A. James O Malley, PhD, Kelly H. Zou, PhD, Julia R. Fielding, MD, Clare M. C. Tempany, MD. Acad Radiol 2001; 8:
|
|
- Victoria Bishop
- 6 years ago
- Views:
Transcription
1 Bayesian Regression Methodology for Estimating a Receiver Operating Characteristic Curve with Two Radiologic Applications: Prostate Biopsy and Spiral CT of Ureteral Stones 1 A. James O Malley, PhD, Kelly H. Zou, PhD, Julia R. Fielding, MD, Clare M. C. Tempany, MD Rationale and Objectives. The authors evaluated two Bayesian regression models for receiver operating characteristic (ROC) curve analysis of continuous diagnostic outcome data with covariates. Materials and Methods. Full and partial Bayesian regression models were applied to data from two studies (n 180 and 100, respectively): (a) The diagnostic value of prostate-specific antigen (PSA) levels (outcome variable) for predicting disease after radical prostatectomy (gold standard) was evaluated for three risk groups (covariates) based on Gleason scores. (b) Spiral computed tomography was performed on patients with proved obstructing ureteral stones. The predictive value of stone size (outcome) was evaluated along with two treatment options (gold standard), as well as stone location (in or not in the ureterovesical junction [UVJ]) and patient age (covariates). Summary ROC measures were reported, and various prior distributions of the regression coefficients were investigated. Results. (a) In the PSA example, the ROC areas under the full model were 0.667, 0.769, and 0.703, respectively, for the low-, intermediate-, and high-risk groups. Under the partial model, the area beneath the ROC curve was (b) The ROC areas for patients with ureteral stones in the UVJ decreased dramatically with age but otherwise were close to that under the partial model (ie, 0.774). The prior distribution had greater influence in the second example. Conclusion. The diagnostic tests were accurate in both examples. PSA levels were most accurate for staging prostate cancer among intermediate-risk patients. Stone size was predictive of treatment option for all patients other than those 40 years or older and with a stone in the UVJ. Key Words. Receiver operating characteristic (ROC) curve; Bayesian regression analysis. Acad Radiol 2001; 8: From the Department of Health Care Policy, Harvard Medical School, 180 Longwood Ave, Boston, MA (A.J.O., K.H.Z.); and the Departments of Medicine (A.J.O.) and Radiology (K.H.Z., J.R.F., C.M.C.T.), Brigham and Women s Hospital, Harvard Medical School, Boston, Mass. Received November 30, 2000; revision requested March 20, 2001; revision received and accepted April 2. Supported in part by grants U01 CA45256 and P01 CA41167 from the National Cancer Institute. Address correspondence to A.J.O. AUR, 2001 Traditionally radiologic diagnostic tests have tended to yield categorical rating data. With more precise measurement tools, however, laboratory diagnostic systems yield a continuous measurement scale for the test outcome. The most important predictor variable for diagnostic accuracy is the underlying truth about disease status (gold standard), which indicates the true disease status of each individual. Additional covariates, such as age and other risk factors, are often collected. In this article, we propose Bayesian regression methods for evaluating the accuracy of such a diagnostic system with covariates incorporated. These new methods are motivated by two radiologic studies. The first study concerns the use of prostate-specific antigen (PSA) levels for staging prostate cancer, with Gleason scores as a covariate (1). The second study concerns the use of ureteral stone size for predicting treatment option, with 713
2 O MALLEY ET AL Academic Radiology, Vol 8, No 8, August 2001 stone location in the ureterovesical junction (UVJ) and patient age as two covariates (2). The accuracy of a diagnostic test can be summarized in terms of a receiver operating characteristic (ROC) curve, which is a plot of sensitivity (or true-positive rate) versus (1 specificity) (or false-positive rate) at all possible decision threshold values (3 5). Notations and assumptions for an ROC curve will be given in the Materials and Methods section. In this article, Bayesian regression methodology will be introduced for ROC analysis derived from continuousoutcome measurement data, with patient covariate information and prior knowledge about the regression coefficients incorporated into the proposed regression models. Our Bayesian regression approach provides a new inferential framework for continuous-outcome data. Note that our approach may be extended to ordinal data, because continuously distributed test outcome data have a natural categorization that preserves all information relevant to ROC-curve fitting (6). Under the non-bayesian framework, an ordinal regression model for categorical outcomes has been developed (7), regression analysis methods for general outcomes have recently been proposed (8 10), and hierarchical random-effects models for ordinal categorical data have been discussed (11,12). Bayesian methodology has previously been developed for ordinal rating data (13 15). Whereas previous studies focused on the implementation of Bayesian methods, we also consider the sensitivity of the results to the choice of prior. Bayesian approaches are different from the more conventional frequentist approaches. In essence, given the observed data, once the prior distribution and the statistical model are specified, Bayes theorem is applied to derive posterior distributions of the parameters in the model or functions of these parameters (16). The motivations behind adopting a Bayesian approach are threefold. First, prior knowledge or pilot information can easily be incorporated into the models. In radiologic studies prior information about diagnostic tests can often be extracted from pilot studies, meta-analysis of relevant literature, or scientific theory. The incorporation of prior information often results in inferences that are more precise than obtained with frequentist methods if such prior information and experimental data cohere. Second, the parameters in the model or complex functions of them, such as summary ROC measures, may be simulated directly via methods for the exploration of posterior distributions. The most frequently used method is the Markov-chain Monte Carlo method via the Gibbs sampler and the Metropolis-Hastings algorithm (16,17). Recent advances in numerical methods and the availability of free software programs, such as BUGS (18), allow for faster and more efficient Bayesian computation. Finally, our Bayesian approaches for continuous-outcome data can also be extended to more complex designs and other scenarios. The benefit of Bayesian hierarchical approaches for ordinal data from multireader, multimodality studies has been illustrated previously (15). Note that the Bayesian treatment of the regression model in this article is a simple case of a hierarchical regression analysis. MATERIALS AND METHODS Clinical Example 1: PSA Level in Prostate Cancer As part of a multicollaborative Radiologic Diagnostic Oncology Group trial, magnetic resonance imaging was performed in 213 patients with prostate cancer, before which PSA levels and Gleason scores were obtained in 180 cases. PSA level was treated as the outcome variable. Radical prostatectomy was performed in all cases to provide the gold standard, and patients were classified into two groups, those with local disease (stage A or B) and those with advanced disease (periprostatic invasion of tumor and spread of disease to the seminal vesicles and lymph nodes) (1). Gleason score was treated as an additional covariate, with patients classified as low risk (Gleason score 6), intermediate risk (Gleason score 7), or high risk (Gleason score 8) (19). Clinical Example 2: Spiral CT of Ureteral Stones One hundred unenhanced spiral computed tomographic (CT) scans were obtained to evaluate flank pain in patients with obstructing ureteral stones documented by means of chart review (2,20). A standard protocol was used (280 ma; 12 kvp; pitch, ). The imaging thickness was 5 mm, with images reconstructed at 5-mm increments. Two radiologists initially reviewed the CT scans independently and blindly, to detect several imaging findings. The size of the stone was treated as the outcome variable. Treatment option, the gold standard, included spontaneous passage and required surgical intervention. The ROC analysis conducted previously suggested that in-plane stone size (in millimeters), followed by stone location (for simplicity, in or not in the UVJ), were the features most predictive of the need for intervention. Therefore, additional covariates were stone location in the UVJ and patient age. 714
