This study uses data made available by the COVID-19 Research Database, a cross-industry collaborative contributing real world, de-identified data to researchers wishing to study issues related to COVID-19. For this analysis we used a clearinghouse database which spanned 184 million claims from almost 30 million patients in 2019 and 94 million claims from nearly 20 million patients in the first 6 months in 2020. The claims were all from members who resided in a consistent set of 18 states, with roughly half living in California in both study years. Lastly, while the claims were associated with all payer types, they were predominately from the commercially insured: 86% commercial (ESI and MA), 12% Medicare, and 2% Medicaid, on average across both years.

Claims clearinghouses are responsible for scrubbing and transmitting medical claims to insurance carriers. Once a response is received from the issuer, the clearinghouse transmits the denial or acceptance of the claim back to the provider. While some clearinghouses may transmit payment information back to the provider, typically only charge data are reliably available from the clearinghouse. Additionally, accurate enrollment files may not be available since they are maintained by the issuer, which presents a limitation for this type of analysis.

We defined our services of interest from physician and other professional service charges (i.e. non-facility charges) as the count of unique claims from the charge submissions for the following HCPCS/CPT codes:

The table below shows the total count of claims used for this analysis by service and by year, from January 1 to October 2 in each year.

Percent Change for each service on a given day of the year is calculated as the difference between the 2020 and 2019 values divided by the 2019 value, then multiplied by 100. That is, Percent Change = (([2020]-[2019])/[2019])×100. 

We attempted to stratify the data by gender and age, however, demographic data for claims submitted in 2020 was significantly more incomplete than the demographic information on claims from 2019. While more specificity in the demographic profile of these services is important, the preventative screenings, tests, and procedures, selected (with the exception of colonoscopies and HIV tests), are utilized primality by a single sex, and immunizations are primarily utilized by children.

Trends were calculated using a 7-day rolling average, and we include a 60-day run out window from the date of data extract to allow for sufficient claims maturity. Note that inferences made comparing 2019 utilization to 2020 utilization may be influenced by factors not related to patients delaying or foregoing care. These factors include but are not limited to claims lag beyond our 60-day window, and the possible change in the mix of providers utilizing the clearinghouse from which these data are derived or a change in the mix of the types of patients being served.

We differentiate diagnostic services from screenings using their respective HCPCS/CPT codes. There may be limitations to our results related to how payers reimburse these services as some payers may require different or additional mechanisms for documenting diagnostic or screening services other than the HCPCS/CPT code.

We also do not discuss impacts of the pandemic on the utilization of telehealth services or other service substitutions, however, with the exception of HIV testing and pregnancy testing, none of these services are customarily performed without direct care from a health professional. While we can’t fully attribute the dips in preventative care to patient’s decisions to delay/forego care (e.g. many providers may have been closed during shutdown orders), this does represent a meaningful shift in the delivery of services.

Finally, the data from these findings are overly representative of the western U.S. and not generalizable to the entire population of insured persons.

The data, technology, and services used in the generation of these research findings were generously supplied pro bono by the COVID-19 Research Database partners, who are acknowledged at https://covid19researchdatabase.org/.

Note on adjusting for claim lags:

Research using claims data must contend with the fact that there is a lag between when a service is performed and when the provider submits a claim. This lag in claims maturity is highly variable, with some claims requiring a full year to fully mature.  The immediate needs for health services information during the pandemic must be balanced against this slow claims data submission process.  For this reason, we allow for a minimum sixty-day runout period for claims to be submitted before we analyze the data. Although traditional research in health services will use claims data with at least a year of maturity, the goal of this analysis is to provide data in real time, necessitating this trade-off.  However, since claims sometimes take longer than sixty days to be submitted, our data is still incomplete, especially towards the more recent service dates.

In order to get a better sense of what the data may look like after all the claims are completed, we attempted to make predictions for the additional number of claims that will be recorded, based on the number of claims that were added between 8/26/20 and 2/01/21 (the earliest and most recent dates on which we extracted the data, respectively).  Across services, the number of claims that appear in the second data pull but are missing in the first data pull consistently rose as the service date neared the date of the first data pull, and we observed that this trend more resembled exponential decay as opposed to a linear trend.

We built a simple regression model to predict the claims on any given day based on the time relative to the data pull. Because of the trend in claims relative to the date, we log-transformed the outcome variable.  We also allowed this trend to vary by service. To do so, we included an interaction term with the time relative to the data pull and service-specific dummy variables. The model can be written as:

ln(Y_it) = β₀ + β₁X₁ + β₂X_2,i + β₃X₁X_2,i + ε_it

Where:

Y_it  = Completion factor for service i on day t, i.e. the number of additional claims outstanding, expressed as a percentage of recorded claims
X₁  = The number of days between the service date t and the date the data is pulled
X_2,i  = Indicator variable for each specific service i

The model was fitted to our data from 8/26/20, with the additional claims that were added by 2/01/21 serving as the outcome variable. The model had an adjusted R² of 0.81, and will be continually tested and re-fitted with future rounds of data.  

In order to get a prediction of the total claims lag (i.e. not just the additional claims that show up after 159 more days of runout, but all the additional claims yet to be recorded), we added together the predicted percentage of additional claims for each day t with those for the day 159 days prior  (i.e. Y_t + Y_t-159 ).  For example, for the most recent service date, if a provider submitted a claim within 0-60 days, that claim is already recorded in the data. To adjust for the number of claims from that date that take longer than 60 days to be submitted, we added together the predicted number of claims with a 61-229 day lag and a 230-388 day lag, creating an estimate of the claim count after practically all the claims are recorded. Finally, we added the predicted number of additional claims to the recorded claims from the most recent round of data (extracted 2/01/2021) to create lag-adjusted claim counts and to calculate the percent change in service utilization from 2019 to 2020. The chart below shows how some of our adjusted numbers compare to the unadjusted ones.  An interactive version of this chart with all services can be found here.

 The table below shows a subset of our calculated completion factors (the predicted additional claims as a fraction of recorded claims) and adjusted claim counts for total vaccinations on selected dates.