Many researchers are interested in how cost sharing affects healthcare utilization, cost, and patient outcomes. This is especially true as high-deductible health plans (HDHPs) have become more common in the US. In 2023, 29% of covered workers in the US had an HDHP.
One type of data useful for analyzing HDHPs is claims data. However, there are challenges when using this data:
the first type [of claims data] It includes detailed information about the plan structure, but often has little externality validity, as it typically comes from a single health insurer or a small subset of enrollees. The second type has improved external validity by grouping all insurers together, but generally does not include the plan structure variables necessary to distinguish between HDHPs and lower deductible plans, nor interpret what the binary variables of the “HDHPs” represent. HDHP”.
an article from Cliff et al. (2024) aims to examine how well plan deductibles can be predicted using claims data using data from Optum Labs. Optum data has information on plan deductibles that is used as a “gold standard.” Four different imputation approaches are used:
- Parametric prediction with expenditure (expenditure regression method). The enrollee’s annual deductible spending is regressed on their total annual spending (plan plus out-of-pocket expenses), common demographic covariates (gender and age), and fixed effects for each plan. Use regression to predict deductibles conditional on a set spending level with plan fixed effects. . Using the coefficients from the best-fitting regression model, deductibles for each plan are predicted at a fixed amount of total spending (which the authors set at $10,000 to exceed most deductibles).
- Parametric prediction with imputation and plan characteristics (Regression method on imputed deductibles). This approach uses two stages. First, deductibles are entered for a subset of plans where they are easily identified. For the second stage, a set of covariates describing observed deductible spending and plan characteristics and collapse data are created from the individual level to the plan level. Using the subset of plans with an imputed deductible, the imputed deductible amounts are regressed on the set of covariates; The generated coefficients are then used to predict the deductibles of the plans that cannot be imputed in the first stage.
- Deductible modal expense (modal method). This simple method enters the highest (non-zero) modal deductible spending amount among enrollees in a plan and applies this deductible to all enrollees in that plan.
- 80th percentile of deductible expense (80th percentile method). Following Rabideau (2021), individual spending is tracked month over month, and people whose spending increases in a given month but deductible spending does not change are assumed to have met their deductible. The individual-level data are then reduced to the plan level and the deductible for all plan enrollees is set at the 80th percentile of the plan’s annual deductible expense.
To evaluate the accuracy of the imputation approaches, the authors calculate the sensitivity, specificity, and positive/negative predictive value (PPV/NPV) of each method for classifying enrollees into HDHP versus non-HDHP plans.
Surprisingly, the simple “modal method” performed best in terms of classifying individuals into HDHP versus not. It also performed well in terms of predicting deductible expense.
The mode method works best; 72% of plans are correctly classified in each category and 69% of plans have an imputed deductible within $250 of the actual deductible. For this method, limiting imputation to groups with more than 50 enrollees improved sensitivity to 85% of plans correctly classified by category and reduced the average difference between the imputed and actual deductible from $700 to $496.
You can read the full article. here.
