Abstract
Background:
Illuminating heterogeneity of treatment effect (HTE) within trials is important for identifying target populations for implementation.
Objectives:
To examine HTE in a trial of group medical visits (GMVs) for patients with type 2 diabetes and elevated body mass index (BMI).
Research Design and Measures:
Participants (n=263) were randomized to GMV-based medication management plus low carbohydrate diet-focused weight management (WM/GMV; n = 127) or GMV-based medication management alone (GMV; n = 136) for diabetes control. We used QUalitative INteraction Trees (QUINT), a tree-based clustering method, to identify subgroups with greater improvement in hemoglobin A1c (HbA1c) and weight from either WM/GMV or GMV. Subgroup predictors included 32 baseline demographic, clinical, and psychosocial factors. Internal validation was conducted to estimate bias in the range of mean outcome differences between arms.
Results:
QUINT analyses indicated that for patients who had not previously attempted weight loss, WM/GMV resulted in better glycemic control than GMV (mean difference in HbA1c improvement = 1.48%). For patients who had previously attempted weight loss and had lower cholesterol and blood urea nitrogen, GMV was better than WM/GMV (mean difference in HbA1c improvement = 1.51%). No treatment-subgroup effects were identified for weight. Internal validation resulted in moderate corrections in mean HbA1c differences between arms; however, differences remained in the clinically significant range.
Conclusions:
This work represents a novel step toward targeting care approaches for patients to maximize benefit based on individual patient characteristics.
Keywords: QUINT, heterogeneity of treatment effect, diabetes, weight management, health services research
Introduction
Clinical trials report average treatment effects across the study population as their main finding, often neglecting potential variability in individual responses to treatment. Exploration of this variability, or heterogeneity of treatment effect (HTE), is needed to better match patients with more personalized, optimal treatment approaches. Doing so could have important implications for the management of many disease processes, particularly multifactorial, difficult-to-treat conditions such as type 2 diabetes and obesity.
Weight loss and medication management remain the sentinel strategies for achieving glycemic control among patients with diabetes. Despite the presence of nationwide campaigns targeting weight loss (1–5) and effective medications to treat diabetes, rates of obesity and suboptimal glycemic control continue to rise (6, 7). Joint guidelines from the American Diabetes Association (ADA) and the European Association for the Study of Diabetes (EASD) have recommended a patient-centered, personalized approach for the treatment of diabetes, incorporating demographic factors, comorbidities, risk of hypoglycemia, and patient preferences, among others (8, 9). Identification of optimal target populations for particular clinical interventions links this personalized approach to the world of clinical trials. To make this linkage valid, principled statistical approaches are needed to better understand HTE.
Exploratory analyses of clinical trials have demonstrated that responses to diabetes and obesity medication management and lifestyle interventions may differ by individual baseline factors, such as hemoglobin A1c (HbA1c), body mass index (BMI), duration of diabetes and obesity, and patient self-efficacy (10–15). However, these traditional approaches often use regression analysis with interaction terms or analysis of variance with paired comparisons, requiring a priori hypotheses and limiting the number of moderator variables capable of being examined (16, 17). With multifactorial diseases such as diabetes and obesity, a priori selection of a restricted number of potential effect moderators may neglect important, understudied factors or combinations of factors that influence responses to an intervention.
To overcome these limitations, new data-driven methods based on recursive partitioning have been developed to handle large numbers of potential moderators with no clear a priori hypotheses, allowing for comprehensive exploration of multidimensional subgroups exhibiting HTE (18–22). The methods are recursive because each subgroup may be partitioned again until some stopping criteria is met. One particular recursive partitioning method of interest is QUalitative INteraction Trees (QUINT), a tree-based approach for determining if qualitative subgroup interactions are present in studies with two active treatments. A qualitative subgroup interaction identifies subgroups based on baseline characteristics where one treatment may work better for one subgroup, while another treatment may work better for another subgroup (18). The goal of data-driven methods such as QUINT is to optimize treatment assignment criteria.
The present study used data from the Jump Start Shared Medical Appointments for Diabetes with Weight Management trial (Jump Start) to identify qualitative subgroup interactions. Jump Start was a noninferiority trial comparing the effectiveness of group medical visits (GMVs) focused on intensive weight management in addition to medication management versus medication management alone on glycemic control for patients with type 2 diabetes and body mass index ≥27kg/m2. The two arms showed similar meaningful improvements in hemoglobin A1c (HbA1c) over 48 weeks (GMV: −0.8%, WM/GMV: −0.9%), but the weight management arm experienced greater weight reduction (GMV: −0.4 kg, WM/GMV: −4.1 kg) (23). The aim of the present study was to apply QUINT to identify characteristics of individuals who are more or less likely to experience improved weight and glycemic control from GMVs focused on low-carbohydrate weight management in addition to medication management versus medication management only. This exploratory work represents a direct application of a broadly applicable statistical method capable of advancing our understanding of HTE across many disease processes and treatments.
Methods
Study design and study population
Data from the “Jump Start Shared Medical Appointments for Diabetes with Weight Management” (Jump Start) trial (Clinicaltrials.gov NCT01973972) were used for these analyses (23, 24). This study was approved by the Durham Veterans Affairs (VA) Medical Center Institutional Review Board. Study details have been published (23, 24).