3 Academic Radiology, Vol 8, No 8, August 2001 BAYESIAN REGRESSION METHODOLOGY FOR ROC CURVES Assumptions and Notations for an ROC Curve We assume that subjects are drawn from one of two distinct populations (ie, nondiseased and diseased), classified according to the gold standard. Furthermore, at each distinct covariate value, the diagnostic outcome measurements, or their transformed versions via the same monotone transformation, have a normal distribution. We now give mathematical notations for ROC analysis: Let Y i denote the test outcome for the ith subject (i 1,...,n). We now denote Y as a generic outcome variable. There are two categories of covariates or predictors. First, associated with each subject is the gold standard (or disease status) labeled D, where D 1 if this subject is diseased and 0 otherwise. This is the most important predictor for discriminating between nondiseased and diseased populations. Second, each patient also has an additional set of covariates labeled X ij with j 1,...,k, possibly extracted from his or her medical record. These covariates, denoted generically as X, consist of all relevant information known about this subject. Given any threshold (or cutoff value), t, let F(t D, X) be the probability that the outcome measurement is less than or equal to t, for a subject with gold standard D and covariates X. The corresponding sensitivity (or true-positive rate, q) of the test at this threshold value is defined as F t D 1, X 1 F t D 1, X. Likewise, the corresponding (1 specificity) (or falsepositive rate, p) is defined as F t D 0, X 1 F t D 0, X. At any given t, the resulting point on the ROC curve, a function of both p and q, is given by p t, q t F t D 0, X, F t D 1, X, for t R. Or equivalently, at any given p, with q viewed as a function of p, p, q p p, F F 1 p D 1, X, for p 0, 1. Two Regression Models The test outcome variable, Y, is assumed to have conditionally independent normal distributions. Its underlying mean, DX, depends on the gold standard (D) and additional covariates (X). For simplicity, the underlying variances, the 2 s, depend only on D. Two regression models are assumed to relate the expected value of the outcome to the gold standard and all other covariates. The full-regression model contains all interaction terms between disease status and covariates, as well as interactions among covariates. In comparison, the partial-regression model omits disease-covariate interaction terms. (These models are referred to hereafter as the full model and the partial model.) In the PSA example, the regression equation for the partial model is D,X * 0 D 1 x 1 2 x 2 (1) and the regression equation for the full model is D,X * 0 D 1 x 1 2 x 2 3 Dx 1 4 Dx 2, (2) where the two covariates are as follows: x 1 1ifthe Gleason score is 7 (intermediate risk) and 0 otherwise; x 2 1 if the Gleason score is 8 or higher (high risk) and 0 otherwise, as recommended by D Amico et al (19). Note that x 1 0 and x 2 0 correspond to the Gleason score of 6 or lower (low risk). Also note that in the full model, an interaction term between x 1 and x 2 is absent because each of these covariates is a dichotomized indicator variable of Gleason score. In the ureteral stone example, the regression equation for the partial model is D,X * 0 D 1 x 1 2 x 2 3 x 1 x 2 (3) and the regression equation for the full model is D,X * 0 D 1 x 1 2 x 2 3 x 1 x 2 4 Dx 1 5 Dx 2 6 Dx 1 x 2, (4) where the two covariates are x 1 1 if the kidney stone is located in the UVJ and 0 otherwise; x 2 is the age of the patient (a continuous covariate). There is a difference between the above regression models and binary regression models. The former assume that the expected value of Y depends on (D, X), whereas the latter assume that the expected value of D depends on (Y, X). As these models serve different inferential pur- 715
4 O MALLEY ET AL Academic Radiology, Vol 8, No 8, August 2001 poses, numerical results should not be compared directly (14). Choice of Prior Distributions Diffuse prior distributions. Little prior information about the parameters is first considered. This level of knowledge is introduced by specifying a prior distribution with very large values for its variance. The prior distributions for all regression coefficients, the s, are assumed to have independent normal N(0, 10 6 ) distributions. The prior distributions for the variance terms, the 2 s, are assumed to be independent inverse gamma IG(0.001, 0.001) distributions. For certain simple problems like those considered here, Bayesian inferences under a diffuse prior will be indistinguishable from those based on maximum-likelihood estimation. This will not be the case when the posterior distribution is asymmetric and a Bayesian estimator other than the posterior mode (eg, the posterior mean) is used. Informative prior distributions. The more informative the prior knowledge, the sharper is its distribution, resulting in a possibly greater effect on the posterior distribution. In this article we consider different informative prior distributions for the main factor mean effect, 0,ofthe gold standard. In our partial model (Eq [3]), 0 1,X 0,X, corresponding to the amount that the data distribution mean changes with disease status. The hyperparameters (mean and variance) of the prior distribution for 0 are assigned with realistic values. In the PSA example, the prior distributions for 0 were N(0.70, 0.05), N(1.15, 0.05), N(1.60, 0.05), and N(1.60, 0.50). In the ureteral stone examples, the prior distributions were N(0.20, 0.01), N(0.50, 0.01), N(0.80, 0.01), and N(0.80, 0.10). See the Results section for justifications for the choice of these prior distributions. Summary ROC Measures For each clinical example the posterior means and standard deviations (SDs) of all model parameters are computed. Characteristics and summary ROC measures, functions of these parameters, are also computed. These include the following: area under the ROC curve (A), partial area (A ) between 50% and 100% specificity, sensitivity (q) at 90% specificity, and specificity p MIS corresponding to maximum improvement of sensitivity (MIS, or q MIS ) over chance. See Appendix A, as well as the ROC literature (3 5) (A, A, q only), for a description of these ROC measures. Curve Fitting via Markov-Chain Monte Carlo Methods Consider the diagnostic data in the form of a triple (Y, D, X). Let ƒ denote the joint probability distribution function of any random variables. With prespecified regression models and observed data, the posterior distribution of is obtained from Bayes theorem (16): where and f, 2 Y, D, X f Y D, X,, 2 f, 2, (5) f Y D, X *, 0, 1,..., k, 2 1 2, 2 2, f Y D, X f Y D, X,, 2 f, 2 d d 2 is the marginal distribution of the data. The posterior mean of any function of the parameters g(, 2 ) is given by the following: E g, 2 Y, D, X g, 2 f, 2 Y, D, X d d 2. (6) For example, by setting g(, 2 ) 0, we obtain the posterior mean of the main factor diagnostic effect in the ROC regression models. The posterior mean sensitivity at a given specificity is obtained by simply setting g(, 2 ) equal to the corresponding ROC curve value (a function of the model parameters and 2 ). The posterior variance of the function h(, 2 ) may be evaluated by setting g(, 2 ) {h(, 2 ) E[h(, 2 ) Y, D, X]} 2, and the posterior SD is the square root of the posterior variance. The integrals involved in Equations (5) and (6) cannot generally be evaluated explicitly. We developed self-written C programs to implement Markov-Chain Monte Carlo methods via the Gibbs sampler (16,17). We used a burn-in of 2,000 iterations and a main simulation of 10,000 iterations. Furthermore, we verified our results by using the free standard software program 716
