diagnoses, because detailed data are unavailable from discharge data sets. For some procedures, problems in surgical technique may be the crucial factor, while for other procedures, inadequate postoperative monitoring may cause poor outcomes.

Even if physician volume is most important, hospital volume is likely to play a role. For example, a hospital with several high-volume and several low-volume surgeons may develop monitoring methods and standard procedures for the staff that catch errors and institute corrective actions. Thus, a low-volume surgeon may be "protected" in a high-volume hospital. Likewise, a surgeon with a high volume across several institutions but low volumes in each may achieve good results. The empirical testing of such hypothetical relationships is difficult because of the need to track data on the same physicians across hospitals.

Volume may not matter at all, but instead may serve as a marker for hospitals or physicians with special skills whose better-than-average performance attracts a disproportionate share of the referrals. This "selective-referral" hypothesis holds that any inverse relationship between volume and outcome arises from the attraction of more patients to physicians and hospitals with better outcomes. The idea that patients in some instances may look for hospitals or physicians with the best results seems implausible to some, who claim that the variation in mortality by disease or procedure is too small to influence patients' choice (218). If complications are correlated with mortality, however, variations in outcomes may be large enough to be noticed by patients' primary physicians who choose specialists for referral. Although it is difficult to identify an individual hospital or physician as having significantly worse than average death rates (396), referral patterns may be based on a simpler set of decision rules. If primary physicians switch referrals after even one "bad outcome," patients eventually are directed away from providers whose outcomes are worse than

average.

Furthermore, even if the majority of patients go to the nearest hospital or otherwise make decisions independently of perceived outcomes, a minority seeking or referred to the "best provider

in town" (or referred away from "poor-quality providers") will result in a selective referral pattern for specific diagnoses and procedures. As a result, hospitals with better outcomes would have higher-than-expected volumes. The question, therefore, is whether some patients are influenced in their choice of physicians and hospitals by relative performance, not whether all patients are so influenced.

Another principally empirical objection to the selective-referral hypothesis is that some studies show little relationship between outcomes and hospital characteristics traditionally considered to be markers of good performance, such as teaching status or board certification of physicians (217,393). However, these measures are rather blunt and unvalidated indicators of special expertise. It is common for a teaching hospital to be outstanding in the treatment of one diagnosis or procedure (e.g., cardiovascular surgery) but not to be particularly distinguished in another (e.g., neurosurgery).

When one attempts to test in a simultaneousequation model both the effects of volume on outcomes and the effect of outcomes on volume, one may observe statistically significant effects for only one causal path. Even if the results indicate just an effect of outcome on volume in such a model, a simple test of volume as a function of outcome alone would probably show a relationship. There is not yet enough work to clearly indicate which causal paths are truly valid.

In designing an experiment, one should undertake a power test (ideally ahead of time) to determine the likelihood of detecting an effect if one truly exists. A power test is based upon the overall likelihood of the outcome's being measured and the sample size. There are substantial differences across studies in the number of patients involved and the average poor outcome (or mortality) rate.

To provide a sense of the issue at hand, consider the research findings from the 11 studies that reported on the hospital-volume/outcome relationship for the total hip replacement procedure. Eight studies showed a relationship between worse outcomes and low-volume hospitals, while three studies found no effect of volume on outcome (see table 8-2 and figure 8-3). The three studies that

showed no effect had smaller sample sizes-under 1,500 patients in two studies and under 10,000 patients in the other study-than the sample sizes of from 13,700 to 33,000 patients in the eight studies that did find an effect. The three studies that had findings inconsistent with the hypothesized volume-outcome relationship probably had insufficient power to detect an effect unless it was very large. The mixed results for total hip replacement are not surprising given the design of the studies.

In summary, the available studies reviewed by OTA provide rather substantial evidence that worse outcomes occur at lower volumes for most of the procedures and diagnoses that have been studied. However, the volume-outcome relation

ship is not universal. For stomach operations and fractures of the femur, the evidence of a relationship is quite mixed, with the majority of studies indicating that volume has no effect on outcome. With the exception of the findings for stomach operations and femur fractures, all the other findings that suggest the lack of a relationship between volume and outcome either have low statistical power; are part of larger analyses in which a physician volume effect is found; or suggest a causal linkage from outcome to volume. Thus, although a relationship often exists, there is not yet enough evidence to distinguish effects due to physicians from effects due to hospitals or to have much confidence in the relative importance of the causal linkages.

