What Is The Impact of a Continuous Coverage Mandate?
John Graves (Vanderbilt University)
The objective of this notebook is to provide analytic insights into the possible impact of a continuous coverage mandate as a replacement for the ACA’s individual mandate. Our goal here is to estimate the number of currently insured people who might retain or obtain some type of non-group coverage under a continuous coverage mandate.
There are several challenges in estimating the impact:
Late enrollment penalties mean that the “bite” of the provision is likely to be small for healthy people who are only uninsured for a few months. Moreover, there is likely to be a 30-60 day grace period (as under HIPPA) which makes this “bite” even softer.
A one-time, fixed value late enrollment penalty might mitigate this a bit, but its not clear whether current proposals include this, or whether the provision will feature a penalty that is a function of the number of months without coverage (as under Medicare Part D).
More generally, Medicare calculates similar late enrollment penalties differently:
- For Part A the penalty is 10% of the monthly premium, applied for twice the amount of time the individual could have signed up for coverage but didn’t.
- For Part B the penalty is 10% for each 12 month period an indiviudal fails to sign up. It applies as long as the individul remains enrolled in Part B.
For Part D the penalty is 1% of the benchmark premium multipled by the number of uncovered months (though Part D or a creditable alternative). It also applies for as long as the individual remains covered through a Part D plan.
The long-term uninsured tend to be lower-income folks, so making coverage more expensive to them raises questions on affordability and whether it will spur them to seek coverage rather than become uninsured.
Prior to the ACA 70% of applicants to the non-group market were underwritten with standard rates, so unless there is an explicit late enrollment penalty, just allowing underwriting for the non-continuously insured may not result in price changes for the healthy. In fact, one could argue that insurers would be explicitly incentivized to not apply the late enrollment penalties to individuals they rate as healthy.
The fastest way to model this is to abstract a bit and think about the effectiveness of the continuous coverage provision. That is, for indiviudals who face loss of coverage in a given year, how effective would the provision be in steering them to another insurance plan? For this exercise we’ll allow this effectiveness to vary based on the duration of time folks would otherwise spend uninsured.
Coverage Transitions in the United States
To answer the above question we first need to understand the number and frequency of insurance transitions in a given year. For this, we’ll draw primarily upon the analytic strategy and data described in Graves and Nikpay (2017).
The following table provides the coverage transition matrix for a 12-month period defined from January 2013-December 2013. For example, the first row of this matrix tells us that 90.5% of employer-sponsored insurance (ESI) policyholders still retain that policy at the end of a year, while 6.8% have became uninsured for at least one month and 1.6% left that plan and became insured through a family member’s ESI plan.
| ESI-Own | ESI-Dependent | Non-Group | Public | Uninsured | |
|---|---|---|---|---|---|
| ESI-Own | 90.5 | 1.6 | 0.5 | 0.6 | 6.8 |
| ESI-Dependent | 5.0 | 89.2 | 0.2 | 0.5 | 5.0 |
| Non-Group | 10.2 | 4.6 | 80.0 | 0.4 | 4.8 |
| Public | 3.2 | 1.5 | 0.5 | 79.2 | 15.6 |
| Uninsured | 10.9 | 5.5 | 1.5 | 9.6 | 72.4 |
Validating the Transition Probability Matrix Against Cross Sectional Estimates
The transition probability matrix above provides estimates of transitions among each major coverage type over a 12-month period. However, we have only modeled the first transition observed among MEPS respondants; to the extent people experience multiple transitions over 12 months (e.g., ESI, uninsured for one month, back to ESI under a new job) we will not capture that here.
However, given the relatively short time horizon (12 months) we can test whether the modeled transitions do a decent job at estimating coverage at the end of 2013. The table below compares the cross sectional estimates of the number and pecent of adults 18-63 in each coverage category (first 4 columns) as well as the modeled estimate based on a Markov chain using January 2013 population estimates (\(p\)) and the transition probability matrix (\(T\)):
\[ p* = p'T \]
| Jan-13 (Observed) | Jan-13 (%) | Dec-13 (Observed) | Dec-13 (%) | Dec-13 (Modeled) | Dec-13 (%; Modeled) | |
|---|---|---|---|---|---|---|
| ESI-Own | 75,989,898 | 39.9 | 77,697,859 | 40.8 | 77,196,714 | 40.6 |
| ESI-Dependent | 39,337,389 | 20.7 | 39,402,594 | 20.7 | 39,545,558 | 20.8 |
| Non-Group | 4,536,967 | 2.4 | 5,022,880 | 2.6 | 4,929,280 | 2.6 |
| Public | 21,753,689 | 11.4 | 21,708,947 | 11.4 | 22,570,508 | 11.9 |
| Uninsured | 48,682,470 | 25.6 | 45,283,448 | 23.8 | 46,058,353 | 24.2 |
This appears reasonably close for now, though future work could incorporate multiple transitions and/or a calibration exercise to match the December 2013 numbers exactly.
