Estimating Anesthesia and Surgical Procedure Times from Medicare Anesthesia Claims

Medicare data for patients 65 yr and older is the most representative healthcare data for the elderly in the United States because Medicare is an entitlement program. The only significant group of elderly citizens not represented in the Medicare claims is those who opted out of the Medicare fee-for-service arrangement and joined a Medicare-approved prepaid health maintenance organization. As part of the Surgical Outcomes Study,16–18we obtained the Medicare Inpatient (Part A), Outpatient Standard Analytic Files, and Physician Part B files for all admissions in general and orthopedic surgical Diagnosis Related Groups in Pennsylvania during 1995 and 1996. These files represent the fee-for-service Medicare population, comprising approximately 90% of all beneficiaries for 1995–1996.19,20We created a longitudinal record by including all inpatient and outpatient claims and physicians' claims during that time interval for each patient. Data also included the Medicare Vital Status File, American Hospital Association Annual Survey for 1996, and the Pennsylvania Health Care Cost Containment Council Hospital Discharge Database for similar years, which included the MedisGroups© (MediQual, Inc., Marlborough, MA) severity score to supplement the Medicare record.21Medicare electronically stores all bills submitted by hospitals, physicians, and all other caregivers, and all payments made. These billing data include limited diagnosis codes collected using International Classification of Diseases (ICD9-CM) coding and procedure specific codes using both ICD9-CM coding and the Healthcare Common Procedure Coding System (HCPCS).

Medicare claims files can be obtained through the Research Data Assistance Center, which is a contractor that provides free assistance to academic, government, and nonprofit researchers interested in using Medicare and/or Medicaid data for their research. The Research Data Assistance Center is staffed by a consortium of epidemiologists, public health specialists, health services researchers, biostatisticians, and health informatics specialists from the University of Minnesota.#

A patient's hospital bill from the Inpatient Standard Analytical file can be linked to the bills submitted by the physicians and other providers who took care of the patient during the hospital stay. Bills submitted by anesthesia providers are identified by the variable “provider specialty.” For example, when this variable is “5,” it identifies an anesthesiologist, and “43” identifies a Certified Registered Nurse Anesthetist. We use the variables “from date” and “through date” in the Inpatient File (Part A) and “expense date” in physician claims file (Part B) to link the information from the hospital stay to the physician bills submitted for services provided during the same stay. The inpatient file provides the ICD9-CM procedure codes for up to five procedures. For this study, we selected the principal procedure as determined in retrospect by Medicare. This procedure was identified in Part B through its corresponding HCPCS code, and its “expense date” in Part B was matched with the anesthesia HCPCS that had the same expense date as the HCPCS corresponding to the principal procedure.

There is an anesthesia time unit, defined as a 15-min interval, associated with each anesthesia provider claim.**††The time units are identified by the variable “mile/time/units/services indicator code.” When this variable equals “2,” it identifies anesthesia time units. In the documentation received with the electronic claims file, time units are reported as integers but should be interpreted as having one decimal. For example, a time unit value of 25 implies 2.5 time units × 15 min, or 37.5 min billed by the anesthesia provider.

According to the Medicare Claims Manual, Section 50G,**the anesthesia time “starts when the anesthesia practitioner begins to prepare the patient for anesthesia services in the operating room or an equivalent area and ends when the anesthesia practitioner is no longer furnishing anesthesia services to the patient, that is, when the patient may be placed safely under postoperative care.” Time units, unlike base units, are not modified by the performance and direction of concurrent anesthesia procedures.**Base units are assigned to each anesthesia HCPCS code and reflect the relative difficulty of the anesthesia procedure.‡‡Base units were not used for this analysis. Both the time and base units are used in determining payment from Medicare.

