2.1 Policy context

The workers' compensation program aims to deliver necessary medical care and replace lost earnings when workers are injured on the job, without the uncertainty, delay, and expense of litigation. A typical workers' compensation claim begins when a worker is injured on the job and notifies his or her employer or insurer about the injury. The insurer then investigates the claim and, if the claim is accepted, pays for medical care and indemnity benefits for lost income while workers are recovering from their injuries and are unable to work.11 Often, workers fully recover from their injuries and return to work. In some cases, workers may have permanently disabling injuries, in which case insurers provide PPD or, in a small share of cases, permanent total disability (PTD) payments.

However, disputes can arise, and when these are perceived by workers as challenging their claim or benefits, workers may seek help from an attorney. For example, Victor and Savych (2010) found that workers who reported that their claim was initially denied (even though it was later paid) were more likely to seek help from an attorney, as were workers who reported that their supervisor did not think the injury was legitimate. Moreover, even in the absence of disputes, workers may seek help from an attorney to advocate on their behalf, navigate the complexities of workers' compensation cases, and inform them about the benefits to which workers may be entitled.

2.2 Descriptive information

The preceding discussion helps explain why simple associations between indemnity benefits and attorney involvement can be misleading as to the causal effect of hiring an attorney, because attorneys get involved in different types of cases having to do with the extent of disputes, complexity, or other features of the injury or claim. Descriptive evidence indicates that, in fact, attorney involvement on behalf of workers varies across different types of workers' compensation claims.12

As shown in Table 1, 34 percent of workers who had more than 7 days of lost time after a work-related injury had an attorney represent them during the claim process (within 36 months of maturity).13 Attorney involvement rates differed widely across different types of claims. In the majority of cases, where workers received only temporary disability benefits and did not have any PPD/LS payments, 14 percent of workers were represented by an attorney.14 In contrast, 64 percent of claims that ended up with payments for PPD benefits and/or LS settlements were represented by an attorney. This is expected since claims with PPD/LS payments are open longer and are more likely to develop disagreements about different aspects of the claim. Lump-sum settlements are used to close out future liability on workers' compensation claims and are often applied for claims that have permanent disability.15 Unadjusted summary statistics make clear that claims with attorney involvement have higher indemnity benefits (Table 2). Median payments are lower than average payments, reflecting a long right tail of the distribution. The average indemnity benefit on claims that we examine was around $19,000, although the amount varied across different types of claims—with higher total indemnity benefits for claims with PPD/LS payments.

2.3 Injury characteristics associated with attorney involvement

Table 3 highlights characteristics of injuries and workers related to attorney involvement that are potentially predictive of both the likelihood of disputes and the resulting indemnity benefits. Attorneys are more likely to be involved in claims with neurologic spine pain and in claims with inpatient stays. Attorney involvement typically decreases with workers' tenure at the employer, and it varies across industries, with the highest rates of attorney involvement among workers in construction. Our analyses control for measures describing injury and worker characteristics that we capture in the data. However, we are not able to capture all relevant worker and injury characteristics (including the behavior of insurance companies). This is the main challenge that we address with our IV approach.

3.1 Attorney involvement

Our treatment of interest is whether workers were represented by an attorney at any time during the claim process. This measure is collected when the insurer receives a notice of representation or notice of appearance informing the insurer that the claimant is represented by an attorney.

3.2 Indemnity benefits

Our outcome of interest is the total indemnity benefits that workers receive within 36 months after an injury. These include payments for temporary disability benefits (either temporary total or temporary partial disability payments), permanent disability benefits (either PPD or PTD payments), and settlements that workers may have received.19 By focusing on total indemnity benefits, we avoid concerns that the definition and determination of different components of indemnity benefits vary across states.20

3.3 Other controls

We control for a rich set of covariates that could affect the indemnity benefits that workers receive after their injuries. For example, indemnity benefits increase with age (Savych & Ruser, 2019) and differ across occupations and industries, which vary regarding workers' ability to go back to work (Galizzi & Boden, 1996). We control for the following worker characteristics in our regressions: age, gender, marital status, tenure at the time of injury, and preinjury weekly wages. We also control for the following workplace characteristics: firms' payroll size and indicator variables for industry/occupation.21

Since indemnity costs vary across injuries (Table 3), we include dummy variables for injury type, as follows: fractures; lacerations and contusions; neurologic spine pain (e.g., discs, peripheral neuropathy); spine (back and neck) sprains, strains, and nonspecific pain; inflammations; other (non-back) sprains and strains; upper extremity neurologic pain (carpal tunnel); and a residual category of other injuries.22 We also control for whether the worker had an inpatient stay during the recovery period.

