Maren K. Olsen

Multiple Imputation for Multivariate Missing-Data Problems: A Data Analyst's Perspective

Analyses of multivariate data are frequently hampered by missing values. Until recently, the only missing-data methods available to most data analysts have been relatively ad1 hoc practices such as listwise deletion. Recent dramatic advances in theoretical and computational statistics, however, have produced anew generation of flexible procedures with a sound statistical basis. These procedures involve multiple imputation (Rubin, 1987), a simulation technique that replaces each missing datum with a set of m > 1 plausible values. The rn versions of the complete data are analyzed by standard complete-data methods, and the results are combined using simple rules to yield estimates, standard errors, and p-values that formally incorporate missing-data uncertainty. New computational algorithms and software described in a recent book (Schafer, 1997a) allow us to create proper multiple imputations in complex multivariate settings. This article reviews the key ideas of multiple imputation, discusses the software programs currently available, and demonstrates their use on data from the Adolescent Alcohol Prevention Trial (Hansen & Graham, 199 I).

A low-carbohydrate, ketogenic diet versus a low-fat diet to treat obesity and hyperlipidemia: a randomized, controlled trial.

Context Low-carbohydrate weight reduction diets are popular despite a dearth of data on long-term efficacy and adverse effects. Contribution Community-dwelling hyperlipidemic persons were randomly assigned to either a low-carbohydrate, ketogenic diet or a low-fat, low-cholesterol, reduced-calorie diet for 24 weeks. Compared to the low-fat group, patients in the low-carbohydrate group lost more weight, had a greater decrease in triglyceride levels, and had higher high-density lipoprotein cholesterol levels. Levels of low-density lipoprotein cholesterol remained stable in both groups. Side effects were more common in the low-cholesterol group but were generally mild. Cautions While the study suggests the efficacy and relative safety of the low-cholesterol diet, the high dropout rate, self-directed adherence to the diet, and relatively short observation period challenge the generalizability of the findings. The Editors As the prevalence of obesity has increased over the past 20 years (1), the difficulties faced by overweight patients and their health care practitioners have become apparent. Fewer than 25% of Americans who attempt to lose weight actually reduce caloric intake and increase exercise as currently recommended (2). Persons who successfully lose weight have difficulty maintaining their weight loss (3). Therefore, it is not surprising that consumers spend