We obtained detailed chart abstraction data in a subset of patients from the same pool of patients for which we had claims data. The charts were abstracted as part of the Surgical Outcomes Study.16–18,22–24 

Four landmark times were abstracted from each available chart: induction, incision, closure, and entrance to recovery room. These abstraction definitions are not exactly those used by the nomenclature convention of the American Association of Clinical Directors,**nor are they consistent with “anesthesia-controlled time,”3,11,25because the intent of this manuscript was not to calculate operating room throughput and overall efficiency. Our landmarks were chosen because they reflect how reimbursement from Medicare is determined and because the abstractors for Medicare had used these landmarks in previous work auditing Medicare charts. We defined “anesthesia chart time” (n = 1,931 patients) as time of recovery room entrance minus time of induction and expressed this in minutes. We defined “surgical chart time” (n = 1,888 patients) as the time of closure minus the time of incision, also expressed in minutes. We excluded patients from the anesthesia time analyses if either induction or recovery room time was missing; we also excluded patients from the surgical time analysis if either incision or closure time was missing.

Each bill for anesthesia services for a patient for a surgical procedure date may include none, one, or more than one claim from the Medicare Part B files, an associated Current Procedural Terminology or HCPCS code, and a variable identifying the specialty of the anesthesia provider. We tested numerous definitions for variable construction concerning the total claim minutes used in the predictive model. We defined the claim time to equal the longest time billed by any anesthesia provider on the day of the surgery. If two anesthesia bills were found for the same date as the principal procedure, the longer bill was used in the modeling. We considered several other possible definitions of the claim time, but they worked poorly. For instance, using a summation of the bills, or modeling two bills as two variables, did not improve the simple rule of using the longest anesthesia provider bill. When we included nurse anesthetist bills separately, by adding a second variable to the modeling process, we found almost no change in fit (R²), and any coefficients on the second variable that were significant were extremely small, producing no appreciable difference in our estimates. Hence, we used a simple rule for defining claim time—the longest bill by an anesthesia provider on the same day as the principal procedure in question.

We used Kendall's τ as a measure of correlation between claim and chart times. Kendall's τ is useful because magnitudes of τ other than -1, 0, and 1 can be given a practical interpretation.26Consider two patients. For these two patients, the bill and the chart are said to be concordant if the patient with the longer chart time also had the longer bill time. The probability θ that two patients will be concordant is related to Kendall's rank correlation τ by the formula θ= (τ+ 1)/2. A perfect correlation yields a τ= 1, yielding a concordance probability of θ= (τ+ 1)/2 = (1 + 1)/2 = 1. Zero correlation, τ= 0, yields random or coin-flip concordance of θ= (τ+ 1)/2 = (0 + 1)/2 = 0.5, or a random chance that a longer claim time will also coincide with a longer chart time. We used the claim time to predict the chart time because researchers with data from Medicare will have the claim times and our algorithm for predicting the chart times from the claim times. The question then is how well that formula will perform. Associated with Kendall's τ is a form of simple linear regression yielding Theil's slope estimate.27When reporting results for individual hospitals or overall results with just a single claim variable, we used Theil's slope estimate, which performs well even in small samples, such as those for individual hospitals, and is insensitive to one or two wild observations in either x or y. When performing multiple regression, we used Huber's robust m-estimation as implemented in SAS Version 9 (SAS Institute, Inc., Cary, NC) using the bisquare weight function.28–30In least-squares regression, R²is the square of the Pearson correlation between observed and predicted values of y = chart time; however, because we performed robust, outlier-resistant regression, the use of the outlier-sensitive Pearson correlation is not appropriate. Therefore, in our robust regressions, we report as R²(or rank R²) the square of the Spearman rank correlation between the observed and expected y's, which is analogous to the square of the Pearson correlation between the observed and predicted ranks of y = chart time. This prevents one or two peculiar claims from increasing or decreasing the R².

There were 2,259 abstracted charts available for review from the Surgical Outcomes Study. The Surgical Outcome Study was a case-control study (two controls for each case) of mortality after surgery, so one third of the patients had died within 2 months of surgery. Of these, 2,007 have an anesthesia bill from Medicare Part B data. We excluded 30 charts with missing chart anesthesia time or negative chart anesthesia time and 40 charts with surgery time longer than anesthesia time. We excluded five charts for operations that were less than or equal to 10 min in length according to the anesthesia claim; we also excluded one operation with claim anesthesia bill time greater than 900 min. Of the 1,931 charts remaining, all included both induction and recovery room times, and 1,888 charts included incision and closure times as well as induction and recovery room times.