With the transition probability matrix estimated, we can now turn to counterfactual exercises to see how effective certain policies could be.
Modeling a Continuous Coverage Mandate
Let’s turn now to thinking through how a continuous coverage mandate might affect these transition probabilities.
The first thing to think about is who would be most affected by the mandate. Under both the ACA’s individual mandate and the continuous coverage mandates being debated, there are grace periods of one or more months. That is, even if the affordability exemption does not apply, the mandate itself does not apply unless an indiviudal was uninsured for 3 or more months during the year. Similarly, continuous coverage provisions historically (e.g., COBRA, HIPPA) and in proposals under consideration also allow for a grace period of 30 or 60 days.
Let’s suppose that people are forward thinking (i.e., that they know how long they will be uninsured) and that a continuous coverage mandate will be least effective (say, 10% effective), among people who become uninsured for 1 months or less. And from there, we’ll allow the effectiveness to increase linearly to 75% among people who become uninsured for 12 months or more. We’ll relax these assumptions and allow non-linearity in the response a bit later on (e.g., to make the provision less effective among the longest-term uninsured).
To examine how these initial assumptions will affect coverage we first need an estimate of how long people are uninsured based on the type of coverage they are losing. We’ll turn to that estimation next.
How Long Are Coverage Gaps in the US?
We can draw upon the same MEPS data underlying the above analyses to isolate the “flow” sample of people becoming uninsured. We can then apply survey-weighted Kaplan-Meier methods to estimate the duration of uninsured spells, stratified by the type of coverage the individual had just prior to becoming uninsured.
| Lost ESI-PH | Lost ESI-DEP | Lost Non-Group | Lost Public | |
|---|---|---|---|---|
| After 3 Months | 65.0 | 69.9 | 85.1 | 86.8 |
| After 6 Months | 47.2 | 50.4 | 69.2 | 73.6 |
| After 12 Months | 21.6 | 22.4 | 27.4 | 49.4 |
The complement of this table provides our estimate of the percent of people who experience uninsured spells less than the given month duration:
| Lost ESI-PH | Lost ESI-DEP | Lost Non-Group | Lost Public | |
|---|---|---|---|---|
| After 3 Months | 35.0 | 30.1 | 14.9 | 13.2 |
| After 6 Months | 52.8 | 49.6 | 30.8 | 26.4 |
| After 12 Months | 78.4 | 77.6 | 72.6 | 50.6 |
Possible Impact of a Continuous Coverage Mandate
We can now integrate across the distribution of number of months uninsured to come up with an overall effectiveness average among people who lose each coverage type. We then apply this effectiveness estimate to each row of the transition probability matrix to “move” people from becoming uninsured to becoming non-group insured instead.
# Effectiveness Estimate
E## esiph esidep ng public
## 0.4212378 0.5754678 0.5989371 0.5730947Let’s now plug this into the transition probability matrix to see what the impact on coverage would be.
| ESI-Own | ESI-Dependent | Non-Group | Public | Uninsured | |
|---|---|---|---|---|---|
| ESI-Own | 90.5 | 1.6 | 3.4 | 0.6 | 4.0 |
| ESI-Dependent | 5.0 | 89.2 | 3.1 | 0.5 | 2.1 |
| Non-Group | 10.2 | 4.6 | 82.9 | 0.4 | 1.9 |
| Public | 3.2 | 1.5 | 9.5 | 79.2 | 6.7 |
| Uninsured | 10.9 | 5.5 | 1.5 | 9.6 | 72.4 |
| Without Auto-Enrollment | With Auto-Enrollment | Change | |
|---|---|---|---|
| ESI-Own | 77,196,714 | 77,196,714 | 0 |
| ESI-Dependent | 39,545,558 | 39,545,558 | 0 |
| Non-Group | 4,929,280 | 10,338,673 | 5,409,393 |
| Public | 22,570,508 | 22,570,508 | 0 |
| Uninsured | 46,058,353 | 40,648,960 | -5,409,393 |
So in summary, a continuous coverage mandate that is 10% effective among people uninsured for 1 month or less, ranging linearly up to 75% effective for people uninsured 12 months or more, would stem off about 5.7 million uninsured people per year.
Dealing with Parameter Uncertainty
Thus far we have considered a single case with effectiveness parameters ranging linearly from 10% to 75%. But these parameters, while plausible, are not estimated; they are assumed. It would be useful, then, to think the sensitivity of our finding when we allow these parameters to vary over a plausible range rather than take on explicit values.
We’ll perform this sensitivity analysis using a latin hypercube design. This design allows us to optimally select points within a grid defined by the two parameters above: \(\alpha\), the effectiveness of a continuous coverage mandate among the short-term uninsured; and \(\beta\), the effectiveness for the long-term uninsured.
