Framework for drawing causal inference from observational studies:

- Treatment groups
- Patient-level outcome
- Summary measure of outcome
- Difference in outcome among groups
- Attribution to group membership
- Factors to control
- Factors we should not control

]]>

We offer two new estimates: projected mortality and attributable deaths.

How did we argue our case that early surgery improves survival after hip fracture?

- We state we are concerned with a policy question, not an associational question. We ask what mortality should be expected if timing policy for hip fracture surgery would change. An associational question is whether mortality varies in patients with various surgical times.
- We acknowledge that answering the policy question with observational data involves counterfactuals, such as expected mortality should all patients had early surgery.
- We stratify observations on factors that block biasing influences from known factors. We use the back-door criterion to find stratification factors in a evidence-informed causal diagram.
- We observe stratification factors have constant values within each stratum. We reason then difference in mortality by timing is unbiased in each strata, because conditioning on statification factors prevents other factors from producing covariation in mortality and timing.
- We average stratum-specific risk differences between timing alternatives to find difference in mortality attributable to a policy change.
- We conclude that if all hip fracture surgeries in Canada were within 2 days, then 8 more patients would survive to hospital discharge, for every 1000 surgeries.

Thank-you!

]]>

In public health, policy makers are often interested in risk reduction that would be achieved if exposure were reduced to an alternative level. Similarly, policymakers involved in health care are really interested to know whether patients would benefit from an alternative treatment, which is more suitable to clinical profiles of patients.

One immediate question in comparing alternative strategies for coronary revascularization (PCI v. CABG) is What would be the impact on the system if some PCI patients would be treated with CABG? For example, what proportion of repeat revascularizations could be avoided had patients suitable for CABG had it instead of PCI?

Another policy question is What would be the impact on the patient population if all CABG patient were undergoing PCI? For example, what proportion of stroke could be avoided had patients suitable for PCI undergone PCI instead of CABG?

In addition to attributable proportion, known to epidemiologists, Pearl offered other quantities that help answer the variety of policy questions arising from CER.

Probability of disablement refers to hypothesis that the absence of exposure would prevent outcome, when we observe that in fact outcome did occur—proportion of observed cases in the total patient population that would not have occurred had the population been entirely unexposed.

Probability of necessity refers to expectations that the absence of exposure would prevent outcome, when we observe that both exposure and outcome did occur.

If interested in the opposite situation we turn to the other two probabilities. Probability of enablement refers to hypothesis that the presence of exposure would produce outcome, given that in fact outcome did not occur.

Probability of susceptibility refers to expectations that the presence exposure would produce outcome, when we observe that neither exposure nor outcome occurred.

If we go back to coronary revascularization, there are two interesting examples of Pearl’s approach. If interested in attribution, we can estimate the proportion of PCI patients who would have survived had they undergone CABG. That is, we consider patients who underwent PCI and died, and ask how many of deaths would be avoided if they undergone CABG instead.

The second example focuses on patients who underwent CABG. If interested in susceptibility, we can estimate the proportion of CABG patients who would have died had they undergone PCI. That is, we consider patients who were suitable and underwent CABG, and ask what would be risk of death if they undergone PCI instead.

These are just two examples that help understand the types of questions that could be asked. Pearl showed that answering causal questions is not limited to randomized trials. Rather it is a function of our ability to incorporate the existing knowledge of causal pathways into identifying the set of factors for adjustment.

]]>Let be a random variable representing a binary outcome after a binary treatment is forced to value ; and let and denote events and , and their complements.

Probability that the absence of event would prevent event , given that in fact and did occur. — attribution

Probability that the presence of event would produce event , given that in fact and did not occur. — susceptibility

Probability that the absence of event would prevent event and the presence of event would produce event . — necessary and sufficient cause

Probability that the absence of event would prevent event , given that in fact did occur. — proportion of observed cases in the total population that would not have occurred had the population been entirely unexposed (disablement)

Probability that the presence of event would produce event , given that in fact did not occur. — enablement

Probability that the presence of event would produce event , given that in fact did not occur. — Effect of treatment on the treated

]]>

Policy makers are often interested in the extent of reduction in incidence that would be achieved had the population been entirely unexposed, compared with its existing risk pattern. Note that this reasoning compares the observed risk with the counterfactual risk . In this paper, the article proposes to distinguish attributable “caseload” from attributable proportion on the basis of a simple rule that proportions have the numerator included in the denominator.

The first measure, *attributable caseload*, can be interpreted as a reduction/increment in observed cases had the population been entirely unexposed.

Notably, the numerator is not included in the denominator, and this measure ranges from to 1. In most cases, the attributable caseload could be one of the most useful measures in public health should an intervention (e.g., vehicle emission control) be implemented to make everyone in the population unexposed.

The second measure, *attributable proportion*, is a proportion of observed cases in the total population that would not have occurred had the population been entirely unexposed

The last probability is exactly what Pearl calls the probability of disablement.

The proportion of cases among those happened to be exposed that would not have occurred had the exposed been entirely unexposed is defined as follows:

The last probability is exactly what Pearl calls the probability of necessity. Under the assumption of monotonicity it could be presented as follows

The first term is equal to an excess risk relative to the exposed risk.

]]>The premise is that variations in health care delivery, clinical practice patterns, socio-economic and demographic factors, and human behavior modify effectiveness across large segments of the populations under investigation. Another aspect relates to the use of observational databases, such as Medicare/Medicaid data, Surveillance, Epidemiology, End Results data, and prospective clinical registries.

]]>The User’s Guide identifies key considerations and best practices for designing observational studies and standardizes the review of study protocols with checklists that are provided in each chapter. Topics in this new User’s Guide include developing study objectives and questions, study design, data sources, and analysis. The User’s Guide has been drafted with a goal of strengthening the overall quality of observational CER research for informing important health care decisions by patients and other stakeholders.

One of the chapters gives a nice intro and overview of causal diagrams, as well as some aspects of the counterfactual theory of causal effects.

]]>