Local labor market characteristics may contribute to differences in the indemnity benefits. We control for the county unemployment rate and whether workers lived in a rural area. We also control for three other population characteristics: the county-level percentage of residents who were disabled, the percentage of working age adults without health insurance, and the state-industry specific share of workers who were members of a union.23 These variables may capture important characteristics of local areas and other factors that are correlated with workers' ability to go back to work and, therefore, with indemnity benefits.

Since determinants of indemnity benefits vary across states, we also include fixed state effects. These account for differences across states in time-invariant workers' compensation policies and other factors that are constant over time. For example, across most states, the policies that determine the generosity of various components of indemnity benefits did not change over the sample period; these include approaches for determining PPD benefits, as well as limits on the duration of temporary disability and PPD benefits.24 With state fixed effects included, we compare indemnity benefits within each state—that is, for cases where the same rules and procedures for determining indemnity benefits and resolving disputes apply. We also include fixed year effects to capture factors that are changing across all states at the same time, such as federal rules and regulations.

4.1 Baseline specification

Y_ijst is indemnity benefit payments for worker i, in area j (explained below) of state s, in year t. ATTORNEY_ijst is the indicator variable capturing whether a worker was represented by an attorney. X_ijst is a vector of the control variables discussed above. YR_t is a vector of year fixed effects, and ST_s is a vector of state fixed effects. ε_ijst is an error term.25

As we discussed above, ordinary least squares (OLS) estimation of Equation (1) does not account for the endogenous or selective nature of attorney involvement. It is true that we control for some factors that can affect both indemnity benefits and whether a worker has legal representation, which can help isolate the association of attorney involvement with indemnity benefits for cases that are similar with regard to the characteristics we can measure in our data. However, these associations may not reflect the causal impact of attorney involvement because they may also reflect unobserved factors related to both attorney involvement and indemnity benefits.

Several kinds of unmeasured characteristics of workers/injuries that may not be captured in our controls could lead to biased estimates of Equation (1). Our controls do not capture severity variation, but injury severity may directly affect indemnity benefits and also prompt a worker to get legal representation. In this example, we would expect the OLS estimate of β in Equation (1) to be biased upward. Workers may also hire attorneys when they feel that they are getting inadequate medical care, are being forced back to work too quickly, or are unsatisfied with settlement offers. In such cases, the bias could go in either direction. On the one hand, such cases are likely to be associated with more severe injuries entailing higher indemnity benefits, leading to upward bias. On the other hand, if the indemnity benefits still end up being lower in these cases—because attorneys do not fully offset, for example, workers having to return to work more quickly—the bias could be negative.

Workers may also engage attorneys to avoid the burden of the workers' compensation system. We do not know how this unmeasured variation in worker preferences to hire an attorney is related to indemnity benefits, so it is hard to know the direction of the bias. If, for example, workers who are less confident about dealing directly with the workers' compensation system (and therefore hire an attorney) tend to have lower earnings potential (and therefore lower indemnity benefits), then the OLS estimate can understate the effect of attorney involvement on the amount of indemnity benefits. Alternatively, high-income workers who are confident about dealing with the system may still hire an attorney, in which case the bias would be in the other direction—overstating the effect of an attorney on indemnity benefits (because indemnity benefits are related to preinjury income).

4.2 Instrumental variables approach

The variables ATTYLOCAL_ijst and DELAY_ijst are the two IVs that we use. This regression includes the same controls as in Equation (1), represented by X_ijst. The absence of direct effects of the IVs on indemnity benefits is the key assumption underlying this strategy. We present the justification for this assumption, for each of the IVs, below.

4.2.1 Local area variation in attorney involvement

One IV is the local area attorney involvement pattern (ATTYLOCAL in Equation (2)). For each individual observation, we determine the local area attorney involvement rate based on claims for fractures, lacerations, and contusions without PPD/LS payments in an area (indexed by j).26 The attorney involvement rate for injured worker i in area j in year t is based on all workers in area j with fractures, lacerations, or contusions, excluding individual i (we denote the number of observations in area j [in state s and year t] as N_jst). Defining ATTORNEY_ijst as equal to one when the worker is represented by an attorney, and zero otherwise, the formula for ATTYLOCAL_ijst is

$$${\text{ATTYLOCAL}}_{{\text{ijst}}}=\frac{{\sum }_{{i}^{^{\prime} }\ne i}({\text{ATTORNEY}}_{{i}^{{^{\prime} }}{\text{jst}}})}{{N}_{{\text{jst}}}-1}.$$$(3)

Excluding individual i avoids creating a mechanical correlation between the instrument and attorney involvement in an individual case—variation that would not be purged of unmeasured injury severity or expected indemnity payments.

