The research team conducted this investigation using data from the Quality Improvement in Depression (QID) collaborative. The QID collaborative consisted of four linked group-randomized trials that collected data from primary care clinicians and their depressed patients to test the effect of quality improvement interventions on the process and outcomes of depression care. The design and methods used in the QID studies are described in detail elsewhere.19
Consecutive patients were recruited from within practices. Patients who screened positive for current depression symptoms at the index visit brief CIDI depression screen52
and met DSM-IV criteria for a major depressive disorder on a structured interview52
were invited to enroll. The QID collaborative was approved by Institutional Review Boards at each participating institution, and study participants provided written informed consent.This analysis includes 1,023 depressed patients for whom we had baseline and 6-month follow-up data treated by 158 clinicians who had four or more patients in the study.
Clinician Burden Measures
Since one of our objectives was to develop a valid measure of clinician burden from data more readily available than direct clinician interview, we first tested the association of five candidate measures against a clinician interview we had available in the QID dataset for a subset of 87 clinicians. The first candidate measure of clinician burden was simply the score from the MOS Chronic Disease Survey,6
a validated and fully field-tested patient-completed instrument that queries for the presence of common chronic diseases. For each clinician, a score was derived from the average of all enrolled patients’ Chronic Disease scores. The remaining composite measures characterized patient mix and visit types by measuring for each clinician the total number of visits per week as well as the percentage of total weekly visits that were new patient visits, follow-up visits, or urgent/emergent care visits.
Eighty-seven clinicians in the QID studies had completed a depression-specific practice burden questionnaire consisting of seven questions from the Physician Belief Scale,53
modified specifically for depression.54
This scale is scored from 0–100; the excellent psychometric properties of this scale are reported elsewhere,54
as is its use in other studies.22
The instrument consists of statements regarding the practice burden associated with treating depression, such as “evaluating and treating depression problems will cause me to be more overburdened than I already am.”
Patient and clinician covariates were selected from those factors suggested by previous studies that may be associated with depression treatment. Clinician-level covariates included clinician age, gender, race/ethnicity, and medical specialty. Patient-level covariates included patient age, gender, race/ethnicity, educational level, marital status, number of chronic medical comorbidities,6
and baseline depression severity. Severity of depressive symptoms at baseline was measured using a 23-item adaptation of the Center for Epidemiological Studies Depression Scale (CES-D)55
developed by Ford et al.19
Continuous patient-level variables were centered at the grand mean prior to analysis.
Although basic provision of antidepressant medication (or psychotherapy) for depressed patients is a minimum quality goal for primary care clinicians, the AHCPR Depression Guideline Panel56
states that further clinical management, which includes providing support, advice, reassurance, side-effect monitoring, dosage adjustments, and adequate follow-up, is important. A set of quality indicators, based on AHCPR guidelines and expert panel review, has been defined for primary care57
and used in the QID dataset. Based on these standards, the outcome variable for this analysis, depression treatment intensity, was operationalized using patient-reported responses to six questions about depression treatment. These represent four key criteria for quality depression treatment in primary care (also see Table ): (a) PCP-initiated referral to MH professional for counseling; (b) PCP-initiated antidepressant therapy and/or antidepressant medication adjustment; (c) discussion of side effects of medications or encouragement to stay on antidepressants; and (d) adequate PCP follow-up—three or more visits. Based on the answers to these questions, intensity of treatment was categorized into four levels, creating an ordinal response variable: (1) no depression treatment—negative on criteria a and b, (2) antidepressant therapy or referral for counseling—positive response on criteria a or b or both and negative on criteria c and d, (3) treatment augmented with communication about medication or adequate follow-up—positive response on criteria a or b or both plus positive response on either c or d, and (4) treatment augmented with both communication about medication and adequate follow-up—positive response on criteria a or b or both and positive on both c and d.
Outcome: Depression Treatment Intensity. Patient-reported: Baseline to 6-month Follow-up
Statistical analysis was performed using SAS, version 9.1 for personal computers.58–59
Descriptive statistics and frequency distributions were generated for clinician and patient characteristics; t-tests and chi-square tests were used to identify differences between (1) patients and clinicians in the analysis and those excluded because of clinicians having fewer than four patients enrolled in the QID collaborative and (2) eligible patients with 6-month assessments who did not respond to the questions used to create the depression treatment intensity score and those who responded.
Two main analyses were carried out in order to: (1) test associations between dataset measures of practice burden and the clinician-reported perceived burden measure and (2) test associations between practice burden and depression treatment intensity, adjusting for relevant clinician and patient covariates. To test whether composite measures were associated with direct report of clinician burden, we examined bivariate associations using Spearman’s correlation coefficients, since this non-parametric measure of association does not assume normality of the underlying variables. Next, to test whether burden measures predicted depression treatment intensity, we used multilevel ordinal regression with clinician as a random effect.60–61
Conceptually, this approach is similar to a series of logistic regressions modeling the log odds of being in (1) response category 1 vs. 2, 3, or 4, (2) categories 1 or 2 vs. 3 or 4, and (3) categories 1, 2, or 3 vs. 4 (Statistical Appendix
available online). Patient- and clinician-level covariates considered clinically significant or with p-values
.2 were included in all models, as well as the intervention group variable to control for the impact of intervention condition on depression treatment intensity. Sample size varied slightly across the models due to missing data.
Multiple imputation was used for item non-response in the QID dataset, as described previously,19
resulting in five replicates of the dataset for both patients and clinicians. We performed analyses on all five imputed datasets and combined the results (mean) or pooled estimates (multilevel analyses) using standard methods62–63
to obtain pooled variance estimates that incorporate both within and between dataset variance.