Recursive Partitioning Techniques Research Articles

Detecting Differential Item Functioning in Multidimensional Graded Response Models With Recursive Partitioning.

Differential item functioning (DIF) is a common challenge when examining latent traits in large scale surveys. In recent work, methods from the field of machine learning such as model-based recursive partitioning have been proposed to identify subgroups with DIF when little theoretical guidance and many potential subgroups are available. On this basis, we propose and compare recursive partitioning techniques for detecting DIF with a focus on measurement models with multiple latent variables and ordinal response data. We implement tree-based approaches for identifying subgroups that contribute to DIF in multidimensional latent variable modeling and propose a robust, yet scalable extension, inspired by random forests. The proposed techniques are applied and compared with simulations. We show that the proposed methods are able to efficiently detect DIF and allow to extract decision rules that lead to subgroups with well fitting models.

Perioperative Changes in Serum Transaminase Levels: Impact on Postoperative Morbidity Following Liver Resection of Hepatocellular Carcinoma.

To define how dynamic changes in pre- versus post-operative serum aspartate aminotransferase (AST) and alanine aminotransaminase (ALT) levels may impact postoperative morbidity after curative-intent resection of hepatocellular carcinoma (HCC). Hepatic ischemia/reperfusion can occur at the time of liver resection and may be associated with adverse outcomes following liver resection. Patients who underwent curative resection for HCC between 2010-2020 were identified from an international multi-institutional database. Changes in AST and ALT (CAA) on postoperative day (POD) 3 versus preoperative values () were calculated using the formula: based on a fusion index via Euclidean norm, which was examined relative to the comprehensive complication index (CCI). The impact of CAA on CCI was assessed by the restricted cubic spline regression and Random Forest analyses. A total of 759 patients were included in the analytic cohort. Median CAA was 1.7 (range, 0.9 to 3.25); 431 (56.8%) patients had a CAA<2, 215 (28.3%) patients with CAA 2-5, and 113 (14.9%) patients had CAA ≥5. The incidence of post-operative complications was 65.0% (n=493) with a median CCI of 20.9 (IQR, 20.9-33.5). Spline regression analysis demonstrated a non-linear incremental association between CAA and CCI. The optimal cutoff value of CAA=5 was identified by the recursive partitioning technique. After adjusting for other competing risk factors, CAA≥5 remained strongly associated with risk of post-operative complications (Ref. CAA<5, OR 1.63, 95%CI 1.05-2.55, P=0.03). In fact, the use of CAA to predict post-operative complications was very good in both the derivative (AUC 0.88) and external (ACU 0.86) cohorts (n=1137). CAA was an independent predictor of CCI after liver resection for HCC. Use of routine labs such as AST and ALT can help identify patients at highest risk of post-operative complications following HCC resection.

Vehicle Driving Safety of Underground Interchanges Using a Driving Simulator and Data Mining Analysis

In the process of driving in an underground interchange, drivers are faced with many challenges, such as being in a closed space, visual changes alternating between light and dark conditions, complex road conditions in the confluence section, and dense signage, which directly affect the safety and comfort of drivers in an underground interchange. Thus, driving simulation, building information modeling (BIM), and data mining were used to analyze the impact of underground interchange safety facilities on driving safety and comfort. Acceleration disturbance and steering wheel comfort loss values were used to assist the comfort analysis. The CART algorithm, classification decision trees, and neural networks were used for data mining, which uses a dichotomous recursive partitioning technique where multiple layers of neurons are superimposed to fit and replace very complex nonlinear mapping relationships. Ten different scenarios were designed for comparison. Multiple linear regression combined with ANOVA was used to calculate the significance of the control variables for each scenario on the evaluation index. The results show that appropriately reducing the length of the deceleration section can improve driving comfort, setting reasonable reminder signs at the merge junction can improve driving safety, and an appropriate wall color can reduce speed oscillation. This study indicates that the placement of traffic safety facilities significantly influences the safety and comfort of driving in underground interchanges. This study may provide support for the optimization of the design of underground interchange construction and internal traffic safety facilities.

