Variable selection for zero-inflated and overdispersed data with application to health care demand in Germany.

In health services and outcome research, count outcomes are frequently encountered and often have a large proportion of zeros. The zero-inflated negative binomial (ZINB) regression model has important applications for this type of data. With many possible candidate risk factors, this paper proposes new variable selection methods for the ZINB model. We consider maximum likelihood function plus a penalty including the least absolute shrinkage and selection operator (LASSO), smoothly clipped absolute deviation (SCAD), and minimax concave penalty (MCP). An EM (expectation-maximization) algorithm is proposed for estimating the model parameters and conducting variable selection simultaneously. This algorithm consists of estimating penalized weighted negative binomial models and penalized logistic models via the coordinated descent algorithm. Furthermore, statistical properties including the standard error formulae are provided. A simulation study shows that the new algorithm not only has more accurate or at least comparable estimation, but also is more robust than the traditional stepwise variable selection. The proposed methods are applied to analyze the health care demand in Germany using the open-source R package mpath.

An economic evaluation of colorectal cancer screening in primary care practice.

INTRODUCTION: Recent colorectal cancer screening studies focus on optimizing adherence. This study evaluated the cost effectiveness of interventions using electronic health records (EHRs); automated mailings; and stepped support increases to improve 2-year colorectal cancer screening adherence. METHODS: Analyses were based on a parallel-design, randomized trial in which three stepped interventions (EHR-linked mailings ["automated"]; automated plus telephone assistance ["assisted"]; or automated and assisted plus nurse navigation to testing completion or refusal [navigated"]) were compared to usual care. Data were from August 2008 to November 2011, with analyses performed during 2012-2013. Implementation resources were micro-costed; research and registry development costs were excluded. Incremental cost-effectiveness ratios (ICERs) were based on number of participants current for screening per guidelines over 2 years. Bootstrapping examined robustness of results. RESULTS: Intervention delivery cost per participant current for screening ranged from $21 (automated) to $27 (navigated). Inclusion of induced testing costs (e.g., screening colonoscopy) lowered expenditures for automated (ICER=-$159) and assisted (ICER=-$36) relative to usual care over 2 years. Savings arose from increased fecal occult blood testing, substituting for more expensive colonoscopies in usual care. Results were broadly consistent across demographic subgroups. More intensive interventions were consistently likely to be cost effective relative to less intensive interventions, with willingness to pay values of $600-$1,200 for an additional person current for screening yielding ≥80% probability of cost effectiveness. CONCLUSIONS: Two-year cost effectiveness of a stepped approach to colorectal cancer screening promotion based on EHR data is indicated, but longer-term cost effectiveness requires further study.

The central role of comorbidity in predicting ambulatory care sensitive hospitalizations.

BACKGROUND: Ambulatory care sensitive hospitalizations (ACSHs) are commonly used as measures of access to and quality of care. They are defined as hospitalizations for certain acute and chronic conditions; yet, they are most commonly used in analyses comparing different groups without adjustment for individual-level comorbidity. We present an exploration of their roles in predicting ACSHs for acute and chronic conditions. METHODS: Using 1998-99 US Medicare claims for 1 06 930 SEER-Medicare control subjects and 1999 Area Resource File data, we modelled occurrence of acute and chronic ACSHs with logistic regression, examining effects of different predictors on model discriminatory power. RESULTS: Flags for the presence of a few comorbid conditions-congestive heart failure, chronic obstructive pulmonary disease, diabetes, hypertension and, for acute ACSHs, dementia-contributed virtually all of the discriminative ability for predicting ACSHs. C-statistics were up to 0.96 for models predicting chronic ACSHs and up to 0.87 for predicting acute ACSHs. C-statistics for models lacking comorbidity flags were lower, at best 0.73, for both acute and chronic ACSHs. CONCLUSION: Comorbidity is far more important in predicting ACSH risk than any other factor, both for acute and chronic ACSHs. Imputations about quality and access should not be made from analyses that do not control for presence of important comorbid conditions. Acute and chronic ACSHs differ enough that they should be modelled separately. Unaggregated models restricted to persons with the relevant diagnoses are most appropriate for chronic ACSHs.

