Estimation and application in log-Fréchet regression model using censored data

This article introduces a new location-scale regression model based on a log-Fréchet distribution. Maximum likelihood and Jackknife methods are used to estimate the new model parameters for censored data. Martingale and deviance residuals are obtained to check model assumptions, data validity, and detect outliers. Moreover, global influence is used to detect influential observations. Monte Carlo simulation study is provided to compare the performance of the maximum likelihood and jackknife estimators for different sample sizes and censoring percentages. The empirical distribution of the martingale and deviance residuals of the proposed model is examined. A real lifetime heart transplant data is analyzed under the log-Fréchet regression model to illustrate the satisfactory results of the proposed model.

A Prognostic Score of Overall Survival In Adults with Acute Myeloid Leukemia

Abstract 2730 Background: Acute myeloid leukemia (AML) is a malignant myeloid disorder with heterogeneous outcomes. A number of factors have been shown to be prognostic; age, white blood cell (WBC), prior malignancy, performance status (Eastern Cooperative Oncology Group; ECOG) and cytogenetics. Methods: We planned to develop and validate a prognostic score for the 5-year (y) overall survival (OS) of adults with AML receiving intensive induction chemotherapy. We used Cox model to estimate the regression coefficients and Kaplan-Meier to estimate the 5-y OS. We used Cox-Snell, Schoenfeld and deviance residuals for model diagnostics and bootstrap validation to estimate the performance measures; Harrell's concordance and deviance residuals. Results: We retrospectively analyzed 779 patients treated between 1998–2008, using a prospectively collected database. The median age was 58 y. Most patients had intermediate risk cytogenetics (61%) and good performance status (ECOG 0–1: 79%). The median follow up for the surviving patients was 26.7 months (95% CI 18.8–32.9 months). The 5-y OS was 26% (22- 30%). All variables were statistically significant in the multivariable Cox regression model; age (y) (Hazard Ratio, HR 1.02; 95% CI 1.018–1.034), WBC (1*10^9/L) (HR 1.004; 1.002–1.006), prior malignancy (HR 1.58; 1.26–2.00), ECOG (ECOG 2 HR 1.41; 1.06–1.88, ECOG 3–4 HR 9.99; 4.72–21.18) and cytogenetics (intermediate risk HR 2.49; 1.41–4.39, poor risk HR 4.74; 2.65–8.50). The score divided patients into four risk groups; good (n=47), intermediate (n=129), poor (n=198) and extremely poor (n=87). The estimated 5-y OS was 0.70 (95% CI: 0.53–0.81), 0.37 (0.28–0.46), 0.15 (0.10–0.21) and 0.03 (0.01–0.10) respectively. The model showed good discrimination with large differences between survival curves and good Harrell concordance of 0.69. It showed good calibration using Cox-Snell and deviance residuals. In the intermediate risk cytogenetic group, the model showed good discrimination with over 45% difference in 5-y OS between the good and extremely poor groups. Conclusions: Our study confirmed the prognostic impact of the 5 variables reported in the literature. Using these factors, we developed a score to predict long term OS that showed good discrimination and calibration. The score added further discrimination in the intermediate risk cytogenetic group. Prospective external validation of the score is needed. Disclosures: No relevant conflicts of interest to declare.