The Goal of Intraoperative Blood Loss in Major Hepatobiliary Resection for Perihilar Cholangiocarcinoma: Saving Patients From a Heavy Complication Burden.

To determine the goal of intraoperative blood loss in hepatectomy for perihilar cholangiocarcinoma. Although massive bleeding can negatively affect the postoperative course, the target value of intraoperative bleeding to reduce its adverse impact is unknown. Patients who underwent major hepatectomy for perihilar cholangiocarcinoma between 2010 and 2019 were included. Intraoperative blood loss was adjusted for body weight [adjusted blood loss (aBL)], and the overall postoperative complications were evaluated by the comprehensive complication index (CCI). The impact of aBL on CCI was assessed by the restricted cubic spline regression. A total of 425 patients were included. The median aBL was 17.8 (interquartile range, 11.8-26.3) mL/kg, and the CCI was 40.6 (33.7-49.5). Sixty-three (14.8%) patients had an aBL<10 mL/kg, nearly half (45.4%) of the patients were in the range of 10 ≤aBL<20 mL/kg, and 37 (8.7%) patients had an aBL >40 mL/kg. The spline regression analysis showed a nonlinear incremental association between aBL and CCI; CCI remained flat with an aBL under 10 mL/kg; increased significantly with an aBL ranging from 10 to 20 mL/kg; grew gradually with an aBL over 20 mL/kg. These inflection points of ~10 and 20 mL/kg were almost consistent with the cutoff values identified by the recursive partitioning technique. After adjusting for other risk factors for the postoperative course, the spline regression identified a similar model. aBL had a nonlinear aggravating effect on CCI after hepatectomy for perihilar cholangiocarcinoma. The primary goal of aBL should be <10 mL/kg to minimize CCI.

Inductive open data study on teleworking dissatisfaction in Spain during the Covid-19 pandemic

PurposeThe research paper aims to study dissatisfaction of teleworking employees in Spain during the Covid-19 health pandemic in order to propose three models: sociodemographic profile of the teleworking dissatisfied employee; advantages and disadvantages for the teleworking dissatisfied employee and advantages for the teleworking dissatisfied employee.Design/methodology/approachThis study uses official open data obtained from the Spanish National Statistical Institute (INE, 2022) through Decision Trees statistical multivariable models implementing Classification and Regression Trees and Recursive Partitioning and Regression Trees techniques to determine the variables that can influence the satisfaction or dissatisfaction of the subjects.FindingsThis investigation offers three models with two sociodemographic profiles of dissatisfied teleworking employee, who is a high/middle-level manager/employee around 45 years old, and she/he lives with the partner. Regarding the most important advantage of teleworking, employees consider “use/saving of time” and as disadvantage “worse organization and coordination of work”.Originality/valueThis research provides empirical evidence with inductive reasoning on understanding the challenges of teleworking dissatisfied employees in Spain not only in turbulent times but also in “normalcy” to improve overall teleworker well-being and accomplish company’s and organization’s long-term objectives for better productivity and effectivity. The study has high practical value due to the integral approach incorporating dissatisfaction as a driver that can trigger negative behaviours towards the organizations and that is seldom addressed in the literature. Additionally, this paper could provide some new ideas for accomplishing “Spain Digital 2025” and “Europe’s Digital Decade: 2030” plans on institutional level.

Identifying heterogeneity using recursive partitioning: evidence from SMS nudges encouraging voluntary retirement savings in Mexico.

Individuals regularly struggle to save for retirement. Using a large-scale field experiment ( ) in Mexico, we test the effectiveness of several behavioral interventions relative to existing policy and each other geared toward improving voluntary retirement savings contributions. We find that an intervention framing savings as a way to secure one's family future significantly improves contribution rates. We leverage recursive partitioning techniques and identify that the overall positive treatment effect masks subpopulations where the treatment is even more effective and other groups where the treatment has a significant negative effect, decreasing contribution rates. Accounting for this variation is significant for theoretical and policy development as well as firm profitability. Our work also provides a methodological framework for how to better design, scale, and deploy behavioral interventions to maximize their effectiveness.