We use non-PPD/LS claims for fractures, lacerations, and contusions for several reasons. First, attorney involvement for these claims is more likely to be driven by the supply of attorneys and workers' preferences toward hiring attorneys in certain locales, rather than by state system features that lead to high rates of attorney involvement, but also influence benefits. It would not be informative to derive this measure from a subset of claims that always attract attorneys; for example, Boden and Victor (1994) showed that nearly all of the low back claims with PPD/LS payments had an attorney in some states. Second, focusing on a subset of injuries avoids concerns about variation in local attorney involvement rates that are driven by injury composition. Third, these injuries are common across all workers' compensation jurisdictions, reducing concerns about the impact of industry and occupation mix in determining local area attorney rates.

We define local areas using commuting zones (CZs).27 CZs aggregate counties to better define local economies. The boundaries of these areas are drawn to minimize commuting flows across CZs and maximize them within CZs (Tolbert & Sizer, 1996). For CZs that span multiple states, we treat each part of the CZ located in a different state as a separate area. In creating our instrument, we only use observations that are in areas with at least 15 observations within each CZ-and-year combination to increase precision of the local pattern estimate.

Attorney involvement rates at the CZ level vary within states. Figure 1 provides information on the distribution of the attorney involvement rates (among fractures, lacerations, and contusions without PPD/LS) for our analysis sample. The figure shows the minimum, 25th percentile, 75th percentile, and maximum across CZs for each state in our sample. We observe variation in the rates of attorney involvement both across and within states; given that we include state fixed effects in our model, the within-state variation is critical to our analysis.28

5.1 OLS estimates

First, we review OLS estimates (Table 6). These estimates reveal the association between attorney involvement and indemnity benefits while adjusting for injury, worker, workplace, and location characteristics. We show estimates using as the dependent variable either the level or the log of benefits. In the first case, we find that claims with an attorney have $30,500 more in payments than claims without an attorney (similar to the unadjusted difference of $33,000 shown in Table 2). In the second case, the estimated impact of 197.4 in logs similarly implies a large increase in indemnity benefits paid when workers are represented by an attorney; the log reduces the impact of very large values of the outcome.33

Using either the level or log specification, the estimates in Table 6 point to a strong positive relationship between attorney involvement and indemnity benefits. However, to obtain estimates of the causal effect of attorney involvement, we turn to our IV approach.

5.2 First-stage estimates

We first present estimates of our first-stage regressions (Equation (2)) showing relationships between whether an individual worker was represented by an attorney, on the one hand, and local attorney involvement patterns and delays in payments (our IVs), on the other, in Table 7. The results indicate that each IV separately and both IVs together have high predictive power for attorney involvement, meeting the first requirement for IVs to be valid. The first model, using only the local pattern to predict attorney involvement, implies that a 10 percentage point increase in the baseline local area attorney involvement rates is associated with a 3.2 percentage point higher likelihood that injured workers have an attorney. The F-statistic on the first-stage estimate is very large (69). The second estimate is for the delay in payments IV. The estimate implies that the likelihood of being represented by an attorney is 18 percentage points higher for claims that had more than 30 days from the time of disability to the first indemnity payment. The F-statistic on the first-stage regression is even larger (1289). Finally, the estimates at the bottom of the table include both IVs; each one of them continues to have a large impact in predicting variation in attorney involvement, and the joint F-statistic is very high (643).

5.3 IV estimates

Table 8 presents the IV estimates of the effect of attorney involvement. We show estimated effects from three specifications: using only local variation in attorney involvement as an IV, using delay in payments as an IV, and using both IVs. The key result is that all the estimates indicate a positive and statistically significant effect of attorney involvement on indemnity benefits. Comparisons with Table 6 indicate that the IV estimates are smaller than the OLS estimates, which is consistent with positive bias in the OLS estimates. A positive bias means that attorney involvement is positively related to unmeasured factors that are associated with higher indemnity benefits.

The magnitude of these estimates in levels, depending on the IVs used, ranges from $7700 to $12,400. Estimates using only the local area variation in attorney involvement suggest that attorneys increase indemnity benefits by $12,400. We view the smaller estimates associated with using the delay IV as likely to be more reliable, in part because the IV is a stronger predictor of attorney involvement, as reflected in the much more precise estimates when using this IV (i.e., the much smaller standard errors)—whether or not combined with the local patterns IV.34 The estimates for log indemnity benefits also suggest that attorney involvement increases indemnity benefits. The estimated coefficient is 169–180, depending on the specification. Again, these are smaller than the OLS estimate. The IV log estimates imply that attorney involvement boosts indemnity benefits four- to fivefold. While large, these estimates are not broadly out of line with levels estimates of a $7700 effect; the median of indemnity benefits is $5500 (Table 1), so a $7700 increase represents more than double the indemnity benefits. The log specification puts more weight on smaller claims, indicating higher percentage gains in benefits among smaller claims.

