Some Economics of Treatment Disparities in Healthcare

1
Some Economics of Treatment Disparities in
Healthcare

• Amitabh Chandra
• Harvard University and the NBER

Douglas Staiger Dartmouth College and the NBER
2

• There is a MASSIVE literature in medicine and
  public health on treatment disparities in
  healthcare.
• The Institute of Medicines (IOM) report Unequal
  Treatment summarizes the key findings of this
  literature.

3

• Racial and ethnic minorities tend to receive a
  lower quality of healthcare than non-minorities,
  even when access-related factors, such as
  patients insurance status and income, are
  controlled. The sources of these disparities are
  complex, are rooted in historic and contemporary
  inequities, and involve many participants at
  several levels, including health systems, their
  administrative and bureaucratic processes,
  utilization managers, healthcare professionals,
  and patients. Consistent with the charge, the
  study committee focused part of its analysis on
  the clinical encounter itself, and found evidence
  that stereotyping, biases, and uncertainty on the
  part of healthcare providers can all contribute
  to unequal treatment.

Smedley, B. D., A. Y. Stith, and A. R. Nelson,
eds. 2003. Unequal treatment Confronting racial
and ethnic disparities in health care.
Washington, DC National Academies Press.
4
Lets look at some examples from the literature
5
Jha, A. K. et al. N Engl J Med 2005353683-691c
6
Acute Myocardial Infarction
7
Primer on Cardiac Catheterization
8
Lets look at some facts from our own tabulations
of AMI Treatments
Every first heart-attack in Medicare since
1992. Approximately 210,000 such patients per
year. Each AMI is matched to Part A claims data
at 30 days and 1 year after admission, and if
relevant, death certificate data.
9
(No Transcript)
10
(No Transcript)
11
(No Transcript)
12
Question

• Do these disparities represent prejudice against
  women and minorities, or statistical
  discrimination?
• Under statistical discrimination, physicians are
  trying to maximize benefit to patient, but
  gender/race are statistically related to the
  benefit (because of biology, compliance, or
  quality of provider).

13
Two Different Views of the World
Prejudice Patients with identical benefit
treated differently in one group.
14
Two Different Views of the World
EFFICIENT ALLOCATION
Statistical discrimination patients with
identical benefit treated the same, but benefits
higher for one group
Prejudice Patients with identical benefit
treated differently in one group.
15
Key Idea

• Take two patients with the same propensity to
  get the treatment (so who are treated the same)
  and then test whether net benefits are the same
  (statistical discrimination).

16
Model

• B Benefit from treatment
• H Hurdle that Benefit must exceed
• Benefit Xb0 femaleb1 e
• Hurdle h0 femaleh1 v
• Pr(Treatment1) Pr (Benefit gt Hurdle)
• Pr (Xb0 femaleb1 e gt h0 femaleh1
  v)
• Pr (Xb0 female(b1-h1) a0 gt v e)
• Pr (Igtv-e)

h1gt0 reflects prejudice (females must overcome
larger hurdle on average to get treatment)
17
Model

• B Net Benefit of treatment
• H Hurdle that B must exceed to receive care
• (B)enefit Xb0 femaleb1 e
• (H)urdle h0 femaleh1 v
• Pr(Treatment1) Pr (Benefit gt Hurdle)
• Pr (Xb0 femaleb1 e gt h0 femaleh1
  v)
• Pr (Xb0female(b1-h1)h0 gt v e)
• Pr (I gt v-e)
• But we want treatment effect on the treated (TT)
• E(Benefit Treatment1) Xb0 femaleb1
  E(e Igtv-e)
• E(Benefit Treatment1) I h0 femaleh1
  E(e Igtv-e)
• g(I) femaleh1

h1gt0 reflects prejudice (females must overcome
larger hurdle on average to get treatment)
Implication 1 In the absence of prejudice
(h10), two people with the same propensity to
get treatment (same I) will have the same
expected net benefit from treatment.
Implication 2 If there is prejudice (h1gt0), then
higher net benefit (conditional on I) in minority
group.
18
Measurement
For two people with the same propensity
(I) E(Benefit T1,male,I) E(?Sm
T1,male,I) g(I) E(Benefit T1,female,I)
E(?Sf T1,female,I) g(I) h1 Estimate
average difference in benefit, h1E(?Sf -
?SmT1) as Survival a0 a1Treat
a2(Treatfemale) Xa3 e, a1 E(?SmT1) and
a2h1
19
Measurement
For two people with the same propensity
(I) E(Benefit T1,male,I) E(?Sm
T1,male,I) g(I) E(Benefit T1,female,I)
E(?Sf T1,female,I) g(I) h1 Estimate
average difference in benefit, h1E(?Sf -
?SmT1) as Survival a0 a1Treat
a2(Treatfemale) Xa3 e, a1 E(?SmT1) and
a2h1
But its not so simple!
Because men and women have different
distributions of I, the above strategy integrates
over different distributions of I fmale(I) ?
ffem(I) We want the distributions of I to be the
same. Alternatively, we need to know g(I). We can
reweight fmale(I), the male distribution of I, to
make it look like ffemale(I) by putting more
weight on men who look like women.
20
For two people with the same propensity
(I) E(Benefit T1,male,I) E(?Sm
T1,male,I) g(I) E(Benefit T1,female,I)
E(?Sf T1,female,I) g(I) h1 Let f(I) be
the pdf of I for women. Then integrating both of
the above over f(I) and taking the difference
between women and men gives h1(?Sf - ?Sm) (g
g), where
21
What about Estimation?