An automated intervention with stepped increases in support to increase uptake of colorectal cancer screening: a randomized trial.

BACKGROUND: Screening decreases colorectal cancer (CRC) incidence and mortality, yet almost half of age-eligible patients are not screened at recommended intervals. OBJECTIVE: To determine whether interventions using electronic health records (EHRs), automated mailings, and stepped increases in support improve CRC screening adherence over 2 years. DESIGN: 4-group, parallel-design, randomized, controlled comparative effectiveness trial with concealed allocation and blinded outcome assessments. (ClinicalTrials.gov: NCT00697047) SETTING: 21 primary care medical centers. PATIENTS: 4675 adults aged 50 to 73 years not current for CRC screening. INTERVENTION: Usual care, EHR-linked mailings ("automated"), automated plus telephone assistance ("assisted"), or automated and assisted plus nurse navigation to testing completion or refusal ("navigated"). Interventions were repeated in year 2. MEASUREMENTS: The proportion of participants current for screening in both years, defined as colonoscopy or sigmoidoscopy (year 1) or fecal occult blood testing (FOBT) in year 1 and FOBT, colonoscopy, or sigmoidoscopy (year 2). RESULTS: Compared with those in the usual care group, participants in the intervention groups were more likely to be current for CRC screening for both years with significant increases by intensity (usual care, 26.3% [95% CI, 23.4% to 29.2%]; automated, 50.8% [CI, 47.3% to 54.4%]; assisted, 57.5% [CI, 54.5% to 60.6%]; and navigated, 64.7% [CI, 62.5% to 67.0%]; P < 0.001 for all pair-wise comparisons). Increases in screening were primarily due to increased uptake of FOBT being completed in both years (usual care, 3.9% [CI, 2.8% to 5.1%]; automated, 27.5% [CI, 24.9% to 30.0%]; assisted, 30.5% [CI, 27.9% to 33.2%]; and navigated, 35.8% [CI, 33.1% to 38.6%]). LIMITATION: Participants were required to provide verbal consent and were more likely to be white and to participate in other types of cancer screening, limiting generalizability. CONCLUSION: Compared with usual care, a centralized, EHR-linked, mailed CRC screening program led to twice as many persons being current for screening over 2 years. Assisted and navigated interventions led to smaller but significant stepped increases compared with the automated intervention only. The rapid growth of EHRs provides opportunities for spreading this model broadly.

Expected estimating equations via EM for proportional hazards regression with covariate misclassification.

In epidemiological and medical studies, covariate misclassification may occur when the observed categorical variables are not perfect measurements for an unobserved categorical latent predictor. It is well known that covariate measurement error in Cox regression may lead to biased estimation. Misclassification in covariates will cause bias, and adjustment for misclassification will be challenging when the gold standard variables are not available. In general, statistical modeling for misclassification is very different from that of the measurement error. In this paper, we investigate an approximate induced hazard estimator and propose an expected estimating equation estimator via an expectation-maximization algorithm to accommodate covariate misclassification when multiple surrogate variables are available. Finite sample performance is examined via simulation studies. The proposed method and other methods are applied to a human immunodeficiency virus clinical trial in which a few behavior variables from questionnaires are used as surrogates for a latent behavior variable.

Do primary care physicians lose contact with their colorectal cancer patients?

