Does Sorry Work the Impact of Apology Laws on Medical Malpractice JRU final 2

beyond payment size. For each claim, there is information regarding the year the incident occurred,13 the nature of the allegation (e.g., diagnosis related, anesthesia related, surgery related, etc.), the outcome of the incident (e.g., emotional injury, minor temporary injury, major permanent injury, death, etc.) ,14 the practitioner’s graduation year and age group, the practitioner’s work and licensing state, and whether the payment was for a judgment or a settlement. This dataset has been widely used in many studies related to medical malpractice (see Baicker & Chandra, 2005; Chandra, Nundy, & Seabury, 2005; Durrance, 2009; Matsa, 2007). We restrict our analysis to the reports in which adverse events occurred after 1991 due to the incomplete reporting in the earlier years. Since the NPDB only receives information about an offense/omission when the payment is made, the dataset is truncated for offenses/omissions that occurred late in the dataset but have yet to be resolved. In Figure 1 we present a histogram of time to resolution for cases that occurred in 1992 so that we can be reasonably certain that this represents a fairly complete distribution of cases. [Insert Figure 1 About Here] Panel A of Table 2 provides summary statistics at the individual level. There are a total of 225,319 payment reports in our sample. Note that the average time to resolution was 3.86 years with a in the NPDB. Chandra, Nundy, and Seabury (2005) compare data from the NPDB with other sources of malpractice information and while they find approximately 20% underreporting, they find that underreporting is not systematically different across states. Therefore, for our analysis, which is extracting information at the state level, there is no obvious reason why the corporate shield loophole would bias the effects of the apology legislation. It is also important to note that the NPDB dataset has been used for most recent influential studies of medical malpractice reform (Currie & MacLeod, 2008) 13 Since the finest date information we have about the case are years, we cannot use any finer definition of date (such as months, quarters) to look at the cases that took place right before the law passed and the cases that took place right after the law passed. 14 The outcome variable only became mandatory for recording in 2004. The categories of injuries are reported by the entities that make payments to the patients. 16 standard deviation of 2.15. Longer resolution times are associated with cases that involve more severe injuries. [Insert Table 2 About Here] In order to understand the aggregate effect of the laws on number of claims and total malpractice compensation payouts, the NPDB was also used to generate an aggregate dataset where each observation is at the state-year level. At the aggregate level, several predictions that follow from the model relate to the rate of litigation and the rate of resolution; however, this dataset does not contain information on open claims and closed claims without payouts and therefore these rates cannot be computed. Instead, we consider two other measures at the state level. The first is the number of payouts made by practitioners working in a given state for offenses committed in a given year.15 The second is the value of malpractice payments made by medical practitioners in a given state for offenses committed in a given year.16 Panel B provides these statistics at the state level. With 51 states (including the District of Columbia) reporting over a 17-year period (1991-2007), there are 867 observations in the state-level dataset.17 In 2000, the median number of incidents that had been resolved by 2008 per state was 184 incidents and the median total malpractice payments was $35.7 million. 15 Another way to construct the state-level dataset is by the total number of settlements/payments made in a given year. Our goal is to analyze the impact of apology laws, which intend to encourage practitioners to apologize and communicate more openly with their patients. The impact on the settlement is hinged upon the apology. While the model in Section 3 cannot distinguish the timing of the apology, the apology is likely to be most effective soon after the incident occurs, not a few years later. Therefore, we aggregate it by the year of incident instead of the year of settlement. 16 We adjust the payment by CPI. Therefore, all payments are in Y2000 dollars. 17 We have excluded all cases that occurred in 2008 since only less than 100 cases which occurred in 2008 had been settled by 2009. 17 4. Empirical Specification and Results The effect of apology laws on medical malpractice outcomes is estimated using a difference-indifference method. The validity of this specification rests on the assumption that the states that have passed apology laws would have otherwise followed the same trend as those states that have not passed apology laws. Therefore, we perform various checks to examine the validity of this assumption. Another crucial assumption that merits attention is that the passage of apology laws is not correlated with any other event that would affect medical litigation—an obvious possibility being the passage of other tort reform or malpractice laws. To ensure that a correlation with other malpractice laws does not drive our results, we include controls for other tort reforms as studied by Currie and MacLeod (2009) in each of our specifications.18 Furthermore, in each specification, we cluster standard errors by state to avoid problems of serial correlation (Bertrand, Duflo, & Mullainathan, 2004). State­Level Analysis There are two main outcome variables. The first variable is the number of already-resolved malpractice cases for incidents that occurred in year t and state s, while the second variable is the value of the total payments made in state s for incidents that occurred in year t. We aggregate the data by the year the incident occurred, not the year the incident settled, because our assumption is that apologies would be most effective when expressed soon after the incident. For example, if a malpractice case took place in 2000 and the law was passed in 2004, and the physician issues an apology only after the law passes, this apology is not going to be as effective in reducing the patient’s intention to litigate as an apology that is issued in 2000. Essentially, we assume that the law has a very limited impact on the malpractice cases that 18 The other law measures for which we have controlled the timing in our study include the existence of noneconomic cap, punitive cap, laws on full information disclosure, joint and several liabilities, and collateral source rule. The information on the existence of the laws (excluding information disclosure laws) is from the annual produced by the American Tort Reform Association (2009). The information on the disclosure laws is from Gibson and Del Vacchio (2006). 18 occurred before the law is passed. If cases that occurred before the law being passed are also affected by the apology law, our estimate would be an underestimate of the true impact.19 We first employ OLS to estimate the following: log Yst  controlsst   apologyst   t Yeart   s States   st , t s where Yst is the outcome variables and apology is a dummy variable which is one if an apology law was in effect in state s during year t and otherwise is zero. Our main coefficient of interest is β, which represents the percentage change in the number of closed claims with positive payout due to the adoption of the apology law. Before proceeding to the regression results, we first note how the fact that the dataset consists only of closed claims with positive payouts affects our findings. If the laws are working as intended (WAI), such that they increase doctors’ apologies and decrease patient anger then we would expect: Scenario WAIa . Cases are now settled faster ( increases) Scenario WAIb. Some cases that would have been litigated before the law are dropped ( decreases). If on the other hand, the unintended consequences (UNC) of the laws dominate (due to devalued apologies, increased patient information, increased moral hazard, changes in judicial standards, etc.), then we would expect the alternative hypothesis to hold: Scenario UNC. More cases are now litigated ( The net effect on the probability of litigation increases) is therefore ambiguous. In our closed claim dataset, we only observe a case from a given year if the case was litigated ( and if the case had settled ( )20 by that time. Therefore, our measure of closed claim frequency gives us . Scenario WAI therefore gives ambiguous predictions on closed claims since the probability of litigation decreases while the probability of settlement increases, while Scenario UNI unambiguously predicts closed claims 19 20 Legally, it is unclear whether apology laws would apply to cases that have occurred before the law passed. Here, we abuse notation a bit, equating the idea of probability of settlement with speed of settlement. 19 frequency would increase. However, in the long run, once enough time has passed, the probability that any claims that were open have settled approaches 1, then the closed claim frequency gives us the probability the patients litigated, , directly. So if the law was working as intended, then for cases that originated in recent years (2008 for example) we expect ambiguous predictions, while for cases that originated early on (2001 for example), the Scenario WAI predicts a lower claim frequency. Short Run Long Run Scenario WAIa increases # Closed Claims ↑ Scenario WAIb decreases # Closed Claims ↓ Scenario UNI increases # Closed Claims ↑ # Closed Claims Unchanged # Closed Claims ↓ # Closed Claims ↑ The basic regression results for closed claim frequency are presented in Table 3. Columns 1 and 4 are presented without controls, columns 2 and 5 add a full set of social policy changes (the existence of a noneconomic cap, a punitive cap, joint and several liabilities and collateral source rule, and a law on full information disclosure), while columns 3 and 6 include a set of time-varying state demographics including the number of physicians in the state, racial compositions, population, and percentage of population that are 65 or above. 21 The results show a consistent 14–15% increase in closed claim frequency with positive payouts.22 The results for the total compensation payout also show an increase of 20–27%. The fact that the percent increase shown in columns 4-6 is larger than those in columns 1-3, suggests that the payment per case increases after the law is enacted. The same analysis has also been performed on settlements excluding all cases that went to trial and the results are similar. Later in this section, we discuss various robustness checks performed, to rule out the possibility that the results are driven by spurious relationships, such as differential time trends from the states with the law. 21 Population, age, and racial composition data are retrieved from census data. Data on number of physicians per state was retrieved from the American Medical Association. 22 The coefficient in Column 3 is at the borderline of the 10% level of significance. The magnitude of coefficient does not change from Column 1 or Column 2, but rather the standard errors have gone up. 20 [Insert Table 3 About Here] Recall that this increase could be explained by either Scenarios WAI or UNC. Therefore, Tables 4 and 5 further addresses this question. If these results are due to the fact that cases are settling faster, as in Scenario WAIa, we should expect the upward bias on the effect of the laws to be more pronounced for the most recent cases. For example, if we examine the difference between states having the law to states without the law, we would expect that the upward bias to be bigger when examining cases occurring in 2008 than cases occurring in 2007, and bigger effects for cases occurring in 2007 than cases occurring in 2006, and so on. More time allows more cases to be resolved. In Table 4 we restrict our sample to the four states that passed an apology law before 2002 and those states that have never passed an apology law. For each column, the dependent variable is the natural log of the number of closed cases that occurred before that year. We include the full set of controls including other tort reforms. Column 1 considers only the number of cases that occurred before 2001 and thus gives us an estimate of the apology laws’ net effect after 8 years have elapsed. Column 2 considers the effect on all cases that occurred before 2002, thereby giving an estimate of the apology laws’ net effect after 7 years have elapsed. Similarly, column 8 considers the effect on cases before 2008 and thus gives us a sense of the apology laws’ net effect on current cases. One notable trend from this table is that these coefficients increase from left to right and become more statistically significant. It indicates that the positive coefficients in Table 3 Columns 1 to 3 are driven by the cases that occur in more recent years. If it were due to Scenario UNC (more people are litigating), it would be difficult to explain this increasing trend. Table 4 suggests that the apology laws’ net effect is zero (or possibly negative) in the long run.23 [Insert Table 4 About Here] 23 One might think, using the aforementioned logic, that the dependent variable should be the cases that occurred in 2002 in column 1, those that occurred in 2003 in column 3, etc. However, we would only have 21 observations in each regression and would not be able to capture any general state or year trends. 21 Table 5 considers the composition of cases in terms of the type of injury. The dependent variable is the log frequency of closed claims in each medical injury category (e.g., insignificant injury, “somewhat” significant injury, and major permanent injury/death).24 Restricted by the sample of cases that occurred after 2002, 25 Table 5 shows that the overall increase in closed claims observed is due entirely to the closed claims for major/permanent injury and death. For insignificant injuries, which normally settle quickly enough to see the apology laws’ full effect, we see a net reduction of 16.7–18.5% in the number of cases. Consider again the possible scenarios that could explain this result. Suppose that doctors are exerting less effort after the laws are implemented (Scenario UNC), it is difficult to explain the pattern in Table 5. The results in Table 5 suggest that after passing the law, there is a short-term increase in the number of cases that normally take many years to resolve,26 but an overall decrease in the number of cases involving the least significant injuries. [Insert Table 5 About Here] Individual­Level Analysis Since we have data on the year in which the incident occurred and the year it was settled, hazard models provide a natural framework for modeling settlement probabilities (e.g., Hannan & McDowell, 1984). Let t be the time elapsed from the time of incident to the time of settlement, relevant explanatory variables, and function as | , | , be a vector of be a vector of coefficients. Denoting the cumulative density and the density function as | , , the hazard function which indicates the probability of settlement at period t, conditional upon no settlement by time is defined as 24 | , /1 1, . There are nine categories of injuries in the NPDB, which we group into three categories for the ease of analysis and presentation (see Table 2 for subcategories). 25 The severity of injuries is only available for cases reported after 2002. For a similar analysis grouped by the size of payment, see Online Appendix on the author’s website. 26 From this dataset we can observe that it is true that cases involving more severely injured patients usually take longer to resolve than insignificant injury cases. 22 We use Cox proportional hazard model, and the advantage of the Cox model is that there is no assumption about the functional form of | , . The hazard ratios from Cox model are reported in Table 6. To interpret the coefficient, one needs to be cautious that the reported hazard ratios are conditional on resolution. Depending on which unintended consequence dominates, the theory is unclear as to how Scenario UNC would affect the hazard ratio as it would depend on the composition of the cases that are affected. We find that the hazard rate changes marginally for insignificant cases and the hazard ratio increase due to the apology laws being positively associated with the severity of the medical outcomes. For a case involving a major/permanent injury, conditional on resolution, the probability it resolves in any given year is increased by 19% when the apology law is in effect.27 It is notable that Kachalia et al. (2010) examine the impact of University of Michigan’s information disclosure program, and find the program increased the rate of settlement by 27%. Although our estimates are slightly smaller, it is encouraging that these two estimates are statistically the same. In Table 7 (Panel A), we consider the impact the law has on the size of malpractice claim payouts using a difference-in-difference model. In Panel A, claim severity is the dependent variable. Since one might worry that the result could be driven by outliers, Panel B presents the result of regressions using the log of payments. Before moving to the results, consider again how each scenario might affect the findings for this table. If Scenario WAIa dominates (cases are settled faster), then we might see an increase of claim severity since the more costly cases are only resolved where the apology laws are adopted. If Scenario WAIb dominates (fewer cases are litigated), the case composition might change and the average claim severity might increase—for example, all the cases with potential payments below a certain cutoff were never litigated due to the passage of the apology law, with this truncation, the cases that remain are those 27 We have also attempted a maximum likelihood estimate of the unconditional hazard ratio using a proportional hazards model that accounts for right truncation from Finkelstein, Moore, and Schoenfeld (1993). However, due in part to their model being weakly identified, the procedure does not converge. 23 with higher payouts. Lastly, if apologies simply reduced anger, apology laws should lower the claim severity. Looking at Table 7, for “insignificant injuries,” we find evidence that claim severity increases marginally. As explained above, this increase should be attributed to the change of case composition. As for “somewhat significant injuries,” we find marginally insignificant results. Finally, for “major permanent injury,” we find that after the law is adopted, malpractice payouts decrease by about 14-17%. Given the effect of the composition shift, this is a lower bound estimate of the true effect of apology laws. In levels, after the law is adopted, claim payout would be reduced by approximately $17,000–27,000 for somewhat severe cases and $55,000–73,000 for the most severe cases.28 [Insert Table 6 About Here] [Insert Table 7 About Here] Taken together, Tables 5, 6, and 7 suggest that apology laws are consistent with the symmetric information model presented above as well as the legislators’ intent; the apology laws’ combined effect is to increase apologies and decrease expected settlement time, and should in the long term speed up settlements and reduce the total number and value of malpractice payments. We see that the apology laws reduce the total number of the insignificant injury cases that tend to settle quickly as well as reducing the payment size and increasing the settlement speed of cases involving major injuries/death. Threat to Validity: Robustness Check It could be worrisome if the effect of the states is spurious to the structure of the data or the time period upon which we estimate the data. Therefore, we perform various robustness checks. For brevity, 28 Regressing the same specification on different payment size quantiles finds that the law has the largest effects on the 3rd quantile and no effect on the 1st and 4th quantile. The lack of effect on 4th quantile payments could be due to the fact that apologies are likely to be less important in cases worth millions of dollars, or that the largest cases take many years to resolve and thus cases of this size have yet to be resolved in most states in which apology laws have been passed. 24 these robustness check results can be found on the authors’ website. First, we randomly assign half of the states as having adopted the law between 2000 and 2005 and estimate the same difference-in-difference regression. Moreover, for all the states that have adopted apology laws, we subtract 3 years from the year of adoption and perform the same analysis to capture any possible spurious effect attributable to properties of the states in question rather than to the laws themselves. The coefficients remain insignificant. Lastly, as we intend to interpret the result as a causal interpretation, we need to check to see if the increase in settlements came after the adoption of the apology laws. Therefore, we include in our differences-in-differences specification a series of lead dummy variables, which specifies whether apology laws will be adopted in that state 1 year, 2 years, 3 years, 4 years, or 5 years into the future. We find that all coefficients on the lead dummies are not statistically different from zero, suggesting that the effects that we find do not predate the passing of the apology laws. Now, knowing that the results are robust and not due to spurious effect, we need to find the hypotheses that could explain the seemingly surprising results. Given that both main outcomes of interest are non-negative, we can also reanalyze the main outcomes using a Poisson model which makes assumptions about the distribution. We still find that most of the increases in the case frequencies are from the medical cases involving the most severely injured patients and that there is little change in the frequency of the cases involving minor injuries. We examine the sensitivity of our results with a number of alternative specifications. First, we omit in turn each of the 36 states and each of the available years, and the regressions yield similar results, suggesting that the results are not driven by a single outlier or a particular year. Furthermore, median regression and population weighted least squares yield similar positive significant results. Our regressions above are analyzed based on the year the offense occurred because the apology laws largely apply only to apologies that were made for adverse events that occurred after the law was passed. Also, arguably apologies are most effective shortly after the adverse event occurred, so medical errors that occurred years before the law was passed should be unaffected. As a robustness check, we provide a difference-in-difference analysis in which the dependent variable is the natural log of 25 settlements grouped by the year of settlement, and we find that laws have no impact on the incidents that occurred before the passing of the apology laws. One of our theory’s predictions is that the probability a settlement is reached should increase. It may seem natural to look at the ratio between settlements over judgments across states over time. However, given that only 3% of the cases in this dataset are judgments, and that cases that have gone to court usually take much longer to resolve and thus would not enter our dataset, this ratio cannot be meaningfully estimated. 5. Conclusion We have presented a simple model of apologies and medical malpractice and estimated the effects of state level apology laws on medical malpractice claims. We find that in the short run the law increases the number of resolved cases, while decreasing the average settlement payment for cases involving more significant and permanent injuries. While having an insignificant impact on the settlement payments for cases involving minor injuries, the apology laws do reduce the total number of such cases. While the short term increase in malpractice settlements could be a surprise to policymakers and advocates of apology laws, we believe this is an artifact of data limitations. Our findings suggest that apology laws reduce the amount of time it takes to reach a settlement in what would normally be protracted lawsuits, leading to more resolved cases in the short run. In the long run, the evidence suggests there could be fewer cases overall. 26 References American Tort Reform Association. (2009). ATRA tort reform record. Electronic Version: ATRA. Baicker, K., & Chandra, A. (2006). The labor market effects of rising health insurance premiums. Journal of Labor Economics, 24(3), 609-634. Bertrand, M., Duflo, E., & Mullainathan, S. (2004). How much should we trust differences-in-differences estimates?*. Quarterly Journal of Economics, 119(1), 249-275. Boothman, M., Blackwell, A., Campbell, D., Commiskey, E., & Anderson, S. (2009). A better approach to medical malpractice claims? the University of Michigan experience. Journal of Health Life Science Law, Jan(2), 125-59. Born, P., Viscusi, W. K., & Baker, T. (2009). The effects of tort reform on medical malpractice insurers' ultimate losses. Journal of Risk and Insurance, 76(1), 197-219. C.G.S. §52-184d(b), as Amended by 2006 Pub. Act no. 06-40, §2(b) Chandra, A., Nundy, S., & Seabury, S. A. (2005). The growth of physician medical malpractice payments: Evidence from the national practitioner data bank. Health Affairs, 240. Currie, J., & MacLeod, W. B. (2008). First do no harm? Tort reform and birth outcomes*. Quarterly Journal of Economics, 123(2), 795-830. Daughety, A., & Reinganum, J. (2000). On the economics of trials: Adversarial process, evidence, and equilibrium bias. J Law Econ Organ, 16(2), 365-394. Durrance, C. P. (2009). Noneconomic damage caps and medical malpractice claim frequency: A policy endogeneity approach. J Law Econ Organ. Division of Practitioner Data Banks, Bureau of Health Professions, Health Resources and Services Administration. (2004). National practitioner data bank public use data file. Farber, H. S., & White, M. J. (1994). A comparison of formal and informal dispute resolution in medical malpractice. The Journal of Legal Studies, 23(2), 777-806. Finkelstein, D. M., Moore, D. F., & Schoenfeld, D. A. (1993). A proportional hazards model for truncated AIDS data. Biometrics, 49(3), 731-740. Hannan, T. H., & McDowell, J. M. (1984). The determinants of technology adoption: The case of the banking firm. The Rand Journal of Economics, 15(3), 328-335. "The Health Care Quality Improvement Act." Title 42 U.S. Code, Sec. 11101-11152. 1998 Ed. Hickson, G. B., Clayton, E. W., Githens, P. B., & Sloan, F. A. (1992). Factors that prompted families to file medical malpractice claims following perinatal injuries. Journal of American Medical Association, 267(10), 1359-1363. Ho, B. (2011). Apologies as signals: With evidence from a trust game. Forthcoming: Management Science. 27 Ho, B. & Liu, E. (2011). What’s An Apology Worth? Difference-in-Differences Analysis of State Apology Laws on Medical Malpractice Payouts. Working Paper. Kachalia, A., Kaufman, S. R., Boothman, R., Anderson, S., Welch, K., Saint, S., et al. (2010). Liability claims and costs before and after implementation of a medical error disclosure program. Annals of Internal Medicine, 153(4), 213-221. Kraman, S. S., & Hamm, G. (1999). Risk management: Extreme honesty may be the best policy. Annals of Internal Medicine, 131(12), 963-967. Lamb, R. M., Studdert, D. M., Bohmer, R., Berwick, D. M., & Brennan, T. (2003) A. Hospital disclosure practices: Results of a national survey. Health Affairs. Liebman, C. B., & Hyman, C. S. (2004). A mediation skills model to manage disclosure of errors and adverse events to patients. Health Affairs, 23(4), 22-32. Liebman, C. B., & Hyman, C. S. (2005). Medical error disclosure, mediation skills, and malpractice litigation: A demonstration project in Pennsylvania. Working Paper. Mastroianni, A. C., Mello, M. M., Sommer, S., Hardy, M., & Gallagher, T. H. (2010). The flaws in State 'apology' and 'disclosure' laws dilute their intended impact on malpractice suits. Health Affairs, 29(9), 1611-1619. Matsa, D. A. (2007). Does malpractice liability keep the doctor away? evidence from tort reform damage caps. Journal of Legal Studies, 36(S2), S143-S182. Mazor, K.M., Simon, S.R., Yood, R.A., Martinson, B. C., Gunter, M. J., Reed, G. W., et al. (2004). Health plan members' views about disclosure of medical errors. Annals of Internal Medicine, 140(6), 409-418. McDonnell, W. M., & Guenther, E. (2008). Narrative review: Do State laws make it easier to say “I'm sorry?”. Annals of Internal Medicine, 149(11), 811-815. Ohbuchi, K., Kameda, M., & Agarie, N. (1989). Apology as aggression control: Its role in mediating appraisal of and response to harm. Journal of Personality and Social Psychology, 56(2), 219. Polinsky, A. M., & Rubinfeld, D. L. (1988). The deterrent effects of settlements and trials. International Review of Law and Economics, 8(1), 109-116. Robbennolt, J. K. (2003). Apologies and legal settlement: An empirical examination. Michigan Law Review, 102(3), 460-516. Sloan, F. A., & Hsieh, C. R. (1995). Injury, liability, and the decision to file a medical malpractice claim. Law & Society Review, 29(3), 413-435. Spier, K. E. (1992). Incomplete Contracts and Signalling. The Rand Journal of Economics, 23(3), pp. 432-443. Spier, K. E. (2002). Settlement with Multiple Plaintiffs: The role of insolvency. Journal of Law, Economics, and Organization, 18(2), 295-323. 28 Vincent, C., Phillips, A., & Young, M. (1994). Why do people sue doctors? A study of patients and relatives taking legal action. The Lancet, 343(8913), 1609. Viscusi, W. K., & Born, P. H. (2005). Damages Caps, Insurability, and the Performance of Medical Malpractice Insurance. Journal of Risk and Insurance, 72(1), 23-43. Wu, A. W. (1999). Handling hospital errors: Is disclosure the best defense? Annals of Internal Medicine, 131(12), 970-972. Wu, A., Huang, I., Stokes, S., & Pronovost, P. (2009). Disclosing medical errors to patients: It’s not what you say, it’s what they hear. Journal of General Internal Medicine, 24(9), 1012-1017. 29