Flexible tree-structured regression models for discrete event times

Discrete hazard models are widely applied for the analysis of time-to-event outcomes that are intrinsically discrete or grouped versions of continuous event times. Commonly, one assumes that the effect of explanatory variables on the hazard can be described by a linear predictor function. This, however, may be not appropriate when non-linear effects or interactions between the explanatory variables occur in the data. To address this issue, we propose a novel class of discrete hazard models that utilizes recursive partitioning techniques and allows to include the effects of explanatory variables in a flexible data-driven way. We introduce a tree-building algorithm that inherently performs variable selection and facilitates the inclusion of non-linear effects and interactions, while the favorable additive form of the predictor function is kept. In a simulation study, the proposed class of models is shown to be competitive with alternative approaches, including a penalized parametric model and Bayesian additive regression trees, in terms of predictive performance and the ability to detect informative variables. The modeling approach is illustrated by two real-world applications analyzing data of patients with odontogenic infection and lymphatic filariasis.

Assessing layered HIV prevention programming: optimizing outcomes for adolescent girls and young women.

To assess how exposure to multiple, layered interventions predicts HIV-related outcomes among adolescent girls (15-19 years) and young women (20-24 years) in Kenya. Survey data from adolescent girls and young women (n = 736) with 14-16 months of engagement with DREAMS, a comprehensive HIV prevention program that provides a range of health education, life skills, social protection, and social and behaviour change interventions. Nonparametric recursive partitioning technique - classification and regression tree (CART) - to identify the best predictors (DREAMS interventions) for achieving the desired HIV-related outcomes (consistent condom use and no transactional sex or sexual violence). Among adolescent girls, schooling support reduced the likelihood of engaging in transactional sex, whereas schooling support and exposure to parenting program reduced the likelihood experiencing sexual violence. Likelihood of consistent condom use increased among adolescent girls with exposure to preexposure prophylaxis (PrEP), schooling support, and the violence prevention programming. Among young women, multiple pathways reduced the likelihood of engaging in transactional sex: exposure to the male sexual partner program; exposure to the youth fund program; exposure to the violence prevention program; or exposure/engagement with schooling support, parenting programming, and the youth fund program. For young women, consistent condom use increased with schooling support and male partner engagement. Additionally, engagement in violence prevention program and male partner engagement increased the likelihood of not experiencing sexual violence among young women. Exposure to a combination of DREAMS interventions predicted outcomes that can reduce HIV risk among AGYW, though the pathways differed by outcome and age group.

The Profile of Female Labor Force Participation in Malaysia Based on Recursive Partitioning Analyses

In determining the level of the Malaysian economy, one of the factors contributing to the economic development of the country is the availability of labour. Studies on the availability of female labour force by identifying female profiles in labor force participation were made using data from Labour Force Survey (LFS), Malaysia for reference year 1990, 2000, 2010 and the latest 2018. Referring to the latest year 2018, the recursive partitioning (RP) technique showed that four subgroup profiles of working women have been created equal to 12%, 43%, 57% and 78%, respectively (percentage of working women at the four subgroups). Majority of the working women are identified in the group of aged 20-59 and the highest certificate obtain was tertiary education (Certificate, Diploma and Bachelor’s Degree) with 78%. Comparison between year 2018 with others reference year, there was a difference in the profile of working women whose working age had increased from 54 to 59 years. There is also a difference in the highest qualification obtained from obtaining only UPSR or SRP qualification in year 1990 to Diploma and Bachelor’s degree certificate in year 2018. Therefore, efforts need to be enhanced by providing initiatives for this group so that those who are in this group or will join this group later will be motivated to join labour market.

