BACKGROUND: A previously published clinical trial of epidural-supplemented versus general anesthesia, Veterans Affairs Cooperative Study No. 345, showed no difference in 30-day mortality and morbidity rates between the two treatments. We hypothesized that long-term postoperative survival would be increased by epidural anesthesia/analgesia supplementation during colon cancer resection.METHODS: We studied long-term survival after resection of colon cancer in a trial of general anesthesia with and without epidural anesthesia and analgesia supplementation for resection of colon cancer in Veterans Affairs Cooperative Study No. 345. Cox and log-normal survival models were used to test the effects of pathological stage, type of anesthesia and other covariates on survival in 177 patients.RESULTS: The presence of distant metastases had the greatest effect on survival. Thus, analyses were performed separately for patients with and without metastases. For those without metastasis, the hazard ratio for the treatment effects changed at 1.46 years. Before 1.46 years, epidural supplementation was associated with improved survival (P = 0.012), while later, the type of anesthesia did not appear to affect survival (P = 0.27). Hypertension was associated with poorer survival (P = 0.029), as was alcoholism in patients who received epidural anesthesia (P = 0.014). Survival of patients with metastases was unaffected by type of anesthesia. There was a significant age by hypertension interaction (P = 0.002). Patients survived longer if they were hypertensive, but had reduced survival if they were older than 66 years and hypertensive.CONCLUSION: Epidural supplementation was associated with enhanced survival among patients without metastases before 1.46 years. Epidural anesthesia had no effect on survival of patients with metastases. Additional studies to confirm or refute these findings are warranted..

17015 Background: VA Cooperative Trial 345 randomized patients having abdominal surgery to unsupplemented general anesthesia (UGA) or epidural-supplemented general anesthesia (ESGA). The long-term clinical significance of the type of anesthesia for patients having surgery for colon cancer has not been well studied. We compared survival of patients with colon cancer randomized to ESGA to survival of those randomized to UGA. Methods: Survival of colon cancer patients was not a primary trial end point. We conducted a post hoc analysis of the colon cancer patients in the trial using TNM staging data from patient records. A Cox survival model was used to test the effects of pathological stage, type of anesthesia and other covariates on survival. Results: 177 patients were evaluable. In both trial arms, survival was similar (P = 0.23) until 4.56 years. After 4.56 years, patients assigned to UGA had significantly better survival (P = 0.01). In both trial arms, the survival in patients without metastases was similar initially, but those who received UGA had significantly better survival later. A similar statistically significant pattern was evident in patients with metastases. The long-term survival benefit of UGA was even more pronounced among patients with metastases and a diagnosis of alcoholism. Conclusions: This study provides the first evidence that ESGA may worsen long-term survival in colon cancer patients, especially those with metastases and alcoholism. Epidural anesthesia affects visceral blood flow, immunological parameters, post-operative pain and other physiological variables, which could explain these findings. Additional studies to confirm or refute these unexpected findings are warranted. No significant financial relationships to disclose.

A new view of the maximum likelihood estimator (MLE) of exponential scale for censored data is presented. This is done by adapting Reid's (Ann. Statist. 9 (1981) 78) approach for obtaining the two influence functions (IF) for the Kaplan–Meier estimate of the survival function; one for uncensored and one for censored data, respectively. The MLEs two IFs are derived. Via this analysis, we propose a new robust estimator, the scaled α-Winsorized estimator (WE). Under Type II censoring, the WE is the MLE and, hence, is asymptotically efficient in that case. Its two IFs are bounded; hence,WE is B-robust. Its breakdown point is α. A comparison is made with respect to asymptotic bias and mean square error at contaminated exponential and Weibull survival models.

The method of least trimmed absolute deviations (LTAD) is formally treated. This method, like the LMS and LTS methods of Rousseeuw (1987), uses a criterion function that is calculated over half-samples. A population analogue is defined and its properties are discussed. The asymptotic properties of the LTAD estimator are derived. It is shown to be consistent and asymtotically normal. It is also shown to be, with asymptotic certainty, the midpoint of the interval spanned by the selected half-sample. Further, it is shown how to construct a distribution-free tolerance interval for predicting the next observation. Finally, comparisons of the asymtotic variances for the LTS and LTAD are made and recommendations are given.

The influence functions for Rousseeuw's (1987) least trimmed squares (LTS) estimator and for Tableman's (1994) least trimmed absolute deviations (LTAD) estimator are derived in the univariate case. The half-sample estimators which possess, by construction, the 50% breakdown point property satisfy three of the four robustness criteria defined by Hampel et al. (1986). They have bounded influence functions, finite gross-error sensitivity, and finite rejection point. However, they have infinite local-shift sensitivity. Hence, these estimates can be highly sensitive to small perturbations in the data. Small shifts in centrally located data (inliers) can cause their values to change by relatively large (though bounded) amounts.

The Krasker-Welsch (1982) approach to bounding influence is merged with rank regression. I propose a one-step estimator that is analogous to Bickel's (1975) one-step M estimator of Type 1 but uses weights that depend on the design vector and the residuals to reduce the influence of outliers. In fact, the standardized sensitivity of the estimator is made equal to a prechosen constant. It is based, however, on a second-derivative approximation to a dispersion surface that is not convex. This one-step variant avoids the problem of multiple roots. The estimator is shown to be consistent and asymptotically multivariate normal. An example shows that it yields results similar to those of the Krasker-Welsch estimator.

SUMMARY This paper discusses the two-sample test of location based on the comparison of two distribution-free one-sample confidence intervals derived from sign statistics. This test procedure, first introduced by Hettmansperger (1984), rejects the null hypothesis of equal population medians when the two intervals are disjoint. He presents three different ways to select the two one-sample intervals and one choice leads to Mood's test. All solutions have the same Pitman efficiency. This paper shows that the choices can be distinguished on the basis of Bahadur's efficiency. We formulate the problem in terms of (asymptotically) fixed width confidence intervals. In this context various median tests (including Mood's test) arise as special cases and they yield different performance. The solution that specifies equal asymptotic lengths for the one-sample intervals (which is different from Mood's test) is recommended.

A two-sample test is studied which rejects the null hypothesis of equal population medians when two Wilcoxon distribution free confidence intervals are disjoint. A confidence interval for the difference in population medians is constructed by subtracting the endpoints of two one-sample confidence intervals. Two different ways to select the one-sample intervals are presented. A solution that specifies equal confidence coefficients for the one-sample intervals is recommended. All solutions are shown to have the same asymptotic (Pitman) efficiency as the Mann-Whitney two-sample test.

On presente une methode generale simple pour construire un intervalle de confiance de longueur bornee pour la difference des medianes de population. On obtient un intervalle a deux echantillons