• Nothing to see here: On a Medicaid cancer study

    Last night, I looked at the Medicaid cancer study Sarah Kliff  posted about yesterday. You can click through to read Kliff’s summary or read the abstract to get a quick take on the findings. Very briefly, it shows that non-elderly patients living in Ohio with one of eight types of cancer enrolled in Medicaid have less favorable survival and higher five-year mortality rates than those not enrolled in Medicaid. Though there is nothing deeply wrong with it or the way the results are characterized, it’s not a terribly informative study. (That’s true of many studies, including some of my own.) I worry it will be misinterpreted by some to mean that Medicaid causes harm. It doesn’t.

    One issue with the study is that the non-Medicaid comparison group is a mix of who-knows-what: some uninsured and some with various forms of insurance (including private coverage and Medicare, for those qualifying due to disability). So, it’s not really clear to what Medicaid is being compared.

    A second problem, which the authors acknowledge, is that the controls are weak. Some key ones (education and income) are at the census tract level, not specific to the individual. There are no controls for health or disability beyond variables pertaining to the cancer itself.

    Even studies like this, of which there are probably hundreds, that use a much richer set of controls find that Medicaid outcomes are worse than being uninsured. That’s because almost no feasible set of controls can overcome the unobservable and powerful reasons individuals self-select into Medicaid. Sicker and less advantaged individuals enroll in Medicaid and observable controls don’t fully account for that. This fact was covered in my NEJM paper with Aaron, Harold, and  Uwe Reinhardt, as well as in many posts on TIE. (Here are links to two in particular that nail the key points and link to many others.)

    All the Medicaid cancer study Kliff posted on tells us is something we already know, that Medicaid enrollees have characteristics that are associated with poorer outcomes, even controlling for observable characteristics. Now this well-established fact is documented for eight cancers too, but that’s not a big contribution. (Plenty for a publication in Cancer but very far from a big deal.) The paper does not show us, and cannot show us, that Medicaid causes poor outcomes.

    As I have covered many times (in posts linked to above), there are other publications out there that show Medicaid improves health, relative to being uninsured. A recent paper that provides evidence consistent with just that is the Oregon Health Study (see also Aaron’s post on it). It would take a study showing Medicaid causes harm with methods at least as strong as the Oregon Health Study before I would begin to question the value of Medicaid to the health of those who rely on it.

    AF

    Share
    Comments closed
     
    • “That’s because almost no feasible set of controls can overcome the unobservable and powerful reasons individuals self-select into Medicaid. Sicker and less advantaged individuals enroll in Medicaid and observable controls don’t fully account for that.”

      Amen. Waiting patiently for the day when people automatically apply the same level of analytical scrunity to the regional variations catalogued by the Dartmouth studies.