Adapting Causal Inference Methods to Improve Identification of - - PowerPoint PPT Presentation

Adapting Causal Inference Methods to Improve Identification of Healthcare Disparities

Benjamin Cook, PhD MPH Director, Health Equity Research Lab Cambridge Health Alliance/Harvard Medical School healthequityresearch.org @cmmhr

June 26, 2017

Identifying Health Disparities and Pathways Amenable for Interventions to Reduce Disparities

2

3

4

Quantifying Disparities and How They Arise

Jones CP et al. J Health Care Poor Underserved 2009

5

Overview

Identifying healthcare disparities: applying concepts from a causal inference framework

• Brief background on race and causal inference
• Measuring disparity using the notion of the

“counterfactual” to measure healthcare disparities.

6

Overview

Identifying healthcare disparities: applying concepts from a causal inference framework

• Brief background on race and causal inference
• Measuring disparity using the notion of the

“counterfactual” to measure healthcare disparities.

7

Adapting counterfactual methods in disparities studies

α= E(Y | X=x1) - E(Y | X=x2)

8

Dabady, M., Blank, R. M., & Citro, C. F. (Eds.). (2004). Measuring racial discrimination. National Academies Press. Holland 1986; 2003; Rubin 1974, 1977, 1978; Pearl 2000.

The causal effect α is a difference in

• utcome Y between treatment (X=x1)and

control (X=x2) Difference between an individual receiving treatment and the same individual not receiving treatment. Because the individual can only take one of these values, one of these is a counterfactual.

Adapting counterfactual methods to disparities studies

α= E(Y | X=x1) - E(Y | X=x2)

9

X Z U Y X Z U Y

Randomization

Dabady, M., Blank, R. M., & Citro, C. F. (Eds.). (2004). Measuring racial discrimination. National Academies Press. Holland 1986; 2003; Rubin 1974, 1977, 1978; Pearl 2000.

The causal effect α is a difference in

• utcome Y between treatment (X=x1)and

control (X=x2) Randomized experiments and quasi-experiments at the population level allow us to calculate average treatment effects that estimate this causal effect.

Adapting counterfactual methods to disparities studies

α= E(Y | X=x1) - E(Y | X=x2)

10

X Z U Y X Z U Y

Randomization

Randomization breaks the link between X and all other observables (Z) and unobserved variables (U) except the outcome (Y) By randomizing at the population level, we are able to infer the difference between the outcome if an individual received the treatment and the outcome if the same individual did not receive the

• treatment. Remember that one of

these is a counterfactual.

Dabady, M., Blank, R. M., & Citro, C. F. (Eds.). (2004). Measuring racial discrimination. National Academies Press. Holland 1986; 2003; Rubin 1974, 1977, 1978; Pearl 2000.

The causal effect α is a difference in

• utcome Y between treatment (X=x1)and

control (X=x2)

Adapting counterfactual methods to disparities studies

11

VanderWeele, T. J., & Robinson, W. R. (2014). On the causal interpretation of race in regressions adjusting for confounding and mediating variables.Epidemiology, 25(4), 473-484. see Krieger letter to editor and response.

• For causation to occur, manipulability of the

potential causal variable is required (Holland 2003)

• Is race manipulable?
• “Racial categories, differential perceptions and

treatment of racial groups, and associations between race and health outcomes are modifiable.”

Race??? Z U Y

Randomization

12

• In disparities studies, minority race is the “treatment”
• f interest.
• Ideally, the counterfactual group is a group identical

in all aspects to the minority group except for minority race status.

• “Balancing” can be achieved (i.e., videos with actors

(Schulman 1999), job applications given names typical

• f blacks and whites (Bertrand and Mullainathan 2004)).
• Implementing the IOM definition of healthcare

disparities requires a hypothetical group with counterfactual distributions of health status variables (Cook et al. 2009)… Adapting counterfactual methods to improve identification of healthcare disparities

Overview

Identifying healthcare disparities: applying concepts from a causal inference framework

• Brief background on race and causal inference
• A framework that uses the notion of the

“counterfactual” to measure healthcare disparities.

