# «ABSTRACT We present evidence that ﬁnancing frictions adversely impact investment in workplace safety, with implications for worker welfare and ...»

To get a sense of the relative variation of injury occurrence in our sample, we calculate between- and within-group variances at the establishment-, ﬁrm-, and industry-level. Table IV reports these variances.

One takeaway from this table is that injury rates are not simply an industry-speciﬁc eﬀect. Rather, they actually vary more on a within-industry basis than across industries, though we note that our industry deﬁnitions are fairly coarse. As Table III shows, there is considerable variation across industries as well. Another takeaway is that, while there is more variation in injury rates across establishments than within establishments, there is considerable variation within an establishment over time. The standard deviation of injuries per employee within an establishment is 2.0% (for comparison, the mean number of injuries per employee over the full sample is 4.13%). This level of variation gives us power to identify the determinants of injuries even after controlling for establishment ﬁxed eﬀects.

** III. Empirical Determinants of Injury Rates**

In this section, we examine the determinants of workplace injury rates. We do so primarily by estimating ﬁxed eﬀects regressions of Injuries/Hour on establishment- and ﬁrm-level characteristics, with a focus on those characteristics previously shown to relate to a ﬁrm’s ability to ﬁnance investment. Our main regression speciﬁcation is Injuries/Hourit = αi + δjt + γst + βxit + (1) it, where t indexes year, i establishment, j industry, and s the state in which the establishment is located. The vector xit consists of establishment-year and ﬁrm-year variables, including those relating to ﬁnancing. Establishment ﬁxed eﬀects αi account for any time-invariant omitted establishment characteristics, industry-year ﬁxed eﬀects δjt for any time-variation in injury risk at the industry level (perhaps due to the evolution of production technology or industry growth), and state-year ﬁxed eﬀects γst for any state-level time-variation in injury 14 risk (perhaps due to changes in the local labor environment or safety laws).

Table V presents results from estimating equation (1). Standard errors clustered at the ﬁrm level are shown in parentheses below each point estimate in this and all subsequent tables presenting regression results. Column (1) presents results without any ﬁxed eﬀects, column (2) results with only establishment ﬁxed eﬀects, and column (3) results with all of the ﬁxed eﬀects. Note that the number of observations decreases from 43,371 in column (1) to 25,053 in columns (2) and (3) because the inclusion of establishment ﬁxed eﬀects limits us to establishments appearing in the sample at least twice.

— Insert Table V here — Column (1) shows that, without ﬁxed eﬀects, injury rates are positively related to Debt/Assets, signiﬁcant at the 5% level, and negatively related to Cash/Assets and CashF low/Assets, signiﬁcant at the 5% and 10% levels, respectively. These relations are all consistent with a greater ability to ﬁnance investment leading to greater workplace safety and hence a lower risk of injury. Columns (2) and (3) show that, once establishment ﬁxed eﬀects are included, injury rates cease to show a relation with Cash/Assets and CashF low/Assets. However, the coeﬃcient on Debt/Assets continues to be positive at the 5% level of statistical signiﬁcance and increases in size.

The 0.0063 coeﬃcient on Debt/Assets in column (3) implies that a one-standard-deviation increase in leverage (0.219) is associated with a 0.0014 unit increase in Injuries/Hour.

This represents a modest though nontrivial increase of 5.6% relative to the sample mean Injuries/Hour of 0.0247. For perspective, this eﬀect is larger than existing estimates of the impact of either penalty-imposing OSHA inspections or plant unionization on injury rates (Mendelhoﬀ and Gray (2005)). While the coeﬃcients on Cash/Assets and CashF low/Assets are not statistically signiﬁcant in columns (2) and (3), the coeﬃcients in column (1) imply that a one-standard-deviation increase in these variables is associated 15 with a 6.34% and 3.01% increase in Injuries/Hour relative to the sample mean, respectively.

In all three columns, injury rates are negatively related to both Dividends/Assets and Log(Assets). Previous papers argue that a ﬁrm’s propensity to pay dividends and its size are inverse proxies for the degree to which the ﬁrm is ﬁnancially constrained. While more indirect, these relations also provide support for ﬁnancing constraints adversely impacting investment in workplace safety.

