Study design and patients
We conducted a retrospective cohort study of patients who were randomly selected from a large outpatient general internal medicine practice affiliated with an academic health center in New York City. This practice has an electronic medical record, which captures all appointments, test orders, laboratory results and progress notes. There was no systematic screening program for diabetes in place at the time of this study, nor did the electronic medical record have specific flags for episodes of diabetes screening. The Institutional Review Board of the Weill Medical College of Cornell University approved this research.
We generated electronically a list of patients who had an initial visit in 1999 with a full-time attending physician (i.e. one who spent ≥80% effort on patient care per week). The rationale for restricting our sample to those with initial, or first, visits was that initial visits often have more complete data for diabetes risk factors, including family history. The alternative approach of tracing existing patients back to their initial visits was considered less feasible (due to transitions from paper to electronic medical records) and more prone to errors in data collection. We randomly sampled from the list of initial visits and reviewed medical records until we found at least 300 patients who met our inclusion criteria: we required that patients have at least 2 additional visits by the end of 2002, and we excluded patients who were younger than 20 years of age, had known diabetes at the initial visit, or were pregnant at any time during the study period. Our target sample size, based on a prior study of self-reported diabetes screening [5], provided 80% power to determine a 15% absolute difference between the rate of glucose testing among patients with 0–1 risk factors and the rate of glucose testing among patients with 2 or more risk factors [9].
Data collection
Two trained abstractors collected data manually from patients' electronic medical records. Data collected from the initial visit included demographic data, medical history, and physical exam findings. Each patient was then followed over time for all glucose tests ordered in the outpatient setting, until a new diagnosis of diabetes was made or until December 31, 2002, whichever came first. Thus, most patients were followed for at least 3 years, a strategy which is consistent with the screening interval recommended by the American Diabetes Association (ADA) [2] and which was chosen to maximize the study's ability to capture glucose testing for patients. If a primary care physician documented that a patient had been diagnosed with diabetes by an outside physician or during a hospitalization since the last visit, we counted that as a diagnosis and did not require the primary care physician to do further diagnostic testing. We recorded the number of visits and the number of different primary care physicians seen by each patient over the observation period. For each glucose test ordered, we recorded: the type of glucose test; the patients' fasting state during the test; polyuria or polydipsia reported at the time of the test; the physician's stated reason, if any, for ordering the test; the glucose value; the physician's interpretation of the glucose value; and the physician's subsequent action.
Inter-rater reliability
Using duplicate data collection for a random 10% sub-sample of the patients, we calculated percent agreement between data abstractors. Percent agreement was high for glucose orders (94%), type of glucose test (98%), fasting state (98%), presence of polyuria or polydipsia (98%), and action following glucose testing (100%). Percent agreement was lower for the reason for testing (66%) and for glucose interpretation (74%).
Statistical analysis
We used descriptive statistics based on complete data to characterize the patients in our sample, the patterns of glucose testing and rates of follow-up. We considered different cutoffs for clinically significant glucose values, based on the ADA's interpretation of fasting values (≥126 mg/dl for diabetes [2], 110–125 mg/dl for the old definition of impaired fasting glucose [10], and 101–125 mg/dl for the new definition of impaired fasting glucose [2, 11]). If glucose values were random or were not documented as fasting, we still considered values in these ranges to be potentially abnormal, because they should trigger repeat testing to confirm that a patient's fasting values are in the normal range. When calculating rates of follow-up for abnormal values, we conducted a sensitivity analysis to assess the effect of continuity, restricting the sample to only those who had all visits with the same physician.
We considered 10 potential patient predictors of glucose testing: age, gender, ethnicity, insurance type, family history of diabetes, hypertension, high cholesterol, body mass index (BMI), the number of visits during the study period, and continuity (whether all visits were with the same physician). Only 1 patient had a documented history of gestational diabetes and no patients had a documented history of impaired fasting glucose; thus, these variables were not considered in prediction models. For each variable included in the model, values were missing for <10% of patients, except for ethnicity (missing for 17%) and height (missing for 59%). To address this, we used multiple imputation to generate 5 complete versions of our dataset, with some variation in the imputed values across versions [12, 13]. We then analyzed each imputed dataset and combined estimates of odds ratios (OR) and 95% confidence intervals (CI) by applying Rubin's formula [14].
We used bivariate and multivariate logistic regression to measure associations between patient characteristics and the odds of having at least 1 glucose test ordered over the study period. Because the proportion of data missing for height was large, we generated 2 different multivariate models: the first model included weight alone (without height) and the second model included BMI (reflecting both height and weight). Results were similar for the bivariate models and for the first multivariate model whether we used complete or imputed data. We therefore display results based on imputed data.
In addition, we generated a composite variable for the number of diabetes risk factors each patient had [age 45 years or older, non-white ethnicity, family history of diabetes, hypertension, high cholesterol, and overweight or obese (BMI ≥ 25)]. We calculated the odds of glucose testing for each additional risk factor, and we separately calculated the odds of glucose testing for 2 or more risk factors vs. 0–1 risk factors.
We determined the proportions of patients who would be eligible for screening based on the criteria set by national guidelines. Patients were considered eligible for screening by the ADA if they were 45 years of age or older, or were overweight or obese and had 1 or more of the other diabetes risk factors listed above [2]. Patients were considered eligible for screening by the Centers for Disease Control and Prevention (CDC) if they were 25 years of age or older [3], and by the U.S. Preventive Services Task Force (USPSTF) if they had hypertension and/or hyperlipidemia [4, 15]. Among those eligible for screening by each guideline, we calculated the proportion that had at least 1 glucose test ordered.
We considered p-values ≤ 0.05 to be significant. Data analysis was performed using Stata 8 (College Station, TX).