For each analysis in table 1, we let the anesthesia claim time represent the independent variable and predict anesthesia chart time as abstracted as part of the Surgical Outcome Study. We present five models: Model I uses only the claim time to predict chart time. Model I estimates chart times from the claim using a formula of the form: chart time =α+β (claim time). Model I is estimated in two ways. The first uses Theil's method,26the second uses Huber's m-estimation.31The remaining models are intended to verify the performance of Model I; they are not intended for practical use. If the bill time is to be used in lieu of the chart time, one would prefer that other patient attributes not alter the conversion formula. Models II-V use m-estimation, which allows multiple regression28,29and can produce a Wald test32on groups of coefficients analogous to an F test in linear regression. Model II uses claim time and patient procedures; Model III uses claim time and patient comorbidities; Model IV uses claim time and hospital indicators; and Model V includes all variables. Model I is the simplest model and performs as well as models that include many other predictors. Theil's estimate yields the equation (chart time) = y =α+βx = 0.85 + 0.97x = 0.85 + 0.97 (bill time), or less than 1 min added to the bill time. Hence, this equation suggests that using the claim-derived anesthesia time unit alone, without the coefficients in the Theil model, would produce very similar estimates of the desired chart time.

Furthermore, as can be seen in table 1, patient characteristics such as comorbid disease or procedure type did not significantly alter the predictions made by the equation that used bill time alone, and the anesthesia time units from Medicare are excellent estimates of the chart abstracted time. The R²statistics are unchanged to two digits regardless of the model used, and the slopes used to convert the anesthesia claims to chart minutes are almost identical. There is statistically significant variation in the bill-to-chart relationship among hospitals, but its impact on predictions for individual patients is small: the improvements in R²and the median absolute error of prediction were very small, and the estimated slope was essentially unchanged. The median absolute error of prediction without the hospital indicators was 5.49 min, and this decreased to 5.28 min using the hospital indicators, a difference that is statistically significant given the sample size but perhaps of little clinical relevance. When we replaced individual hospital indicators with an indicator variable for teaching status, we found that teaching was associated with a clinically small, but statistically significant, difference between claim and chart (−0.7 min, P < 0.0001). However, this difference did not influence the overall R²of the model. We also explored whether designation of urban or rural hospitals influenced the difference between chart time and claim time as determined from the anesthesia time unit information, and it did not. There were also no differences in errors when we compared procedures that were longer than the median procedure time with those that were shorter. These results suggest that using only Medicare anesthesia time units in the context of our algorithm for assigning bills to procedures is probably adequate for most comparisons focused on patients but may be more problematic when comparing, for example, two hospitals to determine which one performs a given operation more quickly.

Figure 1displays the overall data set with Theil's regression line fit through the points. The fit was very good, corresponding to a probability of concordance of 0.93. If one compares two patients, 93% of the time, the patient with the longer bill time also has the longer chart time.

Figure 2displays a box plot of the 1,931 converted bill times (anesthesia time units × 15 min) compared with the abstracted values of anesthesia time. Most errors are very small; in fact, the median absolute error was just 5 min, and 95% confidence interval (CI) for the median absolute error was 5-5.5 min. To the right of the box plot is the corresponding normal probability plot. Errors that were normally distributed would produce a straight line, and here we see that the data are symmetric but far from normally distributed, with long tails, as also evidenced by the fact that the Shapiro-Wilk test W for normality26was 0.69 (P < 0.0001). The lack of normality in these errors suggests that, when using the algorithm we suggest with the Medicare Part B claims, investigators should use robust regression techniques instead of ordinary least squares.28 

Table 2displays a set of regressions similar to those in table 1, but the dependent variable is anesthesia bill time minus chart time alone, instead of chart time, and the right side of the equations do not include bill times. In this set of analyses, we are investigating whether there are systematic errors in the billing of time data as a function of hospital characteristics such as teaching status or for-profit status. As shown in table 2, we did not see any appreciable differences among hospitals.

