Intrastate Regional Cost Differences

Andrew Makarov, Ashley Lowenberg, Damon Raben, and Katie Leveen

Posted Date: December 2025

Key Insights

Regional differences in frequency, severity, and overall costs exist within NCCI’s three largest states by population: Florida, Illinois, and Texas.
Neither frequency nor severity were identified as consistent drivers of the loss ratio relativities across regions.
Differences in results were found for cohorts of policies that span single versus multiple regions within a state.

Introduction

NCCI collects workers compensation (WC) data from 38 jurisdictions and conducts numerous analyses at the state level, our mantra being that every state has a story.

Differences in the WC system manifest across jurisdictions for many reasons: population, employment, industry mix, injury mix, compensability and benefit structures, treatment guidelines, medical fee schedules, and attorney involvement, just to name a few. The volume of data collected also varies across jurisdictions, with larger jurisdictions typically collecting significantly more data than smaller ones. Prior NCCI research has provided insight on many of these differences between jurisdictions.

This report provides a collection of actuarial metrics to compare differences across geographic regions within NCCI’s three largest jurisdictions by population—Florida, Illinois, and Texas. These metrics provide insight into the extent of regional differences in frequency, severity, and overall costs across these jurisdictions. In addition to the report, a downloadable file is available for a deeper dive in the data. By examining these metrics, we expand on the individual story for each state.

Background

Stakeholder interest in understanding regional cost differences within NCCI jurisdictions prompted this research. A robust analysis of regional cost differences requires a considerable amount of data at the regional level, which varies significantly across NCCI jurisdictions. Florida, Illinois, and Texas are NCCI’s three largest states both by population and data volume and therefore serve as natural case studies for exploring intrastate regional cost differences.

Key data elements for this analysis—exposure and loss—are not reported for each individual business location and therefore were estimated by NCCI. For the estimation, NCCI relied on third-party data. While aggregate checks for reasonability were performed on the third-party data, verifying the data on a granular, address-by-address level was not feasible.

Despite these limitations, this research demonstrates the viability of evaluating intrastate regional cost differences in large NCCI states.

Data

The three primary data sources underlying this report are Policy Data, the Unit Statistical Plan, and the Medical Data Call. Exposure and losses are not reported for each individual business location in these data sources; therefore, estimation was necessary to allocate total policy experience across business locations.

We licensed a data set containing employee counts from a third party, Dun & Bradstreet (D&B), to facilitate the allocation of total policy exposure across business locations. We used ZIP Codes reported on paid medical transactions to identify the business location to which the loss amounts should be allocated. The details of these allocation steps are provided in the Appendix.

The volume of data included in the analysis is shown at the right, expressed as a percentage of the total standard premium reported to NCCI. While most policy experience was included in Florida, Illinois, and Texas, some experience was excluded. Policy exclusions were due to an inability to link across the relevant data sources, or other data reporting issues.

Actuarial Metrics

Each of the actuarial metrics that follow are based on data from Policy Years 2021–2023. All losses are incurred at first report and are limited to $500K.

In the loss ratio and frequency calculations, premium is derived by extending exposures by the loss costs in effect during the policy effective period. The experience rating modification (mod) for each risk is included in the premium calculation because the mod may already indirectly reflect some regional cost differences.

To control for class mix, the statewide components of the relativities in the following sections are first calculated at an individual class level and then weighted together using each region’s mix of classes. As a final step, the loss ratio, frequency, and severity relativities are balanced to unity.

Metrics are shown on a combined basis and broken out for two cohorts of policies:

Single-region policies—all business locations are within the same region; therefore geo-location is not necessary
Multi-region policies—business locations span at least two regions; therefore geo-location is necessary

Single- and multi-region policies have different characteristics that may impact the actuarial metrics (see the Appendix for examples). Therefore, analyzing actuarial metrics for each individual cohort—as well as combined—may enhance the regional comparisons.

Loss Ratio Relativities

The graph below shows a point estimate for the loss ratio relativity of each region on top of a range produced by alternative exposure allocation methodologies (as described in the Appendix). Ranges that lie wholly above (or below) unity may signal confidence in the estimated directional impact for each region, while those that cross unity may signal uncertainty. Ranges that are wide imply that regional differences are sensitive to the exposure allocation procedure.

Single- Versus Multi-Region Policies

The following graph enables a comparison of the regional loss ratio relativities for all included policies to those of the single- or multi-region cohorts.

