We performed a retrospective cohort analysis of SMR website use linked to administrative health data at Group Health Cooperative (GHC), a mixed-model health delivery system. GHC provides health insurance and comprehensive care to ~500,000 residents in the northwestern U.S. In 20 clinics operated by GHC, patients choose a PCP who guides and coordinates their care. Beginning August 2003, patients enrolled in these clinics were able to access their SMR via the MyGroupHealth patient website, which was linked to the ambulatory electronic medical record (EpicCare, Verona, WI). A detailed description of the implementation and use of the patient website was reported previously (4
). In brief, the SMR features implemented included the following: secure messaging with health care providers; requesting medication refills and in-person appointments; and viewing test results, after-visit summaries, medical problem lists, allergies, and immunizations. During our study period, GHC physicians were expected to use the SMR to communicate with patients via secure messaging. Documentation of messages, after-visit summaries, and test results were available to patients and GHC health care providers via the shared electronic medical record. To help ensure information security, patients were required to verify their identity before using these features.
Participants included 6,185 individuals aged ≥65 years continually enrolled during the study period who were identified at GHC as having diabetes for at least 1 year before the implementation of the SMR (August 2003); they were followed to the end of the observation period (December 2007), or death. Study procedures were approved by the GHC institutional review board.
Outcomes were initial and subsequent use of the SMR. We defined initial SMR use as the date of the first use of at least one of the eight SMR features during the study period, following initial registration and subsequent postal verification. Initial use of SMR was treated as a binary outcome. Rates of continued SMR use were measured as the number of days/month in which patients used any of the listed features, accounting for censoring due to death.
Predictors were measured in the following manner. Administrative data provided age at baseline and sex. Neighborhood SES status was derived from the zip code for each patient in combination with census block information on median education and income levels derived from the 2000 U.S. census (13
). Distance to the clinic was calculated between the patient's home address and clinic location of the PCP; we chose a distance ≥27 km from the clinic to approximate ≥30 min of traveling time. We used the Johns Hopkins Adjusted Clinical Groups case mix system to measure each individual's overall morbidity burden (15
). This algorithm groups ICD-9 codes by similar expected amounts of care, taking into account acute and chronic conditions. Based on age, sex, and ICD-9 codes identified over the previous 12 months, Adjusted Clinical Groups software assigns a level of overall morbidity between 1 (none) and 6 (very high). Enrollees using insulin were identified using pharmacy data. The PCP's use of the SMR communication feature was measured as the proportion of secure messaging exchanges (or “threads”), divided by the number of threads plus the total number of in-person visits each PCP had with his or her panel in the same period (19
We treated age as categories (65–69, 70–74, and ≥75 years). Sex, driving distance >30 min, and low neighborhood SES were treated as binary variables. We grouped Adjusted Clinical Groups levels into three morbidity categories: “very high,” “high,” or “moderate and lower.” The PCP's proportion of secure messaging was treated as a group-linear variable, using deciles of use (e.g., 0–10%, 11–20%, etc., up to 100%).
Variables were generated as baseline and changing characteristics. We calculated baseline variables on 1 August 2003. For time-varying clinical characteristics of morbidity category, insulin use, and assigned PCP, we calculated a baseline measure and then determined if and when each enrollee changed status. Insulin use and assigned PCP were tracked each month; for overall morbidity, we calculated change variables each quarter. To emphasize larger morbidity changes, we designated morbidity change if the enrollee moved from “moderate or lower” to a “high” or “very high” morbidity level. Insulin initiation was determined when the first prescription for insulin was filled. Change to a PCP with a higher rate of secure messaging was indicated if 1) the patient changed PCP and 2) the new PCP's rate was at least 10 percentage points higher than the previous PCP.
We selected a Cox proportional hazard analysis with robust standard errors to examine the relationship of baseline predictors and time to initial SMR use. After unadjusted analyses for each predictor, we fit an adjusted model with all of the above predictors. To evaluate the effect of changing clinical variables over time, we used an alternate Cox model and added time-dependent covariates updated each month (secure messaging, insulin use) or quarter (morbidity category). To assess whether the effect of worsening morbidity, starting insulin, or a change to a PCP with higher messaging was temporary or lasting, we fit a series of models where the effect of change was allowed to last varying lengths of time: 1 month, 3 months, 6 months, and lastly from the time of the change to the end of the study period.
To examine rates of SMR use among ever-users, we compared rates of use after initial use (days of SMR use/month) among various subgroups determined by the predictors. We then fit a multivariate model with similar covariates as used in the base Cox model to compare adjusted rates of use. To account for over-dispersion of count data, we used a negative binomial regression model with robust standard errors. Variables were reviewed for missingness. We performed sensitivity analyses to assess the effect of excluding participants who died during the study and tested proportional hazards assumptions for the base Cox model. All analyses were run on Stata IC version 10.1 (College Station, TX). P < 0.05 was considered statistically significant.