A Two-Part Random-Effects Model for Semicontinuous Longitudinal Data

33 billion yearly on weight loss products and services in search of effective therapies (2). Because many weight loss interventions are unproven and untested, practitioners often lack information with which to recommend a certain therapy or to monitor a patient once a therapy is chosen. One approach to weight loss that has gained recognition in the face of modest supportive scientific evidence is the low-carbohydrate diet. A popular version of this diet recommends extreme restriction of carbohydrate intake to less than 20 g/d initially (4). This level of carbohydrate restriction can induce serum and urinary ketones and weight loss (5, 6). However, until recently, available data on low-carbohydrate diets came from small studies of short duration, most of which were uncontrolled (5, 7-10). We examined body weight, body composition, serum lipid levels, and adverse effects over 24 weeks in hyperlipidemic persons who were randomly assigned to follow a low-carbohydrate, ketogenic diet or a low-fat, low-cholesterol, reduced-calorie diet commonly used to induce weight loss and decrease serum lipid levels. Methods Participants Generally healthy persons were recruited from the community. Inclusion criteria were age 18 to 65 years, body mass index of 30 to 60 kg/m2, desire to lose weight, elevated lipid levels (total cholesterol level > 5.17 mmol/L [>200 mg/dL], low-density lipoprotein [LDL] cholesterol level > 3.36 mmol/L [>130 mg/dL], or triglyceride level > 2.26 mmol/L [200 mg/dL]), and no serious medical condition. Exclusion criteria were use of any prescription medication in the previous 2 months (except for oral contraceptives, estrogen therapy, and stable thyroid medication), pregnancy or breastfeeding, use of any weight loss diet or diet pills in the previous 6 months, and baseline ketonuria. All participants provided written informed consent, and the institutional review board of Duke University Health System approved the study. Participants received no monetary incentive. Interventions By using a computer-generated simple randomization list, participants were allocated to receive the low-carbohydrate diet or low-fat diet. The intervention for both groups included group meetings, diet instruction, and an exercise recommendation. Group meetings took place at an outpatient research clinic twice monthly for 3 months, then monthly for 3 months. These meetings typically lasted 1 hour and consisted of diet instruction, supportive counseling, questionnaires, and biomedical measurements. During the study, participants selected their own menus and prepared or bought their own meals according to the guidelines presented to them. Participants were encouraged to exercise for 30 minutes at least 3 times weekly, but no formal exercise program or incentives were provided. Low-Carbohydrate Diet Using a popular diet book published by a lay press and additional handouts, trained research staff instructed participants to restrict intake of carbohydrates to less than 20 g/d (4). Participants were permitted unlimited amounts of animal foods (meat, fowl, fish, and shellfish), unlimited eggs, 4 oz of hard cheese, 2 cups of salad vegetables (such as lettuce, spinach, or celery), and 1 cup of low-carbohydrate vegetables (such as broccoli, cauliflower, or squash) daily. Participants were encouraged to drink 6 to 8 glasses of water daily. When participants were halfway to their goal body weight (determined at the week 10 visit with assistance from research personnel), they were advised to add approximately 5 g of carbohydrates to their daily intake each week until they reached a level at which body weight was maintained. To simulate the practice of the study sponsor, the low-carbohydrate diet group also received daily nutritional supplements (multivitamin, essential oils, diet formulation, and chromium picolinate; for a list of the composition of these supplements, see the Appendix) (6). Low-Fat Diet Using a commonly available booklet and additional handouts, a registered dietitian instructed participants in a diet consisting of less than 30% of daily energy intake from fat, less than 10% of daily energy intake from saturated fat, and less than 300 mg of cholesterol daily (11, 12). The recommended energy intake was 2.1 to 4.2 MJ (500 to 1000 kcal) less than the participants calculated energy intake for weight maintenance (body weight in pounds 10) (13). Primary Outcome Measure Body weight and body mass index were the primary outcome measures. At each visit, participants were weighed on the same calibrated scale while wearing lightweight clothing and no shoes. Body mass index was calculated as body weight in kilograms divided by height in meters squared. Secondary Outcome Measures Adherence Adherence to the diet was measured by self-report, food records, and, for the low-carbohydrate diet group, urinary ketone assessment. Diet Composition All participants completed a 24-hour recall of food intake at baseline and take-home food records (5 consecutive days, including a weekend) that were collected at each meeting during the study. Participants were instructed on how to document food intake and were given handouts with examples of how to complete the records. A sample of participants (13 in the low-carbohydrate diet group and 7 in the low-fat diet group) who completed the study was selected for food record analysis by the research staff on the basis of adequacy of detail in their records. A registered dietitian analyzed the food records by using a nutrition software program (Nutritionist Five, version 1.6 [First DataBank, Inc., San Bruno, California]). Ketonuria Restriction of dietary intake of carbohydrates to less than 40 g/d typically results in ketonuria that is detectable by dipstick analysis, which can be used to monitor adherence to the low-carbohydrate diet (14, 15). At each return visit, participants provided a fresh urine specimen for analysis. The following semi-quantitative scale was used to categorize ketone content: none, trace (up to 0.9 mmol/L [5 mg/dL]), small (0.9 to 6.9 mmol/L [5 to 40 mg/dL]), moderate (6.9 to 13.8 mmol/L [40 to 80 mg/dL]), large80 (13.8 to 27.5 mmol/L [80 to 160 mg/dL]), and large160 (>27.5 mmol/L [>160 mg/dL]). Body Composition Body composition was estimated by using bioelectric impedance (model TBF-300A [Tanita Corp., Arlington Heights, Illinois]) at approximately the same time of day (afternoon or evening) at each return visit. In a subset of 33 participants, the percentage of body fat as measured by bioelectric impedance had excellent correlation with the percentage as measured by dual-energy x-ray absorptiometry (r = 0.93 [95% CI, 0.87 to 0.97]). Vital Signs Blood pressure and pulse rate were measured in the nondominant arm by using an automated digital cuff (model HEM-725C [Omron Corp., Vernon Hills, Illinois]) after the participant had been sitting for 3 minutes. Two measurements were taken at each visit and averaged for the analysis. Serum Lipids and Lipoproteins Serum specimens for lipid measurement were obtained in the morning after at least 8 hours of fasting at the screening visit and at 8, 16, and 24 weeks. Other Metabolic Effects Serum tests for sodium, potassium, chloride, urea nitrogen, creatinine, calcium, phosphorus, total protein, albumin, uric acid, total bilirubin, alanine aminotransferase, aspartate aminotransferase, alkaline phosphatase, thyroid-stimulating hormone, iron, hemoglobin, leukocyte count, and platelet count were obtained at the screening visit and at 8, 16, and 24 weeks. The glomerular filtration rate was estimated by using an equation that included age; sex; race; and serum levels of albumin, creatinine, and urea nitrogen (Modification of Diet in Renal Disease Study equation) (16). Adverse Effects At all return visits, participants completed an open-ended questionnaire on side effects. At the 20- and 24-week visits, participants completed a checklist of the side effects that were most often mentioned during the study. Statistical Analysis Analyses were performed by using S-PLUS software, version 6.1 (Insightful Corp., Seattle, Washington), or SAS software, version 8.02 (SAS Institute, Inc., Cary, North Carolina). For categorical outcomes, groups were compared by using the chi-square test or Fisher exact test, as appropriate. For all primary and secondary continuous outcomes, linear mixed-effects models (PROC MIXED procedure in SAS software) that included fixed and random effects were used to determine expected mean values at each time point and to test hypotheses of group differences. In most body weight and b