An Exploration of Stress: Leveraging Online Data from Crowdsourcing Platforms

Background: Early detection of community health risk factors such as stress is of great interest to health policymakers, but representative data collection is often expensive and time-consuming. It is important to investigate the use of alternative means of data collection such as crowdsourcing platforms. Methods: An online sample of Amazon Mechanical Turk (MTurk) workers (N = 500) filled out, for themselves and their child, demographic information and the 10-item Perceived Stress Scale (PSS-10), designed to measure the degree to which situations in one’s life are appraised as stressful. Internal consistency reliability of the PSS-10 was examined via Cronbach’s alpha. Analysis of variance (ANOVA) was utilized to explore trends in the average perceived stress of both adults and their children. Last, Rasch trees were utilized to detect differential item functioning (DIF) in the set of PSS-10 items. Results: The PSS-10 showed adequate internal consistency reliability (Cronbach’s alpha = 0.73). ANOVA results suggested that stress scores significantly differed by education (p = 0.024), employment status (p = 0.0004), and social media usage (p = 0.015). Rasch trees, a recursive partitioning technique based on the Rasch model, indicated that items on the PSS-10 displayed DIF attributable to physical health for adults and social media usage for children. Conclusion: The key conclusion is that this data collection scheme shows promise, allowing public health officials to examine health risk factors such as perceived stress quickly and cost effectively.

Stratification of Postoperative Prognosis by Invasive Tumor Thickness in Perihilar Cholangiocarcinoma.

The pathological tumor classification of distal cholangiocarcinoma in the American Joint Committee on Cancer (AJCC) Cancer Staging Manual 8th edition is based on invasive depth, whereas that of perihilar cholangiocarcinoma (PHCC) continues to be layer-based. We aimed to clarify whether invasive depth measurement based on invasive tumor thickness (ITT) could help determine postoperative prognosis in patients with PHCC. We enrolled 184 patients with PHCC who underwent hepatectomy plus extrahepatic bile duct resection or hepatopancreatoduodenectomy with curative intent. ITT was measured using simple definitions according to the sectioning direction or gross tumor pattern. The median ITT was 5.8mm (range 0.7-15.5). Using the recursive partitioning technique, ITT was classified into grades A (ITT < 2mm, n = 9), B (2mm ≤ ITT < 5mm, n = 68), C (5mm ≤ ITT < 11mm, n = 81), and D (11mm < ITT, n = 26). The median survival times (MSTs) in patients with grade B, C, or D were 90.8, 44.6, and 21.1months, respectively (patients with grade A did not reach the MST). There were significant differences in postoperative prognosis between ITT grades (A vs. B, p = 0.027; B vs. C, p < 0.001; C vs. D, p = 0.004). Through multivariate analysis, regional node metastasis, invasive carcinoma at the resected margin, and ITT grade were determined as independent prognostic factors. ITT could be measured using simple methods and may be used to stratify postoperative prognosis in patients with PHCC.

Recursive Quad-Tree Block Partitioning for Data Embedding in Images

In this paper, we propose to embed messages in raw images using a recursive block partitioning technique adopted from HEVC video coding technology. This work introduces a quad-tree partitioning solution in which square blocks of pixels are partitioned recursively according to a homogeneity criterion. As such, large block sizes are allocated for homogenous areas of the image and smaller block sizes are used elsewhere. For this to work, the variable block sizes should cover the whole image without gaps. Additionally, the criterion used for determining the block partitioning should work for both message embedding and message extraction. As such, the extractor process can reproduce the same variable block sizes. Matrix encoding is used for message embedding and extraction, as such, a maximum of one-pixel value is changed per block. The proposed solution is assessed in terms of percentage of changed pixels, PSNR, SSIM, histogram changes and blind steganalysis. Comparison with existing work reveals that the proposed solution reduces both the image distortions and pixel change rates. With embedding rates ranging from 2.5 to 38%, the average PSNR of the cover images ranged from 72 to 60 dB, and the average normalized histogram differences ranged from 3 to 12 pixels. It is also shown that the proposed solution is less detectible when tested with blind steganalysis in comparison to existing solutions where the average detection accuracy is 49.8%.

Model-Based Recursive Partitioning of Patients' Return Visits to Multispecialty Clinic During the 2009 H1N1 Pandemic Influenza (pH1N1).

