or intervention? Be specific about timing, intensity, and measurement.

b
Define the outcome precisely

What is being measured as the result? Over what time horizon? How is it coded?

c
State the claimed effect

What is the magnitude of the association? Is it a relative risk, odds ratio, regression coefficient?

2 Map the Selection Process

a
How did units receive treatment?

Was it randomized? Self-selected? Assigned by policy? Provider decision? Understanding the mechanism is crucial.

b
What factors influenced treatment receipt?

List every factor you can think of that affected who got treatment and who didn't.

c
Which of these factors also affect outcomes?

Any factor that affects both treatment and outcome is a potential confounder.

3 Inventory Measured Confounders

a
List all confounders you can measure

These are variables in your dataset that affect both treatment and outcome.

b
Assess measurement quality

Are these confounders measured well? Error in confounders leads to residual confounding even after adjustment.

c
Check for colliders

Are you controlling for variables caused by the treatment? This can introduce bias rather than remove it.

4 Identify Unmeasured Confounders

a
What factors can't you measure?

List variables that likely affect both treatment and outcome but aren't in your data.

b
How strong is each unmeasured confounder?

Estimate the association between each unmeasured confounder and the outcome.

c
How imbalanced is each across treatment groups?

Estimate how different the prevalence of each confounder is between treated and untreated.

5 Conduct Sensitivity Analysis

a
Apply a bounding formula

Use established methods (like the E-value or Rosenbaum bounds) to quantify required confounder strength.

b
Compare to measured confounders

Is the required unmeasured confounder stronger than any you've already controlled for?

c
Evaluate plausibility

Would such a strong, prevalent, unmeasured confounder be plausible given what you know about the domain?

6 Consider Alternative Designs

a
Is there exogenous variation?

Can you find a policy change, natural experiment, or instrument that assigns treatment independently of confounders?

b
Would a different comparison group help?

Are there "almost-treated" groups that share confounders but didn't receive treatment?

c
What would convince a skeptic?

Imagine someone who doesn't believe your finding. What evidence would change their mind?

Ready to apply the checklist?

The next panel provides an interactive assessment tool using a realistic scenario.

Interactive Assessment

Apply the checklist to a scenario. Adjust the sliders and selections to see how your assessment of confounding threat changes. This tool demonstrates the logic; real sensitivity analysis requires more formal calculations.

Scenario: Diabetes Prevention Program

A county-level study finds that communities with diabetes prevention programs have 15% lower hospitalization rates. The study adjusts for income, age distribution, and insurance coverage. You need to assess whether unmeasured confounding could explain this finding.

1

How strong is the observed effect?

Relative Risk Reduction 15%

Moderate effect

2

How completely does the study measure selection into treatment?

3

How strong are measured confounders?

Strength of measured confounders Moderate

Income and age each explain 10-20% of outcome variation

4

Can you identify plausible unmeasured confounders?

5

Would unmeasured confounding need to be stronger than what you measured?

Required strength of unmeasured confounder 2x measured

Unmeasured confounder would need to be twice as strong as income to explain the effect

What does this all add up to?

The final panel summarizes the key insight that distinguishes a design-based approach to confounding assessment.

Key Insight

The confounding assessment checklist shifts the question from "did we adjust for confounders?" to "what would confounding need to look like to matter?" This reframing is central to how economists evaluate observational evidence.

Adjustment is necessary but not sufficient

Controlling for measured confounders is important, but it does not eliminate the threat of unmeasured confounding. The question is whether the remaining threat is large enough to change conclusions.

Quantification creates accountability

Vague statements like "residual confounding may exist" are unhelpful. Specifying that confounding would need to have RR=2.5 to explain the finding creates a concrete standard others can evaluate.

Comparison to measured confounders is informative

If the required unmeasured confounder is stronger than anything you measured, it may be implausible. If it is weaker, concern is warranted.

Design matters more than adjustment

Finding exogenous variation (policy changes, natural experiments, discontinuities) is more powerful than ever-more-sophisticated adjustment. This is the economist's preferred solution.

Concepts Demonstrated in This Lab

Sensitivity analysis: Exploring how conclusions change under different assumptions about unmeasured confounding
E-value: The minimum strength a confounder would need to fully explain away an observed effect
Selection mechanism: The process that determines who receives treatment and who doesn't
Identification: Finding sources of treatment variation independent of confounders

Key Takeaway

No amount of statistical adjustment can fix a flawed comparison. The economist's contribution to confounding assessment is not just listing more confounders, but systematically characterizing what unmeasured confounding would need to look like to explain the findings. When the required confounder is implausibly strong, findings are more credible. When it is plausible, the solution is not more adjustment but better research designs that provide exogenous variation. This is what economists mean by "identification."