Frequency and Severity Relativities

The graph below retains the point estimates of regional loss ratio relativities from the previous section and adds point estimates of regional relativities for frequency, indemnity severity, and medical severity.

Comparing the component relativities to unity shows whether frequency (or severity) in the region is higher than the class-mix-adjusted statewide average. In some regions, the component relativities are all above (or below) unity and therefore are influencing the loss ratio relativity in the same direction. In other regions, however, the component relativities fall on opposite sides of unity, suggesting that offsetting effects may exist.

Comparing the component relativities to the loss ratio relativity of the region shows whether frequency, severity, or both are the predominant driver of the regional loss ratio relativity. No component relativity was identified as a consistent driver of the loss ratio relativities across regions.

Single- Versus Multi-Region Policies

The graph below displays the regional frequency and severity relativities for the single- and multi-region policy cohorts.

Claim Ratio Deviations

The metrics in this section are provided as supplementary information. These descriptive statistics may or may not be material drivers of the regional loss ratio relativities. The metrics displayed are:

The share of total claims that are lost-time (LT)
The share of LT claims that remain open
The share of LT claims with a permanent injury type
The rate of attorney involvement on LT claims

When graphing these metrics, regional deviations (additive effects) are shown instead of the previously displayed regional relativities (multiplicative effects). The regional deviations are from the statewide averages adjusted for class mix. The overall statewide averages with no mix adjustment are displayed in the legend for each metric.

Conclusion

The regional differences in frequency, severity, and overall costs expand the individual state story for Florida, Illinois, and Texas. Looking forward, NCCI is considering periodic updates to this analysis. As future data collection expands and evolves, NCCI may revisit this analysis to reduce reliance on estimation and explore opportunities to research additional states.

Appendix

The methodology to create a geo-located dataset involves three major steps:

Exposure allocation—third-party data is needed to supplement NCCI data
Loss allocation—the location of the claim is estimated using NCCI’s Medical Data Call
Defining intrastate regions—regions should be large enough to be credible, while still reflecting real cost differences

A Simple Example—Click to Expand

Consider a business operating at three distinct addresses in Florida, as reported in NCCI’s Policy Data. Each location falls into a major city that may differ in wage levels, work environments, workforce composition, access to care, and many more relevant characteristics. Conducting a regional analysis could add insight into how these differing characteristics may be impacting key actuarial metrics.

NCCI’s Policy Data includes the addresses at which the business operates, but not the exposure associated with each address. To estimate the exposure at each address, a third-party dataset was licensed from D&B. This dataset contains employee counts by location, which can be linked to NCCI data and used to allocate total policy exposure among addresses. NCCI’s Policy Data is further linked to NCCI’s Unit Data so that audited exposures and losses can be used.

Because payroll by business location is estimated using employee counts, regional differences in wage levels are not reflected in the data. Instead, regional wage adjustments calculated from US Bureau of Labor Statistics (BLS) wage data are required.

Claims must also be assigned to a business location. This is accomplished with NCCI’s Medical Data Call, which includes a three-digit ZIP Code on each paid medical transaction.

In this example, the claim has three transactions all within the same three-digit ZIP Code associated with Fort Lauderdale. Therefore, the claim is assigned to the region containing Fort Lauderdale.

Regional boundaries must also be defined. Once defined, all losses and claims within each region are aggregated and actuarial metrics can be derived. Finally, regional relativities to statewide are computed and balanced such that the weighted average is unity.

While not relevant to this simple example, the impacts of experience rating and class mix are also considered when allocating exposure to regions.

Exposure Allocation

To estimate the amount of Unit Data exposure that should be allocated to each business address reported on Policy Data, the following steps are taken:

Compile a dataset of Unit Data exposure with Policy Data details (e.g., addresses, names, FEIN, and telephone number).
Use reported Policy Data employee counts when reasonable. For the remaining policies, link D&B business details to NCCI data to estimate employee counts at each address.
Account for regional wage differences using wage relativities derived from employee counts and wage data from BLS’s Occupational Employment and Wage Statistics.

\[ \text{wageRel}_{\text{reg}} = \frac{\sum_{occ} \text{avgWage}_{\text{reg},occ} / \text{avgWage}_{\text{sw},occ} \cdot \text{emplCnt}_{\text{reg},occ}}{\sum_{occ} \text{emplCnt}_{\text{reg},occ} } \]

Allocate total policy exposure using the wage-adjusted employee count share of each address.

\[ \text{exposure}_{\text{addr}} = \text{exposure}_{\text{pol}} \cdot \frac{\text{emplCnt}_{\text{addr}} \cdot {\text{wageRel}_{\text{reg}}}}{\sum_{addr} \text{emplCnt}_{\text{addr}} \cdot {\text{wageRel}_{\text{reg}}}} \]

Reintroduce unmatched single-region policies.