Background During the 2009 H1N1 influenza pandemic (pH1N1), the proportion of outpatient visits to emergency departments, clinics and hospitals became elevated especially during the early months of the pandemic due to surges in sick, 'worried well' or returning patients seeking care. We determined the prevalence of return visits to a multispecialty clinic during the 2009 H1N1 influenza pandemic and identify subgroups at risk for return visits using model-based recursive partitioning technique. Methods This study was a retrospective analysis of ILI-related medical care visits to multispecialty clinic in Houston, Texas obtained as part of the Houston Health Department Influenza Sentinel Surveillance Project (ISSP) during the 2009 H1N1 pandemic influenza (April 2009 - March 2010). The data comprised of 2680 individuals who made a total of 2960 clinic visits. Return visit was defined as any visit following the index visit after the wash-out phase prior to the study period. We applied nominal logistic regression and recursive partitioning models to determine the independent predictors and the response probabilities of return visits. The sensitivity and specificity of the outcomes probabilities were determined using receiver operating characteristic (ROC) curve. Results Overall, 4.56% (Prob. 0.0%-17.5%) of the cohort had return visits with significant variations observed attributed to age group (76.0%), type of vaccine received by patients (18.4%) and Influenza A (pH1N1) test result (5.6%). Patients in age group 0-4 years were 9 times (aOR: 8.77, 95%CI: 3.39-29.95, p<0.0001) more likely than those who were 50+ years to have return visits. Similarly, patients who received either seasonal flu (aOR: 1.59, 95% CI 1.01-2.50, p=0.047) or pH1N1 (aOR: 1.74, 95%CI: 1.09-2.75, p=0.022) vaccines were about twice more likely to have return visits compared to those with no vaccination history. Model-based recursive partitioning yielded 19 splits with patients in subgroup I (patients of age group 0-4 years, who tested positive for pH1N1, and received both seasonal flu and pH1N1 vaccines) having the highest risk of return visits (Prob.=17.5%). The area under the curve (AUC) for both return and non-return visits was 72.9%, indicating a fairly accurate classification of the two groups. Conclusions Return visits in our cohort were more prevalent among children and young adults, and those that received either seasonal flu or pH1N1 or both vaccines. Understanding the dynamics in care-seeking behavior during pandemic would assist policymakers with appropriate resource allocation, and in the design of initiatives aimed at mitigating surges and recurrent utilization of the healthcare system.

Oblique Survival Trees in Discrete Event Time Analysis.

One of the main objectives of survival analysis is to predict the failure time that is usually considered as a continuous variable. In longitudinal studies, the data are often collected at every certain period of time, for example, monthly, quarterly, or yearly. Such data require appropriate techniques to handle the discrete time values that often have incomplete information about the failure occurrence-so-called "censored cases." Tree-based models are common, assumption-free methods of survival prediction. In this paper, the author proposes three recursive partitioning techniques able to cope with discrete-time censored survival data, which, in contrast to already-existing models limited to univariate trees, allow splits to have a form of any hyperplane. The performance of proposed methods, expressed as a mean absolute error, was examined on the basis of both synthetic and real data sets available in the literature and compared with existing tree-based models. To demonstrate the applicability of the methods in identifying subgroups of patients with a similar survival experience and to assess the influence of covariates on the risk of failure, a Veteran's Administration lung cancer data set was used. The results confirm proposed models to be good prediction tools for discrete-time survival data.

Impact of tumour budding grade in 310 patients who underwent surgical resection for extrahepatic cholangiocarcinoma.

Tumour budding is a risk factor for poor prognosis in various cancers. Tumour buds may present an epithelial-mesenchymal transition (EMT) morphological phenotype. This study aimed to elucidate the prognostic impact of tumour budding grade and its association with clinicopathological and EMT-related features in perihilar cholangiocarcinoma (PHCC) or distal cholangiocarcinoma (DCC). Subjects included 195 PHCC and 115 DCC patients. The numbers of tumour buds in different patients were stratified for postoperative survival using the recursive partitioning technique. Consequently, the numbers of tumour buds in PHCC patients were classified into three grades; namely, low (0-4 buds); intermediate (5-11 buds); and high (≥12 buds); those of DCC patients were classified into two grades; namely, low (0-4 buds) and high (≥5 buds). In both PHCC and DCC patients, high tumour budding grade was associated with poor histological differentiation, higher pT factor, presence of lymphatic, venous, perineural invasion and regional lymph node metastasis. In PHCC patients, residual invasive tumour in the resected margin was also associated with high tumour budding grade. For both PHCC and DCC patients, high tumour budding grade was an independent adverse prognostic factor in multivariate analysis (P<0001 and P=0.046, respectively). Immunohistochemical examination revealed that the number of tumour buds increased in patients with tumours showing a mesenchymal profile (negative for E-cadherin and positive for vimentin). Higher tumour budding grade is associated with invasive clinicopathological features, adverse postoperative prognosis and EMT status in extrahepatic cholangiocarcinoma.

