Data source
The Chang Gung Research Database (CGRD) was utilized. It comprises de-identified individual EMRs of disease diagnoses, medical visits (outpatient, inpatient, and emergency room), pharmacy records, examination reports, and laboratory data from seven medical institutes throughout Taiwan, covering 1.3 million individuals (about 6% of Taiwan’s total population) [16]. Its validity for real-world pharmacoepidemiological studies is documented elsewhere [16-20].
Study subjects
As shown in Figure 1, we identified patients diagnosed with T2D (International Classification of Diseases, Ninth Revision [ICD-9] disease code of 250.X0 or 250.X2; ICD-10 disease code of E11) during 2015-2019. Then, we included patients aged 18 years or older and newly initiated on dulaglutide or liraglutide in 2016-2018. The first prescription date of dulaglutide or liraglutide was the index date. All patients were followed up from the index date until the end of 2019. The study subject identification period of 2016-2018 for GLP-1RAs users allowed for at least one-year follow-up period. Patients were excluded if they were diagnosed with type 1 diabetes or gestation diabetes in the year before the index date, or treated with exenatide during the study period. To obtain sufficient data for patient baseline conditions, all study patients were required to have at least one clinic visit and one HbA1c record in the year before the index date.
Study effectiveness outcomes
The primary outcome was the comparison of dulaglutide versus liraglutide on the HbA1c change from the index date (baseline) to 3, 6, 9, and 12 consecutive months in the one-year follow-up. We identified HbA1c records within each 3-month interval and the HbA1c value closest to each corresponding assessment time point was used in the analyses. Body weight, systolic blood pressure (SBP), and liver (alanine aminotransferase [ALT]) and renal (estimated glomerular filtration rate [eGFR]) functions were also measured from baseline and every three months in the follow-up. We implemented multiple imputations using the Markov chain Monte Carlo method with an expectation maximization algorithm and combined 10 simulations to deal with missing data in the follow-up [21].
Statistical analyses
Analytic procedures are outlined in Figure 1. To minimize potential selection bias and enhance the comparability of study subjects between two treatment groups, we applied a propensity score (PS) matching procedure [22]. Propensity scores of study subjects were estimated using a multivariable logistic regression model based on various patient baseline characteristics listed in Table 1. We used the nearest-neighbor 1:1 PS matching with a caliper of 0.05 on the PS scale with 8-digit greedy matching [23].
Table 1. Characteristics of study patients before and after propensity score matching (PSM)
|
Before PSM
|
After PSM
|
|
Dulaglutide
(n = 1,513)
|
Liraglutide
(n = 1,512)
|
SMD
|
Dulaglutide
(n = 983)
|
Liraglutide
(n = 983)
|
SMD
|
Demographics
|
|
|
|
|
|
|
Age at the index date
|
57.612.6
|
57.613.6
|
< 0.01
|
57.013.0
|
57.113.3
|
< 0.01
|
Sex (male)
|
47.1%
|
50.6%
|
0.06
|
48.3%
|
47.6%
|
0.01
|
Biochemical tests in the year before the index date
|
|
|
|
|
|
|
Weight (kg)
|
77.7±18.2
|
77.1±17.1
|
0.02
|
77.8±17.6
|
77.5±16.8
|
0.01
|
SBP (mmHg)
|
140.1±20.2
|
139.8±20.6
|
0.01
|
140.7±19.9
|
140.2±20.2
|
0.02
|
DBP (mmHg)
|
78.5±12.0
|
77.1±11.9
|
0.10
|
78.8±12.0
|
78.2±12.0
|
0.04
|
HbA1c (%)
|
9.3±1.6
|
9.5±1.7
|
0.14
|
9.3±1.6
|
9.3±1.5
|
0.02
|
Fasting plasma glucose (mg/dL)
|
177.9±62.3
|
179.6±69.8
|
0.02
|
179.3±63.2
|
178.6±67.4