Two self-management interventions to improve hypertension control: a randomized trial.

A semicontinuous variable has a portion of responses equal to a single value (typically 0) and a continuous, often skewed, distribution among the remaining values. In cross-sectional analyses, variables of this type may be described by a pair of regression models; for example, a logistic model for the probability of nonzero response and a conditional linear model for the mean response given that it is nonzero. We extend this two-part regression approach to longitudinal settings by introducing random coefficients into both the logistic and the linear stages. Fitting a two-part random-effects model poses computational challenges similar to those found with generalized linear mixed models. We obtain maximum likelihood estimates for the fixed coefficients and variance components by an approximate Fisher scoring procedure based on high-order Laplace approximations. To illustrate, we apply the technique to data from the Adolescent Alcohol Prevention Trial, examining reported recent alcohol use for students in grades 7–11 and its relationships to parental monitoring and rebelliousness.

Association Between Bariatric Surgery and Long-term Survival

In this trial, 636 patients with hypertension were randomly assigned to receive usual care; a telephone-delivered, nurse-administered behavioral self-management intervention; home blood pressure se...

Expectations and Outcomes of Prolonged Mechanical Ventilation

IMPORTANCE Accumulating evidence suggests that bariatric surgery improves survival among patients with severe obesity, but research among veterans has shown no evidence of benefit. OBJECTIVE To examine long-term survival in a large multisite cohort of patients who underwent bariatric surgery compared with matched control patients. DESIGN, SETTING, AND PARTICIPANTS In a retrospective cohort study, we identified 2500 patients (74% men) who underwent bariatric surgery in Veterans Affairs (VA) bariatric centers from 2000-2011 and matched them to 7462 control patients using sequential stratification and an algorithm that included age, sex, geographic region, body mass index, diabetes, and Diagnostic Cost Group. Survival was compared across patients who underwent bariatric surgery and matched controls using Kaplan-Meier estimators and stratified, adjusted Cox regression analyses. EXPOSURES Bariatric procedures, which included 74% gastric bypass, 15% sleeve gastrectomy, 10% adjustable gastric banding, and 1% other. MAIN OUTCOMES AND MEASURES All-cause mortality through December 2013. RESULTS Surgical patients (n = 2500) had a mean age of 52 years and a mean BMI of 47. Matched control patients (n = 7462) had a mean age of 53 years and a mean BMI of 46. At the end of the 14-year study period, there were a total of 263 deaths in the surgical group (mean follow-up, 6.9 years) and 1277 deaths in the matched control group (mean follow-up, 6.6 years). Kaplan-Meier estimated mortality rates were 2.4% at 1 year, 6.4% at 5 years, and 13.8% at 10 years for surgical patients; for matched control patients, 1.7% at 1 year, 10.4% at 5 years, and 23.9% at 10 years. Adjusted analysis showed no significant association between bariatric surgery and all-cause mortality in the first year of follow-up (adjusted hazard ratio [HR], 1.28 [95% CI, 0.98-1.68]), but significantly lower mortality after 1 to 5 years (HR, 0.45 [95% CI, 0.36-0.56]) and 5 to 14 years (HR, 0.47 [95% CI, 0.39-0.58]). The midterm (>1-5 years) and long-term (>5 years) relationships between surgery and survival were not significantly different across subgroups defined by diabetes diagnosis, sex, and period of surgery. CONCLUSIONS AND RELEVANCE Among obese patients receiving care in the VA health system, those who underwent bariatric surgery compared with matched control patients who did not have surgery had lower all-cause mortality at 5 years and up to 10 years following the procedure. These results provide further evidence for the beneficial relationship between surgery and survival that has been demonstrated in younger, predominantly female populations.

