Tuesday. July 30, 2019

The United States spends more on healthcare than any other rich country and gets less for it — 39 nations have longer life expectancies (Gapminder). That gap makes the system a prime target for analysis, and lately the data has caught up: CMS now publishes both the payments drug companies make to physicians (Open Payments) and what doctors actually prescribe (Medicare Part D).

ProPublica already surfaced the uncomfortable headline: doctors who take money from drug makers — even just a meal — prescribe more brand-name drugs, and are two to three times as likely to be among the heaviest brand-name prescribers in their specialty. I wanted to know what else drives those choices, and what might be quietly inflating drug prices and feeding the opioid crisis.

CMS publishes two Part D files a year: one with a row for every doctor–drug pair (about 25 million rows in 2017), and a summary with one row per doctor — brand and generic costs and claims, plus their patients’ plans, incomes and conditions. I was after the big picture, so I combined the 2015–2017 summaries in SQLite. The highlights:

• Brand-name drugs were only 18% of claims but 80% of spending.
• Had every brand-name claim been filled with a generic, Medicare would have spent about $85 billion less a year. That’s a ceiling: many brand-name drugs have no generic.
• The number of prescribers grew almost 3% a year.
• Brand-name claims fell almost 4% a year and generic claims rose 3%, yet total spending climbed 8% a year, while inflation ran under 2%.

To capture each doctor’s habits, I used the share of their prescriptions that were brand-name, the number my models would try to predict. For most doctors it was 5–25% (18% on average); a large group prescribed none, and about 3,500 prescribed nothing else:

To compare practices of different sizes fairly, I turned raw totals into ratios — cost per claim, cost per patient and claims per patient — for every category in the data: plan type, low-income subsidy, opioids, antibiotics and more. The gap they reveal is stark: across the country, a brand-name claim costs on average 17 times as much as a generic one.

On almost every measure, practices cluster near zero with a long tail of outliers: doctors whose claims and costs far exceed everyone else’s. I pulled the outliers into one group. Here’s what they practice, where, and what made them outliers:

Among these outliers, six of the ten features most correlated with brand-name share were ones I’d engineered. The top three, by Pearson’s r:

• Cost per claim for patients without a low-income subsidy (nonlis_cost_per_claim, r = 0.58): the higher it is, the more brand-name drugs the doctor prescribes.
• Cost per claim for patients in stand-alone drug plans (pdp_cost_per_claim, r = 0.48): same direction.
• Long-acting opioids’ share of opioid claims (la_opioid_prescriber_rate, r = −0.45): the higher it is, the fewer brand-name drugs.

To see how the features interact, I fit a single decision tree (full resolution):

Its first question is whether a practice’s cost per claim for patients without a low-income subsidy is above about $82. The darkest leaf, at about 66% brand-name, is a small group of ophthalmologists with high costs per claim across plans; on the way there, the tree asks one question about specialty and four about patients’ subsidy and plan.

A single tree is easy to read but a blunt predictor (this one has only 32 possible answers), so I also tried linear regression, a random forest and a voting ensemble of models. The ensemble didn’t beat the random forest alone, and performance held with just the top 20 features. My engineered features on their own scored best of all, but their residuals showed clear patterns, a sign the model was missing something.

The features the random forest ranked highest tell the same story as the tree: costs and claims by patients’ subsidy and plan. The common thread among heavy brand-name prescribers seems to be who picks up the bill.

An analyst’s job is to widen the options, not to choose among them. Still, my read of this data is that Part D’s structure feeds cost growth, and the drug makers and insurers who gain from it aren’t the ones paying. Patients and taxpayers are.

Share this post