Psychosocial distress in oncology: using the distress thermometer for assessing risk classes.

Cancer patients often suffer for psychological distress, which can compromise their quality of life. Our study aimed to recognize risk classes for the development of psychosocial distress. Three hundred seventy-two adult cancer patients were assessed by the Distress Thermometer (DT) and Problem List at the National Cancer Research Centre "Giovanni Paolo II" of Bari. We also compiled a socio-medical and clinical-medical record survey ad hoc for collecting socio-demographic information and clinical variables. To examine the interplay among the different variables and distinguish internally homogeneous subgroups of patients with diverse risks of distress, the RECursive Partitioning and Amalgamation (RECPAM) technique was used. Most of patients were female and the most frequent diagnosis was breast cancer, followed by gastro-intestinal cancer and hematological cancer. Distress was present in 43% of the sample, with a total of 156 patients with a DT > 5. The RECPAM analysis identified three distinct and homogeneous patient subgroups (RECPAM classes) with different risks of distress: diagnosis, marital status, and Eastern Cooperative Oncology Group Performance Status. The use of the distress thermometer allows clinicians to identify patients with certain characteristics that may increase the risk of developing psychosocial distress. This evaluation can allow timely psychological intervention and improve the patient's therapeutic program.

A Conditional Inference Tree Model for Predicting Sleep-Related Breathing Disorders in Patients With Chiari Malformation Type 1: Description and External Validation.

The aim of this study is to generate and validate supervised machine learning algorithms to detect patients with Chiari malformation (CM) 1 or 1.5 at high risk of the development of sleep-related breathing disorders (SRBD) using clinical and neuroradiological parameters. We prospectively included two independent datasets. A training dataset (n = 90) was used to obtain the best model, whereas a second dataset was used to validate it (n = 74). In both cohorts, the same clinical, neuroradiological, and sleep studies were carried out. We used two supervised machine learning approaches, multiple logistic regression (MLR) and the unbiased recursive partitioning technique conditional inference tree (URP-CTREE), to detect patients at high risk of SRBD. We then compared the accuracy, sensitivity, and specificity of the two prediction models. Age (odds ratio [OR] 1.1 95% confidence interval [CI] 1.05-1.17), sex (OR 0.19 95% CI 0.05-0.67), CM type (OR 4.36 95% CI 1.14-18.5), and clivus length (OR 1.14 95% CI 1.01-1.31) were the significant predictor variables for a respiratory disturbance index (RDI) cutoff that was ≥ 10 events/h using MLR. The URP-CTREE model predicted that patients with CM-1 who were age 52 years or older and males with CM-1 who were older than 29 years had a high risk of SRBD. The accuracy of predicting patients with an RDI ≥ 10 events/h was similar in the two cohorts but in the URP-CTREE model, specificity was significantly greater when compared to MLR in both study groups. Both MLR and URP-CTREE predictive models are useful for the diagnosis of SRBD in patients with CM. However, URP-CTREE is easier to apply and interpret in clinical practice.

Variability in HbA1c, blood pressure, lipid parameters and serum uric acid, and risk of development of chronic kidney disease in type 2 diabetes.