Regardless, when we compare workers with and without an attorney using only the variation in attorney involvement that is predicted by different local patterns across parts of the state and/or by initial delays in payments, we find considerably smaller effects of an attorney on the amount of benefits than what is implied by an OLS regression. Nonetheless, the IV estimates imply large positive effects of total indemnity benefits from legal representation of workers. Moreover, the log specification is easier to interpret when asking whether the estimates imply that workers who hire an attorney are better off once we account for fees paid to the attorney. Our understanding is that attorneys are usually paid a given percentage of the indemnity benefits received—typically 20 to 33 percent.35 Our estimates of effects of well over 200 percent suggest that workers benefit substantially from hiring an attorney.36

Next, we report IV estimates for four separate groups of injuries (using both IVs): (1) fractures, lacerations, and contusions; (2) low back pain cases; (3) inflammations and non-back sprains and strains; and (4) other injuries, not included in the other groups. We do this to check that our results are not affected by potential issues with the delay in payments IV associated with it being correlated with compensability. In particular, similar results for fractures, lacerations, and contusions, where compensability is not at issue, would bolster our findings. The IV estimates in Table 9 indicate that attorneys have a strong effect on indemnity benefits across all groups of injuries. Attorneys increase indemnity benefits by $4900 among workers with fractures, lacerations, and contusions. The estimated effect is larger among workers with other injuries—a $7800 increase for low back pain cases, and an $8300 increase for inflammations and non-back sprains and strains—but average indemnity benefits are larger for these injuries (Table 1). The estimated coefficients for log indemnity benefits range from 147 to 210 across different injury groups, implying three- to sevenfold increases in indemnity payments due to attorney involvement.

5.4 Threats to identification

One possible concern about the IV based on local area variation in attorney involvement is that differences across areas may reflect other important within-state factors that affect both attorney involvement and benefits. For instance, areas with high attorney involvement may be dominated by one or two large employers or workers' compensation insurance companies that may be particularly aggressive in challenging claims. This would create a greater demand for attorneys, causing them to locate in the area and making it easy for workers to find an attorney. At the same time, more aggressive challenging of claims would mean longer time to resolve the claims, which could lead to either higher or lower indemnity benefits depending on changes in claim composition and approaches for resolving claims. In this case, the omitted employer behavior would violate the exclusion restriction for valid IV estimation. We address this concern in two ways. First, we estimate the local area variation in attorney involvement while leaving out the worker's employer in the year when the worker has a workers' compensation claim. Second, we estimate local area variation in attorney involvement ignoring claims in that year from the insurer who is handling a specific claim. These are two related ways of ensuring that the local area variation in attorney involvement does not reflect the demand for attorneys driven by employer and insurer practices but rather reflects the supply of attorneys across different areas.

The estimates in Table 10 show that our conclusions about the effects of an attorney on indemnity benefits do not change when we use these two alternative approaches for measuring local area variation in attorney involvement, although the estimates that use only the local attorney pattern IV are somewhat higher when we measure local area variation excluding either the worker's employer or the insurer, but much less so when using the two alternative approaches for measuring the local attorney involvement IV and also using the delay IV. The estimates in logs display similar patterns.

Another concern about the local attorney involvement rate is that insurers may change their claim-handling behavior in areas where attorney involvement is higher, because that higher involvement acts as an implicit threat of legal conflict. Thus, to try to preclude costly legal battles, insurers may offer a more generous initial settlement for all claims when they perceive a higher threat of attorney involvement in an area. This may create a direct relationship between local attorney involvement rates in an area and workers' indemnity payments, violating the exclusion restriction. In the appendix, we provide evidence against this concern based on overidentification tests and the robustness of the results to using only the delay IV and not excluding the local attorney involvement measure from the second-stage regression (Equation (1)).

A potential concern about the delay in payments IV is that it may reflect employer or insurer approaches toward different types of claims, which may, in turn, be related to indemnity benefits (again, leading to violation of the exclusion restriction). We address this concern by estimating specifications that control for insurer and employer fixed effects.37 The estimates in Table 11 show that the estimates are not sensitive to including insurer or employer fixed effects, whether using levels or logs. These specification checks highlight the robustness of our results to multiple potential concerns about the IVs.

