Meta‐analysis of factors associated with antidiabetic drug prescribing for type 2 diabetes mellitus

There is a lack of consensus on prescribing alternatives to initial metformin therapy and intensification therapy for type 2 diabetes mellitus (T2DM) management. This review aimed to identify/quantify factors associated with prescribing of specific antidiabetic drug classes for T2DM.


| INTRODUCTION
Diabetes mellitus (DM) is a chronic progressive disorder characterised primarily by persistent hyperglycaemia 1 ; according to the International Diabetes Federation, in 2021, around 537 million adults were diagnosed with DM worldwide. 2 More than 90% of people with DM have type 2 DM (T2DM) which is characterised by chronic hyperglycaemia and insulin resistance, contributing to the development of diabetes-related life-threatening complications. 3 These complications can be prevented/attenuated by achieving adequate glycaemic control following an appropriate nonpharmacological and pharmacological care plan. 4,5 Several groups of antidiabetic drugs (ADDs) with different effectiveness and safety profiles are currently available. The most commonly used ADDs are metformin, sulfonylurea (SU), thiazolidinedione (TZD), dipeptidylpeptidase-4 inhibitors (DPP4-I), sodium glucose transporter-2 inhibitors (SGLT2-I), glucagon-like peptide receptor-1 agonists (GLP1-RA) and insulins. [4][5][6] All clinical guidelines have agreed on metformin as first-line therapy for patients newly diagnosed with T2DM. [7][8][9] However, the choice of intensifying therapy or alternative initial therapy in the presence of contraindications to metformin, is more variable, and prescribing decision could be influenced by several factors relevant to patients and drugs characteristics. [7][8][9] Several observational studies have evaluated the association of multiple factors with ADD prescribing (ADP) in clinical practice, including, for example, patient's age, sex, ethnicity, socioeconomic status, body mass index (BMI), glycaemic control (HbA1c), renal function and history of microvascular/macrovascular complications. [10][11][12][13][14] Nevertheless, no previous studies extensively quantified the impact of these different factors on prescribing decisions of different ADDs, which would be of interest especially after the introduction of newer ADDs which provided prescribers not only with wider options for T2DM management, but with ADDs that may have independent cardiac and renal protection effects. [15][16][17] Generally, factors associated with drug prescribing may indirectly reflect prescriber's adherence to guideline recommendations and specific drug features. This highlights the importance of studying these factors in a systematic way to assess the process of patient care by investigating which and how factors contribute to the decision-making in clinical practice. 18 Therefore, this systematic review (SR) and meta-analysis (MA) aimed to summarise and quantify factors associated with ADP at both the initiation and intensification stages.

| Data sources and search strategy
The search strategy was developed using three main concepts: participants (patients with T2DM), intervention (antidiabetic drugs) and outcome (factors associated with ADP). Medline/PubMed, Embase, Scopus and Web of Science were searched for studies published between January 2009 and January 2021 (the date of starting data synthesis). Additional searches were conducted to ensure literature saturation on ProQuest, Open Grey database and the reference lists of included articles. The search strategy was independently reviewed by experienced researchers and an academic librarian. As an example, the full Medline search strategy is available in the Appendix S2.

| Eligibility criteria
Only quantitative observational studies reporting factors associated with ADP among adults with T2DM in primarycare/outpatient settings and published in English were included (Table 1). Literature was searched from 2009 onwards to ensure the inclusion of newer ADDs (GLP1-RA, DPP4-I and SGLT2-I), which have been introduced into market since 2009. Only adults (≥18 years of age) were included to ensure that all people were subject to the same treatment recommendations since different treatments are recommended for children with T2DM.

| Study selection, data extraction and quality assessment
Two stages of study selection were conducted using Covidence 20 : title/abstract screening; and full-text screening. Relevant data was extracted from included studies using an MS Excel extraction form that was initially piloted on a random 10% of included studies to assess whether it captures all relevant data. All identified factors were mapped into four categories: 1 -demographic factors; 2 -clinical factors; 3 -socioeconomic factors; and 4 -prescriber-related factors, which were initially developed based on the literature around factors affecting physician's prescribing decision and modified as appropriate to fit the current research question. 21 Along with data extraction, the included studies were evaluated for the risk of bias using the Newcastle-Ottawa scale (NOS) for cohort studies 22 and the adapted NOS for cross-sectional studies. 23,24 More details on the quality assessment and study scoring are available in the Appendix S5. At each step of screening, extraction and quality assessment, a total of 20% of included studies were validated by two independent reviewers.