BACKGROUND: Primary care physicians and patients perceive that they lose contact with each other after a cancer diagnosis. The objective of this study was to determine whether colorectal cancer (CRC) patients are less likely to see their primary care physicians after cancer diagnosis. METHODS: This was a longitudinal cohort study using 1993 to 2001 Surveillance Epidemiology and End Results (SEER)-Medicare claims data. Eligible patients were those with stage 0 to 1 and 2 to 3 CRC aged 67 to 89 years at diagnosis. Main measures included the proportion of individuals with a face-to-face primary care visit and mean annual primary care visits per patient at baseline and during 5 years after treatment. RESULTS: Fewer than half of the cancer patients visited with a primary care physician at baseline. In the first year after treatment, patients with stage 0 to 1 CRC (48.9% vs 53.3%; P ≤ .001) and stage 2 to 3 CRC (43.6% vs 53.4%; P ≤ .001) significantly increased their likelihood of visiting a primary care physician from baseline. The proportion of patients with stage 0 to 1 CRC with a primary care visit remained relatively stable, and the proportion of patients with stage 2 to 3 CRC decreased somewhat between the first and fifth year after treatment. The findings for mean annual primary care visits per patient roughly paralleled those for the proportion of individuals with a primary care visit. CONCLUSIONS: Elderly patients with CRC, especially stage 2 to 3 CRC, increase rather than decrease contact with primary care providers after diagnosis. More work is needed to understand the care that different physician specialties provide cancer patients and to support their collaboration.

No pain, but no gain? The disappearance of angina hospitalizations, 1992-1999.

BACKGROUND: Hospitalization for angina is commonly considered an ambulatory care sensitive hospitalization and used as a measure of access to primary care. OBJECTIVE: To analyze time trends in angina-related hospitalizations and seek possible explanations for an observed, marked decline during 1992 to 1999. RESEARCH DESIGN: We analyzed Medicare claims of SEER-Medicare control subjects for occurrence of angina hospital discharges, using the Agency for Healthcare Research and Quality Prevention Quality Indicator (PQI) definition, along with occurrence of related events including angina admissions with revascularization, angina admissions discharged as coronary artery disease (CAD) or myocardial infarction, and overall ischemic heart disease discharges. SUBJECTS: Approximately 124,000 cancer-free Medicare beneficiary/ies, with subjects contributing data for 1 to 8 years. RESULTS: Angina PQI hospital discharges declined 75% between 1992 and 1999. CAD hospital discharges rose in a reciprocal pattern, while angina discharges with revascularization declined and discharges for myocardial infarction and ischemic heart disease were relatively constant during this time period. CONCLUSIONS: The marked decline in angina PQI hospital discharges during 1992-1999 does not appear to represent improvements in access to care or prevention of heart disease, but rather increased coding of more specific discharge diagnoses for CAD. Our findings suggest that angina hospitalization is not a valid measure for monitoring access to care and, more generally, demonstrate the need for careful, periodic re-evaluation of quality measures.

Special populations recruitment for the Women's Health Initiative: successes and limitations.

The Women's Health Initiative (WHI) is a study designed to examine the major causes of death and disability in women. This multi-arm, randomized, controlled trial of over 160,000 post-menopausal women of varying ethnic and socioeconomic backgrounds and a goal of 20% of the study participants from minority populations is perhaps one of the most challenging recruitment efforts ever undertaken. Of the two main study arms, the Clinical Trial (CT) and the Observational Study (OS), the CT arm recruitment goal was to randomize 64,500 postmenopausal women 50-79 years of age. Women enrolled in the study will be followed for a period of 8-12 years. Ten clinical centers, out of a total of 40 throughout the United States, were selected as minority recruitment centers on the basis of their history of interaction with and access to large numbers of women from certain population subgroups. WHI enrollment began in September 1993 and ended in December 1998, resulting in the randomization and enrollment of a total of 161,856 (17.5% minority) women participants (68,135 (18.5% minority) in the CT and 93,721 (16.7%) in the OS). Within the CT arm, WHI achieved 101.7% of the goal of 48,000 participants in the Dietary Modification (DM) component, and 99.4% of the goal of 27,500 in the hormone-replacement component (HRT), with 11.8% overlap between DM and HRT. Of those who expressed initial interest in WHI, African Americans had the highest randomization yields in the DM component and Hispanics had the highest in the HRT component (15.2% and 10.2%, respectively). Overall, mass mailing was the greatest source of randomized participants. In addition, minority clinics found community outreach, personal referrals, and culturally appropriate recruitment materials particularly effective recruitment tools. For minority recruitment, our findings suggest that the key to high yield is reaching the target population through appropriate recruitment strategies and study information that get their attention. Also, once minority subjects are reached, they tend to participate.