5 Academic Radiology, Vol 8, No 8, August 2001 BAYESIAN REGRESSION METHODOLOGY FOR ROC CURVES Table 1 PSA Example: Posterior Distributions of the Regression Parameters for Diffuse Prior Distributions Model * Partial Full Note. Values are means SDs. regression coefficients, where * intercept, 0 main effect of the diagnostic standard, 1 effect of intermediate risk based on Gleason scores, 2 effect of high risk based on Gleason scores, 3 effect of (diagnostic standard) (intermediate-risk category) interaction, 4 effect of (diagnostic standard) (high-risk category) interaction, 1 2 error variance for the nondiseased measurements, and 2 2 error variance for diseased measurements. Independent N(0, 10 6 ) priors were used for all regression coefficients. Independent IG(0.001, 0.001) priors were used for the variance parameters. Table 2 PSA Example: Posterior Distributions of Summary Measures of Diagnostic Accuracy for Diffuse Prior Distributions Model A A q p MIS q MIS Partial (overall) Full Low risk Intermediate risk High risk Note. Values are means SDs. A area under ROC curve, A partial area under ROC curve between specificities of 50% and 100%, q sensitivity at specificity 90%, p MIS 1 specificity corresponding to maximum improvement of sensitivity, and q MIS maximum improvement of sensitivity. BUGS run on the UNIX platform (18). Appendix B provides the BUGS code for fitting the partial model and estimating the area under the ROC curve in the ureteral stone example. RESULTS Clinical Example 1: PSA Level in Prostate Cancer Overall results and summary statistics. Sixty-six patients (36.7%) had local disease, and 114 (63.3%) had advanced disease. Their PSA levels (outcome variable) ranged from 0.1 to 58.0 ng/ml (mean SD, ng/ml 9.37). This variable was transformed by using a Box-Cox transformation to normality with coefficient of 0.33; see reference 20 for justification for a similar clinical example. On the basis of Gleason scores, 88 subjects (48.9%) were classified as low risk, 51 (28.3%) as intermediate risk, and 41 (22.8%) as high risk. Results under full versus partial models. For both regression models (Eqq [1, 2]), Table 1 presents posterior means and SDs of all model parameters with diffuse priors. In both models the posterior mean of the gold standard main effect, 0, is above 0, reflecting the predictive value of PSA levels for advanced disease stage. In the full model, the posterior mean of the (gold standard) (intermediate-risk category) interaction term, 3, is also positive, suggesting that the test is most predictive for the intermediate-risk group. Table 2 and Figure 1 present the fitted ROC curves, along with characteristics and summary measures. The overall ROC curve under the partial model is displayed, along with three curves (for the low-, intermediate-, and high-risk groups) under the full model. PSA levels are most accurate for the intermediate-risk group (A 0.769) and least accurate for the low-risk group (A 0.667). All sensitivity values, the q s, at p 10% (90% specificity), are quite low. The maximum improvement of sensitivity over chance occurs at about p MIS 40% (60% specificity). For example, under the partial model, p MIS (56% specificity), q MIS 0.314, and the sum of these two quantities yields the corresponding sensitivity q MIS In contrast, the sensitivity q (improvement of over chance) at p 10% (90% specificity). The highest maximum improvement of sensitivity occurs for the intermediate-risk group. Results under different prior distributions for 0. Table 3 provides posterior means and SDs under the partial model and five different prior distributions for 0. Diffuse priors are still assumed for all other parameters. The posterior mean and variance of 0 under the 717
6 O MALLEY ET AL Academic Radiology, Vol 8, No 8, August 2001 diffuse prior are and , respectively. To investigate the robustness of the choice of prior, we first construct a realistic prior distribution N(1.15, 0.05) for 0, because it contains information consistent with the data. We then vary the mean and variance of this prior in order to obtain the following alternative prior distributions: N(0.70, 0.05), N(1.15, 0.05), N(1.60, 0.05), and N(1.60, 0.50). Table 4 and Figure 2 present the fitted ROC curves, along with characteristics and summary measures. The simulated posterior density functions of the area A under these prior distributions are plotted in Figure 3. These ROC curves do not vary much, with A ranging from to This suggests that the models are robust with respect to the prior distributions. When the mean of the normal prior distribution for 0 is increased (eg, from 0.70 to 1.60) while the variance is fixed, the ROC curves move from conservative (eg, smaller area) to anticonservative (eg, greater area). When the variance of the normal prior distribution is increased while the mean is fixed (eg, from 0.05, 0.50, to 10 6 ), the results tend toward those under the diffuse prior. Clinical Example 2: Spiral CT of Ureteral Stones Overall results and summary statistics. Seventy-one patients passed stones spontaneously, and 29 required interventional therapy. The stone size (outcome variable) ranged from 1 to 16 mm (mean SD, 5.03 mm 2.69). This variable was transformed by using a log transformation (20). Thirty-nine subjects had stones located in the UVJ. The age range was years (mean SD, years 12.03) in the spontaneous passage group and (mean SD, years 13.24) in the surgical intervention group. Results under full versus partial models. For both regression models (Eqq [3, 4]), Table 5 presents posterior means and SDs of all model parameters with diffuse priors. In both models the posterior mean of the gold standard main effect, 0, is above 0, reflecting the predictive value of stone size for treatment option. In the full model, the posterior mean of the (gold standard) (UVJ) interaction term 4 and the (gold standard) (UVJ) (age) term 6 are different from 0, suggesting a complex structure. Table 6 and Figure 4 present the fitted ROC curves, along with characteristics and summary measures. The overall ROC curve under the partial model is displayed, along with four curves (for the combinations of UVJ status age 30 or 40 years) under the full model. Stone Figure 1. In the PSA example, the overall ROC curves (partial model) and the curves for three levels of risks based on the Gleason scores (full model). Figure 2. In the PSA example and under the partial model, the ROC curves based on different prior distributions. size is least predictive (A 0.577) for stones in the UVJ of patients 40 years of age. The predictive values for all other groups (A, ) were similar to that for 718
7 Academic Radiology, Vol 8, No 8, August 2001 BAYESIAN REGRESSION METHODOLOGY FOR ROC CURVES Table 3 PSA Example, Partial Model: Posterior Distributions of Regression Parameters Based on Different Prior Distributions Prior ( 0 ) * N(1.15, 10 6 ) N(0.70, 0.05) N(1.15, 0.05) N(1.60, 0.05) N(1.60, 0.50) Note. Values are means SDs. See Table 1 for explanation of regression parameters. Table 4 PSA Example, Partial Model: Posterior Distributions of Summary Measures of Diagnostic Accuracy Based on Different Prior Distributions Prior ( 0 ) A A q p MIS q MIS N(1.15, 10 6 ) N(0.70, 0.05) N(1.15, 0.05) N(1.60, 0.05) N(1.60, 0.50) Note. Values are means SDs. See Table 2 for expanded abbreviations. Figure 3. In the PSA example and under the partial model, the posterior distributions of the area under the ROC curve based on different prior distributions. the partial model (A 0.774). The maximum improvement of sensitivity occurs at 75% specificity, approximately. The greatest improvement ( q MIS 0.410) is for a 40-year-old subject with a stone not located in the UVJ. Results under different prior distributions for 0. Table 7 provides posterior means and SDs under the partial model and five different prior distributions for 0. Diffuse priors are still assumed for all other parameters. The posterior mean and variance of 0 under the diffuse prior are and , respectively. A realistic prior distribution that could have been assumed is N(0.50, 0.01). We then vary the mean and variance of this prior in order to obtain the following alternative prior distributions: N(0.20, 0.01), N(0.50, 0.01), N(0.80, 0.01), and N(0.80, 0.10). Table 8 and Figure 5 present the fitted ROC curves, along with characteristics and summary measures. The simulated posterior density functions of A under these prior distributions are plotted in Figure 6. These ROC curves vary, with A ranging from to 0.838, suggest- 719
8 O MALLEY ET AL Academic Radiology, Vol 8, No 8, August 2001 ing that the models are sensitive to the choice of prior distributions. The effect of mean and variance on the summary measures has similar tendencies as in the PSA example. Moreover, the posterior SDs vary depending on the precision of the prior distribution. Remarks on Markov-chain Monte Carlo Computations Convergence of the Markov-chain Monte Carlo algorithm was assessed by using CODA (21). For each model and data set the rate of convergence was rapid. The chains were monitored for convergence by using trace plots (14) and diagnostic measures such as Gelman and Rubin s (22) shrink factor. The estimated 50th percentiles of Gelman and Rubin s shrink factor were less than 1.05 in all cases, indicating that 10,000 iterations sufficed to achieve convergence. To assess whether the models were appropriate, we examined the residuals associated with the fitted posterior means of the outcomes with no evidence of lack of fit. In addition, the partial and full models were formally compared by using the deviance information criterion (DIC), a likelihood ratio test (LRT), and the pseudo-bayes factor (PSBF) (23,24). In the PSA example, the models performed similarly (the difference between the DIC of the partial and full models was 2.64, 2log-likelihood ratio 1.29, and PSBF 