As has been discussed, there is frequently a relationship between volume and outcome. The general pattern is that better patient outcomes are associated with higher inhospital volumes. However, because there is hardly ever a perfect relationship, there are always some low-volume hospitals with apparently good outcomes and some high-volume ones with poor outcomes. This situation raises the obvious question, "How useful is volume as an indicator of the quality of care?” Since mortality data on Medicare patients are routinely available, why bother with volume data?

There will always be some chance component to a hospital's reported death rate in any single year, even after all adjustments for patient characteristics have been included. Various statistical calculations are designed to provide measures of this chance component and thus the degree of confidence one should have in the observed results for a particular hospital. It is inherent in the nature of small samples that one must expect much more variability in observed outcomes in hospitals with low volumes. One death among 10 or 20 patients may produce a mortality rate well above the average, but it is likely to be a chance occurrence. Similarly, even if the true or long-run mortality rate for that hospital is worse than average, with few patients in any particular year, there will often be years in which there are no deaths.

To get a better estimate of the true performance of the outcomes in a low-volume hospital, one might aggregate data over several years, if they are available. Unfortunately, this technique makes it impossible to determine whether outcomes are improving or getting worse.

Combining data on volume and outcome is an alternative way of organizing a given amount of data to reduce the influence of chance and provide useful information. By aggregating data across hospitals within volume categories or using a regression to smooth out hospital-specific variability, the volume-outcome studies provide much more stable estimates of the performance of a class of hospitals. Although average results for all lowvolume hospitals may not apply to a particular low-volume hospital, it is important to remember that, because of chance variability, last year's mortality rate for a particular hospital is not a very reliable indicator either. The two pieces of information, however, may be used together to guide a decision about a particular hospital.

The situation is different for high-volume hospitals, because the role of chance is smaller the larger the number of patients. Of course, hospitalspecific mortality results will still be sensitive to unmeasured differences in patient characteristics that may not be adequately captured in the available data. If a high-volume hospital with worse

than-average outcomes claims that unmeasured patient-related factors account for the poor results, that claim may be worth more detailed investigation.

If high volumes for a particular procedure or diagnosis are primarily the result of superior outcomes, then the argument for volume data is even stronger. Since published hospital mortality data have only recently become available (see ch. 4), a relationship between volume and outcome implies that physicians (and possibly patients) have been able to use informal qualitative measures to guide more patients to physicians and hospitals with better results. Primary care physicians may consider both the mortality and other complications of their patients referred to certain specialists. Observations in the operating room or at the bedside may also alter one's confidence in the quality of care provided by specific physicians. Although such methods may be somewhat haphazard, they allow for a wide range of implicit but important criteria that may be valuable in the identification of which providers to seek out and which ones to avoid. It would be impossible to collect and make available such data, but if selective referral occurs, then the observation of a higher than expected volume of patients with diagnosis X in a hospital may be a valuable indicator of better-than-average quality.

It is important to note, however, that to use volume as an indicator of the quality of care, one must control for the various factors that influence volume. Large hospitals, for example, tend to have more patients of most diagnoses than small hospitals, irrespective of their relative quality. Public hospitals tend to treat a disproportionate share of diagnoses common among poor people. Selective contracts between certain payers and hospitals will also alter volumes. In much the same way that hospital-specific mortality rates are meaningless as outcome indicators until adjusted for case mix and certain other factors, hospital volumes are meaningless until adjusted for factors such as size of hospital, ownership, medical staff, and selective contracts. Although analyses with such adjustments have not yet been undertaken, they may be worth pursuing, especially for diagnoses and procedures for which there is evidence of selective referrals.

