health care disparities: impossible but essential Alan Zaslavsky, - - PowerPoint PPT Presentation

Causal inference for health and health care disparities: impossible but essential

Alan Zaslavsky, Ben Lê Cook Harvard Medical School

Outline

• Three faces of “disparities”
• Health disparities and healthcare disparities
• “IOM definition” of health care disparities
• Regression modeling to apply the IOM definition
• Was this causal inference? Causal effect of what?

– “No causality without intervention”?

• Nonetheless …

What is a disparity? (“inequality”, …)

• Difference between groups in treatment or
• utcomes
• Socially/ethically/morally unjust/unacceptable

Three faces of disparities research

• Descriptive/predictive:

– Which groups are affected? – Correlates, mediators of intergroup differences – Partitioning of variation attributable (predictively) to different factors/actors

• Normative:

– Which differences are objectionable?

• Causal:

– What factors would change outcomes if modified? – … and which of these could be modified?

Health disparities & health care disparities

Health disparities

• Effects of exposures and

experiences over entire life course (and before)

• Cumulative
• Past is continuous with

present

• Recognize broad social

responsibility, then specific actors

• Both general and specific

mechanisms Health care disparities

• Effects of interactions with

specific systems in specific episodes

• Incremental (mostly)
• Can define and control for

pre-treatment status

• Identify specific responsible

parties/subsystems, then broader patterns

• Specific causal factors and

mechanisms

“IOM definition” of Healthcare Disparity

Quality of care

Whites Blacks

Difference Clinical Need & Appropriateness & Patient Preferences Healthcare Systems & Legal / Regulatory Systems Discrimination: Bias, Stereotyping, and Uncertainty The difference is due to: Disparity IOM, Unequal Treatment 2002

Operationalization of IOM Definition

• Disparities do not include differences related to health

status (clinical appropriateness and need), and patient preferences

• Disparities do include differences due to SES (differential

impact of healthcare systems and the legal/ regulatory climate), and discrimination.

Problematical aspects of IOM definition

• Patient preferences

– Shaped by past personal and group experiences – “Tuskegee effect” – Legally forced segregation in South until Medicare – Indistinguishable from effects of inadequate communication, etc.

• Discriminatory effect of health care priorities

– Which diseases, conditions get R&D? – Ancillary resources essential to health care – Regional disparities with racial/ethnic effects

“IOM definition” of Healthcare Disparity (modified)

Quality of care

Whites Blacks

Difference Clinical Need & Appropriateness Socioeconomic correlates

• f race

Direct racial/ethnic responses: Bias, Stereotyping, and Uncertainty Disparity IOM, Unequal Treatment 2002

Examples: Implementing the IOM Definition

• Example 1: Difference overestimates disparity

– Hispanics are on average younger and therefore use less medical care. This is not an “unfair” difference.

• Example 2: Difference underestimates disparity

– African-Americans are on average less healthy than Whites but may have very similar rates of utilization. – If Blacks were made to be as healthy as Whites, we would see much less use for Blacks compared to Whites - an “unfair” difference.

Operationalizing the IOM Definition

(1) Fit a model (2) Transform distribution of health status (not SES) (3) Calculate predictions for minorities with transformed health status

• Average predictions by group and estimate

disparities

Oaxaca-Blinder decomposition

• Linear model in groups g=A,B
• Apply to compare groups A, B

– First term is difference predicted by covariate difference – Second term is difference predicted by difference in coefficients (single or interacted group effect)

Nonlinear Oaxaca-Blinder decomposition

• Nonlinear model in groups g=A,B
• Apply to compare groups A, B

– First term is difference predicted by covariate difference – Second term is difference predicted by difference in coefficients (single or interacted group effect)

Apply to disparities calculation

• Objective: adjust for differences in allowable

variables (health status) but not disparity mediators

• Estimate intergroup difference if

(counterfactually):

– Group B had group A distribution of health status – But retained group B distribution of race, SES, etc.

• In nonlinear model, construct a joint

distribution of race, SES, health status with given margins.

Rank and replace

• Separate linear predictor into “allowed” (health

status) and “disallowed” terms

– Aggregate all health status covariates into combined effect – Observations in the A and B samples separately ranked by the linear predictor from the health status variables – Match by their respective rankings. – (similar to “Fairlie method of non-linear decomposition”)

• Replace A health status with matched B health status
• Calculate adjusted comparison

Adjust Need (HS) “Index” (Rank and Replace)

100 White Black

Transform Distribution of Health Status

• 1. Fit a model
• 2. Transform HS distribution
• 3. Calculate predictions

• 10%

0% 10% 20% 30% 40% 50% 60% White Black Hispanic

Access to Mental Health Care Among those in Need (PHQ-2>=3 or K6>=13) – 2004-2013 MEPS

IOM-Concordant Prediction Results: Any mental health care

47.3% 29.4% 32.7%

• 10%

0% 10% 20% 30% 40% 50% 60% 70% 80% White Black Hispanic

HbA1c Check Among those with Diabetes – 2004-2013 MEPS

IOM-Concordant Prediction Results: HbA1c Check in Last Year

66.4% 52.7% 56.8%

What did we just do with regression?

• “Prediction” in regression

– What distribution for Y if: – (1) postulated values/distribution for X – (2) relationships are maintained – Does not require belief in scientific generality of model – Gives substantive interpretation of covariates

• Prediction may be factual or counterfactual
• If counterfactual, may be

– Matched (observed values) – Interpolated (within range of observed data) – Extrapolated (beyond range of observed data)

Can this be called a causal effect?

• Rubin: causality only meaningful for a modifiable

factor

– If unmodifiable, no experiment/intervention possible – What might be modifiable is the system response to race/ethnic ID or appearance – Descriptive inference still useful;

• Which is causal?

– The doctor refused pain meds because the patient was Black – The doctor refused pain meds because she was told that Black patients were more likely to abuse

Regression prediction → Causal inference?

• Threats to validity

– Extrapolation without strong conceptual basis – Relationships differ in another setting

• Are effects the same for given variable with …

– Natural variation – Natural variation with selection (observational study) – Experimental intervention – Program implementation – Social change

Generalizability (“External validity”)

Health disparities

• Mechanisms variable across

settings, subgroups

• Desired outcomes involve

major extrapolation from existing conditions

• Natural variation in major

psychosocial factors hard to identify

• Lifetime effects, manifest

and subtle Health care disparities

• Mechanisms based in

invariant clinical processes (sometimes, somewhat)

• Effects in natural variation,

trials, program implement- ation may be similar

• Natural variation in

treatment by geography, providers, etc.

• Can control for relevant

background

No causality without …

• Strong version:

“No causality without intervention”

– Need intervention to deduce causality

• Weak version:

“No causality without relevant variation”

– Establish basis for generalizability – Causal inference should inform us regarding effects of possible intervention – Conversely, intervention should recognize nature of underlying variation

Effects of race? Effects of racism?

• “Effects of racism” as an ultimate objective

– Is there accessible variation?

• “Effects of race” looks at interaction
• To generalize, need to:

– Examine generalizability of studied variation – Recognize when measures may be off the causal pathway – Consider relevance to plausible interventions

Thank you

• And thanks to previous speakers and chair for

their contributions

• Responses, questions and discussion?