3.10). In the ureteral stone example, the partial model appeared more appropriate (difference of 5.64 for the DIC, 2log-likelihood ratio 0.60, PSBF 72.24). We used posterior means and SDs to summarize the location and spread of the posterior distributions because they have a unimodal and reasonably symmetric shape (Figs 3 and 6). Not surprisingly, our results under the diffuse prior were very similar to those based on maximum-likelihood estimation. If the posterior distributions were quite skewed, then the posterior medians or nonsymmetric credibility intervals, such as highest posterior density regions, would be preferable measures. DISCUSSION We have considered two Bayesian regression models, namely, the full and partial models, for ROC analysis that uses continuous diagnostic outcome data and accounts for covariates. The outcome variable was first transformed via a suitable monotone transformation to ensure that the modeling assumptions were appropriate. The posterior means and SDs of the parameters and functions of these Table 5 Ureteral Stone Example: Posterior Distributions of Regression Parameters for Diffuse Prior Distributions Model * Partial Full Note. Values are means SDs. regression coefficients, where * intercept, 0 main effect of the diagnostic standard, 1 effect of stone in UVJ, 2 effect of age, 3 effect of (UVJ) (age) interaction, 4 effect of (diagnostic standard) (UVJ) interaction, 5 effect of (diagnostic standard) (age) interaction, 6 effect of (diagnostic standard) (UVJ) (age) interaction, 1 2 error variance for the nondiseased measurements, and 2 2 error variance for diseased measurements. All coefficients of terms involving age were multiplied by 100. Independent N(0, 10 6 ) priors were used for all regression coefficients. Independent IG(0.001, 0.001) priors were used for the variance parameters. 720
9 Academic Radiology, Vol 8, No 8, August 2001 BAYESIAN REGRESSION METHODOLOGY FOR ROC CURVES Table 6 Ureteral Stone Example: Posterior Distributions of Summary Measures of Diagnostic Accuracy for Diffuse Prior Distributions Model A A q p MIS q MIS Partial (overall) Full Stone not in UVJ; pt age, 30 y Stone in UVJ; pt age, 30 y Stone not in UVJ; pt age, 40 y Stone in UVJ; pt age, 40 y Note. Values are means SDs. See Table 2 for expanded abbreviations; pt patient. Figure 4. In the ureteral stone example, the overall ROC curves (partial model) and the curves for three levels of risks based on the Gleason scores (full model). parameters were simulated via Markov-Chain Monte Carlo methods. The effect of different prior distributions was also investigated. We selected two real applications to appeal to a wide audience and to illustrate how ROC curves varied for subgroups of patients based on discrete covariates (Gleason score in the PSA example) and continuous covariates (age in the ureteral stone example). We conclude from our overall analysis (under the partial model) that both PSA levels and ureteral stone size had satisfactory accuracy. The full models yielded multiple ROC curves corresponding to different values of the covariates, a consequence of the interactions between the covariates and the gold standard. In particular, the accuracy was the highest for the intermediate-risk group in the PSA example. In the ureteral stone example, the accuracy decreased dramatically with age when a stone was located in the UVJ. On the contrary, the accuracy was fairly constant over age when a stone was not located in the UVJ. The robustness with respect to prior distributions was related to the sample size. The choice of prior distributions had greater influence on the results in the ureteral stone example (n 100) than in the PSA example (n 180). Radiologic data such as illustrated in our examples are often subject to variability and prior knowledge. Our approach has several advantages. First, pilot information and prior belief can be incorporated in our regression models. Such additional information allows for precise inferences if it is consistent with clinical data. When different prior distributions are considered, it appears that the analysis becomes more subjective with varied results. Objectivity can be achieved, however, if these distributions are derived from reliable sources, such as relevant pilot studies. Second, Bayesian inferences use explicit probability statements about the parameters or any other quantities of interest. The uncertainly of these quantities is also expressed in terms of probability distributions, for example the posterior distributions of the area under the ROC curve (Figs 3 and 6). Third, complex inferences can easily be made via direct simulations from the posterior distributions because under the Bayesian paradigm a formal procedure exists for obtaining a solution to all inference problems (25). Finally, computer codes using standard software are readily available. The analysis we performed has several potential limitations. First, we require that the test outcome variable is normally distributed or explicitly transformed to normality before analysis. Alternatively, we can extend the model so that the transformation to normality is incorporated in our model (16,20) or so that only a latent decision vari- 721
10 O MALLEY ET AL Academic Radiology, Vol 8, No 8, August 2001 Table 7 Ureteral Stone Example, Partial Model: Posterior Distributions of Regression Parameters Based on Different Prior Distributions Prior ( 0 ) * N(0.50, 10 6 ) N(0.20, 0.01) N(0.50, 0.01) N(0.80, 0.01) N(0.80, 0.10) Note. Values are means SDs. See Table 5 for explanation of regression parameters. Table 8 Ureteral Stone Example, Partial Model: Posterior Distributions of Summary Measures of Diagnostic Accuracy Based on Different Prior Distributions Prior ( 0 ) A A q p MIS q MIS N(0.50, 10 6 ) N(0.20, 0.01) N(0.50, 0.01) N(0.80, 0.01) N(0.80, 0.10) Note. Values are means SDs. See Table 2 for expanded abbreviations. able is assumed to be normal (6,7). Second, for illustration purposes, we constructed prior distributions based mainly on observed data and then arbitrarily varied the hyperparameters. If clinical studies similar to ours are to be conducted, however, or if further clinical study is called for, our posterior distributions can then be used in the future. Third, when dealing with complex models, exact Bayesian inferences may be computationally intensive. Special attention should be paid in applying our Bayesian approach, especially in constructing prior distributions. The choice of prior distributions should not be arbitrary but rather should depend on the pilot data or plausible prior belief. We recommend investigating the sensitivity of the results to the choice of prior distribution. Several future research topics in this area are under way. In this article the transformation of the outcome variable was provided before analysis. We will incorporate such a transformation in a Bayesian hierarchical generalized linear models framework (15). In addition, we will elicit prior distributions by using formal statistical rules (26). We will also consider prior distributions with valid clinical basis that impose constraints on the parameter space. For example, if it is known for certain that the accuracy of a diagnostic test is higher than chance, the Figure 5. In the ureteral stone example and under the partial model, the ROC curves based on different prior distributions. parameters should be constrained accordingly. Another possible constraint is that the variance of the outcome data is smaller for the nondiseased group than that for the diseased group, or vice versa (11,27). 722
11 Academic Radiology, Vol 8, No 8, August 2001 BAYESIAN REGRESSION METHODOLOGY FOR ROC CURVES Figure 6. In the ureteral stone example and under the partial model, the posterior distributions of the area under the ROC curve based on different prior distributions. APPENDIX A: SUMMARY ROC MEASURES FOR OUR REGRESSION MODELS Let D,X denote the mean of the diagnostic outcome Y at the gold standard disease status D and additional covariates X. For simplicity, variances of the outcomes are 12 and 22 for the nondiseased and diseased groups, respectively. Furthermore, let the ROC parameters be X ( 1,X 0,X )/ 1 and 2 / 1. Note that it can easily be shown in our partial model that X 0 / 1, free of X, where 0 is the regression coefficient for D. Thus, there is a single ROC curve under this model. 1. Sensitivity at given specificity: For any given (1 specificity), p, the underlying sensitivity is q p X 1 p. Figure 7. A hypothetical ROC curve illustrating the maximum improvement of sensitivity over chance at corresponding (1 specificity). 2. Area under the curve: A X / 1 2. Increasing demand on speed and cost-effectiveness are placing restrictions on the size and breadth of studies. Therefore, it is important to use statistical methods that incorporate all available information in the analysis, allowing for fully informed decisions. The Bayesian approach provides a coherent methodology for updating an initial knowledge base with experimental data. It is therefore well suited to the demands of the modern study. The area under the curve is equal to the probability that the outcome for a randomly drawn diseased subject is higher than for a randomly drawn nondiseased subject (28). 3. Partial area between p 1 and p 2 : p 1, p 2 : A p 1, p 2 p1 p 2 q p d p. 723
