Thread by @michelleko2d

Continuing @VPrasadMDMPH @walidgellad @AnilMakam discussion on structural racism #HSR @AcademyHealth: A thread on research critique w. a specific application to:
pnas.org/content/early/…
Let's go! 1/

1) When engaging in thoughtful critique, whether your own or someone else's, read the entire paper & supplement.
1x) Critiques of Greenwood et al that raise issues such as SES, comorbidities, etc, imply that the commenter has not read the entire paper. Specificity is key.

2/

2) When identifying limitations of the methodology, differentiate between a) those not mentioned by the authors; b) those mentioned by the authors but not addressed as fully as desired.

3/

2x)
a) Lack of specificity of attribution- which attending was this?
b) Not accounting for teams, selection effects- discussed by the authors, explanations/sensitivity analyses presented.

4/

3) WRT 2a) When presenting limitations, it is incumbent upon you to present the alternatives & your expected direction of biases based on your hypothesized alternatives. This is #HSR 101. Simply saying there is a flaw w/ no attention to why/relation to findings is inadequate.

5/

3x) Inability to attribute MD role: was this the initial attending upon birth? Was this the attending at some point in hospital stay? (The data are described as admissions, thus discharge is less likely- but you can argue unobserved someone expected this to be dc MD)

6/

Presenting expected bias:
a. This is the MD present at birth OR assigned at some point in stay, w/ limited family input: quasi-random argument presented by authors: bias estimates to null
b. This is MD expected to continue care, w/ family input. bias estimates +
7/

c. Assignment unknown, no systematic underlying process, bias estimates to null.

--> The identified associations cannot be attributed to either MD effects OR selection effects. Limiting your argument to b. is inadequate.

8/

4) When presenting limitations also discussed by authors, it is incumbent to discuss the limitation, what the authors did to address it, & why their approach is insufficient.
THEN your alternatives & again, expected directions of effects.

9/

4x) Two criticisms:
a. Medicine is a team sport: attribution to a single MD does not reflect actual process of care in which many team members interact w pts & families.

How did the authors address this?

a. First: within-MD fixed effects. The underlying assumption that MDs do not work w/ same care teams over time. Biases effects towards null.
OR: you presume that MDs may work w some consistency of teams based on scheduling & preferences. Thus MD-team are correlated.
11/

They also compared teaching vs. non-teaching settings, to attempt to account for resident effects. No effect for residents would be expected- this also does not address other care team members which they acknowledge. The main conclusion is no diff in teaching vs nonteaching
12/

b. Selection bias: concordance reflects unobs factors correlated w pt agency to choose Black MD & infant mortality.
See above- 1st must account for data collection process. 2nd- id what unobs factors do you hypothesize & direction?
13/

Given both FE & stratified analyses partly accounting for income/insurance/geography/quarter, unobs may include:
Education independent of income, social, cultural health capital- bias +
Past experiences of racism in institutions- bias -
More conceptual grounding needed
14/

5) Critiques of methodology need to be grounded in an understanding of the underlying conceptual framework. (Also everyone in #HSR!: your research methods need to be grounded in an underlying conceptual framework!)

15/

5x) The aim of this study is to examine associations betw MD-pt racial concordance & infant in-hospital mortality.

What are we observing when we measure "racial concordance"?

16/

More broadly, how are Black patients "matched" to MDs?
Structural racism in healthcare identifies several processes:
- Structural inequities in income/occ, ins coverage restricts MD pool
-Racial residential segregation --> healthcare segregation further restricts MD pool
17/

-Health system segregation, in which hospitals & practices engage in behaviors to limit their Black patient populations
These major structural drivers need to be acknowledged in any concordance discussion- Black pts do not have the level of "choice" implied by selection

18/

This paper addresses these by hospital/insurance fixed effects- by intent, estimates will be biased towards zero. Criticisms that the observed effect size (and also after physician fixed effects) is small should consider that this is by design- the null hypothesis is *zero*

19/

What about physicians within the context of structural racism in healthcare? Critiques of selection bias reflect assumptions that fall on the patient side, which is problematic given assumptions about patient agency, hence my preference for "endogeneity"