13

1) Racial & ethnic disparities in care associated

with worse outcomes, thus unacceptable

2) Disparities reflect broader inequality

& discrimination in American society

3) Health systems, providers, managers

& patients contribute to disparities

4) Provider uncertainty, stereotyping,

& bias contribute to disparities

5) Small differences in refusal rates

do not explain disparities

Institute of Medicine, 2003

Unequal Treatment

Defining Racial/ethnic healthcare disparities

• Unpacking healthcare “disparity” to make it more

relevant to practice / policy

15

Health care differences are due to many factors:

• African-Americans and Latinos



have lower rates of education and income.



more likely to be uninsured.

• Asians have lower rates of illicit drug and alcohol use than whites
• Latinos are on average younger than whites and more likely to be in age

groups that have higher prevalence of mental illness

• Providers have biases that may lead to discrimination.
• Hospitals and community health centers have had a legacy of racist policies



Simkins v Moses H. Cone Memorial Hospital (1963), challenged the federal government’s use of public funds to expand and maintain segregated hospital care



differential harm from research, detention, involuntary commitment

16

Income Education Rates of Substance Use Age Geography Discrimination Racism Insurance Employment Comorbidities

Should differences due to all of these factors be considered a disparity?

17

 Are these allowable or

justified differences?

 Should the health care

system be held accountable for these differences in care?

 To track progress in a way

that is useful for policy, do we count all these differences?

Differences due to:

Defining Healthcare Disparity: Differences, Discrimination, and Disparity

18

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

In Unequal Treatment, the IOM made a distinction between allowable and unallowable differences

Allowable / Justified

Need for Care (Substance abuse rates) Prevalence of MI Preferences for Care

N

• n
• M

i n

• r

i t y M i n

• r

i t y

Quality of Care

Difference Clinical Need & Appropriateness, Patient Preferences Healthcare Systems & Legal / Regulatory Systems Discrimination: Bias, Stereotyping, & Uncertainty Disparity

Quality of Care

Difference Clinical Need & Appropriateness, Patient Preferences Healthcare Systems & Legal / Regulatory Systems Discrimination: Bias, Stereotyping, & Uncertainty Clinical Need & Appropriateness, Patient Preferences Healthcare Systems & Legal / Regulatory Systems Discrimination: Bias, Stereotyping, & Uncertainty Disparity IOM, 2002

The IOM Definition The IOM Definition

• f Healthcare Disparities
• f Healthcare Disparities

Whites Blacks

Unallowable / Unfair

Discrimination Income Education Employment Insurance

? Comorbidities Geography Legacy of racist care

19

Definition of Racial Disparities: IOM

• 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.

• Different than HHS definition1,2: “All differences

among populations in measures of health and health care.”

1 Healthy People 2010 2 National Healthcare Disparities Report, 2003

Definition of Racial Disparities: IOM

• 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.

Commonly Used Disparities Methods

• Typical method of measuring disparities using a

regression framework from previous studies

1) y=0+ RRACEi+ AAgei+ GGenderi+ε 2) y=0+ RRACEi+ AAgei+ GGenderi + HHealthi+ε 3) y=0+ RRACEi+ AAgei+ GGenderi + HHealthi + IIncomei+ε

• R represents a “residual direct effect”
• Omitted variable bias - R difficult to interpret
• Difficult to track this coefficient (or change in

coefficient) over time and across studies

Operationalizing the IOM Definition

• Adjust for mental health status (clinical

appropriateness/need), but not SES variables (system level variables)

• In a regression framework:

y=0+ RRACEi+HHealthi + SSESi+ε White: yW=0+ RRACEWhite+ HHealthWhite + SSESWhite+ε Black: yB=0+ RRACEBlack+ HHealthWhite+ SSESBlack+ε Disparity: yW-yB

^ ^ ^ ^ ^ ^ ^ ^ ^ ^

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

Step 1: Fit a (Non-linear) Model of MH Care Expenditures

Two-part model Access (Expenditure>0): Probit Prob(y>0) = Ф(x'β) Expenditures: GLM with quasi-likelihoods E(y|x) = μ(x'β) and Var(y|x) = (μ(x'β))λ



with log link function

 and variance proportional to mean (λ=1)

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

Step 2: Transformation of Health Status

• In a linear model, we could adjust at the mean
• In a non-linear model, we must adjust the entire

distribution.

• How do we adjust?
• “Rank and replace” method
• Propensity score weighting balances

racial/ethnic groups on health status only

• (Random replacement – not recommended)
• 1. Fit a model
• 2. Transform HS distribution
• 3. Calculate predictions

Rank and replace

• Each observation in the white subsample and full

black sample is separately ranked by the predicted probabilities generated from the health status portion of the probit (or other non-linear) equation and matched by their respective rankings.