To better explore the differences in the estimation of chart time by claim time within and among hospitals, we estimated the relationship between anesthesia chart time and bill time separately for each hospital with at least 14 patients. Table 3presents Kendall's τ and the associated Theil's slope and intercept for each of the 20 hospitals with 14 or more patients in the Surgical Outcomes Study and for all hospitals together. For 19 of these 20 hospitals (that is, all hospitals except hospital 11), the estimated probability of concordance between chart and bill times was greater than 0.9, so for two patients in these hospitals, more than 90% of the time, the patient with the longer bill time will also have the longer chart time. The 20 slopes and intercepts estimated separately from the 20 hospitals were generally very similar despite the small sample sizes in many of the hospitals. We next investigated how different a claim-based estimate of anesthesia time per hospital would be from the actual abstracted anesthesia chart time. For each hospital in table 3, we computed the difference between the anesthesia time based on the bill using unadjusted anesthesia time units (1 unit = 15 min) and the observed anesthesia time based on chart review. Using the estimated anesthesia times based on claims produced, on average, estimates very similar to the observed chart times. Of the 20 hospitals in table 3, only 2 had median errors greater than 10 min (hospitals 4 and 11). Each of these displayed 19.5-min median errors, or exactly 15 min (1 time unit) more than the median error associated with all hospitals combined. The P  value in the last column of table 3is for a hospital dummy variable in robust regression controlling for comorbidity and procedure. Although the single equation works well for converting bill to chart times, several hospitals do exhibit statistically significant deviations.

We display the nine hospitals with the most study patients in our data set in figure 3. When estimating anesthesia claim time from chart time, we consistently found a coefficient that was approximately 1. As shown, this relationship was quite stable among hospitals. There are some outliers, that is, individual patients who fall far from the lines; therefore, robust statistical methods that give little weight to outliers are needed when working with bill times.

Unlike estimating anesthesia time, in which anesthesia providers submit bills based on minutes of care, we do not have such bills from the surgeon when estimating surgical time. Nevertheless, we chose to explore how well the same anesthesia bill metric, which worked so well at estimating chart anesthesia time, would succeed at estimating surgical chart time. For each analysis in table 4, we let the anesthesia claim time represent the independent variable to predict surgical chart time as abstracted as part of the Surgical Outcome Study. We again present five models: Model I uses only the anesthesia claim time to predict surgical chart time; Model II uses claim time and patient procedures; Model III uses claim time and patient comorbidities; Model IV uses claim time and hospital indicators; and Model V includes all variables. Model I is the simplest model and performs as well as models that include many other predictors. The linear equation to estimate surgical chart time from claim time using Theil's estimate was equal to –23.81 min + 0.82*anesthesia claim time minutes, and using the robust estimate, the equation was –27.63 min + 0.84* anesthesia claim time minutes. The negative constant term, along with a slope less than 1, makes good sense because patients are under anesthesia before and throughout the operation. As shown in table 4, patient characteristics such as comorbid disease, procedure type, and hospital did influence the prediction of chart time after accounting for claim time, but the effect was very small, as evidenced by the similar R²for each model.

To better explore the differences in the estimation of chart time by claim time within and among hospitals, we computed Theil's regression models for each hospital individually and among all hospitals. As in table 3, table 5presents Kendall's τ and the slope and intercept for each hospital with 14 or more patients in the Surgical Outcomes Study and also among all hospitals in the Surgical Outcomes Study. The estimates of chart time using claim time are somewhat less reliable within hospitals for surgery than for anesthesia.

Figure 4displays the overall data set with Theil's regression line fit through the points. Anesthesia claim times are more similar to anesthesia chart time than is anesthesia claim time to surgical chart time, and there was more variability in the overall association between anesthesia claim and surgical chart time than for anesthesia chart time (compare with fig. 1). Finally, in figure 5, we display individual hospitals with their Theil fitted slopes and intercepts. Again, the relationship between claim time and chart time is quite reliable, although there are occasionally large errors in some patient predictions.