Interventions
Jump Start patients were randomized to either: 1) a traditional GMV group focused on medication management and self-management counseling (GMV), or 2) a novel GMV group focused on combining low-carbohydrate diet-focused weight management with traditional medication management (WM/GMV), both for 48 weeks. Sessions in both arms occurred every 2–8 weeks based on preset schedules for a total of 9–13 possible sessions (depending on arm), lasted 90–120 minutes, and consisted of individual data collection, group content delivery, and one-on-one medication management with arm-specific medication goals. Module topics specific to each arm can be found in the Appendix (Tables A1 and A2). In general, GMV covered diabetes self-management topics while WM/GMV primarily covered low-carbohydrate diet and weight management topics. For medication management, the goal of GMV was to intensify antiglycemic and antihypertensive therapy until blood glucose and blood pressure were controlled, whereas the goal of WM/GMV was to avoid hypoglycemia and decrease use of medications promoting weight gain.
Outcome measures
Primary and secondary outcomes were changes in HbA1c and weight from baseline to 48 weeks, respectively. At 48 weeks, 37 participants missed follow-up assessment of HbA1c (WM/GMV n=18, GMV n=19) and 54 participants missed follow-up assessment of weight (WM/GMV n=26, GMV n=28). We used Empirical Best Linear Unbiased Prediction (EBLUP) estimates from linear mixed effects models as a single imputation for missing data in the 48-week outcomes and then calculated change scores such that larger positive changes indicated greater improvement in HbA1c and weight at 48 weeks (25).
Predictor variable selection
Potential moderator variables were collected at baseline, prior to randomization. All variables collected at baseline were considered in the present analyses; additional details can be found in the Appendix (Table A3). Based on a missingness and correlation analysis among the 36 candidate variables, we excluded income category and three variables derived from the International Physical Activity Questionnaires ((IPAQ); minutes of walking, moderate exercise, and vigorous exercise) to select the final list of 32 predictors. The final QUINT cohort excluded n=29 patients due to missing data in included predictors, leaving n=115 in WM/GMV and n=119 in GMV.
Statistical analyses: Overview of QUINT
QUalitative INteraction Trees (QUINT) is a tree-based clustering method designed to identify if qualitative treatment-subgroup interactions are present in a study using a binary partitioning algorithm. If present, the algorithm recursively splits the study sample into three potential subgroups of patients for which either treatment A is better than B, treatment B is better than A, or neither treatment is better (18, 26). The partitioning criterion (used to identify potential splits for subgroups in the study sample) of QUINT optimizes qualitative treatment-subgroup interactions by simultaneously maximizing the absolute difference in treatment outcome between interventions within each subgroup and the total sample size of each subgroup (to avoid trivial interactions based on only small subgroups). QUINT scans all candidate moderator variables and all split points of each moderator to optimize the partitioning criterion and thus the qualitative treatment-subgroup interaction identification. This process is repeated with each subsequent split point of the tree until the value of the partitioning criterion can no longer be increased, or until a prespecified stopping criteria is achieved.
Application of QUINT to the Jump Start study
QUINT analyses were performed using the R-package “quint” (version 2.1.0) (18, 27) in R version 1.2.1335 (R Core Team, 2019). Tuning parameters (e.g. minimum sample size for a subgroup, maximum number of leaves) were determined a priori based on QUINT documentation and published simulation data (27). The sample size in our analyses was relatively small for QUINT, so conservative tuning parameters were preferred. To fit the tree, the difference in means criterion was selected as the type of partitioning criterion because the scales of both our outcomes are readily interpretable. Default values were used for the weights of the difference in treatment outcome and cardinality (sample size) components (equal weighting), and the maximum number of leaves of the tree (10). For the partitioning criteria, we increased the minimum absolute effect size to 0.40 to minimize inferential errors based on results from a published simulation study (27). Given the relatively small sample size in Jump Start, the default value of 10% of treatment group sample sizes was increased to 40 participants (20 per treatment arm) per subgroup for the primary analysis and 30 participants (15 per treatment arm) per subgroup for the sensitivity analysis. For all analyses, we used a bootstrap-based bias-correction procedure with the recommended 200 bootstraps for tree pruning to reduce the risk of finding spurious subgroups (27). Lastly, to provide a bias-corrected estimate of the range in treatment outcome differences, we conducted an internal validation procedure via bootstrap resampling to estimate optimism or bias in the range of mean outcome differences between pairs of arms in the final selected tree (i.e., the apparent range) due to overfitting (28). We followed the steps in Sections B and C.2 of the web appendix of Dusseldorp and Mechelen for tree pruning and internal validation, respectively (18).
Results
Predictor variable selection
Descriptive statistics for the potential baseline predictors, intervention arm assignments, and primary and secondary outcomes were similar between the overall Jump Start sample (n=263) and the participants included in the QUINT analysis (n=234) (Table 1).
Table 1.
