Recapping Exploratory Analysis Part 1

We started with the MLR as our North Star. MLR motivated several longitudinal analyses of procedures, providers, and consumers (pets).

• We should not use the historical claims for modeling the auto-adjudication events in 2023 and 2024.
  • In general, time interactions with procedures are nonexistent.
  • Based on the yearly changes in inter-arrival times and reviews, we can exclude 2021 and 2022 and focus only on 2023 and 2024.
  • 2023 and 2024 provide sufficient retrospective and prospective model improvements.
    • Enables triaging workflows for specific procedures or focus on regional markets where these diagnostics and interventions are performed the most.

Recapping Exploratory Analysis Part 2

• Lab Tests, X-rays, and Surgeries have longer inter-arrival times. The most expensive procedures—X-rays, labs, and surgeries—have the greatest variation in reimbursement and longer times for final claim submission. Mutatis mutundis for the lower-cost preventative procedures.
• Certain providers have higher adjudication acceptance or denial rates; others do not.
• This same behavior is observed in sub-segments of the pet population.

Now What?

• We use the incoming and adjudicated pre-authorization data itself.
  • The rationale for year-to-year model fitting and stratification is to account for the lack of time-dependent main effects in the 2023 and 2024 years.
• We focus on procedures and whether or not a reviewer was involved first.
• We then capture provider and pet-specific effects.
  • This allows us to quantify the amount of variation in the approval probability attributed to each of these market constituents.
  • We also capture the gap that is not being captured and is unobserved in the provided claims.

2023 - Unpacking the Base Model:

Takeaways:

• Not surprisingly, check-ups are estimated always to be approved.
  • The marginal effect for no clinical review is unreliable.
• Clinical Reviewer involvement does not moderate the probability of approval across procedures.
  • The two exceptions are dental cleaning and vaccinations.
• Compared to surgeries and X-rays, dental cleaning PAs have the lowest predicted probability when a clinician reviewer is involved.

Predictions for 2023 PAs that are Pending

• If I had more time, I would conduct an intrinsic evaluation of claims from a provider or pet from 2023.
  • We should ensure predicted probabilities of approval or rejection conform to historical claims of the respective patient and the provider.

2024 - Unpacking the Base Model:

Takeaways:

• As observed in 2023, check-ups are estimated to always be approved.
  • The marginal effect for no clinical review is unreliable.
• Observing the marginal effects:
  • Clinical Reviewer involvement is associated with an increase in the probability of approval in surgical procedures.
  • Clinical Reviewer involvement is associated with a decrease in the probability of approval in X-Ray diagnostics.
  • Clinical reviewer involvement does not moderate the probability of approval for Lab Tests and Dental Cleaning.

Predictions for 2024 PAs that are Pending

• If I had more time, I would conduct an intrinsic evaluation of claims from a provider or pet from 2024.
  • We should ensure predicted probabilities of approval or rejection conform to historical claims of the respective patient and the provider.

2023 Multi-level Model: Provider Random Intercept and Fixed Effects

• The ICC was 0.04, indicating that about 4% of the variance in approval decisions in 2023 is attributable to between-provider differences, while most of the variance occurs at the individual claim level.
• The random intercept variance for providers was 0.13, suggesting modest variation in baseline denial rates across providers.
• The model’s fixed effects explained 62% of the variance in outcomes, increasing slightly to 63.4% when accounting for provider-level clustering.

Visualize Provider Effects

Note the following:

• Positive values mean higher baseline log-odds of approval.

• Negative values mean lower baseline log-odds of approval.

• The provider for lower odds may be performing more surgeries and conducting diagnostics.

  • They may be working in an area with few providers who have the resources to conduct surgeries or diagnostics.
    • These very same providers dominate within the PAs that are denied.
  • The pets these providers service are prime targets for marketing and revised engagement to lower ED usage and expensive procedures.

Provider and Pet Random Effects

• The ICC was 0.18, indicating that 18% of the variance in authorization outcomes is attributable to differences between providers and pets.
• The variance of random intercepts for pet was 0.53, suggesting substantial variability in baseline denial/approval odds between individual pets.
• The variance for provider random intercepts was 0.17, reflecting modest variability between providers.
• The residual variance was 3.29, consistent with logistic mixed model assumptions.

Visualize Pet Random Effects

We should target pets with lower baseline odds of acceptance should be targeted for the programming revisions cited in my response to the business problems question.

Moreover, any risk adjustment modeling should focus on the subset of pets with higher denials—there may be chronic conditions or other environmental factors affecting their rejections. Or the providers may be upcoding these pets based on the intensity of care. These are all hypotheses that must be interrogated.

Visualize ICC breakdown

Here we present the Variance Partitioning (ICC) by Random Effect: