Introduction

‘Defensive medicine’, where physicians supply more or different medical care than what a fully informed consumer would demand in order to deter lawsuits, has received much national attention lately. An example of this is the hypothesized positive association between area lawsuit activity and Cesarean sections (c-sections) (see, e.g., Tussing and Wojtowycz, Reference Tussing and Wojtowycz1997; Localio et al., Reference Localio, Lawthers, Bengtson, Hebert, Weaver, Brennan and Landis1993). It is possible that area lawsuit activity cause c-section rates to be higher than the optimum level.

Independent of the defensive medicine literature, the small area variations literature challenges the proposition that clinical procedures are optimally supplied or demanded (see, e.g. Phelps and Mooney, Reference Phelps and Mooney1993). Rather, practice styles and/or patient preferences often vary widely by geographic area. Consider recent research on medical procedures which are relatively insulated from lawsuits. Skinner et al. (Reference Skinner, Weinstein, Sporer and Wennberg2003) show that the rates of knee anthroplasty among Medicare patients, an elective procedure, has wide regional variation, even after adjusting for confounders. There is also wide variation in clinical procedures among hospitals. For instance, Wennberg et al. (Reference Wennberg, Fisher, Strukel, Skinner, Sharp and Bronner2004) found widespread medical variation even amongst highly respected hospitals in the United States. Therefore, while area lawsuit pressure may increase c-section rates, the assumption that physicians would adhere to best practices (optimal supply) in the absence of lawsuits is dubious. At least some of the variation in the mode of delivery would likely exist in the absence of lawsuits.

This paper examines whether area lawsuit activity in part drives higher area c-section rates through defensive medicine, after controlling for small area variation through multilevel modeling. Small area variation, which has largely been ignored in the defensive medicine literature, may bias or confound the relationship (in either direction) between c-sections and area lawsuit activity.

Rice and Jones (Reference Rice and Jones1997) suggest that multilevel statistical modeling techniques can be used to address small area variations. Health practice variation across regions and hospitals can be explicitly modeled in the error structure by including random effects which capture hospital and regional variation related to practice style and patient tastes. This paper uses a three-level model: Individual, hospital, and the Dartmouth hospital referral region (HRR) (Rabe-Hesketh et al., Reference Rabe-Hesketh, Pickles and Skrondal2001).

The previous literature on defensive c-sections, by ignoring small area variation which occurs independently of lawsuits, may have reported biased estimates of the association between c-sections and lawsuit activity. Controlling for regional- and hospital-level variation may strengthen the association between lawsuit activity and delivery practice/demand, perhaps explaining the mixed nature of the results in the literature. On the other hand, the bias in the association between lawsuits and c-section rates may be in the other direction.

The Healthy People 2010 report recommends that, for first-time deliveries, the c-section rate be reduced to 15% (United States, 2000). If only a small percentage of the unnecessary or unwanted c-sections are due to lawsuit pressure, the cost savings of reducing unnecessary c-sections could be substantial. The Health Insurance Association of America uses charge data to estimate that hospitals charge an average of $4,360 more for c-section deliveries than vaginal deliveries in 1996 (Health Insurance Association of America, 1999).Footnote ¹ Recent legislative efforts to limit awards in malpractice lawsuits would likely reduce defensive medicine. Less defensive medicine could potentially lead to lower health insurance premiums and lower taxes, the latter primarily due to savings in Medicaid. Consumers may be willing to accept restrictions on lawsuits in exchange for paying less in taxes and health insurance premiums, particularly if health outcomes remain unchanged.

Background

Several studies in the health policy and health services literature examine c-section rates as defensive medicine. Dubay et al. (Reference Dubay, Kaestner and Waidman1999) find a positive association between c-section rates and sub-state area malpractice premiums. However, they find that the magnitude of the effect of lawsuits on c-sections is small.Footnote ² Localio et al. (Reference Localio, Lawthers, Bengtson, Hebert, Weaver, Brennan and Landis1993) and Tussing and Wojtowycz (Reference Tussing and Wojtowycz1997) also find a positive relationship between malpractice lawsuits and c-section rates. Baldwin et al. (Reference Baldwin, Hart, Lloyd, Fordyce and Rosenblatt1995) and Sloan et al. (Reference Sloan, Entman, Reilly, Glass, Hickson and Zhang1997) find no relationship between malpractice lawsuit activity and c-sections.Footnote ³

None of these papers explicitly controls for unobserved regional or hospital variation, leaving the possibility of bias in the relationship between lawsuits and c-sections. Dubay et al. (Reference Dubay, Kaestner and Waidman1999) deal with county-level unobserved variation with an econometric fixed effects model, but their study did not account for hospital-level variation. Further, using fixed effects by county is less likely to capture practice styles than approaches based on hospitals and hospital markets. Tussing and Wojtowycz (Reference Tussing and Wojtowycz1997) control for ‘peer effects’ of physician practice patterns by including the county-level c-section rate. While their approach does address some of the practice style issue, Rice and Jones (Reference Rice and Jones1997) point out that multilevel models are more efficient because of spatial clustering effects of individual hospitals and regions. The clustering effects may occur because of differing practice styles at hospitals or HRRs or because of spatial differences in patient tastes.

In Texas, almost half of the births are to Hispanic mothers. Recent research has shown that Hispanics are more likely to deliver via c-section, even after controlling for hospital and physician random and fixed effects (Chung et al., Reference Chung, Garite, Adele, Hollard, Wing and Lagrew2006). The spatial clustering of Hispanic mothers, such as in the border regions, may accelerate the likelihood of c-sections, perhaps due to delivery mode preferences. Multilevel modeling affords the possibility to control for this effect.

The Dartmouth Atlas of Health Care divides the country into 306 hospital referral regions (HRRs) that are typically centered on a tertiary hospital that offers cardiovascular and neurosurgical procedures (Wennberg and Cooper, Reference Wennberg and Cooper1999). HRRs are defined according to hospital markets, where most patients and doctors within the region use hospitals within the region. People are unlikely to be referred out of the HRR for tertiary care. In addition to knee angioplasty, HRR regional variation has also been found for hip fractures, colorectal cancer, and acute myocardial infarction (Fisher et al., Reference Fisher, Wennberg, Stukel, Gottlieb, Lucas and Pinder2003a,b).

Conceptually, the model is similar to other models found in the literature (see, e.g. Dubay et al., Reference Dubay, Kaestner and Waidman1999). C-sections are related to hospital characteristics, maternal socio-economic characteristics including health insurance, maternal clinical indicators, and the lawsuit environment.

In order to control for random hospital and area variation, a three-level multilevel model is estimated. The first level is the individual (the mother), the second is the hospital, and the third is the HRR. A three-level model will address hospital and regional variation by shifting the intercept for each hospital as well as for each HRR (Snijders and Bosker, Reference Snijders and Bosker1999). Consider the following multilevel model

(1)

where n_ijk is the logit predictor of Y_ijk, which is a binary variable indicating whether a c-section was performed or not. Index i corresponds to the mother, index j indicates the hospital, and k indicates the HRR. B_o is the intercept, B_i is a vector of coefficients corresponding to maternal clinical indicators and socio-economic characteristics (including health insurance), B_j is a vector of coefficients corresponding to hospital variables, and B_k is the coefficient for the lawsuit environment for the county within the HRR. u_jk is the random intercept term for the hospital, while u_k is the random intercept term for the HRR. It is assumed that u_jk and u_k are independently normally distributed (Rabe-Hesketh et al., Reference Rabe-Hesketh, Pickles and Skrondal2001).

Among the hospitals with 100 or more births in the Texas Health Care Information Council data, which is described later, the highest hospital c-section rate (previous c-section cases included) is 51.2% and the lowest hospital c-section rate is 13.9%. The coefficient of variation for c-sections across hospitals is 305. Therefore, the hospital c-section rates are quite variable by hospital. Figure 2 shows the c-section rate by HRR. Clearly, there is wide variation even amongst neighboring HRRs. As seen in Figure 2, the highest HRR c-section rate (previous c-section cases included) is 40.2% whereas the lowest HRR c-section rate is 21.8%. Recall that the HRRs are large enough to have at least one tertiary hospital so that very few mothers would travel outside of the HRR for delivery. Therefore, the unadjusted variation between the HRR regions is quite substantial.

There are two main sources of data for this analysis. First, there is the Texas Health Care Information Council (THCIC) database of hospital discharges for the first three quarters of the year 2002 (Texas Health Care Information Council, 2002). The last quarter of 2002 is available, but it is excluded because health insurance data are missing (Texas Health Care Information Council, 2002). These data represent a census of all discharges from all hospitals in the state of Texas. Second, data from the Texas Closed Claim File, which includes all lawsuits resolved for each county for the years 1988–2001, were included. The THCIC data includes birth cases, with c-sections being identified by whether there was a Centers for Medicare and Medicaid Services diagnostic-related group (DRG) number of 370 or 371 reported.Footnote ⁴ Because previous c-section deliveries almost perfectly predict subsequent c-sections, deliveries with prior c-section deliveries were excluded. Finally, Medicare deliveries were excluded.

Hospital variables include the teaching status of the hospital. Teaching hospitals are those defined as such by the Texas Department of Health, Facility Licensing Division. Maternal socio-economic variables include dummy variables for women 17 years of age and younger and women 35 years of age and older, as well as for African-American women, Asian women and Hispanic women. Dummy variables are also included for patients having Medicaid or self-pay status.

The clinical indicators for c-sections are summarized in three dummy variables, constructed via ICD9 codes from the THCIC data. They are organized chronologically: conditions indicating c-sections prior to pregnancy, conditions during pregnancy, but before labor, and emergency labor indicators. Otherwise, they largely follow Localio et al. (Reference Localio, Lawthers, Bengtson, Hebert, Weaver, Brennan and Landis1993). Controls for problems prior to pregnancy include herpes (541) and diabetes (250, 6,480). Controls for problems during pregnancy include illegal drug use (305, 6,483), hypertension (642), prematurity (644), post-term pregnancy (645), sexually transmitted disease (647), multiparity (651), breech presentation (6,522), face or brow presentation (6,524), oblique or brow presentation (6,523), disproportion (653), cervical incompetence (6,545), antepartum hemorrhage (6,413, 6,418, 6,419), and placenta previa (6,410). Finally, the emergency, within labor indicators for c-section are abruptio placenta (64,121), fetal distress (6,563), cord prolapse (6,630), and dystocia (6,600, 6,601, 6,602, 6,603, 6,604, 6,608, 6,609).

From the Texas Closed Claim File, cases where a hospital or a doctor were sued by the parents of a newborn plaintiff, up to one year of age, were kept. Note that choice of OBGYNs and hospitals are thought to be jointly determined (see, e.g., Brown, Reference Brown III2001). These are referred to as ‘natal’ lawsuits. Cases where the indemnity payment, defined as the jury award or the negotiated settlement amount, was less than $25,000 could not be kept because the age of the injured party is suppressed by the state of Texas. Cases yielding $10,000 to $25,000 in settlements represent approximately 43% of the cases in the year 2000. It may be that lawsuit activity, no matter how small in terms of award or settlement, induces defensive medicine.

The Closed Claims File includes resolved lawsuits only, meaning that unresolved lawsuits may be outstanding. Further, lawsuit resolution can take years. For instance, the natal lawsuits included in our 2001 Closed Claim File were resolved in 2001, but most were filed in earlier years. Resolution took 2.22 years on average. Therefore, the data are not useful for precise time series measurements of lawsuit activity. This disadvantage is overcome by averaging lawsuits with indemnities over 14 years, thus creating a stable measure of county-level reputation for litigation. If there have been many lawsuits in a county over the entire 14-year period, one would expect the county to have a reputation as a lawsuit haven. Thus, OBGYNs working there may practice defensive medicine. On the other hand, near-term lawsuits may represent temporary fluctuations, which OBGYNs recognize. Lawsuit rates are created by dividing county aggregate lawsuits by the county number of OBGYNs in 1999 from the Area Resources File (ARF, 2001).

Note that 101 out of 254 counties in Texas had at least one OBGYN. However, the counties that have hospitals typically had OBGYNs and much higher lawsuit activity. Less than 1% of births in the entire database occurred in counties without an OBGYN and/or a natal lawsuit.

Note that malpractice insurance is not fully experience-rated for individual doctors (see, e.g., Sloan and Hassan, 1989). Note further that in Texas, many malpractice insurance firms vary insurance premia by wide areas. For instance, the Medical Protective Company varies its premium for only five regions that are self-defined. Therefore, the short-run effects of increases in malpractice premiums due to lawsuit activity are collective rather than individual OBGYN-specific. Of course, some OBGYNs with many suits are dropped from coverage and shifted to high-risk pools with high premiums in the longer term.

Multilevel analysis requires large computational capacity. Therefore, the analysis was conducted using a random sample, with replacement, of 20,000 births out of the total birth population of 260,292. These were then analyzed using GLLAMM in TATA (Rabe-Hesketh et al., Reference Rabe-Hesketh, Pickles and Skrondal2001).

Results

Table 1 provides the means and proportions for the variables included in the model. Note that they are sample means and proportions. Note that the overall c-section rate in Texas was 19.7% in 2002.

Table 1.

Means and proportions

During 1988–2001, there were on average 0.728 natal lawsuits per OBGYN. Figure 1 shows the per OBGYN natal lawsuit rate by county. There is significant area variation.

Table 2 shows the two models estimated by simple logistic regression. Model 2 is identical to model 1 except that it excludes the indicators during delivery variable.

Table 2.

Logistic regressions

Dependent variable: C-section (1=Yes; 0=No)

Let us examine the results from model 1. The odds ratio estimates for the demographic variables in Table 2 are as expected. Mothers 35 years of age and older are at greater risk for c-section delivery compared to women between 18 and 34 years of age. African-American women have more c-sections than the omitted group, similar to the national c-section rate for African-Americans (Hamilton et al., Reference Hamilton, Martin and Sutton2003). The difference achieves statistical significance. Hispanic women have fewer c-sections.

Of the health insurance variables in Table 2, Medicaid and Self-pay are significantly, negatively related to c-sections. The reference category in health insurance is private insurance.

The clinical indicators in Table 2 are very strongly positively associated with c-sections, especially, the indicators present prior to pregnancy. The primary variable of interest is the 1988–2001 per OBGYN lawsuit rate. It is significantly associated with the likelihood of c-section.

Now let us examine the results in model 2 of Table 2. Tussing and Wojtowycz (Reference Tussing and Wojtowycz1997) show that indicators during delivery, such as fetal distress and dystocia, are potentially endogenous. Therefore, the indicators during delivery variable is excluded. However the odds ratio estimates for the other independent variables change very little.

In Table 3, the multilevel results are given. Both the hospital and HRR-level variances of the error terms are significantly different from zero for all regressions considered. A likelihood ratio test of the significance of the error terms is performed, and the variance of the hospital and HRR region error terms are statistically significant. After controlling for hospital and HRR variation, the hypothesis that lawsuits per OBGYN are associated with defensive medicine in the form of more c-sections is statistically significant. However, the odds ratio estimate is of lower magnitude when multilevel logistic regression is used than when standard logistic regression is used (Table 2). A one standard deviation increase in the lawsuits per OBGYN variable is associated with a 1.073 increase in the risk of having a c-section performed when multilevel modelling is used. This is compared to an increase of 1.14 when standard logistic regression is used. When there is a two unit increase in lawsuits per OBGYN, there is a 1.24 increase in the risk of having a c-section performed when multilevel modelling is used. This is compared to an increase of 1.48 when standard logistic regression is used.

Table 3.

Multilevel logistic regressions

Dependent variable: C-section (1=Yes; 0=No)

Discussion and conclusion

The results support the hypothesis that more market area natal lawsuits against physicians and hospitals lead to more defensive medicine in the form of more c-sections. After controlling for HRR-level (market area) and hospital-level variation in a multilevel model, the association between natal lawsuits and c-sections is reduced even though it remains statistically significant. Failure to extract HRR- and hospital-level variations in the previous literature may have led to estimates of associations between area lawsuits and c-sections which were biased upwards.

In terms of policy, the results show that tort reform may lower costs by preventing defensive medicine. However, the reduced association between lawsuits and physician practice in the multilevel logistic model as compared to the standard logistic model implies that the potential savings may be lower than previously thought.

This research provides support to the view that idiosyncratic practice styles and differing tastes across regions in delivery style pervade many geographic areas and hospitals. While the optimal amount of c-sections performed can vary depending on the tastes of the local population, there remains substantial hospital- and HRR-variation. Both unobserved HRR- and hospital-level variation are important in delivery practice. The importance of the unobserved HRR- and hospital-level variation is shown. This result adds to the literature on regional- and hospital-level variation. Rather than focusing on tort reform exclusively, efforts to lower c-section rates should focus on other sources of small area variations.

A limitation of the study is the lack of information about physicians in the data. As Entman et al. (Reference Entman, Glass, Hickman, Githens, Whetten-Goldstein and Sloan1994) have shown, previous lawsuits against a physician are good predictors of subsequent lawsuits. Further, Entman et al. (Reference Entman, Glass, Hickman, Githens, Whetten-Goldstein and Sloan1994) have shown that a few physicians are the target of the vast majority of lawsuits. Another limitation of the paper is that the lawsuit data does not include unresolved cases and cases with low awards. Still, the results clearly suggest that hospital- and HRR-variation in c-section rates are important.

By failing to control for small area and hospital variation, the previous literature on defensive medicine may have reported biased estimates of the relationship between lawsuits and c-sections. This perhaps explains some of the mixed results reported in the literature.