The most conservative way to do this would be to allow both \(\alpha\) and \(\beta\) to range between 0 and 1; in that way we are not making strong behavioral assumptions about how people will respond to the mandate. However, there are still some assumptions (e.g., linearity in response) underlying this that we’ll deal with later on.
Here are the first few “rows” of the sensitivity analysis (based on 10,000 draws) showing the drawn value of \(\alpha\) and \(\beta\) as well as the estimted impact on the number of uninsured.
| alpha | beta | reduc.uninsured |
|---|---|---|
| 0.5703171 | 0.2234592 | -6.244798 |
| 0.8938484 | 0.5499404 | -8.267539 |
| 0.0514614 | 0.7300244 | -5.449172 |
| 0.7268191 | 0.1773572 | -6.738353 |
| 0.6595345 | 0.7380867 | -7.811704 |
| 0.1295987 | 0.8894751 | -6.129251 |
Density Distribution of Effects, Widest Possible Range (Mean Value at Dotted Red Line)
The results here show that the average value is -6.6 million with a range of -9.6 to -3.5. However, this is likely an implausible range because we have searched over the entire grid of effectiveness parameters as shown below (for a random sample of 1,000 points in the hypercube):
1,000 Parameter Samples in the Latin Hypercube Design
Sensitivity Analyses Over a Plausible Range
Next lets consider results when we allow \(\alpha\), the effectiveness of a continuous coverage mandate for the short-term uninsured to vary uniformly from 5-20%, and \(\beta\), the effectiveness parameter for the longer-term uninsured, to vary uniformly from 60-85%.
Here is what the sampled parameter space looks like:
100 Parameter Samples in the Latin Hypercube Design
Density Distribution of Effects, Widest Possible Range (Mean Value at Dotted Red Line)
In this scenario the average value is -5.7 million with a range of -6.3 to -5.2.
Non-Linear Responses
Given the considerations listed at the top, it is also plausible that the effectiveness of a continuous coverage mandate is not linear in the number of months uninsured. Suppose, instead, that the effectiveness is low for the short-term uninsured (since the bite of the mandate is small), then rises until we hit the median uninsured spell, and then declines after that (since the longer-term uninsured tend to be lower-income and are less likely to take-up more expensve coverage).
To model this we’ll turn to sensitivity analyses based on a non-linear interpolation function with 0-20% effectiveness at 1 month uninsured, rising to 50-100% effectiveness at 8 months uninsured, and then declining to 0-20% effectiveness at 22 months or more uninsured.
Here, for example, is the effectivneness curve (by number of months uninsured) for a model run with 10% effectiveness among the short term uninsured, 75% effectiveness at the median, then declining to 10% effectiveness at 22 months or more uninsured.
Non-Linear Response of Continuous Coverage Mandate
Density Distribution of Effects, Widest Possible Range (Mean Value at Dotted Red Line)
In this scenario the average value is -4.9 million with a range of -6.6 to -3.2. What is most notable is that the estimates seem to vary quite consistently between a reduction of 4 to 6 million uninsured.
How Large Could the Penalties Be and How Does that Compare to the Individual Mandate Penalties?
Let’s suppose the penalty is set at 1% of the premium value for a non-group plan for each month an individual remains uninsured. We’ll allow for a single month grace period, and as a benchmark plan use the median Bronze premium for a 40 year old in the ACA’s federally facilitated marketplaces 345.11.
We can use this estimated penalty information, along with the distribution of uninsured months estimated in the MEPS data above, to come up with an estimate of how large the late enrollment penalty people face might be.
For example, we know from the above analysis that 36% of adults who lose ESI coverage are insured again within 3 months. So for them, the maximum their penalty could be is 3% of the benchmark bronze premium, or 10.4. This would be the penalty that they would carry forward should they be assessed it after their spell without coverage.1
Using the pre-ACA distribution of uninsured spells here is the distribution, by type of coverage lost, of the “penalty” faced:
| Penalty | |
|---|---|
| ESI-Own | 247 |
| ESI-Dependent | 277 |
| Non-Group | 346 |
| Public | 368 |
Conclusions
So what can we conclude from the above exercises? It appears that under a variety of assumption, an effecfive continuous coverage mandate would, in a single year, reduce the number of uninsured by about 5 million.
This analysis only considered effectiveness in terms of a forward-looking person considering only the number of months she would be uninsured. Further analyses are needed to evaluate how effectiveness might vary based on the type of coverage lost. For example, the effectiveness of a continuous coverage provision might be less for people coming off public coverage than private.
References
Graves, John A., and Sayeh S. Nikpay. 2017. “The Changing Dynamics Of US Health Insurance And Implications For The Future Of The Affordable Care Act.” Health Affairs 36 (2): 297–305. doi:10.1377/hlthaff.2016.1165.
Most of these folks obtain ESI again after their short uninsured spell, and it’s not clear that the continuous coverage provision would apply a late enrollment penalty to any other plans other than non-group plans.↩