12 O MALLEY ET AL Academic Radiology, Vol 8, No 8, August 2001 Figure A1. BUGS code for the ureteral stone s example partial model. Partial area is often preferred to A, especially when only a particular range of specificity or sensitivity is of interest (29). 4. Maximum improvement of sensitivity over chance. This is the maximum difference in observed sensitivity and sensitivity at chance (lying on a 45 line in ROC space) over all values of specificity(fig 7). The corresponding (1 specificity), denoted p MIS, is found to be p MIS X X log 1/2 / 2 1 when 1, p MIS ( X /2) when ( X 1, 1), and p MIS {0, 1} when ( X 1, 1). The sensitivity corresponding to the maximum improvement of sensitivity is given by q MIS X 1 p MIS, and the maximum improvement of sensitivity itself by q MIS q MIS p MIS. Note that due to local convexity of the ROC curve when 1, there is always at least one value of p at which sensitivity is strictly less than p. In the ureteral stone example, this is evident on the upper right portion of the ROC curve for a 40-year-old subject with a stone in the UVJ (Fig 4). APPENDIX B Figure A1 depicts the BUGS code for the ureteral stone example, partial model (18). ACKNOWLEDGMENTS Appreciation is extended to Barbara McNeil, MD, PhD, for reviewing the manuscript and to Daryl Caudry, MS, for providing us with the data for the prostate example. We also thank two anonymous referees for their careful review and comments that improved the quality of this manuscript. 724
13 Academic Radiology, Vol 8, No 8, August 2001 BAYESIAN REGRESSION METHODOLOGY FOR ROC CURVES REFERENCES 1. Tempany CM, Zhou X, Zerhouni EA, et al. Staging of prostate cancer: result of radiologic diagnostic oncology group project comparison of three MR imaging techniques. Radiology 1994; 192: Fielding JR, Silverman SG, Samuel S, Zou KH, Loughlin KR. Unenhanced helical CT of ureteral stones: a replacement for excretory urography in planning treatment. AJR Am J Roentgenol 1998; 171: Swets JA, Pickett RM. Evaluation of diagnostic systems. New York, NY: Academic Press, Campbell G. General methodology. I. Advances in statistical methodology for the evaluation of diagnostic and laboratory tests. Stat Med 1994; 13: Shapiro DE. The interpretation of diagnostic tests. Stat Methods Med Res 1999; 8: Metz CE, Herman BA, Shen J. Maximum-likelihood estimation of receiver operating characteristic (ROC) curves from continuous distributed data. Stat Med 1998; 17: Tosteson ANA, Begg CB. A general regression methodology for ROC curve estimation. Med Decis Making 1988; 8: Pepe MS. A regression modeling framework for ROC curves in medical diagnostic testing. Biometrika 1997; 84: Pepe MS. Three approaches to regression analysis of receiver operating characteristic curves for continuous test results. Biometrics 1998; 54: Pepe MS. An interpretation for the ROC curve and inference using GLM procedures. Biometrics 2000; 56: Beam CA. Random-effects models in the receiver operating characteristic curve-based assessment of the effectiveness of diagnostic imaging technology: concepts, approaches, and issues. Acad Radiol 1995; 2(suppl 1):S4 S Gatsonis CA. Random-effects models for diagnostic accuracy data. Acad Radiol 1995; 2(suppl 1):S14 S Peng F, Hall WJ. Bayesian analysis of ROC curves using Markovchain Monte Carlo methods. Med Decis Making 1996; 16: Hellmich M, Abrams KR, Jones DR, Lambert PC. A Bayesian approach to a general regression model for ROC curves. Med Decis Making 1998; 18: Ishwaran H, Gatsonis CA. A general class of hierarchical ordinal regression models with applications to correlated ROC analysis. Can J Stat 2000; 28: Gelman A, Carlin JB, Stern HS, Rubin DB. Bayesian data analysis. London, England: Chapman & Hall, Tanner MA. Tools for statistical inference: methods for the exploration of posterior distribution and likelihood function. 2nd ed. New York, NY: Springer-Verlag, Spiegelhalter DJ, Thomas A, Best NG, Gilks WR. BUGS: Bayesian inference using Gibbs sampling, version Cambridge, England: MRC Biostatistics Unit, D Amico AV, Whittington R, Malkowicz SB, et al. Biochemical outcome after radical prostatectomy, external beam radiation therapy, or interstitial radiation therapy for clinically localized prostate cancer. JAMA 1998; 280: Zou KH, Tempany CM, Fielding JR, Silverman SG. Original smooth receiver operating characteristic curve estimation from continuous data: statistical methods for analyzing the predictive value of spiral CT of ureteral stones. Acad Radiol 1998; 5: Best NG, Cowles MK, Vines SK. CODA manual version Cambridge, England: MRC Biostatistics Unit, Gelman A, Rubin D. Inference from iterative simulation using multiple sequences. Stat Sci 1992; 7: Spiegelhalter DJ, Best NG, Carlin BP. Bayesian deviance, the effect number of parameters, and the comparison of arbitrarily complex models Web site of the MRC Biostatistics Unit, Cambridge, England. Available at: preslid.shtml. Accessed May 21, Gelfand AE, Dey DK, Chang H. Model determination using predictive distributions with implementation via sampling-based methods. In: Bernardo JM, Berger JO, Dawid AP, Smith AFM. eds. Bayesian statistics 4. New York, NY: Oxford University Press, 1992; Lindley DV. Bayesian inference. In: Kotz S, Johnson NL, eds. International encyclopedia of statistics. Vol 1. New York, NY: Wiley, 1982; Kass RE, Wasserman L. The selection of prior distribution by formal rules. JASA 1996; 91: Tandberg D, Deely JJ, O Malley AJ. Generalized likelihood ratios for quantitative diagnostic test scores. Am J Emerg Med 1997; 15: Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology 1982; 143: McClish DK. Analyzing a portion of the ROC curve. Med Decis Making 1989; 9:
Reconstruction of individual patient data for meta analysis via Bayesian approach
Reconstruction of individual patient data for meta analysis via Bayesian approach Yusuke Yamaguchi, Wataru Sakamoto and Shingo Shirahata Graduate School of Engineering Science, Osaka University Masashi
More informationBayesian multivariate hierarchical transformation models for ROC analysis
STATISTICS IN MEDICINE Statist. Med. 2006; 25:459 479 Published online 11 October 2005 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/sim.2187 Bayesian multivariate hierarchical transformation
More informationBayesian Inference. Chapter 1. Introduction and basic concepts
Bayesian Inference Chapter 1. Introduction and basic concepts M. Concepción Ausín Department of Statistics Universidad Carlos III de Madrid Master in Business Administration and Quantitative Methods Master
More informationSample Size Calculations for ROC Studies: Parametric Robustness and Bayesian Nonparametrics
Baylor Health Care System From the SelectedWorks of unlei Cheng Spring January 30, 01 Sample Size Calculations for ROC Studies: Parametric Robustness and Bayesian Nonparametrics unlei Cheng, Baylor Health
More informationMarkov Chain Monte Carlo methods
Markov Chain Monte Carlo methods By Oleg Makhnin 1 Introduction a b c M = d e f g h i 0 f(x)dx 1.1 Motivation 1.1.1 Just here Supresses numbering 1.1.2 After this 1.2 Literature 2 Method 2.1 New math As
More informationBayesian methods for sample size determination and their use in clinical trials
Bayesian methods for sample size determination and their use in clinical trials Stefania Gubbiotti Abstract This paper deals with determination of a sample size that guarantees the success of a trial.
More informationA note on Reversible Jump Markov Chain Monte Carlo
A note on Reversible Jump Markov Chain Monte Carlo Hedibert Freitas Lopes Graduate School of Business The University of Chicago 5807 South Woodlawn Avenue Chicago, Illinois 60637 February, 1st 2006 1 Introduction
More informationeqr094: Hierarchical MCMC for Bayesian System Reliability
eqr094: Hierarchical MCMC for Bayesian System Reliability Alyson G. Wilson Statistical Sciences Group, Los Alamos National Laboratory P.O. Box 1663, MS F600 Los Alamos, NM 87545 USA Phone: 505-667-9167
More informationBayesian Meta-analysis with Hierarchical Modeling Brian P. Hobbs 1
Bayesian Meta-analysis with Hierarchical Modeling Brian P. Hobbs 1 Division of Biostatistics, School of Public Health, University of Minnesota, Mayo Mail Code 303, Minneapolis, Minnesota 55455 0392, U.S.A.