Investment in safety should be more relevant in industries in which production involves physical assets than in more service-oriented industries. Column (4) reports estimates from the main speciﬁcation when we restrict the sample to establishments in industries with above-median average T angibleAssetRatio. The coeﬃcient on Debt/Assets is 50% higher in column (4) than in column (3). As mean Injuries/Hour is 0.0266 in these establishments, the coeﬃcient of 0.0093 on Debt/Assets implies that a one-standard-deviation increase in Debt/Assets is associated with an increase in Injuries/Hour of 7.66% relative to the mean.

Column (5) presents results using DAF W Injuries/Hour — the rate of more serious injuries — is the dependent variable. The coeﬃcient on Debt/Assets in this regression is

0.023 and is statistically signiﬁcant at the 5% level. A one-standard-deviation increase in leverage of 0.219 is associated with an increase in DAF W Injuries/Hour of 0.0005, which represents a 6.5% increase relative to the sample mean DAF W Injuries/Hour of 0.0077.

Thus, more serious injuries appear to increase meaningfully with leverage.

In the next two regressions, we analyze the extensive and intensive margins of the injury process separately. Column (6) presents results from a linear probability model with establishment, year-industry, and year-state ﬁxed eﬀects. The dependent variable is one if the establishment reports a positive number of injuries in a given year and zero if it reports zero injuries. Not surprisingly, larger establishments (those with more employees) are more likely to experience at least one injury in a given year. Among the ﬁrm-level characteristics, the probability of an injury is only statistically signiﬁcantly related to Debt/Assets, with which 16 it has a positive relation.

Column (7) estimates equation (1) for with a positive number of injuries in the given year.

As in the full sample (column (3)), the injury rate is positively related to Debt/Assets and negatively related to Dividends/Assets and Log(Assets) for establishments on the intensive margin. These results help allay concerns that the large number of zeros in the injury data might somehow skew the regression results presented in the ﬁrst three columns.

We further address concerns about the distribution of the injury data by estimating count models, which explicitly account for the nonnegative discrete nature of injuries. The dependent variables in these regressions is the number of injuries rather than the injury rate.

We specify an exposure variable, HoursW orked, to account for the scale of baseline exposure.

Column (8) reports estimates from an establishment ﬁxed eﬀects Poisson model.13 We also include state-year dummies in the regression, though industry-year dummies are omitted because they cause the estimation routine to fail. The coeﬃcient on Debt/Assets is positive and statistically signiﬁcant at the 1% level, consistent with the estimates in Table V.14 Column (9) presents estimates from a negative binomial model. Unlike the Poisson model, the negative binomial model does not admit ﬁxed eﬀects. However, it does relax the assumption of equal conditional mean and variance that the Poisson model imposes. We are able to include both industry-year and state-year dummies in this model. The coeﬃcient on Debt/Assets is again positive and statistically signiﬁcant at the 1% level. The coeﬃcient on Cash/Assets is also negative and statistically signiﬁcant at the 1% level in this regression.15 Overall, Table V presents evidence of a robust positive relation between an establishment’s injury rate in a given year and parent ﬁrm ﬁnancial leverage at the end of the prior year. We further explore this relation by regressing year t Injuries/Hour on various leads and lags of Debt/Assets, controlling for the same variables as in Table V.16 Table VI presents the results.

— Insert Table VI here — 17 Column (1) presents results when we include just the ﬁrst lead and lag of leverage, along with contemporaneous leverage, as explanatory variables. The injury rate is positively related to year t − 1 Debt/Assets. In contrast, it is unrelated to t + 1 Debt/Assets, partially alleviating concerns that the results in Table V might be driven by reverse causality. It is also unrelated to contemporaneous Debt/Assets. In column (2), we include the ﬁrst three leads and lags of Debt/Assets. The coeﬃcients on all three lags of Debt/Assets are positive, with those on the ﬁrst two lags statistically signiﬁcant at the 5% level. This suggests that any impact of leverage on injury risk persists for at least two years. The coeﬃcients on all three leads of Debt/Assets are small in magnitude and statistically insigniﬁcant. Columns (3) and (4) conﬁrm these patterns in Poisson regressions.