• Estimate difference in ?S from
• Survival a0 a1Treat a2(Treatfemale) Xa3
  e, where a1 and a2

• Estimation method
• OLS (very good Xs)
• IV (using diffdist, difdistfemale as IVs)
• Weighting
• Unweighted estimation ? But this produces
  treatment effects integrated over different
  distributions of treatment propensity.
• For testing our model, we need same distribution
  of propensity in both groups.
• Reweight men using Barsky, et al. (JASA, 2002) so
  that distribution of cath propensity is same as
  women
• Find 1st, 2nd, ., 99th percentile of female
  distribution of cath propensity.
• Reweight men by .01 over fraction of men in each
  range

22
Empirical Work

• Test predictions of both models using data from
  the Cooperative Cardiovascular Project (CCP)
• Chart data on 140,000 Medicare beneficiaries
  (over 65) who had heart-attacks matched to Part
  B claims.
• Sample is restricted to fresh-AMIs we exclude
  transfers from another ER, or nursing home
  facilities.
• Use CATH as marker for intensive treatment
• Use DIFFERENTIAL-DISTANCE to CATH hospital as IV
  for Catheterization.

23
Construction of Clinical Appropriateness for
Aggressive Treatments Pr(CATH1X)
24
Table 1 Means by sex and race, CCP data
25
Table 2 Probit Coefficients marginal effects
of the effect of Sex and Race on Catheterization
26
Reweighting f(I)
27

28
(No Transcript)
29
Things you probably want to see

• How good of an instrument is DD?
• Are physicians using the right g(I)? Is survival
  benefit increasing in g(I)?
• Are physicians using the same g(I) function for
  men and women (blacks and whites)?
• What if survival per dollar (instead of survival)
  is equalized?

30
Wald Estimates
31
Do Race-Specific Models Explain Disparities in
Treatments after AMI? Jha, Lee, Staiger and
Chandra (AHJ, 2007)
32
Do Race-Specific Models Explain Disparities in
Treatments after AMI? Jha, Lee, Staiger and
Chandra (AHJ, 2007)
33
Defining Net Benefit

• NB (S)urvival ?.(C)ost,
• where ? is survival per 1000 dollars
• What are BIG and small values for ??
• Some might use ? 0 (physician should ignore
  costs of care infinite value of life)
• BIG value for ? implies small value of life-gt
  Costs matter!
• One survivor at 1 year realizes about 5 years of
  life.
• Minimum value of life year would be 20k,
  implying ? 0.01
• More reasonable value of life year would be
  100k, implying ? 0.002
• Our sense is that reasonable values of ? lie
  between 0.01 and 0.002

34
For two people with the same propensity
(I) E(NB T1,male,I) E(?ST1,male,I)
?.E(?C T1,male,I) g(I) E(NB
T1,female,I) E(?ST1,female,I) ?.E(?C
T1,female,I) g(I) h1 Let f(I) be the
pdf of I for women. Then integrating both of the
above over f(I) and taking the difference between
women and men gives
35
What about Estimation?

• Estimate ?S and ?C from
• S a0 a1Treat a2(Treatfemale) Xa3 e,
  where a1 and a2
• C ß0 ß1Treat ß2(Treatfemale) Xß3 e,
  where ß1 and ß2
• For all ? between 0.0-0.1, we test
• H0 a2- ?ß2 0
• if a2- ?ß2 gt 0 ? h1gt0 ? prejudice against women
• if a2- ?ß2 lt 0 ? h1lt0 ? prejudice against men

36

37
(No Transcript)
38
Conclusions

• If anything, women blacks are getting lower
  returns, even after we adjust for costs.
• Our IV estimates are imprecise, but we plan to
  update with 1992-2003 claims data (about 20x the
  sample).
• Key question is why are the benefits of care
  different?
• Genes? Contentious explanation for race
  differences
• Geography? Cant explain sex differences.
• Follow-up care?