Baseline data and main outcomes of overall study participants and QUINT analysis participants with both treatment arms combined
| Overall (N=263) | QUINT analysis participants (N=234) | |
|---|---|---|
|
| ||
| Intervention arm | ||
| GMV, n (%) | 136 (51.7%) | 119 (50.9%) |
| WM/GMV, n (%) | 127 (48.3%) | 115 (49.1%) |
| Study Outcomes | ||
| HbA1c change at 48 weeks (%), mean (SD) | −0.858 (1.32) | −0.923 (1.31) |
| Weight change at 48 weeks (kg), mean (SD) | −2.24 (5.91) | −2.30 (6.02) |
| Attended ≥75% of sessions, n (%) | 152 (57.79%) | 144 (61.54%) |
| Demographic factors | ||
| Age (years), mean (SD) | 60.7 (8.18) | 60.5 (8.31) |
| Age at diagnosis (years)*, mean (SD) | 47.4 (10.3) | 47.1 (10.2) |
| Gender, male, n (%) | 235 (89.4%) | 208 (88.9%) |
| Race, n (%) | ||
| Black | 143 (54.4%) | 128 (54.7%) |
| White | 112 (42.6%) | 99 (42.3%) |
| Other | 8 (3.0%) | 7 (3.0%) |
| Distance from VA, n (%) | ||
| 0–20 miles | 124 (47.1%) | 107 (45.7%) |
| 21–40 miles | 86 (32.7%) | 80 (34.2%) |
| ≥40 miles | 53 (20.2%) | 47 (20.1%) |
| Education, n (%) | ||
| HS degree or less | 51 (19.4%) | 48 (20.5%) |
| Secondary school | 99 (37.6%) | 85 (36.3%) |
| Undergraduate degree (AS or BS) | 84 (31.9%) | 76 (32.5%) |
| Graduate work | 29 (11.0%) | 25 (10.7%) |
| Marital status*, n (%) | ||
| Not married | 102 (38.8%) | 90 (38.5%) |
| Married/living together | 160 (60.8%) | 144 (61.5%) |
| Income category*, n (%) | ||
| <$29,999 | 71 (27.0%) | 63 (26.9%) |
| ≥$30,000 to ≤$59,999 | 107 (40.7%) | 99 (42.3%) |
| ≥$60,000 | 73 (27.8%) | 64 (27.4%) |
| Clinical factors | ||
| Hemoglobin A1c, mean (SD) | 9.08 (1.29) | 9.08 (1.32) |
| Weight (kg), mean (SD) | 108 (19.6) | 109 (19.5) |
| Waist circumference (in), mean (SD) | 46.2 (5.11) | 46.4 (5.11) |
| Body mass index (kg/m2), mean (SD) | 35.3 (5.10) | 35.4 (5.09) |
| Diastolic blood pressure (mmHg)*, mean (SD) | 79.1 (10.8) | 78.9 (10.7) |
| Systolic blood pressure (mmHg)*, mean (SD) | 130 (17.3) | 130 (17.1) |
| Hypoglycemic events (n)*, mean (SD) | 1.25 (3.23) | 1.30 (3.32) |
| Glucose (mg/dL), mean (SD) | 181 (65.0) | 178 (65.1) |
| Creatinine (mg/dL), mean (SD) | 1.08 (0.231) | 1.07 (0.231) |
| Urea nitrogen (mg/dL), mean (SD) | 16.2 (5.53) | 16.1 (5.69) |
| Cholesterol (mg/dL), mean (SD) | 153 (39.3) | 153 (40.5) |
| High-density lipoprotein (mg/dL), mean (SD) | 40.7 (10.8) | 40.9 (10.8) |
| Low-density lipoprotein (mg/dL), mean (SD) | 91.3 (31.8) | 91.1 (32.6) |
| Medication regimen | ||
| Currently taking diabetes medications, yes (%) | 262 (99.6%) | 233 (99.6%) |
| Type of insulin use, n (%) | ||
| No insulin | 101 (38.4%) | 89 (38.0%) |
| Basal or premixed | 76 (28.9%) | 68 (29.1%) |
| Prandial | 86 (32.7%) | 77 (32.9%) |
| Complexity† of insulin regimen, n (%) | ||
| Simple | 175 (66.5%) | 156 (66.7%) |
| Complex | 88 (33.5%) | 78 (33.3%) |
| Diabetes medication adherence (VOILS MNQ)*, n (%) | ||
| Adherent | 100 (38.0%) | 89 (38.0%) |
| Non-adherent | 158 (60.1%) | 145 (62.0%) |
| Receiving cholesterol medication, yes (%) | 209 (79.5%) | 186 (79.5%) |
| Receiving hypertension medication, yes (%) | 235 (89.4%) | 211 (90.2%) |
| Psychosocial factors | ||
| PAID score*, mean (SD) | 30.5 (21.6) | 30.4 (21.6) |
| PHQ-2 score*, mean (SD) | 1.56 (1.70) | 1.52 (1.68) |
| PROMIS Pain Interference score, mean (SD) | 57.7 (9.93) | 57.8 (9.87) |
| EQ-5D score*, mean (SD) | 0.725 (0.168) | 0.730 (0.168) |
| Diabetes MES, mean (SD) | 2.29 (1.07) | 2.29 (1.08) |
| Prior attempts at weight loss, yes (%) | 218 (82.9%) | 194 (82.9%) |
| IPAQ: Walking exercise (mins)*, mean (SD) | 460 (737) | 473 (766) |
| IPAQ: moderate exercise (mins)*, mean (SD) | 39.9 (147) | 35.3 (143) |
| IPAQ: Vigorous exercise (mins)*, mean (SD) | 130 (349) | 125 (356) |
Abbreviations: VOILS MNQ = VOILS Medication Non-adherence Questionnaire; PAID = Problem Areas in Diabetes Questionnaire; PHQ-2 = Patient Health Questionnaire-2; PROMIS = Patient-Reported Outcomes Measurement Information System; EQ-5D = EuroQol-5D; MES = Medication Effect Score; IPAQ = International Physical Activity Questionnaire
Missing data in overall study population: Age at diagnosis (n=4), marital status (n=1), income category (12), diastolic blood pressure (n=9), systolic blood pressure (n=9), hypoglycemic events (n=5), diabetes medication adherence (n=5), PAID score (n=4), PHQ-2 score (n=11), PROMIS Pain interference score (n=4), EQ-5D score (n=3), IPAQ: walking, moderate, and vigorous exercise (n=32)
Simple insulin regimen is defined as only one type of insulin or no insulin. Complex insulin regimen is defined as multiple types of insulin use.