One additional use of volume as an indicator of the quality of care arises from the possibility of a volume-outcome relationship for physicians. Fewer studies have examined the volume-outcome relationship for physicians than have examined it for hospitals. Furthermore, the results for physicians are less consistent than those for hospitals, although some of the inconsistency may be due to methodological problems that can be overcome with better data and more analysis. Moreover, the problems of chance variation in small numbers of patients would make physician-specific data on mortality rates even less reliable than hospital-specific data. Volume data for physicians, however, may be far less controversial than outcome data. Thus, work on the volume-outcome relationship and familiarity with the use of hospital data could help set the stage for the use of physician volume data as an additional guide for consumers.

In choosing a physician or hospital, consumers should not just "go by the numbers." Instead, if there is good evidence of a volume-outcome relationship for the patient's specific diagnosis or prospective procedure, the patient should discuss the information with a primary care physician. Suppose, for example, that a physician is recommending that a patient have CABG surgery and there are several hospitals in the community with openheart surgery teams. Even if hospital-specific mortality data are available, there may be questions as to how they should be interpreted if none of the hospitals have significantly high or low mortality rates. As proximity is not a major issue if there are several local hospitals and if the mortality rate (3 to 5 percent) is not trivial, the patient may want to find the best, or at least avoid the worst, institution.

Suppose the hospital initially selected had a low (but not significantly so) mortality rate last year, but this rate was based on only a small number of cases. If this hospital also had a low volume, it would be reasonable to press the physician on whether one of the higher volume centers with comparable mortality rates might not be more likely to have a lower true risk of a poor outcome. Such a question may encourage the physician to think further about the referral and perhaps informally seek out additional information about

the best hospital to send the patient. Although this is a rather "soft" use of information, it is probably commensurate with the precision of the available data.

In using information about the relationship between volume and outcome, it is important to know the form of the curve for a particular diagnosis or procedure. In the analysis in this chapter, all findings with dichotomous results and with "downward-sloping," "L-shaped," and "U-shaped" curves were grouped together. If there truly is a "U-shaped" curve, then it is necessary to identify the volume level above which mortality rates begin to worsen. Several studies have estimated "Ushaped" curves, but none have tested whether a "U" was really superior to an "L" or similar form. Nor did the studies find much evidence that very high-volume hospitals actually had worse results. The only exceptions are the studies of outcomes for newborns by Rosenblatt, et al. (538) and Williams (702). In both instances, the authors argued that the apparently worse outcomes for newborns in the very high-volume hospitals could be attributed to the very high-risk infants referred to those hospitals pursuant to perinatal regionalization policies. Unless additional studies provide clear evidence that worse outcomes occur in very high-volume centers, the public need not worry too much about reports of "U-shaped" curves.

Even if outcomes do not get worse in very highvolume hospitals, available volume-outcome studies do not necessarily imply that more is better. In many instances, the rule might be: Avoid the very low-volume setting; once you find a hospital with a volume of X, there is little to be gained by looking for a hospital with higher volume. To make recommendations about specific optimal volumes would require analyzing up-to-date data on specific diagnoses and procedures across a wide range of hospitals. Unfortunately, the available published studies do not present such analyses, but the data are generally available and it would be relatively simple for an experienced research group to undertake the necessary analyses and make public the findings.

To provide a sense of how data might be presented, consider figure 8-4. (Similar data are published in a consumers' guide in the Washington, DC area (693).) The figure indicates age- and sex

adjusted mortality rates for patients undergoing CABG surgery in hospitals with various volumes and also shows the confidence intervals, the ranges in which mortality rates would be expected to fall if volume were not a factor. (Although adjusting for risk factors other than age and sex would improve the quality of the data, the presentation could be similar.) Mortality rates in the very highest volume hospitals are significantly lower than expected; part of the reason is that at higher volumes, the confidence interval narrows. Because hospital-specific mortality data are more reliable at high volumes, however, the volume data for hospitals with high volume are less valuable. Also, patients will be less willing to switch hospitals for the relatively small incremental improvement in expected mortality associated with very high-volume, in contrast to medium- or highvolume, hospitals.

Figure 8-4 also shows that patients undergoing CABG surgery in low-volume hospitals experience significantly higher than expected mortality rates. The difference not only is statistically significant, but it amounts to a half-again higher