Overall, the results in this section provide evidence that ﬁnancing constraints adversely impact workplace safety, with the consistent relation between injury rates and ﬁnancial leverage providing the strongest evidence. Controlling for ﬁrm-level variables such as Capex/Assets and T angibleAssetRatio helps account for diﬀerences in growth and production technology.

The inclusion of establishment, industry-year, and state-year ﬁxed eﬀects requires that any omitted variable driving the relationship be time-varying within an establishment, and not purely through industry- or state-level time-variation. Ultimately, however, given the potentially endogenous nature of the explanatory variables with respect to injury rates, the results should not be interpreted as strong evidence of a causal link between leverage and injury risk.17

In this section, we further explore the eﬀect of ﬁnancing constraints on injury rates using three quasi-natural experiments involving cash ﬂow shocks. The ﬁrst experiment exploits a 18 provision in the American Jobs Creation Act (AJCA) of 2004 allowing ﬁrms to pay a tax rate of 5.25% on repatriated foreign income on a one-time basis instead of the standard corporate tax rate of 35%. This shock represented a signiﬁcant windfall for the domestic coﬀers of ﬁrms with proﬁtable foreign subsidiaries. Firms collectively repatriated $312 billion in response to the AJCA according to IRS estimates. Dharmapala, Foley, and Forbes (2011) and Faulkender and Petersen (2011) study the eﬀects of this shock on investment levels in general.

The second experiment exploits the maturity structure of ﬁrms’ debt at the onset of the ﬁnancial crisis in late 2007. Credit markets seized up in the U.S. starting in August 2007 and remained tight through 2008, making it diﬃcult for ﬁrms to roll over maturing debt.

A ﬁrm with a lot of debt maturing during this period eﬀectively faced a negative cash ﬂow shock. A ﬁrm’s maturity structure as of the beginning of the crisis is plausibly exogenous with respect to factors that might aﬀect injury risk, as it was unlikely that ﬁrms anticipated the crisis when setting maturity schedules in the preceding years. Almeida et al. (2012) study the eﬀect of this shock on investment levels in 2008.

The third experiment exploits substantial ﬂuctuations in oil prices during our sample period. Oil prices increased from around $25 per barrel in 2002 to over $130 per barrel in 2008, before falling to the $40s in 2009. Higher oil prices increase the cash ﬂow generated by oil producers. Because ﬁrms can reallocate capital internally, this increases the cash ﬂow available to any non-oil establishments owned by oil producers. Assuming that oil price movements do not impact these non-oil establishments for other reasons, these shocks can be treated as exogenous with respect to injury rates in these non-oil establishments. Following this logic, Lamont (1997) studies the eﬀect of a 1985 drop in oil prices on investment.

Each of the cash ﬂow shocks involved in the experiments impacted some ﬁrms more than others. The AJCA represented a cash ﬂow shock only for ﬁrms with previously unrepatriated foreign proﬁts. The ﬁnancial crisis represented a larger shock for ﬁrms with high 19 levels of debt maturing in 2008 than for ﬁrms with less debt maturing in 2008. Oil price movements represented a cash ﬂow shock primarily for ﬁrms with oil-producing subsidiaries.

For the AJCA and ﬁnancial crisis experiments, we exploit the diﬀerential exposure to conduct diﬀerence-in-diﬀerences analysis. Speciﬁcally, we identify establishments exposed to the shock in question (“treated” establishments) and those not exposed (“untreated” establishments). We then match each treated establishment to an untreated establishment, which we refer to as a “control” establishment, to form a matched sample and estimate regressions of

**the following form using that sample:**

Injuries/Hourit = βT reatmentt ∗ Exposurei + φXit + αi + δjt + γst + (2) it, where T reatment equals one for observations after the given shock and zero before, and Exposure equals one for treated establishments and zero for control establishments. We include all of the explanatory variables in the previous section as control variables Xit, and also include establishment (αi ), industry-year (δjt ), and state-year (γst ) ﬁxed eﬀects.18 We follow a similar approach for the oil price experiment, though here we set T reatment to the log of the mean oil price in year t, which is continuous.