Exposure Allocation Alternatives

The steps above provide an approximation of exposure at each address. Even though the steps are consistent from policy to policy, the steps still depend on a number of choices and assumptions, any of which could potentially impact the results. We therefore performed sensitivity testing, which showed that the results are generally stable, as shown by the ranges in the loss ratio relativity graphs. One reason is that the geo-location steps only affect policies that span multiple regions, while single-region policies represent the majority of standard premium in our dataset as shown in the chart to the right.

Four of the five alternative exposure allocation procedures analyzed during sensitivity testing impact step 2 of the exposure allocation. In this step, a link is sought between the addresses reported to NCCI in Policy Data and the addresses provided in D&B’s employee count data. One alternative procedure used D&B employee counts in lieu of reported Policy Data employee counts for all addresses. The remaining exposure allocation procedures varied on how to geo-locate unmatched addresses, meaning addresses for which a link to D&B data was not found. Options include imputing the minimum or median employee count across all matching addresses or excluding the unmatched addresses and reallocating the exposure to the remaining addresses.

Displayed to the right is the percentage of unmatched multi-region addresses for each state. All addresses on a policy are impacted when it has any unmatched addresses due to the allocation approach. Therefore, the share of multi-region standard premium with one or more unmatched address is also displayed. Unmatched addresses impacted a majority of multi-region premium in all three states.

State	Multi-Region Premium %	Unmatched Address %	Estimated Premium % (Total Policy)
Florida	29.2%	24.9%	55.0%
Illinois	20.8%	30.5%	73.3%
Texas	35.2%	31.7%	71.7%

The final alternative exposure allocation procedure analyzed during sensitivity testing adds a class mix adjustment to step 4. In this alternative, all exposure reported for Class Code 8810 is assigned to the headquarters address of the business in cases where the D&B data identifies that address. Otherwise, an equal class mix at all locations within a policy is assumed.

To validate the exposure allocation process in general, and to select a particular exposure allocation procedure, the Quarterly Census of Employment and Wages (QCEW) is used. The QCEW is an external data source that provides total wage figures by county. Regional shares of exposure are calculated for each allocation methodology and compared to the regional shares from the QCEW.

The methodology that uses NCCI employee counts when available and excludes unmatched addresses offers the closest match to the QCEW wage distribution, estimated regional exposure shares generally align with QCEW across each study state and all regions. The headquarters adjustment for exposure reported for Class Code 8810 did not, in general, have a material impact on the actuarial metrics.

Loss Allocation

To align exposures and losses by region, each claim reported to NCCI via Unit Data is assigned to a business location. To estimate the address to which each claim should be assigned, the following steps are taken:

Link each Unit Data claim to the associated medical transactions reported on the Medical Data Call, when possible.
Geo-locate each claim using the collection of three-digit ZIP Codes reported across the medical transactions, excluding medical service categories that are unlikely to be local (e.g., telehealth).
Assign each geo-located claim to the nearest reported business location.

Following this procedure, some claims remain unassigned to a region due to data reporting issues (e.g., invalid ZIP Codes or no link between Unit Data and the MDC). These claims are assigned to regions randomly, in proportion to the policy premium allocated to each region. As an additional rule, the total experience for any policy with an over-reliance on this loss allocation method was excluded.

The single-address policy cohort was analyzed to calibrate the geo-location procedure for claims. In particular, research was conducted to determine which medical transactions provide the most insight into estimating the matching exposure location. For example, ZIP Codes reported for medical treatments provided on the same day of the accident are given more weight than ZIP Codes reported for subsequent medical treatments. This cohort was also used to identify and exclude medical transaction locations that are likely to be billing addresses rather than medical provider locations.

Region Definition

The regions used in this analysis are a variation of state government-defined economic development regions.

Economic development regions are designed with economic cohesion in mind. Economic cohesion may identify regions with different labor markets, thereby impacting wages and unemployment rates that could influence both exposures and claimant behavior. Different economic environments may also result in different exposures to risk even within the same class code. For example, variation in regional traffic density or weather patterns may lead to regional differences in claims experience.

