A response to AER re evaluating by Excel

I recently read an interesting but I think incomplete and less than pragmatic paper in the American Economic Review by Fang and Gong. In that paper they use Medicare Part B claims data to advance what they argue is a good first pass claims fraud detection methodology. They seek to identify individual National Provider Identification (NPI) numbers which perform more than 100 hours of billable work per week over the course of a year of claims. They then contend that these individuals are likely fraud investigation targets. From there, they build an interesting model that looks at time utilization and cost.

I have several major objections to this paper. The first is a pragmatic objection which I think renders the data as they interpreted it to be less than useful. The second objection is more of a critique of their methodological choices to model fraud. But first, I want to highlight something I wrote when the Center for Medicare and Medicaid Services (CMS) started to release Medicare billing data:

• Claims rolling up to a provider’s NPI or Medicare ID. Non-MD/non-D.O. clinicians such as Certified Nurse Practitioners, Physician Assistants, Master and Doctorate level Physical Therapists etc. often will roll their billing up to a doctor’s Medicare billing number. This means we can’t do a simple time management bullshit detection study based solely on “This provider is claiming he is doing 17 Medicare Part B procedures a day. Each of these procedures takes 30 minutes… IMPOSSIBLE”. That type of first level analysis might identify odd situations, but most will be explained by seeing three or four CRNPs/PAs doing most of the work that the doctor than bills for.
• Medicare Advantage is not in this data set. Some regions have lots of Medicare Advantage enrollment. Others don’t Some docs have a lot of Medicare Advantage patients. Others don’t. We can’t generalize too well to the entire Medicare population from the CMS data set.
• No way to determine medical neccessity/particular skill. This is pure counting data, it is not quality data…

Fong and Gong only look at the provider NPI. They neglect to examine the possibility that there is significant number and time of services being performed by non MD/DO providers who are performing services incidental to the supervising physician’s practice and billing under the MD/DO NPI.

There is nothing in the paper that addresses non-MD/DO providers rolling up their billing to an MD/DO NPI. This is quite common and allowed. Some master’s level clinicians in some states routinely bill independently. This information is captured in this study and noted in footnote 5. However many non-MD/DO providers will bill via their supervising MD/DO. The Wall Street Journal noted this arrangement in 2014 for an orthopedic practice in the middle of an article about the problems of analyzing outliers by Excel without contextual knowledge. The physical therapists that practiced underneath the surgeons billed with the surgeon’s NPI. This is legal and common. I think this paper is missing the non MD/DO provider universe as implied in Footnote #6;

For expositional simplicity, we will refer to all individual providers as physicians even if a small fraction of them are nurses or physician assistants

Nowhere in the paper did I see the phrase “incidental” or “modifier code” or any other indicator that a significant portion of primary care and first level specialty work is now being performed by non-MD/DO providers who routinely roll up their billing to an MD/DO provider.

The Office of Inspector General at CMS has examined very high implied time billing practices in the past. What they have found is the services were being legitimately performed although at times they were being performed by clinicians who were not authorized by either license or law to perform a set of services as services incidental to the MD/DO who supervised them. Here the better assumption is that this is either waste or abuse rather than deliberate fraud unless there is extraordinary evidence to argue for fraud.
If there is limited to no inclusion of master’s level clinician service rendering then the analysis performed by Fang and Gong will lead to numerous false positives which renders their tool an inefficient use of time and resources to detect actual incidents of fraud.

My second major concern with this paper is their modeling choice of time utilization.

I think it is flawed and biases their estimates of potentially fraudulent activity. If we assume that there is significant fraud via overclaiming of time, then the methodology that is used in this paper which the authors claim is conservative will overstate the hours worked. Actual fraud will go undetected while legitimate and appropriate services will be flagged.
The authors methodology:

Our idea is to use the time requirement for timed codes described above to estimate the time requirement for all other codes. In order to do this, we construct the expected time needed for each code based on the “typical time needed” suggested by the AMA guideline.10 This is important because the actual time to furnish a service code may vary both across and within physicians. We construct the “expected time needed” from the “typical time needed” as in AMA guideline as follows. Assuming the time needed follows a uniform distribution, we take the simple average of the minimum and maximum time allowed for each code to get the expected time. Specifically, some codes may have an explicit range of time needed, such as “5-10 minutes of medical discussion.” For such codes, the expected time needed is simply the average of the lower and upper bounds. For codes that do not have such a range, physicians are supposed to file the code whose typical time needed is closest to the actual time spent. For example, between codes 99202 and 99203 as described in Table 1, a physician who spent 23 minutes should file the code 99202 instead of 99203. Following this logic, the expected time needed we will assign to HCPCS codes 99201 through 99205 are 7.5, 20, 31.25, 45, and 60 minutes, respectively. To see this, consider HCPCS code 99201 for example. Note that physicians who spends 0 to 15 minutes with a new patient is supposed to file HCPCS code 99201 if they follow the AMA guideline. Thus, under the plausible assumption that the actual time spent with patients follows a uniform distribution, the simple average of the minimum (0 minute) and maximum (15 minutes) time allowed for filing 99201 is 7.5 minutes

We know in other fields (auto mechanics, HVAC repair) there is a significant disparity between the time that is recommend for an action to occur and how long it takes for a skilled individual who is optimizing for revenue maximization to complete the task. Under most circumstances, the performance time is significantly less than the recommended time. I would have a hard time believing that some MD/DO providers who have no one else rolling claims up to them have not adapted a system where they consistently perform time recommended tasks near the bottom of the recommended range.

If we are to assume that a provider is intentionally attempting to maximize Medicare Part B revenue through aggressive billing for services then this metric is optimizing on cost per service. I do not think that is the appropriate metric. Instead the greatest constraint is physician time so the relevant unit of analysis should not be revenue per procedure code but revenue per procedure code per minute. For instance, in 2014, the National Payment Amount for 99201 is $43.35 at a non-facility. The per minute rate ranges from $43.35 for a one minute visit to $2.89 per minute for a fifteen minute visit. A 7.5 minute visit averages $5.78. 99202 has a national payment amount in 2014 of $74.51. At a sixteen minute length of service, this is a rate of $4.65 per minute. Using the author’s estimates $3.72 per minute is the estimated revenue per minute for 99202.

The clear incentive if the physician is actively attempting to maximize revenue with a strong time constraint is to minimize the amount of time spent with a patient and maximize the number of patients that are seen. An estimated 99201 visit of 7.5 minutes is more valuable on this metric than an author estimated 99202 visit of 20 minutes if the provider is able to see more patients. At this point, it is a clinical and pragmatic need to examine how many patients are seen and how often are they seen.

These pragmatic concerns of not addressing non-MD/DO billing as well as the wrong choice and thus the wrong modeling of provider revenue considerations as a function of time instead of a function of service codes leads me to believe that what should have been an interesting paper is not actively advancing pragmatic knowledge of the field.