QUINT results: HbA1c
Figure 1 displays the final bootstrapped, pruned tree from our primary QUINT analysis of HbA1c improvement where minimum subgroup size was n=40 study participants (i.e., a minimum of n=20 in each arm within a subgroup). The final tree contained four leaves with three total split points, identical to the unpruned tree. The first split of the tree was based on the dichotomous response to, “Have you ever tried to lose weight before?” Those who answered “no” (n=40) represented a subgroup of study participants who derived greater HbA1c improvement from WM/GMV compared to GMV (dark gray, mean difference in HbA1c improvement between arms = 1.48%). For those who answered “yes”, further splits of the tree involved baseline total cholesterol (split point of 169 mg/dL) and with a second split for those with total cholesterol ≤ 169 mg/dL and blood urea nitrogen (BUN) (split point of 13.5 mg/dL) for a total of 3 additional subgroups. The subgroup of study participants (n=49) who derived greater benefit from GMV compared to the WM/GMV (light gray, mean difference in HbA1c improvement between arms = 1.51%) was defined by prior attempts at weight loss, cholesterol (≤169 mg/dL), and BUN (≤13.5 mg/dL). The final two subgroups of participants derived similar benefits in response to each intervention (white). The first of these subgroups (n=88) was defined by prior attempts at weight loss, cholesterol (≤169 mg/dL), and BUN (>13.5 mg/dL) (subgroup mean BUN = 19.6 +/− 4.9 mg/dL), while the second (n=57) was defined by prior attempts at weight loss and cholesterol (>169 mg/dL) (subgroup mean cholesterol = 201.9 +/− 27.8 mg/dL). It is important to note that while no intervention was preferred for each of these subgroups, both subgroups still experienced clinically meaningful improvements in response to each arm (WM/GMV: 0.65% versus GMV: 0.87% and WM/GMV: 1.42% versus GMV: 0.74%, respectively).
Figure 1.
Result of QUINT primary analysis for improvement in hemoglobin A1c with minimum sample size per subgroup of n=40 total participants (≥20 per arm). In the light gray leaves, GMV is more beneficial than WM/GMV while for the dark gray leaves the reverse is true. In the white leaves, there are no differences between GMV and WM/GMV. The scale (−2 to 2) represents the difference in mean HbA1c improvement between arms (GMV – WM/GMV).
Figure 2 displays the final bootstrapped, pruned tree from our sensitivity QUINT analysis where minimum subgroup size was n=30 study participants (i.e., a minimum of n=15 in each arm within a subgroup). The final pruned tree contained five leaves with four total split points; results were identical to the unpruned tree and similar to the primary analysis results. For the two subgroups that derived greater benefit with WM/GMV than GMV (dark gray), mean differences in HbA1c improvement were 1.48% and 1.04% higher in WM/GMV compared to GMV. One subgroup (n=40), characterized by participants who had never previously attempted weight loss, was identical to the primary analysis. The second subgroup (n=39) contained participants who had previously attempted weight loss with higher BUN (>11.5 mg/dL) and low-density lipoprotein cholesterol (LDL-C, >106.5 mg/dL). For the two subgroups that derived greater benefit with GMV than WM/GMV (light gray), mean differences in HbA1c improvement were 1.22% and 0.98% higher in GMV compared to WM/GMV. One subgroup (n=39) contained participants who had previously attempted weight loss with lower BUN (≤11.5 mg/dL) while the other subgroup (n=44) contained participants who had previously attempted weight loss and had higher blood urea nitrogen (>11.5 mg/dL), lower LDL-C (≤106.5 mg/dL), and higher systolic blood pressure (>134.5 mmHg). One subgroup of participants (n=72) derived similar benefits in response to each intervention (white), with a HbA1c improvement of 0.60% in the WM/GMV arm and 0.58% in the GMV arm.