Variability in HbA1c and blood pressure is associated with the risk of diabetic kidney disease (DKD). No evidence exists on the role of variability in lipids or serum uric acid (UA), or the interplay between the variability of different parameters, in renal outcomes. Within the AMD Annals database, we identified patients with ≥5 measurements of HbA1c, systolic blood pressure (SBP) and diastolic blood pressure (DBP), total-, high-density lipoprotein (HDL)- and low-density lipoprotein (LDL)-cholesterol, triglycerides, and UA. Patients were followed-up for up to 5 years. The impact of measures of variability on the risk of DKD was investigated by Cox regression analysis and recursive partitioning techniques. Four-thousand, two-hundred and thirty-one patients were evaluated for development of albuminuria, and 7560 for decreased estimated glomerular filtration rate (eGFR; <60 mL/min/1.73 m2 ). A significantly higher risk of developing albuminuria was associated with variability in HbA1c [upper quartile hazard ratio (HR) = 1.3; 95% confidence interval (CI) 1.1-1.6]. Variability in SBP, DBP, HDL-C, LDL-C and UA predicted the decline in eGFR, the association with UA variability being particularly strong (upper quartile HR = 1.8; 95% CI 1.3-2.4). The concomitance of high variability in HbA1c and HDL-C conferred the highest risk of developing albuminuria (HR = 1.47; 95% CI 1.17-1.84), while a high variability in UA (HR = 1.54; 95% CI 1.19-1.99) or DBP (HR = 1.47; 95% CI 1.11-1.94) conferred the highest risk of decline in eGFR. The variability of several parameters influences the development of DKD, having a different impact on albuminuria development and on the decline in GFR.

Item-focussed Trees for the Identification of Items in Differential Item Functioning.

A novel method for the identification of differential item functioning (DIF) by means of recursive partitioning techniques is proposed. We assume an extension of the Rasch model that allows for DIF being induced by an arbitrary number of covariates for each item. Recursive partitioning on the item level results in one tree for each item and leads to simultaneous selection of items and variables that induce DIF. For each item, it is possible to detect groups of subjects with different item difficulties, defined by combinations of characteristics that are not pre-specified. The way a DIF item is determined by covariates is visualized in a small tree and therefore easily accessible. An algorithm is proposed that is based on permutation tests. Various simulation studies, including the comparison with traditional approaches to identify items with DIF, show the applicability and the competitive performance of the method. Two applications illustrate the usefulness and the advantages of the new method.

Abstract 16399: Use and Customization of Risk Scores for Predicting Cardiovascular Events Using Electronic Health Record Data

Background: The Framingham Risk Score (FRS) and the ACC/AHA Pooled Cohort Score (PCS) are widely used in clinical practice to guide individual patient care decisions. However, these risk scores have been estimated and validated mostly using data from longitudinal cohort studies and their performance when applied to patient data extracted from electronic health records is less well-established. Methods: Risk factor and outcomes data were obtained from the electronic medical records and insurance claims of 84,130 adults aged 40-79 receiving care at a large health care delivery and insurance organization from 2001-2011. We assessed calibration and discrimination for four risk scores: the published versions FRS and PCS, and versions obtained by re-fitting the FRS and PCS Cox regression models using a subset of the available data. Population subgroups where the various models gave highly divergent risk predictions were identified using recursive partitioning techniques. Results: The original FRS was well-calibrated (Calibration statistic K &lt; 5.7, p = 0.13), but the original PCS was not ( K = 30, p &lt; 0.001). The original PCS yielded somewhat better discrimination than the original FRS (C-statistic C = 0.757 vs. 0.738). The refitted FRS ( K = 1.4, p = 0.71 and C = 0.754) and refitted PCS ( K = 8, p = 0.046 and C = 0.756) were both well-calibrated and showed good discrimination. The original and refitted FRS disagreed substantially on the risks of individuals within a number of population subgroups, most of which encompassed individuals under the age of 60; the original and refitted PCS differed substantially in their risk predictions for African-American males; and the refitted FRS and PCS produced similar risk predictions across subgroups. Conclusions: Both the FRS and PCS are appropriate for use in clinical decision support systems which rely on electronic health data, though it may be advisable to refit the statistical models they are based on using available data from the target population to ensure acceptable calibration and discrimination performance.