,
24 tweets,
7 min read
Read on Twitter
Re: the recent report on racial differences in AD biomarkers in CSF (bit.ly/2T7hZoC) & the superb commentary by @beyoung40 (bit.ly/2FPYkG4). This thread details another issue in this study: analytical decisions in #healthdisparities research ...
@melodem_group
@melodem_group
1/19 @ManlyEpic posted a terrific comment in @alzforum abt the study 👆, highlighting, w/ data, its vulnerability to selection bias & its unsettlingly small number of African American participants (which is a fast, though not a guaranteed way to end up w selection bias).
2/19 Looking at the analysis description, one additional feature of this study stands out: what question(s) was it answering?
3/19 I realize this sounds glib (or is it flip?), but the variables we choose to adjust for in our analyses should fall logically from the question we want to answer. It’s not rocket science—not even brain surgery.
4/n But in #healthdisparities research, we do a disservice to understanding and righting these disparities if we toss a bucket of variables in a regression model and hope for the best. Here’s the run down:
5/n First, here was their question, as shown in their abstract. Not too complicated, right? 
6/n By one perspective, this can be viewed as the public health-oriented question. It asks “how well” self-identifying black and white populations are doing relative to each other without any mechanistic (aka mediator) explanations for those differences.
7/n It does not ask, for example, after you account for the fact that our society historically and currently screws African Americans in the realms of education and housing, what are the differences? (more on mediators later) It asks, Houston, do we have a problem at all?
8/n The analytical translation? Maybe adjust for age and sex, if the distributions are off between the groups, but that’s about it.
It almost seems too easy. Who ever published anything from a regression model with 3 variables in it?
It almost seems too easy. Who ever published anything from a regression model with 3 variables in it?
9/n Yet, these analyses tell us: if I take a self-identifying white and a self-identifying Afr Am person of the same age and sex, how do their AD biomarker levels in CSF compare? Adjusting for age and sex means that the differences can't be explained by them.
10/19 Note that age and sex are not *results* of race (or racism). And perhaps the same could be true of APOE4 carriership. The key here is that these analyses adjust for nothing caused by race/racism.
11/n The DAG below shows all of the variables adj for in the paper's analysis & how they are likely to be related to race and the CSF biomarkers. (The "AD" variable [geen] was not included but is shown to illustrate how some of the variables are linked to each other.)
12/n And yet, consider variables such as education and BMI, which they also adjusted for. What are they doing in there? No way do education or BMI determine race. How would they be sources of confounding?
13/n Or perhaps the question at hand wasn't public-health-oriented so much as it was “post-explanatory,” i.e., is there a direct effect of race/racism on CSF biomarkers after accounting for likely mediators, ie, factors that are both caused by racism and affect the biomarkers?
14/19 Perhaps the authors were thinking of these variables as mechanistic explanations for any race-biomarker link and by adjusting for them, wanted to "adjust them away."
15/n If that was the question, what’s the point? Is it to ask if African Americans have some inherently different risk [which sounds a wee bit like scientific racism ... but maybe that's me]? Even if not, why just those variables & not a huge suite of other explanatory variables?
16/n Plus, there are much better methods for #mediationanalysis, but I digress. The last set of problematic variables. First, clinical status. This is not just a possible result of race/racism, but also the CSF biomarker (via the backdoor path thru AD). Why is it in the model?
17/19 Second, BMI. Yeah, I know I called it a potential mediator earlier, but if it was measured late enough in life, it was probably also affected by clinical status. Two strikes against BMI in this model.
18/19 Wrapping up, this was not hard, yes? But so much research comparing health outcomes between African Americans & whites ignores this.
19/19 So, not only should we do better to addressing inclusion & selection bias in this research, as articulated by @beyoung40 & @ManlyEpic, but we also need to match our analyses to our research questions (& possibly involve people on our team to push us in this direction 🙂).
IOW: the "total effect" of being African American (vs being white) on CSF biomarker level, through all mechanistic pathways, known and not.
Translation for DAG newbies: Arrows leading *from* race and eventually to the CSF biomarker are tracing a putative mechanistic pathway. We don't want to adjust for variables on these pathways in estimating the total effect of race [the social construct] on CSF biomarker.
And variables that have an arrow pointing into race and also an arrow pointing into a separate pathway to CSF biomarker? Those are potential sources of confounding of the estimate. We want to adjust for these.
Closing credits: thanks to @ManlyEpic @ER_Mayeda @MariaGlymour, and Jay Kaufman the conversation that led to this thread. (Funny how Jay manages to have a massive twitter presence even without emitting a single tweet.)