Figure 2.
Result of QUINT sensitivity analysis for improvement in hemoglobin A1c with minimum sample size per subgroup of n=30 total participants (≥15 per arm). In the light gray leaves, GMV is more beneficial than WM/GMV while for the dark gray leaves the reverse is true. In the white leaves, there are no differences between GMV and WM/GMV. The scale (−2 to 2) represents the difference in mean HbA1c improvement between arms (GMV – WM/GMV).
QUINT results: weight
Using similar tuning parameters as the HbA1c analyses, we also examined weight change using QUINT. Results demonstrated no treatment-subgroup interactions. In the QUINT cohort, mean weight reduction was 3.93 kg greater in participants originally randomized to WM/GMV (n=115) compared to GMV (n=119) (4.3 kg versus 0.37 kg, respectively). This suggests there are no subgroups of study participants for which we would expect to see a greater weight loss in response to GMV compared to WM/GMV, thus the average treatment effect generalizes to all patients.
QUINT internal validation
For the primary analysis, the apparent range, defined as the difference between the largest positive difference in mean HbA1c between arms in a single subgroup (1.51%) and largest negative difference in mean HbA1c between arms in a single subgroup (−1.48%), was 2.99% (Appendix Table A4). After performing an internal validation procedure for the primary results and correcting for estimated optimism, the apparent range was reduced to 1.80%; however, this result still represented a clinically meaningful difference in HbA1c change. For the sensitivity analysis, we found a similar reduction in the apparent range from 2.70% to 1.38% when corrected for estimated optimism. Again, this represents a clinically meaningful difference in HbA1c change.
Discussion
This study applied a novel, principled approach for identifying moderators of treatment effect for a combined low-carbohydrate weight management-focused GMV (WM/GMV) versus a medication management-focused GMV (GMV) in patients with type 2 diabetes and elevated BMI. Results from the primary qualitative subgroup interaction analyses suggest that individuals new to diet and weight loss efforts may receive more benefit from adding carbohydrate restriction and dedicated weight loss counseling to GMVs. Conversely, for individuals who have previously attempted weight loss and have lower cholesterol and BUN, a medication intensification focus may be better. While exploratory in nature, these results demonstrate the potential value of using data-driven methods to explore HTE and highlight specific patient characteristics to prioritize in future hypothesis-driven research for diabetes and obesity.
Findings in the context of current literature
According to a cross-sectional study of over 48,000 adults from the National Health and Nutrition Examination Survey, from 1999–2016 there has been a significant upward trend in the proportion of the population who have attempted weight loss (34.3% to 42.2%) – paralleling the nationwide rise in BMI and weight – with reduced overall food consumption, increased consumption of fruits and vegetables, and exercise cited as some of the most common strategies (29). This prevalence of weight loss attempts is even higher among those with obesity, with reports of 60–70% attempting weight loss per year (30, 31). Despite these trends, it is unclear why our study participants with previous weight loss attempts responded less to the weight management intervention. Given these individuals still met BMI and diabetes inclusion criteria for this study, they might represent a subgroup inherently refractory or unlikely to adhere to a low-carbohydrate diet or other weight loss intervention. On the other hand, because individuals without previous weight loss attempts responded more to the weight management intervention, identifying such individuals may be important in clinical practice because they may particularly benefit from increased weight management resources and education. Current evidence suggests that lower income, lack of employment, and incorrect self-perception of weight status are associated with a lower likelihood of prior weight loss attempts and poor weight management behaviors (32–34).
Our study also demonstrated that among individuals who have previously attempted weight loss, those with lower cholesterol and BUN disproportionately benefitted from GMV medication intensification. Given that lower levels of cholesterol and BUN is consistent with less insulin resistance (35–38), this subgroup may be particularly responsive to intensification of medications that act through insulin (i.e., insulin, sulfonylureas, glinides). Additionally, lower cholesterol and lower BUN may be surrogates for responsiveness or adherence to other medications, particularly statins and renoprotective antihypertensives. Studies have shown that prior medication adherence and prescription filling characteristics predicts future adherence to additional pharmacotherapy (39–41). Further, among patients with diabetes and other chronic conditions, polypharmacy has not been linked to medication nonadherence and may even be associated with improved medication compliance (42–44). By corollary, intensification and optimization of antiglycemic medications in this study subgroup may lead to enhanced glycemic control given a predilection for medication response and adherence. It is important to note that a majority of individuals above the split points for cholesterol and BUN in our study were within normal to high-normal laboratory ranges. However, individuals with fasting serum LDL-C ≥190 mg/dL, fasting triglycerides ≥600 mg/dL, or kidney disease (serum creatinine >1.5mg/dL in men, >1.3 mg/dL in women) were excluded from Jump Start, possibly precluding involvement of individuals with worse levels of baseline medication responsiveness and adherence and insulin resistance. Even so, our results in conjunction with the current literature (39–44) suggest that physicians should not be deterred from adding on or intensifying medications to achieve optimal glycemic control, particularly among patients with a history of medication responsiveness.