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

• 1. Fit a model
• 2. Transform HS

distribution

• 3. Calculate predictions

100 White Black

Transform Distribution of Health Status

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

• Weight each individual on the propensity of “being white”

conditional on a vector of observed health status covariates.

• Measure P(White)=β0+ β1(HS)+ε = ê (Hi)
• Weight minority individuals by their probability to be White

(ê(Hi)), and White individuals by their probability to be minority (1- ê(Hi)).

• Conditional on the propensity score, the distributions of
• bserved health status covariates are the same for

minorities and Whites (Rubin 1997)

• “Weights up” individuals with ê(Hi) close to 0.5, whose

health status distributions could be either White or Black.

Propensity Score Weighting

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

Propensity Score Weighting

P(White)=β0+ β1(HS) = ê i(z)

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

Before PS weighting After PS weighting

Step 3. Disparity estimation

• Disparity estimation
• White prediction minus counterfactual minority

prediction

White: yW=0+ RRACEWhite+ HHealthWhite + SSESWhite+ε Black: yB=0+ RRACEBlack+ HHealthWhite+ SSESBlack+ε

Disparity: yW-yB

• Variance estimation - resampling methods can be

used:

• Recomputing an estimate on many replicate

samples (bootstrap, balanced repeated replication) instead of having a special variance estimator for each estimand.

^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^

Does the Method Matter?

33

Total Expenditure

Unadjusted Rank and Replace Propensity Score Recycled White - Black White - Latino

Probability of Any Expenditure

0.00 0.02 0.04 0.06 0.08 0.10 White - Black White - Latino

Mean Expenditure Given Nonzero Expenditure

• 400
• 200

200 White - Black White - Latino

Predicted Expenditure

20 40 60 80 120 RDE

Does the Method Matter? (2)

34

Disparity (SE) Disparity (SE) 488.60 (78) 912.63 (101) 1125.72 (98) 1284.72 (411) 1407.32 (482) 1454.25 (491)

Source: Combined Medical Expenditure Panel Survey (MEPS) data from 2003 and 2004

Rank and Replace RDE of Model with No SES Non-linear models Combined Method 841.75 (94) Table 4. Comparison of methods of calculating black-white disparity in total medical expenditure using linear and non-linear models Propensity Score Adjustment for HS and SES (RDE)1 Linear RDE Linear Models Non-Linear RDE Adjustment for HS2 Oaxaca-Blinder Decomposition

Racial/Ethnic Disparities in Initiation and Adequacy of Mental Health Care

* p<0.05: Source: 2004-2010 MEPS ;Cook, Zuvekas, McGuire et al., HSR, 2013 * * * *

40% 16% 24% 10% 27% 11%

0% 5% 10% 15% 20% 25% 30% 35% 40% 45%

Initial care Adequate Whites Blacks Latinos Adequacy of Care Initiation of Care Respondents with K-6>12 or PHQ-2>2

Cook et al. Psychiatric Services 2016

 Mental health care access disparities are at

least 2:1 and have worsened in the U.S.

MH Care Access Disparity Trends over Time

2004-2012 Medical Expenditure Panel Surveys (N= 214, 597) All disparities within year are significant (p<0.05) Starting in 2010, these disparities are significantly greater than 2004

White rates of any MH care: 16.3% in 2004 19.5% in 2012

Summary

• Counterfactual: “What would the rates of healthcare

access be for a group of Black individuals with a white distribution of health status?”

• IOM-concordant methods adjust for health status

(but not SES) in the presence of non-linear models and correlations between health status and SES.

• Similar to non-linear decomposition (Fairlie 2006)

and can be used to “decompose disparities” (Saloner, Carson, Cook 2014)

• The rank and replace method and the modified

propensity score method are “IOM-concordant”

• Both methods had similar estimates and variance

in separate empirical analyses.

37

Summary

• Make a choice about how to define disparity; a race

coefficient may not be what you want

• Consider whether or not to adjust for comorbidities,

preferences, SES, geography, etc.

• The IOM definition provides a framework for how to

measure healthcare disparities accounting for health status, preferences, SES, and discrimination.

• Follow-up decomposition and mediation analyses

help to makes explicit the underlying contributors to

• disparities. This can be used to identify intervention

targets.

Thanks

bcook@cha.harvard.edu @cmmhr www.healthequityresearch.org

39