More informationBayesian Inference in GLMs. Frequentists typically base inferences on MLEs, asymptotic confidence
Bayesian Inference in GLMs Frequentists typically base inferences on MLEs, asymptotic confidence limits, and log-likelihood ratio tests Bayesians base inferences on the posterior distribution of the unknowns
More informationBayesian Inference for the Multivariate Normal
Bayesian Inference for the Multivariate Normal Will Penny Wellcome Trust Centre for Neuroimaging, University College, London WC1N 3BG, UK. November 28, 2014 Abstract Bayesian inference for the multivariate
More information7. Estimation and hypothesis testing. Objective. Recommended reading
7. Estimation and hypothesis testing Objective In this chapter, we show how the election of estimators can be represented as a decision problem. Secondly, we consider the problem of hypothesis testing
More informationThe Bayesian Approach to Multi-equation Econometric Model Estimation
Journal of Statistical and Econometric Methods, vol.3, no.1, 2014, 85-96 ISSN: 2241-0384 (print), 2241-0376 (online) Scienpress Ltd, 2014 The Bayesian Approach to Multi-equation Econometric Model Estimation
More informationSupplement to A Hierarchical Approach for Fitting Curves to Response Time Measurements
Supplement to A Hierarchical Approach for Fitting Curves to Response Time Measurements Jeffrey N. Rouder Francis Tuerlinckx Paul L. Speckman Jun Lu & Pablo Gomez May 4 008 1 The Weibull regression model
More informationTHE SKILL PLOT: A GRAPHICAL TECHNIQUE FOR EVALUATING CONTINUOUS DIAGNOSTIC TESTS
THE SKILL PLOT: A GRAPHICAL TECHNIQUE FOR EVALUATING CONTINUOUS DIAGNOSTIC TESTS William M. Briggs General Internal Medicine, Weill Cornell Medical College 525 E. 68th, Box 46, New York, NY 10021 email:
More informationMarkov Chain Monte Carlo in Practice
Markov Chain Monte Carlo in Practice Edited by W.R. Gilks Medical Research Council Biostatistics Unit Cambridge UK S. Richardson French National Institute for Health and Medical Research Vilejuif France
More informationThe STS Surgeon Composite Technical Appendix
The STS Surgeon Composite Technical Appendix Overview Surgeon-specific risk-adjusted operative operative mortality and major complication rates were estimated using a bivariate random-effects logistic
More informationLehmann Family of ROC Curves
Memorial Sloan-Kettering Cancer Center From the SelectedWorks of Mithat Gönen May, 2007 Lehmann Family of ROC Curves Mithat Gonen, Memorial Sloan-Kettering Cancer Center Glenn Heller, Memorial Sloan-Kettering
More informationA NOTE ON ROBUST ESTIMATION IN LOGISTIC REGRESSION MODEL
Discussiones Mathematicae Probability and Statistics 36 206 43 5 doi:0.75/dmps.80 A NOTE ON ROBUST ESTIMATION IN LOGISTIC REGRESSION MODEL Tadeusz Bednarski Wroclaw University e-mail: t.bednarski@prawo.uni.wroc.pl
More informationPart 8: GLMs and Hierarchical LMs and GLMs
Part 8: GLMs and Hierarchical LMs and GLMs 1 Example: Song sparrow reproductive success Arcese et al., (1992) provide data on a sample from a population of 52 female song sparrows studied over the course
More informationLongitudinal breast density as a marker of breast cancer risk
Longitudinal breast density as a marker of breast cancer risk C. Armero (1), M. Rué (2), A. Forte (1), C. Forné (2), H. Perpiñán (1), M. Baré (3), and G. Gómez (4) (1) BIOstatnet and Universitat de València,
More informationPrinciples of Bayesian Inference
Principles of Bayesian Inference Sudipto Banerjee University of Minnesota July 20th, 2008 1 Bayesian Principles Classical statistics: model parameters are fixed and unknown. A Bayesian thinks of parameters
More informationEstimating Optimum Linear Combination of Multiple Correlated Diagnostic Tests at a Fixed Specificity with Receiver Operating Characteristic Curves
Journal of Data Science 6(2008), 1-13 Estimating Optimum Linear Combination of Multiple Correlated Diagnostic Tests at a Fixed Specificity with Receiver Operating Characteristic Curves Feng Gao 1, Chengjie
More informationMULTILEVEL IMPUTATION 1
MULTILEVEL IMPUTATION 1 Supplement B: MCMC Sampling Steps and Distributions for Two-Level Imputation This document gives technical details of the full conditional distributions used to draw regression
More informationBest Linear Unbiased Prediction: an Illustration Based on, but Not Limited to, Shelf Life Estimation
Libraries Conference on Applied Statistics in Agriculture 015-7th Annual Conference Proceedings Best Linear Unbiased Prediction: an Illustration Based on, but Not Limited to, Shelf Life Estimation Maryna
More informationMarkov Chain Monte Carlo methods
Markov Chain Monte Carlo methods Tomas McKelvey and Lennart Svensson Signal Processing Group Department of Signals and Systems Chalmers University of Technology, Sweden November 26, 2012 Today s learning
More informationConstructing Confidence Intervals of the Summary Statistics in the Least-Squares SROC Model
UW Biostatistics Working Paper Series 3-28-2005 Constructing Confidence Intervals of the Summary Statistics in the Least-Squares SROC Model Ming-Yu Fan University of Washington, myfan@u.washington.edu
More informationBayesian Estimation of Prediction Error and Variable Selection in Linear Regression
Bayesian Estimation of Prediction Error and Variable Selection in Linear Regression Andrew A. Neath Department of Mathematics and Statistics; Southern Illinois University Edwardsville; Edwardsville, IL,
More informationBayesian data analysis in practice: Three simple examples
Bayesian data analysis in practice: Three simple examples Martin P. Tingley Introduction These notes cover three examples I presented at Climatea on 5 October 0. Matlab code is available by request to
More informationMarkov Chain Monte Carlo
Markov Chain Monte Carlo Recall: To compute the expectation E ( h(y ) ) we use the approximation E(h(Y )) 1 n n h(y ) t=1 with Y (1),..., Y (n) h(y). Thus our aim is to sample Y (1),..., Y (n) from f(y).
More information[Part 2] Model Development for the Prediction of Survival Times using Longitudinal Measurements
[Part 2] Model Development for the Prediction of Survival Times using Longitudinal Measurements Aasthaa Bansal PhD Pharmaceutical Outcomes Research & Policy Program University of Washington 69 Biomarkers
More informationNONLINEAR APPLICATIONS OF MARKOV CHAIN MONTE CARLO
NONLINEAR APPLICATIONS OF MARKOV CHAIN MONTE CARLO by Gregois Lee, B.Sc.(ANU), B.Sc.Hons(UTas) Submitted in fulfilment of the requirements for the Degree of Doctor of Philosophy Department of Mathematics
More informationModelling Receiver Operating Characteristic Curves Using Gaussian Mixtures
Modelling Receiver Operating Characteristic Curves Using Gaussian Mixtures arxiv:146.1245v1 [stat.me] 5 Jun 214 Amay S. M. Cheam and Paul D. McNicholas Abstract The receiver operating characteristic curve
More informationA New Confidence Interval for the Difference Between Two Binomial Proportions of Paired Data
UW Biostatistics Working Paper Series 6-2-2003 A New Confidence Interval for the Difference Between Two Binomial Proportions of Paired Data Xiao-Hua Zhou University of Washington, azhou@u.washington.edu
More informationA Note on Bayesian Inference After Multiple Imputation
A Note on Bayesian Inference After Multiple Imputation Xiang Zhou and Jerome P. Reiter Abstract This article is aimed at practitioners who plan to use Bayesian inference on multiplyimputed datasets in
More informationMeta-Analysis for Diagnostic Test Data: a Bayesian Approach
Meta-Analysis for Diagnostic Test Data: a Bayesian Approach Pablo E. Verde Coordination Centre for Clinical Trials Heinrich Heine Universität Düsseldorf Preliminaries: motivations for systematic reviews
More informationRank Regression with Normal Residuals using the Gibbs Sampler
Rank Regression with Normal Residuals using the Gibbs Sampler Stephen P Smith email: hucklebird@aol.com, 2018 Abstract Yu (2000) described the use of the Gibbs sampler to estimate regression parameters
More informationBayesian nonparametric estimation of finite population quantities in absence of design information on nonsampled units
Bayesian nonparametric estimation of finite population quantities in absence of design information on nonsampled units Sahar Z Zangeneh Robert W. Keener Roderick J.A. Little Abstract In Probability proportional
More informationHealth utilities' affect you are reported alongside underestimates of uncertainty
Dr. Kelvin Chan, Medical Oncologist, Associate Scientist, Odette Cancer Centre, Sunnybrook Health Sciences Centre and Dr. Eleanor Pullenayegum, Senior Scientist, Hospital for Sick Children Title: Underestimation
More informationBagging During Markov Chain Monte Carlo for Smoother Predictions
Bagging During Markov Chain Monte Carlo for Smoother Predictions Herbert K. H. Lee University of California, Santa Cruz Abstract: Making good predictions from noisy data is a challenging problem. Methods
More informationIntroduction to Bayesian Statistics and Markov Chain Monte Carlo Estimation. EPSY 905: Multivariate Analysis Spring 2016 Lecture #10: April 6, 2016
Introduction to Bayesian Statistics and Markov Chain Monte Carlo Estimation EPSY 905: Multivariate Analysis Spring 2016 Lecture #10: April 6, 2016 EPSY 905: Intro to Bayesian and MCMC Today s Class An
More informationDefault Priors and Effcient Posterior Computation in Bayesian
Default Priors and Effcient Posterior Computation in Bayesian Factor Analysis January 16, 2010 Presented by Eric Wang, Duke University Background and Motivation A Brief Review of Parameter Expansion Literature
More informationStatistical Inference for Stochastic Epidemic Models
Statistical Inference for Stochastic Epidemic Models George Streftaris 1 and Gavin J. Gibson 1 1 Department of Actuarial Mathematics & Statistics, Heriot-Watt University, Riccarton, Edinburgh EH14 4AS,
More informationSimulation of truncated normal variables. Christian P. Robert LSTA, Université Pierre et Marie Curie, Paris
Simulation of truncated normal variables Christian P. Robert LSTA, Université Pierre et Marie Curie, Paris Abstract arxiv:0907.4010v1 [stat.co] 23 Jul 2009 We provide in this paper simulation algorithms
More informationBayesian modelling. Hans-Peter Helfrich. University of Bonn. Theodor-Brinkmann-Graduate School
Bayesian modelling Hans-Peter Helfrich University of Bonn Theodor-Brinkmann-Graduate School H.-P. Helfrich (University of Bonn) Bayesian modelling Brinkmann School 1 / 22 Overview 1 Bayesian modelling
More informationDivision of Pharmacoepidemiology And Pharmacoeconomics Technical Report Series
Division of Pharmacoepidemiology And Pharmacoeconomics Technical Report Series Year: 2013 #006 The Expected Value of Information in Prospective Drug Safety Monitoring Jessica M. Franklin a, Amanda R. Patrick
More informationStatistical Practice
Statistical Practice A Note on Bayesian Inference After Multiple Imputation Xiang ZHOU and Jerome P. REITER This article is aimed at practitioners who plan to use Bayesian inference on multiply-imputed
More informationPIRLS 2016 Achievement Scaling Methodology 1
CHAPTER 11 PIRLS 2016 Achievement Scaling Methodology 1 The PIRLS approach to scaling the achievement data, based on item response theory (IRT) scaling with marginal estimation, was developed originally
More informationNon parametric ROC summary statistics
Non parametric ROC summary statistics M.C. Pardo and A.M. Franco-Pereira Department of Statistics and O.R. (I), Complutense University of Madrid. 28040-Madrid, Spain May 18, 2017 Abstract: Receiver operating
More informationDistribution-free ROC Analysis Using Binary Regression Techniques
Distribution-free Analysis Using Binary Techniques Todd A. Alonzo and Margaret S. Pepe As interpreted by: Andrew J. Spieker University of Washington Dept. of Biostatistics Introductory Talk No, not that!