|
0.01
|
Cholesterol (mg/dL)
|
175.6±45.7
|
174.9±45.3
|
< 0.01
|
176.4±46.6
|
175.2±43.1
|
0.02
|
HDL-C (mg/dL)
|
43.9±12.3
|
42.6±11.9
|
0.02
|
43.4±12.7
|
43.6±11.6
|
0.02
|
LDL-C (mg/dL)
|
96.3±32.6
|
95.7±34.0
|
0.08
|
97.1±33.8
|
96.2±32.9
|
0.01
|
Triglycerin (mg/dL)
|
207.9±241.5
|
218.9±240.1
|
0.04
|
214.9±272.0
|
213.8±247.5
|
<0.01
|
eGFR (ml/min/1.73 m2)
|
81.5±36.9
|
79.4±38.3
|
0.06
|
82.4±38.2
|
82.1±35.6
|
<0.01
|
ALT (U/L)
|
35.0±29.4
|
34.5±31.8
|
0.01
|
36.1±30.2
|
35.4±32.6
|
0.02
|
Prior comorbidities in the year before the index date
|
|
|
|
|
|
|
aDCSI
|
1.8±2.5
|
2.5±2.9
|
0.24
|
2.0±2.6
|
1.9±2.3
|
0.02
|
CCI
|
1.8±1.8
|
2.1±2.0
|
0.14
|
1.9±1.8
|
1.8±1.8
|
<0.01
|
Hypertension
|
65.0%
|
67.1%
|
0.04
|
65.3%
|
65.8%
|
0.01
|
Dyslipidemia
|
71.4%
|
71.6%
|
< 0.01
|
71.9%
|
72.5%
|
0.01
|
Ischemic heart disease
|
11.4%
|
19.6%
|
0.22
|
13.8%
|
14.3%
|
0.01
|
Heart failure
|
3.8%
|
6.4%
|
0.11
|
4.5%
|
3.9%
|
0.03
|
Cerebrovascular disease
|
6.6%
|
8.3%
|
0.06
|
6.9%
|
8.3%
|
0.05
|
Liver disease
|
18.7%
|
19.2%
|
0.01
|
18.4%
|
18.4%
|
< 0.01
|
COPD
|
2.3%
|
2.1%
|
< 0.01
|
2.2%
|
2.1%
|
< 0.01
|
CKD
|
11.6%
|
17.0%
|
0.15
|
12.4%
|
12.0%
|
0.01
|
Cancer
|
12.2%
|
10.9%
|
0.03
|
11.3%
|
11.2%
|
< 0.01
|
Prior exposure of co-medications in the year before the index date
|
|
|
|
|
|
|
ACEI/ARB
|
60.2%
|
63.6%
|
0.07
|
61.3%
|
61.1%
|
< 0.01
|
Calcium channel blockers
|
22.2%
|
24.8%
|
0.06
|
22.0%
|
23.9%
|
0.04
|
β-blockers
|
28.9%
|
35.0%
|
0.13
|
29.8%
|
30.4%
|
0.01
|
Diuretics
|
15.0%
|
18.6%
|
0.09
|
14.8%
|
15.8%
|
0.02
|
Lipid-lowering agents
|
76.5%
|
76.3%
|
< 0.01
|
75.8%
|
75.6%
|
< 0.01
|
Nitrates
|
8.7%
|
15.1%
|
0.19
|
10.2%
|
10.0%
|
< 0.01
|
Digoxin
|
0.8%
|
0.8%
|
< 0.01
|
0.8%
|
0.6%
|
0.02
|
Antiplatelet
|
31.4%
|
36.4%
|
0.11
|
32.2%
|
32.2%
|
< 0.01
|
Anticoagulant
|
2.0%
|
3.2%
|
0.07
|
2.3%
|
2.3%
|
< 0.01
|
Antidepressant
|
8.4%
|
9.6%
|
0.04
|
9.1%
|
8.7%
|
0.01
|
Antipsychotic
|
4.2%
|
5.8%
|
0.06
|
4.8%
|
4.5%
|
0.01
|
NSAID
|
23.1%
|
24.3%
|
0.02
|
24.2%
|
24.5%
|
< 0.01
|
Concomitant GLAs at the index date
|
|
|
|
|
|
|
Metformin
|
81.2%
|
67.8%
|
0.31
|
78.1%
|
78.0%
|
< 0.01
|
Sulfonylurea
|
70.6%
|
46.0%
|
0.51
|
61.1%
|
62.7%
|
0.03
|
DPP-4i
|
5.5%
|
4.5%
|
0.04
|
5.1%
|
5.1%
|
< 0.01
|
Thiazolidinedione
|
23.5%
|
10.8%
|
0.34
|
15.0%
|
14.8%
|
< 0.01
|
Alpha glucosidase inhibitors
|
18.8%
|
8.1%
|
0.31
|
12.0%
|
11.6%
|
0.01
|
Meglitinide
|
2.5%
|
4.2%
|
0.09
|
3.4%
|
2.7%
|
0.04
|
SGLT-2i
|
4.6%
|
2.4%
|
0.12
|
2.3%
|
3.2%
|
0.05
|
Medical specialty at the index date
|
|
|
0.29
|
|
|
0.05
|
Metabolism and endocrinology
|
81.8%
|
83.0%
|
|
83.3%
|
83.9%
|
|
Cardiology
|
3.9%
|
9.0%
|
|
4.9%
|
5.4%
|
|
Family medicine
|
1.5%
|
1.7%
|
|
1.5%
|
1.8%
|
|
Other
|
12.8%
|
6.3%
|
|
10.3%
|
8.9%
|
|
Hospital level at the index date
|
|
|
0.16
|
|
|
0.02
|
Medical centers
|
46.0%
|
40.7%
|
|
51.0%
|
51.0%
|
|
Region hospitals
|
48.7%
|
55.5%
|
|
31.0%
|
30.3%
|
|
Local hospitals
|
5.3%
|
3.8%
|
|
18.0%
|
18.7%
|
|
Abbreviations: SMD, standardized mean difference, SBP, systolic blood pressure; DBP, diastolic blood pressure; eGFR, estimated glomerular filtration rate; ALT, alanine aminotransferase; CCI, Charlson comorbidity index; aDCSI, adapted Diabetes Complications Severity Index; COPD, chronic obstructive pulmonary disease; CKD, chronic kidney disease; ACEI/ARB, angiotensin-converting enzyme inhibitors/angiotensin receptor blockers; NSAID, non-steroidal anti-inflammatory drugs; GLA, glucose lowering agents; DPP-4i, dipeptidyl peptidase-4 inhibitors; SGLT-2i, sodium-glucose transport protein-2 inhibitors.