For the weight analyses, the lack of qualitative subgroup interactions and the superiority of WM/GMV for weight reduction may be logical. Diabetes medication intensification can result in undesirable weight gain, while a low-carbohydrate diet can reduce diabetes medication needs, providing a direct and indirect pathway for weight reduction (3–5, 45). Unless a low-carbohydrate, weight management focus had detrimental effects on weight for a certain type of individual, there would be no well-defined subgroups of patients for which we would expect to experience more weight loss in the medication management focus relative to the weight management focus.
Implications of findings beyond Jump Start
Considering the heterogeneity of patients with type 2 diabetes and elevated BMI and evidence suggesting differential patient responses to a wide variety of targeted interventions (10–15), recent advancements in data driven methods such as QUINT may play a key role in better understanding HTE more broadly across other diabetes and weight treatment options. Next steps would be to repeat QUINT analyses in other diabetes and weight loss trials to finalize combinations of potential effect moderators, and then test these hypotheses in new clinical trials using stratified sampling schemes based on the moderators of interest. Ultimately, this optimization of treatment selection criteria may improve patient outcomes. Further, given that randomized controlled trials are the gold standard for evidence-based medicine, principled approaches such as QUINT appear to be well-suited for application across a whole host of disease processes and interventions, and thus are of broad clinical interest.
Limitations
The current study has limitations. Our Veteran cohort may limit generalizability of our findings, particularly to females. It is also important to acknowledge that QUINT is a data-driven method; subgroup identification and effect sizes may be spurious associations or overestimations (19), and only patient factors ascertained in the study can be investigated as potential moderators. Additionally, the sample size used in this analysis was slightly smaller than recommended for optimal use of QUINT, increasing the potential risk of type one and type two errors (18). We took several approaches to mitigate potential biases from these limitations. For one, a priori adjustment of our tuning parameters and tree pruning represented a conservative approach, with preference for smaller trees and larger subgroups to minimize inferential errors and overfitting (18, 27, 46). The unpruned, pruned, and final trees for the primary and sensitivity analyses were quite similar, lessening concerns of overfitting and instability of the identified subgroups. We performed an internal validation procedure using bootstrap resampling to provide bias-corrected estimates in our effect sizes (18). While moderate corrections were made, corrected estimates for differences in HbA1c response between arms for the subgroups in our final tree all remained in a clinically significant range, suggesting the results may be applicable beyond this cohort. We have not provided uncertainty estimates of effect size differences as standard uncertainty methods do not account for all the uncertainty due to the data driven process; this is an area of active research in HTE analysis. Furthermore, QUINT does not account for clustering of participants in GMVs, although estimates of the Intraclass Correlation Coefficient (ICC) were low for A1c (<0.01) and weight (not estimable). While these limitations are intrinsic to data-driven methods in general, such work is important for generating hypothesis-driven research and an important first step in optimizing treatment assignment criteria.
Conclusion
This study is the first to use QUINT analyses to examine HTE and identify moderators of treatment effect for diabetes and weight-focused interventions. While these analyses were exploratory in nature, they suggest there are subgroups of patients with type 2 diabetes that respond differentially to weight management and medication intensification interventions. For patients who had not previously attempted weight loss, weight management resulted in better glycemic control, while for patients who had previously attempted weight loss and had lower cholesterol and blood urea nitrogen, medication intensification was better. Additional QUINT analyses should be performed to validate these results among different diabetes and weight-related interventions. As a whole, this work starts to pave the way for more personalized medicine and may better help inform patient-expected outcomes.
Supplementary Material
Acknowledgments:
Sources of Funding: Research reported in this publication was supported by the Department of Veterans Affairs (IIR 13-053) and the Durham Center of Innovation to Accelerate Discovery and Practice Transformation (ADAPT) (CIN 13-410) at the Durham VA Health Care System. EAK is supported by the National Center for Advancing Translational Sciences of the National Institutes of Health (TL1 TR002555). Effort on this manuscript was funded by a Research Career Scientist award (RCS 14-443) to Dr. Voils and a Senior Research Career Scientist award (RCS 10-391) to Dr. Maciejewski from the Health Services Research & Development service of the Department of Veterans Affairs.
Footnotes
Conflicts of Interest: EAK, MJC, AJ, JZ, DEE, CIV, and CJC have no conflicts of interest to disclose. WSY is a scientific advisor for dietdoctor.com. MLM owns Amgen stock due to his spouse’s employment.
NIH trial registry number: NCT01973972
Disclaimer: The content is solely the responsibility of the authors and does not necessarily reflect the position or policy of Duke University, the United States Department of Veteran Affairs, the United States government, or the National Institutes of Health.