Finally, we explored what percentage of anesthesia time could be explained by surgical time. The correlation between anesthesia chart time and surgical chart time was computed for 1,888 charts that had all four landmark times. The Spearman rank correlation was 0.904. Hence, approximately 82% of the variation in anesthesia time can be explained by surgical time.

We found that Medicare anesthesia claims can be used to estimate anesthesia chart time using a simple algorithm that is based on the longest claim in the Medicare bill. Anesthesia and surgical times are often discussed and analyzed in the anesthesia and surgical literature, but except for the Veterans Affairs studies,10few large-scale studies have been reported among institutions. The potential use of easily available claim time variables from Medicare has never been formally tested on a large number of charts before this study. We can now say with some confidence that Medicare claim information, at least in the years 1995 and 1996, provide a useful proxy for anesthesia time as recorded by the anesthesia team in the chart. Given that the median anesthesia time is longer than 130 min in our study, the median absolute errors of approximately 5 min for anesthesia claims predicting anesthesia chart time and approximately 14 min for predicting surgical chart time are quite small. The possible uses of this information are manifold.

In earlier work, Abouleish et al.  1have demonstrated an excellent correlation between minutes as reported from billing departments at four different hospitals and their “surgical time” recorded from the chart (corresponding to our “anesthesia chart time”). They reported a correlation of 0.85 and displayed a figure that shows an apparent slope near 1 relating the two quantities. Our results, using Medicare claims obtained from Medicare and chart review of anesthesia and surgical time, display a similar correlation and slope but are derived from Medicare claims data and not directly obtained from hospital billing departments. As we described earlier, such Medicare claims are readily available through the Centers for Medicare and Medicaid Services as facilitated by the Research Data Assistance Center.#

In our data, we had both anesthesia bill times from Medicare and anesthesia chart times from chart abstraction, whereas in typical applications, only the bill times will be available. In our data, we were able to predict the chart time with good accuracy from the bill time by a simple linear equation. To predict anesthesia chart time, further adjustments for comorbidities and procedures were not needed; they did not improve the prediction compared with using bill times alone. Two cautions about the use of anesthesia bill times for chart times are important. First, the relationship between chart and bill times had a moderate number of outliers, that is, isolated patients with incommensurate chart and bill times. For this reason, when working with bill times in place of chart times, one should use robust statistical procedures, such as m-estimation,28,33,34that limit the influence of individual patients. In an analysis of thousands of bill times from Medicare, it is very likely that one or two patients will have bizarre bill times, which convert to bizarre estimated chart times, even though there was nothing bizarre in the underlying clinical practice. Robust statistical methods will give limited weight to these few bizarre times. Second, we did find some differences among hospitals in the relationship between chart and bill times: these differences were only moderate in size or clinical significance, but they were often statistically significant. This has two consequences. First, if one finds a clinically small but statistically significant difference between two hospitals in their chart times predicted from their bill times, one should not jump to the conclusion that there is a corresponding difference in clinical practice at the two hospitals; it may instead reflect a small but highly systematic difference in billing practices at the two hospitals. Second, in studying patient or clinical issues related to anesthesia time, it is safest to incorporate into the analysis adjustments for hospitals, such as matching patients from the same hospital, stratification on the hospital, suitable modeling, or combinations of these.

In summary, noting these caveats, it seems that the study of Medicare anesthesia claims has the potential to provide new insights into the practice of anesthesiology and surgery and variations among hospitals, providers, and their patients undergoing surgical procedures.

We thank William J. Greeley, M.D., M.B.A. (Anesthesiologist-in-Chief, Department of Anesthesiology and Critical Care Medicine, The Children's Hospital of Philadelphia, Philadelphia, Pennsylvania), and Lee Fleisher, M.D. (Chair, Department of Anesthesiology and Critical Care Medicine, The University of Pennsylvania School of Medicine, Philadelphia, Pennsylvania), for their aid in conducting this research. We thank the anonymous reviewers for their insightful comments and suggestions regarding this manuscript. We also thank Traci Frank (Administrative Coordinator) and Andrea Millman, B.A. (Research Assistant), from the Center for Outcomes Research, The Children's Hospital of Philadelphia, Philadelphia, Pennsylvania, for aid in preparing this manuscript.