20/

Concordance mechanisms: unobserved & bias direction
-pt prefs & unobserved agency & capital that is also correlated w outcomes +
-pt prefs due to prior experiences of better quality of care +
-pt prefs due to prior experiences of institutional discrimination -
21/

Concordance mechs: unobs & bias direction
-communication/trust +
-clinical decision-making +
-Undefined quality of interaction +
--> all of these are applicable to a) the attending MD; b) the "team" attached to an MD; & biased - if imprecision in attending MD/team matching
22/

Concordance: not pt selection, not direct pt interaction- one of the main mechanisms not discussed by authors/critics is the assumption that selection effects x race/ethnicity only apply to pts & not MDs-- despite *also* being a mechanism of structural racism in healthcare
23/

We have ample generations of evidence that a) White MDs actively select out of Black communities; b) Black MDs select into Black communities; and that both of these processes occur irrespective of SES bc of racism.
Observed racial concordance reflects both pt & MD preferences
24/

Racial concordance must therefore be examined in context of the contrast: White MDs-Black pts reflects selection effects- which White MDs serve Black communities & why, as well as proxy effects- observed racial mismatch can also reflect a host of structural inequities
25/

The authors' testing of hospital patient racial composition addresses this to some extent-->adverse selection of White MDs to Black-serving institution, with potentially lower quality of care, increased risk of bias, discriminatory behavior.
26/

The additional teaching vs. non-teaching analysis also obliquely addresses the potential that White MDs in teaching are preferentially serving Black patients & higher quality. Maybe.
27/

-Conversely, Black MDs caring for Black pts may be more motivated to monitor quality of care, vs. White MDs who do not have the same selection motivations. In general: there is unobservable selection, but said selection occurs bidirectionally: MDs cannot be ruled out of this
28/

MD selection behavior means that the assumption: "all else equal, Black pts are choosing solely by race" does not hold- they are not offered the equivalent choice set. MD selection effects (or team or whatever you think this is a proxy for) occur irrespective amt pt interaxn

29/

The conceptual weakness of lasered focus on attribution methodology- pathways via structural racism are complex, "directness" of attribution limits explanatory mechanisms to direct pt interaction only, & selection arguments need to account for simultaneous MD/pt selection
30/

What can we conclude from this paper?
I think it's reasonable to assume some form of racially concordant pt care is happening, whether or not it's the admitting MD, or an attending MD at some point in the process, & their team.

31/

We can assume based on prior lit that racially concordant interactions have, on avg better communication/better trust & discordant interactions have +likelihood of bias, microaggressions, discrimination.

32/

Can we attribute these mechanisms alone to infant outcomes? No.
Can we rule out these mechanisms? No. There is far too much evidence on Black experiences of racism in healthcare.
We can also assume that racial concordance serves as a signal for selection on both pt/MD/team
33/

Can we attribute findings to unobs correlated outcomes w/ health due to pt preferences for racially concordant care? Yes. However, we must also assume systematic attribution by MD race, & we cannot rule out that pt prefs reflect individual advantages + better care experiences
34/

Can we attribute findings to unobs correlations w/ MD racial selection? Yes. Either thru direct quality of care or unobs proxy for the racially structured opportunities available to the pts.
Can we rule out other mechs? No. Most likely all are happening.
35/

One signal of selection effects is the higher comorbidity level of Black infants seen by BlacK MDs, but this could be supported by both pt & MD selection fx
Greater unobs resources--> selection of Black MD
Greater difficulty obtaining tx from White MD--> Black MD
36/

What do we do with all this?
We have positive findings under the H0:0, which supports that one or more mechanisms are in play that we cannot observe. We have some unanswered precision on attribution which may not rule out any of those.

37/

Essentially: there is an association between racial concordance & infant mortality, which can be explained by multiple mechanisms of structural racism and we cannot rule out MD, pt or other health system effects.
38/

Did you read all 38 Tweets? 🤪
This is not the end! Next up, in a *new* thread: Continuing the @AcademyHealth discussion on #StructuralRacism in #HSR, using this critique as an example.
Stay tuned.