References
- 1.Engelgau MM, Geiss LS, Saaddine JB, et al. The evolving diabetes burden in the United States. Ann Intern Med 2004;140:945–950 [DOI] [PubMed] [Google Scholar]
- 2.Flegal KM, Carroll MD, Kit BK, et al. Prevalence of obesity and trends in the distribution of body mass index among US adults, 1999–2010. JAMA 2012;307:491–497 [DOI] [PubMed] [Google Scholar]
- 3.Franz MJ, Powers MA, Leontos C, et al. The evidence for medical nutrition therapy for type 1 and type 2 diabetes in adults. J Am Diet Assoc 2010;110:1852–1889 [DOI] [PubMed] [Google Scholar]
- 4.Kirk JK, Graves DE, Craven TE, et al. Restricted-carbohydrate diets in patients with type 2 diabetes: a meta-analysis. J Am Diet Assoc 2008;108:91–100 [DOI] [PubMed] [Google Scholar]
- 5.Norris SL, Zhang X, Avenell A, et al. Long-term effectiveness of lifestyle and behavioral weight loss interventions in adults with type 2 diabetes: a meta-analysis. Am J Med 2004;117:762–774 [DOI] [PubMed] [Google Scholar]
- 6.Ali MK, McKeever Bullard K, Imperatore G, et al. Characteristics associated with poor glycemic control among adults with self-reported diagnosed diabetes--National Health and Nutrition Examination Survey, United States, 2007–2010. MMWR Suppl 2012;61:32–37 [PubMed] [Google Scholar]
- 7.Carls G, Huynh J, Tuttle E, et al. Achievement of Glycated Hemoglobin Goals in the US Remains Unchanged Through 2014. Diabetes Ther 2017;8:863–873 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Inzucchi SE, Bergenstal RM, Buse JB, et al. Management of hyperglycemia in type 2 diabetes, 2015: a patient-centered approach: update to a position statement of the American Diabetes Association and the European Association for the Study of Diabetes. Diabetes Care 2015;38:140–149 [DOI] [PubMed] [Google Scholar]
- 9.Inzucchi SE, Bergenstal RM, Buse JB, et al. Management of hyperglycemia in type 2 diabetes: a patient-centered approach: position statement of the American Diabetes Association (ADA) and the European Association for the Study of Diabetes (EASD). Diabetes Care 2012;35:1364–1379 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Babenko AY, Savitskaya DA, Kononova YA, et al. Predictors of Effectiveness of Glucagon-Like Peptide-1 Receptor Agonist Therapy in Patients with Type 2 Diabetes and Obesity. J Diabetes Res 2019;2019:1365162. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.da Vico L, Monami M, Biffi B, et al. Targeting educational therapy for type 2 diabetes: identification of predictors of therapeutic success. Acta Diabetol 2013;50:309–317 [DOI] [PubMed] [Google Scholar]
- 12.Delahanty LM, Peyrot M, Shrader PJ, et al. Pretreatment, psychological, and behavioral predictors of weight outcomes among lifestyle intervention participants in the Diabetes Prevention Program (DPP). Diabetes Care 2013;36:34–40 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Duckworth WC, Abraira C, Moritz TE, et al. The duration of diabetes affects the response to intensive glucose control in type 2 subjects: the VA Diabetes Trial. J Diabetes Complications 2011;25:355–361 [DOI] [PubMed] [Google Scholar]
- 14.O’Connor PJ, Crabtree BF, Abourizk NN. Longitudinal study of a diabetes education and care intervention: predictors of improved glycemic control. J Am Board Fam Pract 1992;5:381–387 [PubMed] [Google Scholar]
- 15.Yagi S, Aihara KI, Kondo T, et al. Predictors for the Treatment Effect of Sodium Glucose Co-transporter 2 Inhibitors in Patients with Type 2 Diabetes Mellitus. Adv Ther 2018;35:124–134 [DOI] [PubMed] [Google Scholar]
- 16.Cohen J, Cohen P, West S, et al. Applied multiple regression/correlation analysis for the behavioral sciences. Mahwah, N.J.: L. Erlbaum Associates; 2003 [Google Scholar]
- 17.Shaffer J Probability of directional errors with disordinal (qualitative) interaction. Psychometrika 1991;56:29–38 [Google Scholar]
- 18.Dusseldorp E, Van Mechelen I. Qualitative interaction trees: a tool to identify qualitative treatment-subgroup interactions. Stat Med 2014;33:219–237 [DOI] [PubMed] [Google Scholar]
- 19.Lipkovich I, Dmitrienko A, D’ Agostino S BR. Tutorial in biostatistics: data-driven subgroup identification and analysis in clinical trials. Stat Med 2017;36:136–196 [DOI] [PubMed] [Google Scholar]
- 20.Loh WY, Fu H, Man M, et al. Identification of subgroups with differential treatment effects for longitudinal and multiresponse variables. Stat Med 2016;35:4837–4855 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Loh WY, He X, Man M. A regression tree approach to identifying subgroups with differential treatment effects. Stat Med 2015;34:1818–1833 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Seibold H, Zeileis A, Hothorn T. Model-Based Recursive Partitioning for Subgroup Analyses. Int J Biostat 2016;12:45–63 [DOI] [PubMed] [Google Scholar]