More information(5) Multi-parameter models - Gibbs sampling. ST440/540: Applied Bayesian Analysis
Summarizing a posterior Given the data and prior the posterior is determined Summarizing the posterior gives parameter estimates, intervals, and hypothesis tests Most of these computations are integrals
More informationMcGill University. Department of Epidemiology and Biostatistics. Bayesian Analysis for the Health Sciences. Course EPIB-675.
McGill University Department of Epidemiology and Biostatistics Bayesian Analysis for the Health Sciences Course EPIB-675 Lawrence Joseph Bayesian Analysis for the Health Sciences EPIB-675 3 credits Instructor:
More informationBayesian Methods for Machine Learning
Bayesian Methods for Machine Learning CS 584: Big Data Analytics Material adapted from Radford Neal s tutorial (http://ftp.cs.utoronto.ca/pub/radford/bayes-tut.pdf), Zoubin Ghahramni (http://hunch.net/~coms-4771/zoubin_ghahramani_bayesian_learning.pdf),
More informationPrerequisite: STATS 7 or STATS 8 or AP90 or (STATS 120A and STATS 120B and STATS 120C). AP90 with a minimum score of 3
University of California, Irvine 2017-2018 1 Statistics (STATS) Courses STATS 5. Seminar in Data Science. 1 Unit. An introduction to the field of Data Science; intended for entering freshman and transfers.
More informationSample size determination for a binary response in a superiority clinical trial using a hybrid classical and Bayesian procedure
Ciarleglio and Arendt Trials (2017) 18:83 DOI 10.1186/s13063-017-1791-0 METHODOLOGY Open Access Sample size determination for a binary response in a superiority clinical trial using a hybrid classical
More informationThree-group ROC predictive analysis for ordinal outcomes
Three-group ROC predictive analysis for ordinal outcomes Tahani Coolen-Maturi Durham University Business School Durham University, UK tahani.maturi@durham.ac.uk June 26, 2016 Abstract Measuring the accuracy
More informationLocal Likelihood Bayesian Cluster Modeling for small area health data. Andrew Lawson Arnold School of Public Health University of South Carolina
Local Likelihood Bayesian Cluster Modeling for small area health data Andrew Lawson Arnold School of Public Health University of South Carolina Local Likelihood Bayesian Cluster Modelling for Small Area
More informationBayesian Networks in Educational Assessment
Bayesian Networks in Educational Assessment Estimating Parameters with MCMC Bayesian Inference: Expanding Our Context Roy Levy Arizona State University Roy.Levy@asu.edu 2017 Roy Levy MCMC 1 MCMC 2 Posterior
More informationPlausible Values for Latent Variables Using Mplus
Plausible Values for Latent Variables Using Mplus Tihomir Asparouhov and Bengt Muthén August 21, 2010 1 1 Introduction Plausible values are imputed values for latent variables. All latent variables can
More informationBayesian model selection: methodology, computation and applications
Bayesian model selection: methodology, computation and applications David Nott Department of Statistics and Applied Probability National University of Singapore Statistical Genomics Summer School Program
More informationRecap on Data Assimilation
Concluding Thoughts Recap on Data Assimilation FORECAST ANALYSIS Kalman Filter Forecast Analysis Analytical projection of the ANALYSIS mean and cov from t-1 to the FORECAST mean and cov for t Update FORECAST
More informationNonparametric predictive inference with parametric copulas for combining bivariate diagnostic tests
Nonparametric predictive inference with parametric copulas for combining bivariate diagnostic tests Noryanti Muhammad, Universiti Malaysia Pahang, Malaysia, noryanti@ump.edu.my Tahani Coolen-Maturi, Durham
More informationA noninformative Bayesian approach to domain estimation
A noninformative Bayesian approach to domain estimation Glen Meeden School of Statistics University of Minnesota Minneapolis, MN 55455 glen@stat.umn.edu August 2002 Revised July 2003 To appear in Journal
More informationBayesian Linear Regression
Bayesian Linear Regression Sudipto Banerjee 1 Biostatistics, School of Public Health, University of Minnesota, Minneapolis, Minnesota, U.S.A. September 15, 2010 1 Linear regression models: a Bayesian perspective
More informationQuantile POD for Hit-Miss Data
Quantile POD for Hit-Miss Data Yew-Meng Koh a and William Q. Meeker a a Center for Nondestructive Evaluation, Department of Statistics, Iowa State niversity, Ames, Iowa 50010 Abstract. Probability of detection
More informationStat 542: Item Response Theory Modeling Using The Extended Rank Likelihood
Stat 542: Item Response Theory Modeling Using The Extended Rank Likelihood Jonathan Gruhl March 18, 2010 1 Introduction Researchers commonly apply item response theory (IRT) models to binary and ordinal
More informationInference on the Univariate Frailty Model: An Approach Bayesian Reference Analysis
Inference on the Univariate Frailty Model: An Approach Bayesian Reference Analysis Vera Lucia D. Tomazella and Camila Bertini Martins Universidade Federal de São Carlos, SP-Brasil (vera@ufscar.br; cacamila-bertini@yahoo.com.br)
More informationEstimating Diagnostic Error without a Gold Standard: A Mixed Membership Approach
7 Estimating Diagnostic Error without a Gold Standard: A Mixed Membership Approach Elena A. Erosheva Department of Statistics, University of Washington, Seattle, WA 98195-4320, USA Cyrille Joutard Institut
More informationSampling Methods (11/30/04)
CS281A/Stat241A: Statistical Learning Theory Sampling Methods (11/30/04) Lecturer: Michael I. Jordan Scribe: Jaspal S. Sandhu 1 Gibbs Sampling Figure 1: Undirected and directed graphs, respectively, with
More informationApproximating Bayesian Posterior Means Using Multivariate Gaussian Quadrature
Approximating Bayesian Posterior Means Using Multivariate Gaussian Quadrature John A.L. Cranfield Paul V. Preckel Songquan Liu Presented at Western Agricultural Economics Association 1997 Annual Meeting
More informationPrinciples of Bayesian Inference
Principles of Bayesian Inference Sudipto Banerjee and Andrew O. Finley 2 Biostatistics, School of Public Health, University of Minnesota, Minneapolis, Minnesota, U.S.A. 2 Department of Forestry & Department
More informationA Parametric ROC Model Based Approach for Evaluating the Predictiveness of Continuous Markers in Case-control Studies
UW Biostatistics Working Paper Series 11-14-2007 A Parametric ROC Model Based Approach for Evaluating the Predictiveness of Continuous Markers in Case-control Studies Ying Huang University of Washington,
More informationStatistical Methods in Particle Physics Lecture 1: Bayesian methods
Statistical Methods in Particle Physics Lecture 1: Bayesian methods SUSSP65 St Andrews 16 29 August 2009 Glen Cowan Physics Department Royal Holloway, University of London g.cowan@rhul.ac.uk www.pp.rhul.ac.uk/~cowan
More informationLecture 5: Spatial probit models. James P. LeSage University of Toledo Department of Economics Toledo, OH
Lecture 5: Spatial probit models James P. LeSage University of Toledo Department of Economics Toledo, OH 43606 jlesage@spatial-econometrics.com March 2004 1 A Bayesian spatial probit model with individual
More informationRonald Christensen. University of New Mexico. Albuquerque, New Mexico. Wesley Johnson. University of California, Irvine. Irvine, California
Texts in Statistical Science Bayesian Ideas and Data Analysis An Introduction for Scientists and Statisticians Ronald Christensen University of New Mexico Albuquerque, New Mexico Wesley Johnson University