Identifying, recruiting, and retaining seriously-ill patients and their caregivers in longitudinal research

Objective:To compare prolonged mechanical ventilation decision-makers’ expectations for long-term patient outcomes with prospectively observed outcomes and to characterize important elements of the surrogate-physician interaction surrounding prolonged mechanical ventilation provision. Prolonged mechanical ventilation provision is increasing markedly despite poor patient outcomes. Misunderstanding prognosis in the prolonged mechanical ventilation decision-making process could provide an explanation for this phenomenon. Design:Prospective observational cohort study. Setting:Academic medical center. Patients:A total of 126 patients receiving prolonged mechanical ventilation. Interventions:None. Measurements and Main Results:Participants were interviewed at the time of tracheostomy placement about their expectations for 1-yr patient survival, functional status, and quality of life. These expectations were then compared with observed 1-yr outcomes measured with validated questionnaires. The 1-yr follow-up was 100%, with the exception of patient death or cognitive inability to complete interviews. At 1 yr, only 11 patients (9%) were alive and independent of major functional status limitations. Most surrogates reported high baseline expectations for 1-yr patient survival (n = 117, 93%), functional status (n = 90, 71%), and quality of life (n = 105, 83%). In contrast, fewer physicians described high expectations for survival (n = 54, 43%), functional status (n = 7, 6%), and quality of life (n = 5, 4%). Surrogate-physician pair concordance in expectations was poor (all &kgr; = <0.08), as was their accuracy in outcome prediction (range = 23%–44%). Just 33 surrogates (26%) reported that physicians discussed what to expect for patients’ likely future survival, general health, and caregiving needs. Conclusions:One-year patient outcomes for prolonged mechanical ventilation patients were significantly worse than expected by patients’ surrogates and physicians. Lack of prognostication about outcomes, discordance between surrogates and physicians about potential outcomes, and surrogates’ unreasonably optimistic expectations seem to be potentially modifiable deficiencies in surrogate-physician interactions.

Utility of hemoglobin A1c in predicting diabetes risk

Background: In order to improve the state of science in palliative care, we must increase our ability to document the real-time experience of patients and families as they traverse the end of life. Yet, frequently, prospective measurement is impeded by difficulty with patient identification, recruitment, enrollment, and retention. The palliative care literature is replete with descriptions of studies unable to meet enrollment goals, and that as a result, do not have adequate power to test hypotheses or draw conclusions. Objectives: To review the literature describing difficulties associated with ascertainment, enrollment, and attrition. To outline the successful recruitment methods of a new longitudinal study of patients and their caregivers. Design: A two-year longitudinal study of 240 patients with Stage IV cancer (breast, prostate, colorectal, lung), advanced congestive heart failure (CHF) LVEFB < 40 or advanced chronic obstructive pulmonary disease (COPD) pCO2 > 46, and their caregivers, interviewed monthly for up to two years. Patients were identified using clinical and administrative databases from one geographic region. Results: Representative and successful ascertainment was associated with use of clinical criteria and medical record review versus physician or other provider prognostication, use of recruitment letters from personal physician, recruitment letter content, brochure content, small monetary incentives, refined phone scripts, use of matched ethnicity interviewers, in-home and phone interview strategies, measure selection, patient and caregiver rapport, and on-going staff support (including grief and bereavement). Conclusions: Recruitment to prospective longitudinal studies at the end of life is difficult, but possible. The lessons learned from this study are applicable to future investigators conducting prospective research.