- 23.Yancy WS Jr., Crowley MJ, Dar MS, et al. Comparison of Group Medical Visits Combined With Intensive Weight Management vs Group Medical Visits Alone for Glycemia in Patients With Type 2 Diabetes: A Noninferiority Randomized Clinical Trial. JAMA Intern Med 2019 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Crowley MJ, Edelman D, Voils CI, et al. Jump starting shared medical appointments for diabetes with weight management: Rationale and design of a randomized controlled trial. Contemp Clin Trials 2017;58:1–12 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Littell RC. SAS for mixed models. Cary, N.C.: SAS Institute, Inc.; 2006 [Google Scholar]
- 26.Doove LL, Van Deun K, Dusseldorp E, et al. QUINT: A tool to detect qualitative treatment-subgroup interactions in randomized controlled trials. Psychother Res 2016;26:612–622 [DOI] [PubMed] [Google Scholar]
- 27.Dusseldorp E, Doove L, Mechelen I. Quint: An R package for the identification of subgroups of clients who differ in which treatment alternative is best for them. Behav Res Methods 2016;48:650–663 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Harrell FE, SpringerLink (Online service). Regression Modeling Strategies : With Applications to Linear Models, Logistic and Ordinal Regression, and Survival Analysis. Springer Series in Statistics,. Cham: Springer International Publishing : Imprint: Springer,; 2015:XXV, 582 p. 157 illus., 553 illus. in color. [Google Scholar]
- 29.Han L, You D, Zeng F, et al. Trends in Self-perceived Weight Status, Weight Loss Attempts, and Weight Loss Strategies Among Adults in the United States, 1999–2016. JAMA Netw Open 2019;2:e1915219. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Andreyeva T, Long MW, Henderson KE, et al. Trying to lose weight: diet strategies among Americans with overweight or obesity in 1996 and 2003. J Am Diet Assoc 2010;110:535–542 [DOI] [PubMed] [Google Scholar]
- 31.Martin CB, Herrick KA, Sarafrazi N, et al. Attempts to Lose Weight Among Adults in the United States, 2013–2016. NCHS Data Brief 2018:1–8 [PubMed] [Google Scholar]
- 32.Barebring L, Winkvist A, Augustin H. Sociodemographic factors associated with reported attempts at weight loss and specific dietary regimens in Sweden: The SWEDIET-2017 study. PLoS One 2018;13:e0197099. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Dorsey RR, Eberhardt MS, Ogden CL. Racial and ethnic differences in weight management behavior by weight perception status. Ethn Dis 2010;20:244–250 [PubMed] [Google Scholar]
- 34.Riggs A, Melton B, Bigham L, et al. Body Weight Misconception and Maintenance Among Female African American College Students. Journal of Black Studies 2017;48:698–709 [Google Scholar]
- 35.Gylling H, Hallikainen M, Pihlajamaki J, et al. Insulin sensitivity regulates cholesterol metabolism to a greater extent than obesity: lessons from the METSIM Study. J Lipid Res 2010;51:2422–2427 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Hoenig MR, Sellke FW. Insulin resistance is associated with increased cholesterol synthesis, decreased cholesterol absorption and enhanced lipid response to statin therapy. Atherosclerosis 2010;211:260–265 [DOI] [PubMed] [Google Scholar]
- 37.Spoto B, Pisano A, Zoccali C. Insulin resistance in chronic kidney disease: a systematic review. Am J Physiol Renal Physiol 2016;311:F1087–F1108 [DOI] [PubMed] [Google Scholar]
- 38.Xie Y, Bowe B, Li T, et al. Higher blood urea nitrogen is associated with increased risk of incident diabetes mellitus. Kidney Int 2018;93:741–752 [DOI] [PubMed] [Google Scholar]
- 39.Lauffenburger JC, Franklin JM, Krumme AA, et al. Predicting Adherence to Chronic Disease Medications in Patients with Long-term Initial Medication Fills Using Indicators of Clinical Events and Health Behaviors. J Manag Care Spec Pharm 2018;24:469–477 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40.Shields BM, Hattersley AT, Farmer AJ. Identifying routine clinical predictors of non-adherence to second-line therapies in type 2 diabetes: A retrospective cohort analysis in a large primary care database. Diabetes Obes Metab 2020;22:59–65 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 41.Wiegand P, McCombs JS, Wang JJ. Factors of hyperlipidemia medication adherence in a nationwide health plan. Am J Manag Care 2012;18:193–199 [PubMed] [Google Scholar]
- 42.Grant RW, Devita NG, Singer DE, et al. Polypharmacy and medication adherence in patients with type 2 diabetes. Diabetes Care 2003;26:1408–1412 [DOI] [PubMed] [Google Scholar]
- 43.Hope HF, Binkley GM, Fenton S, et al. Systematic review of the predictors of statin adherence for the primary prevention of cardiovascular disease. PLoS One 2019;14:e0201196. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 44.Rolnick SJ, Pawloski PA, Hedblom BD, et al. Patient characteristics associated with medication adherence. Clin Med Res 2013;11:54–65 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 45.Intensive blood-glucose control with sulphonylureas or insulin compared with conventional treatment and risk of complications in patients with type 2 diabetes (UKPDS 33). UK Prospective Diabetes Study (UKPDS) Group. Lancet 1998;352:837–853 [PubMed] [Google Scholar]
- 46.Hastie T, Tibshirani R, Friedman JH. The elements of statistical learning : data mining, inference, and prediction. New York: Springer; 2009 [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.