5.5 Attorney effects on the distribution of indemnity benefits

Next, we examine how attorney involvement shifts the distribution of indemnity benefits conditional on claim characteristics. Table 12 shows estimates from quantile regressions and IV quantile regressions for the effects of attorney involvement across deciles of indemnity benefits. In every single case in Table 12, the estimates from the IV quantile regressions are smaller than estimates from the quantile regressions, which is consistent with positive bias because attorney involvement is positively related to unmeasured factors that shift the distribution of indemnity payments—exactly paralleling the earlier results.

These quantile regression estimates provide information on whether attorney involvement has different effects in different parts of the distribution of indemnity benefits. For instance, perhaps attorney involvement does not matter much for smaller cases but has a larger impact at the upper tails of the distribution, where the claims are more expensive (and attorney payments larger). These potentially contrasting effects are apparent from the quantile regressions estimated for higher and lower percentiles of the distribution. Estimates from regressions in levels indicate an increasing effect of an attorney on the amount of benefits a worker receives. These vary from an increase of less than $6000 for the bottom three deciles of the distribution of indemnity benefits to increases of $15,000–$18,000 for the top three deciles of the distribution of benefits. It is natural that the level effects are larger higher up in the distribution. Estimates from log specifications make the findings clearer, showing that the percentage effects are larger for cases with smaller payments. Note that this finding is consistent with what we found in the earlier standard regression IV estimates in levels and logs. The log specification puts more weight on the effect of attorney involvement in cases involving lower benefit levels, so the larger estimates from the log specification reported above reflect larger impacts in cases involving relatively lower indemnity benefits.

5.6 Variation in effects of attorney involvement across states

States differ in terms of the percentage of claims with attorney involvement as well as the percentage of claims with PPD/LS payments (for which attorney involvement is more common).38 The nature of disputes and how disputes are resolved also vary across states. In some states, attorneys are involved in nearly every claim that receives PPD/LS payments; in other states, they are only involved in a subset of these claims. States also differ in terms of rules that determine whether lump-sum settlements are used—a common occurrence in some states but not in others. The states also differ in terms of how attorney fees are determined and whether attorney fees are capped, which may contribute to the different incentives for attorneys to choose to represent workers' compensation clients. Given these differences across states, there is a potential concern that the results are driven by particular states, possibly because of workers' compensation system features in those states.

To address this issue, we explore estimates computed separately for each state in our data. These estimates are plotted in Figure 2. For the models in levels, the IV estimates by state vary a lot. In some cases, we find small and statistically insignificant estimates or even a few negative and significant estimates. However, the estimates from log specifications are more stable and always positive. Moreover, in every state but one (North Carolina), the estimate is statistically significant and sufficiently large relative to the percentage paid to attorneys that the estimate implies that workers gain from hiring an attorney.

When state-specific models of the relationship between attorney involvement and indemnity benefits are estimated, the estimates can become imprecise, both because of fewer observations and because (within states) the IVs may not predict attorney involvement very strongly. Indeed, in most states, the first-stage estimates indicate only a weak relationship between local area attorney involvement rates and individual-level attorney rates, although we still find a strong relationship between an initial delay in indemnity benefits and the individual attorney involvement rates. Hence, we need to be mindful of whether a given state-specific estimate is reliable enough to provide useful information. We emphasize the concern about precision of state-specific estimates point by including in Figure 2 (at the top) the overall estimate and confidence interval; this confidence interval is markedly smaller than for most states.

To provide information on what the state-level evidence says for the subset of states for which the data are sufficiently informative, we selected those states for which the estimates are precise enough to be able to identify our aggregated estimates from Table 8 as statistically significant. For the model in levels, we found that (with both IVs) attorneys increase indemnity benefits by $7700, so the standard error needs to be less than 3928 for the estimate to remain statistically significant at the 5 percent level. For the regressions in logs, the standard error needs to be less than 91 for the estimates to remain statistically significant. We apply those thresholds about how informative the state estimates are in Figure 2. In particular, we shade in black the estimates corresponding to states with estimates that meet the precision criterion just specified, and the other less precise estimates are shaded in gray (which, by construction, have larger confidence intervals). For estimates in levels (Panel a), of the 16 states for which the estimates are informative, the estimated effects are positive and statistically significant for 10 states. For estimates in logs (Panel b), all states had standard errors below our threshold of 91 (and hence none are shaded gray), although there were a number of states with fairly large estimated standard errors. This suggests that estimates from the log specifications provide a more reliable view of the relationship between attorney involvement and indemnity benefits at the state level. This could reflect the impact of outliers that become more influential in smaller samples but are less important when we use logs.39 Moreover, nearly all of the estimates for log benefits are positive and statistically significant.