Home Blood Pressure Management and Improved Blood Pressure Control Results From a Randomized Controlled Trial

AbstractBACKGROUND: There is controversy surrounding the issue of whether, and how, to screen adults for type 2 diabetes. Our objective was to measure the incidence of new diabetes among outpatients enrolled in a health care system, and to determine whether hemoglobin A1c (HbA1c) values would allow risk stratification for patients’ likelihood of developing diabetes over 3 years. METHODS: We conducted a prospective cohort study with 3-year follow-up at a single large, tertiary care, Department of Veterans Affairs Medical Center (VAMC). A convenience sample of 1,253 outpatients without diabetes, age 45 to 64, with a scheduled visit at the VAMC, were screened for diabetes using an initial HbA1c measurement. All subjects with HbA1c ≥6.0% (normal, 4.0% to 6.0%) were invited for follow-up fasting plasma glucose (FPG). We then surveyed patients annually for 3 years to ascertain interval diagnosis of diabetes by a physician. The baseline screening process was repeated 3 years after initial screening. After the baseline screening, new cases of diabetes were defined as either the self-report of a physician’s diagnosis of diabetes, or by HbA1c ≥7.0% or FPG ≥7.0 mmol/L at 3-year follow-up. The incidence of diabetes was calculated as the number of new cases per person-year of follow-up. RESULTS: One thousand two hundred fifty-three patients were screened initially, and 56 (4.5%) were found to have prevalent unrecognized diabetes at baseline. The 1,197 patients without diabetes at baseline accrued 3,257 person-years of follow-up. There were 73 new cases of diabetes over 3 years of follow-up, with an annual incidence of 2.2% (95% confidence interval [CI], 1.7% to 2.7%). In a multivariable logistic regression model, baseline HbA1c and baseline body mass index (BMI) were the only significant predictors of new onset diabetes, with HbA1c having a greater effect than BMI. The annual incidence of diabetes for patients with baseline HbA1c ≤ 5.5 was 0.8% (CI, 0.4% to 1.2%); for HbA1c 5.6 to 6.0, 2.5% (CI, 1.6% to 3.5%); and for HbA1c 6.1 to 6.9, 7.8% (CI, 5.2% to 10.4%). Obese patients with HbA1c 5.6 to 6.0 had an annual incidence of diabetes of 4.1% (CI, 2.2% to 6.0%). CONCLUSIONS: HbA1c testing helps predict the likelihood that patients will develop diabetes in the future. Patients with normal HbA1c have a low incidence of diabetes and may not require rescreening in 3 years. However, patients with elevated HbA1c who do not have diabetes may need more careful follow-up and possibly aggressive treatment to reduce the risk of diabetes. Patients with high-normal HbA1c may require follow-up sooner than 3 years, especially if they are significantly overweight or obese. This predictive value suggests that HbA1c may be a useful test for periodic diabetes screening.

Differences in one-year health outcomes and resource utilization by definition of prolonged mechanical ventilation: a prospective cohort study

BACKGROUND To determine which of 3 interventions was most effective in improving blood pressure (BP) control, we performed a 4-arm randomized trial with 18-month follow-up at the primary care clinics at a Veterans Affairs Medical Center. METHODS Eligible patients were randomized to either usual care or 1 of 3 telephone-based intervention groups: (1) nurse-administered behavioral management, (2) nurse- and physician-administered medication management, or (3) a combination of both. Of the 1551 eligible patients, 593 individuals were randomized; 48% were African American. The intervention telephone calls were triggered based on home BP values transmitted via telemonitoring devices. Behavioral management involved promotion of health behaviors. Medication management involved adjustment of medications by a study physician and nurse based on hypertension treatment guidelines. RESULTS The primary outcome was change in BP control measured at 6-month intervals over 18 months. Both the behavioral management and medication management alone showed significant improvements at 12 months-12.8% (95% confidence interval [CI], 1.6%-24.1%) and 12.5% (95% CI, 1.3%-23.6%), respectively-but not at 18 months. In subgroup analyses, among those with poor baseline BP control, systolic BP decreased in the combined intervention group by 14.8 mm Hg (95% CI, -21.8 to -7.8 mm Hg) at 12 months and 8.0 mm Hg (95% CI, -15.5 to -0.5 mm Hg) at 18 months, relative to usual care. CONCLUSIONS Overall intervention effects were moderate, but among individuals with poor BP control at baseline, the effects were larger. This study indicates the importance of identifying individuals most likely to benefit from potentially resource intensive programs. TRIAL REGISTRATION clinicaltrials.gov Identifier: NCT00237692.