Note:
A standardized mean difference (SMD) value of > 0.10 indicates a statistical difference in a given patient characteristic between the two drug groups. Index date refers to the first date of initiation of dulaglutide or liraglutide.
Primary analyses were based on an intention-to-treat (ITT) scenario where the loss to follow-up in the CGRD, death, or end of the 12-month follow-up, whichever came first, was censored. Analyses were divided into two parts. First, we used the paired t-test to estimate changes in clinical effectiveness at 12 months from baseline within each treatment group for assessing the within-group difference, and then used the two-sample t-test to determine the between-group difference in changes of clinical effectiveness at 12 months from baseline. Second, to consider time-varying changes in biomarkers (e.g., HbA1c) that were repeatedly assessed every three months during the follow-up, we performed a mixed-model analysis to consider treatment groups, assessment time points, and the interaction of treatment groups and assessment time points as fixed effects and individual patients as a random effect [24].
A series of sensitivity and subgroup analyses were conducted. First, to account for possible over-estimation of treatment effects in the ITT analyses where non-adherence to treatments was ignored, we performed the as-treated analysis where patients who switched away from or discontinued the use of a study drug were also censored, in addition to the censoring defined in the ITT analyses (Sensitivity 1). Second, to avoid potential confounding from short-term or accidental use of GLP-1RAs, we performed analyses where only stable users were included (Sensitivity 2). Stable users were defined as patients who had at least three consecutive refills of dulaglutide or liraglutide with any gaps between two consecutive refills of less than 90 days [25]. Third, we performed analyses with adjustment for potential healthy user bias (Sensitivity 3). Specifically, the patients who received GLP-1RAs and also used a dipeptidyl peptidase 4 inhibitor (DPP-4i) or sodium glucose cotransporter 2 inhibitor (SGLT-2i) were identified as possible healthy users because the combined use of a GLP-1RAs with a DPP-4i or SGLT-2i is not reimbursed by Taiwan’s National Health Insurance program and patients have to pay out-of-pocket fees. Under this circumstance, patients who are willing to self-pay for more intensive treatments would be likely to be engaged in healthier behaviors. We re-ran the analyses using a subset of patients who did not use a DPP-4i or SGLT-2i in combination with GLP-1RAs to avoid potential healthy user bias. Fourth, to enhance the study generalizability through retaining study cohort patients as many as possible, we applied two PS weighting procedures, inverse probability of treatment weighting (IPTW) and standardized mortality ratio weighting (SMRW) [26] (Sensitivities 4 and 5). Specifically, the patients at the 5th to 95th percentiles of the distribution of PS were first trimmed to minimize potential residual confounding [27]. Then, for IPTW, dulaglutide users were weighted as the inverse of the estimated PS and liraglutide users were weighted as the inverse of 1 minus the estimated PS. For SMRW, dulaglutide users were given a weight of 1 and liraglutide users were given a weight based on the ratio of the estimated PS to 1 minus the estimated PS.
In subgroup analyses, the procedures that were performed in the primary analyses were applied to examine the treatment effects on study outcomes in subgroups according to a series of patient baseline characteristics, including HbA1c (≥ 9%, < 9%), age (≥ 65 years, < 65 years), eGFR (≥ 60 ml/min/1.73 m2, < 60 ml/min/1.73 m2), ALT (> upper normal limit [UNL], ≤ UNL), and body mass index (≥ 27 kg/m2, < 27 kg/m2). A two-tail p-value of less than 0.05 was considered statistically significant. Data were analyzed using SAS Enterprise Guide, version 7.1 (SAS Institute, Cary, NC, USA).
Meta-analysis
We further performed a meta-analysis on clinical effectiveness (i.e., HbA1c, weight and SBP) of liraglutide vs. dulaglutide by pooling the results from prior studies and the present study. Two reviewers (Chang and Shao) independently searched studies from the PubMed and Embase from the inception of database to May 31, 2020 that reported the comparison of liraglutide and dulaglutide. The search strategy and key terms were listed in Appendix Table 1. Effectiveness outcomes abovementioned were measured from 6 and 12 months of follow-up periods. We included both randomized control trials (RCTs) and observational studies without imposing any language restrictions. Data were presented as mean difference with 95% CIs. We conducted the random-effects model meta-analysis using the reverse invariance method. The statistical heterogeneity was assessed by the statistic I2. To minimize the heterogeneity of included studies, we further conducted subgroups analyses for meta-analysis of RCTs or observational studies only. Data were analyzed by Review Manager version 5.3 (Copenhagen: The Nordic Cochrane Centre, The Cochrane Collaboration, 2014).