The economic development regions group counties together. Using counties (rather than ZIP Codes) as the most granular level of aggregation helps to mitigate potential concerns of bias. Guidance teaches that “[w]hen territory boundaries are chosen based on political boundaries such as at the city or county level, territory residents are more likely to be racially diverse. Finer boundaries such as ZIP Code or latitude/longitude produce territories that are more likely to be dominated by one racial group.”¹

¹ Understanding Potential Influences of Racial Bias on P&C Insurance, Members of the 2021 CAS Race and Insurance Research Task Force

Rural economic development regions were further consolidated to increase their data volume, while keeping regions with major population centers separate. Using larger regions also increases the share of single-region policies, which in turn reduces the reliance on the exposure and loss allocation procedures. As the share of premium from single-region policies increases within each region, there are fewer instances of three-digit ZIP Codes spanning multiple regions. Consequently, workers are more likely to have sought medical treatments in the same region as the address of their associated exposure.

To convey the data volume of each region, lost-time claim counts from Policy Years 2021–2023 are displayed in the maps below.

Lost-Time Claim Counts
Policy Years 2021–2023

Lost-Time Claim Counts
Policy Years 2021–2023

Lost-Time Claim Counts
Policy Years 2021–2023

Defining Actuarial Metrics

The loss ratio for class $c$ is defined as

\[ \text{lr}_c = \frac{\text{indloss}_c + \text{medloss}_c}{\text{manual premium}_c} = \frac{\text{L}_c}{\text{P}_c} \text{.} \]

The loss ratio for a region (or state) is defined as the premium-weighted average of the class loss ratios

\[ \text{lr}_{reg} = \frac{\sum_c L_c}{\sum_c P_c} = \frac{\sum_c P_c \cdot \text{lr}_c}{\sum_c P_c} \text{,} \]

where the sums are taken over all classes in the region (or state). Regional premium weights are used for both the region and state loss ratios.

At this stage, the regional loss ratio is adjusted for the impact of experience rating. This is accomplished by adjusting the premium denominator of each regional loss ratio to reflect the premium-weighted average experience modification factor of the region, $Emod_{reg}$, via

\[ \text{adj_lr}_{reg} = \frac{\text{lr}_{reg}}{\text{Emod}_{reg}} \text{.} \]

Once each regional loss ratio has been adjusted for the impact of experience rating, the loss ratios can be normalized to ensure that the statewide weighted average loss ratio is unity, via

\[ \begin{align} \text{norm_adj_lr}_{reg} &= \frac{\text{adj_lr}_{reg}}{\text{adj_lr}_{state}} \text{, where} \\ \text{adj_lr}_{state} &= \frac{\sum_{reg} \text{P}_{reg} \cdot \text{Emod}_{reg} \cdot \text{adj_lr}_{reg}}{\sum_{reg} \text{P}_{reg} \cdot \text{Emod}_{reg}} \text{.} \end{align} \]

Regional frequency relativities are calculated in a similar manner, beginning with the frequency for class $c$ defined as

\[ \text{freq}_c = \frac{\text{claims}_c}{\text{manual premium}_c} = \frac{\text{N}_c}{\text{P}_c} \text{.} \]

As with loss ratio relativities, premium is adjusted for the impact of experience rating, and the regional frequency relativities are normalized to a statewide average of unity.

Regional severity relativities are calculated in a similar manner; however, as the denominator of severity is claim counts, we switch from premium weights to claim count weights when deriving regional averages from class severities.

Single- Versus Multi-Region Policies

In the graph below, several policy characteristics are displayed for the single- and multi-region cohorts:

Policy Size—Multi-region policies are typically larger than single-region policies, as measured by standard premium. Job responsibilities of individual workers may vary across businesses of differing sizes, even within the same classification. This may result in differences in the actuarial metrics between the two cohorts.
Industry Mix—The industrial composition of single- and multi-region policies is different. The relativity calculations control for class mix by estimating the statewide denominator using the class mix observed within each policy cohort. This ensures the same industrial composition of the numerator and denominator of the relativities applicable to each cohort. However, the industrial compositions of the single- and multi-region cohorts are different, even within the same region.
Average E-Mod—The average experience rating mod of a region is calculated by weighting the mod applicable to each policy by its manual premium. Non-experience-rated risks are included with a mod of unity. About a third of single-region policies meet the premium eligibility thresholds for experience rating, compared to about two-thirds of multi-region policies. Multi-region policies are more likely to span multiple states, meaning the mods often reflect the experience of multiple regions and multiple states.