More informationOptimal rules for timing intercourse to achieve pregnancy
Optimal rules for timing intercourse to achieve pregnancy Bruno Scarpa and David Dunson Dipartimento di Statistica ed Economia Applicate Università di Pavia Biostatistics Branch, National Institute of
More informationBayes Estimation in Meta-analysis using a linear model theorem
University of Wollongong Research Online Applied Statistics Education and Research Collaboration (ASEARC) - Conference Papers Faculty of Engineering and Information Sciences 2012 Bayes Estimation in Meta-analysis
More informationItem Parameter Calibration of LSAT Items Using MCMC Approximation of Bayes Posterior Distributions
R U T C O R R E S E A R C H R E P O R T Item Parameter Calibration of LSAT Items Using MCMC Approximation of Bayes Posterior Distributions Douglas H. Jones a Mikhail Nediak b RRR 7-2, February, 2! " ##$%#&
More informationANALYSIS OF ORDINAL SURVEY RESPONSES WITH DON T KNOW
SSC Annual Meeting, June 2015 Proceedings of the Survey Methods Section ANALYSIS OF ORDINAL SURVEY RESPONSES WITH DON T KNOW Xichen She and Changbao Wu 1 ABSTRACT Ordinal responses are frequently involved
More informationBAYESIAN ESTIMATION OF LINEAR STATISTICAL MODEL BIAS
BAYESIAN ESTIMATION OF LINEAR STATISTICAL MODEL BIAS Andrew A. Neath 1 and Joseph E. Cavanaugh 1 Department of Mathematics and Statistics, Southern Illinois University, Edwardsville, Illinois 606, USA
More information16 : Approximate Inference: Markov Chain Monte Carlo
10-708: Probabilistic Graphical Models 10-708, Spring 2017 16 : Approximate Inference: Markov Chain Monte Carlo Lecturer: Eric P. Xing Scribes: Yuan Yang, Chao-Ming Yen 1 Introduction As the target distribution
More informationBayesian Statistical Methods. Jeff Gill. Department of Political Science, University of Florida
Bayesian Statistical Methods Jeff Gill Department of Political Science, University of Florida 234 Anderson Hall, PO Box 117325, Gainesville, FL 32611-7325 Voice: 352-392-0262x272, Fax: 352-392-8127, Email:
More informationSawtooth Software. CVA/HB Technical Paper TECHNICAL PAPER SERIES
Sawtooth Software TECHNICAL PAPER SERIES CVA/HB Technical Paper Copyright 2002, Sawtooth Software, Inc. 530 W. Fir St. Sequim, WA 98382 (360) 681-2300 www.sawtoothsoftware.com The CVA/HB Technical Paper
More informationBayesian Analysis. Bayesian Analysis: Bayesian methods concern one s belief about θ. [Current Belief (Posterior)] (Prior Belief) x (Data) Outline
Bayesian Analysis DuBois Bowman, Ph.D. Gordana Derado, M. S. Shuo Chen, M. S. Department of Biostatistics and Bioinformatics Center for Biomedical Imaging Statistics Emory University Outline I. Introduction
More information8 Nominal and Ordinal Logistic Regression
8 Nominal and Ordinal Logistic Regression 8.1 Introduction If the response variable is categorical, with more then two categories, then there are two options for generalized linear models. One relies on
More informationMcGill University. Department of Epidemiology and Biostatistics. Bayesian Analysis for the Health Sciences. Course EPIB-682.
McGill University Department of Epidemiology and Biostatistics Bayesian Analysis for the Health Sciences Course EPIB-682 Lawrence Joseph Intro to Bayesian Analysis for the Health Sciences EPIB-682 2 credits
More informationBAYESIAN CLASSIFICATION OF HIGH DIMENSIONAL DATA WITH GAUSSIAN PROCESS USING DIFFERENT KERNELS
BAYESIAN CLASSIFICATION OF HIGH DIMENSIONAL DATA WITH GAUSSIAN PROCESS USING DIFFERENT KERNELS Oloyede I. Department of Statistics, University of Ilorin, Ilorin, Nigeria Corresponding Author: Oloyede I.,
More informationARIC Manuscript Proposal # PC Reviewed: _9/_25_/06 Status: A Priority: _2 SC Reviewed: _9/_25_/06 Status: A Priority: _2
ARIC Manuscript Proposal # 1186 PC Reviewed: _9/_25_/06 Status: A Priority: _2 SC Reviewed: _9/_25_/06 Status: A Priority: _2 1.a. Full Title: Comparing Methods of Incorporating Spatial Correlation in
More informationComputer Vision Group Prof. Daniel Cremers. 10a. Markov Chain Monte Carlo
Group Prof. Daniel Cremers 10a. Markov Chain Monte Carlo Markov Chain Monte Carlo In high-dimensional spaces, rejection sampling and importance sampling are very inefficient An alternative is Markov Chain
More informationStatistics 220 Bayesian Data Analysis
Statistics 220 Bayesian Data Analysis Mark E. Irwin Department of Statistics Harvard University Spring Term Thursday, February 3, 2005 - Tuesday, May 17, 2005 Copyright c 2005 by Mark E. Irwin Personnel
More informationPOSTERIOR ANALYSIS OF THE MULTIPLICATIVE HETEROSCEDASTICITY MODEL
COMMUN. STATIST. THEORY METH., 30(5), 855 874 (2001) POSTERIOR ANALYSIS OF THE MULTIPLICATIVE HETEROSCEDASTICITY MODEL Hisashi Tanizaki and Xingyuan Zhang Faculty of Economics, Kobe University, Kobe 657-8501,
More informationFeature selection and classifier performance in computer-aided diagnosis: The effect of finite sample size
Feature selection and classifier performance in computer-aided diagnosis: The effect of finite sample size Berkman Sahiner, a) Heang-Ping Chan, Nicholas Petrick, Robert F. Wagner, b) and Lubomir Hadjiiski
More informationBayesian Linear Models
Bayesian Linear Models Sudipto Banerjee 1 and Andrew O. Finley 2 1 Department of Forestry & Department of Geography, Michigan State University, Lansing Michigan, U.S.A. 2 Biostatistics, School of Public
More informationTables of Probability Distributions
Tables of Probability Distributions Table A. Discrete distributions Probability Distribution mass function Mean Mode Variance ( ) n Binomial p(y π)= π y ( π) n y nπ (n + )π nπ( π) y Y Bin(n,π) y = 0,,...,n
More informationRichard D Riley was supported by funding from a multivariate meta-analysis grant from
Bayesian bivariate meta-analysis of correlated effects: impact of the prior distributions on the between-study correlation, borrowing of strength, and joint inferences Author affiliations Danielle L Burke
More informationStatistics Boot Camp. Dr. Stephanie Lane Institute for Defense Analyses DATAWorks 2018
Statistics Boot Camp Dr. Stephanie Lane Institute for Defense Analyses DATAWorks 2018 March 21, 2018 Outline of boot camp Summarizing and simplifying data Point and interval estimation Foundations of statistical
More informationSubjective and Objective Bayesian Statistics
Subjective and Objective Bayesian Statistics Principles, Models, and Applications Second Edition S. JAMES PRESS with contributions by SIDDHARTHA CHIB MERLISE CLYDE GEORGE WOODWORTH ALAN ZASLAVSKY \WILEY-
More informationOptimising Group Sequential Designs. Decision Theory, Dynamic Programming. and Optimal Stopping
: Decision Theory, Dynamic Programming and Optimal Stopping Christopher Jennison Department of Mathematical Sciences, University of Bath, UK http://people.bath.ac.uk/mascj InSPiRe Conference on Methodology
More informationSAMPLE SIZE RE-ESTIMATION FOR ADAPTIVE SEQUENTIAL DESIGN IN CLINICAL TRIALS
Journal of Biopharmaceutical Statistics, 18: 1184 1196, 2008 Copyright Taylor & Francis Group, LLC ISSN: 1054-3406 print/1520-5711 online DOI: 10.1080/10543400802369053 SAMPLE SIZE RE-ESTIMATION FOR ADAPTIVE
More informationAssessing the Effect of Prior Distribution Assumption on the Variance Parameters in Evaluating Bioequivalence Trials
Georgia State University ScholarWorks @ Georgia State University Mathematics Theses Department of Mathematics and Statistics 8--006 Assessing the Effect of Prior Distribution Assumption on the Variance
More information