Category
Inclusion criteria

Language English
Publication Studies did not specify the type of antidiabetic groups being studied

| Data synthesis and meta-analysis
All factors related to ADP were identified from included studies. However, only factors reported by more than two studies for their association with the individual antidiabetic class were eligible for MA. Accordingly, MA was applied on five of the identified factors: age, sex, HbA1c, BMI and kidney-related problems. Three-level random-effect models were used to combine the pooled estimates (presented as odds ratio (OR) and 95% confidence intervals (CI)) measuring the association of each factor with ADP; a three-level MA approach was used to address the presence of dependency or correlation of effect sizes arising from reporting more than one effect size per study due to examining the outcomes of more than one antidiabetic group. 25,26 Subsequently, the pooled estimate measuring the association of the individual factor with each type of antidiabetic class was calculated using a two-level random effect model. Studies to be included in the MA had to report the effects of the identified factors as OR from either binary or continuous data; or provide primary baseline data required for OR calculation. Appendix S3 provides details on the method of effect size computation.
Study heterogeneity was measured by conducting Higgins &Thompson's (I 2 ) test over three levels to compute the overall heterogeneity as well as within-study (level-2) and between-studies (level-3) variance, with I 2 > 75% indicating high heterogeneity. 27 Furthermore, the usefulness and performance of the three-level model was evaluated by conducting log-likelihood-ratio test. [27][28][29] Moderator (sub-group) analysis was performed to explore any source of heterogeneity including the potential effect of several variables related to study characteristics on the overall estimate, such as class of ADDs, stage of treatment at which the outcome was assessed (initiation, intensification or not specified stage), quality of study, type of analysis used (adjusted vs. un-adjusted), study design and year of publication. 28 A p-value of <.05 indicates a significant result. Some factors (age, BMI, HbA1c) were reported as a binary variable in some studies and as a continuous variable in other studies. The pooled estimate of those factors was initially computed including all studies presenting the outcome as binary or continuous data following the approach described by Cochrane guideline. Additional sub-group analysis based on the type of reported data was performed to assess whether there was a significant difference in the pooled estimates according to the data type.
Publication bias was assessed with the funnel plot and extended Eggers' test. 30,31 Moreover, the number of outliers was measured and plotted as histogram for each MA and a sensitivity analysis was conducted to explore the influence of outliers on the pooled estimate. An effect size was considered as an outlier when its CI does not overlap with the CI of the pooled estimate. 32 Cook's distance (D) test was also performed with the results presented as scatter plots to explore the influential impact of included studies. 32 A Cook's-D value of ≥4/k (k: the number of effect sizes) indicated an influential impact of a study on the overall estimate. 33 All statistical tests were performed using RStudio; for full R-syntax, please refer to Appendix S4.

| Meta-analysis results
The following factors were identified in the included studies as factors associated with ADP (Table 3): demographic factors (patients' age, sex, ethnicity, smoking status, family history of diabetes and educational level), clinical factors (obesity, glycaemic status (HbA1c), kidney function, having macrovascular/microvascular complications or other comorbidities and diabetes duration), socioeconomic factors (deprivation level, income level, employment status, having insurance, area of living and type of medical facility) and prescriber-related factors (prescriber age, sex, speciality and practice experience).
However, it was possible to perform the MA on only five factors: age, sex, HbA1c, BMI and kidney-related problems.

| Sex
Out of the 40 eligible studies, 36 (90%) reported on sex association with ADP and all except one 50 were included in the MA, contributing to a total of 96 effect sizes. Heintjes et al. 50

| Age
Age was evaluated in a total of 38 studies; 31 studies were included in the MA, contributing to a total of 88 effect sizes. Seven studies were excluded since they did not present the outcome as OR and did not provide the required data for OR computation. 37

| Baseline BMI
The influence of BMI on ADP was evaluated in 21 studies. All except one 50 were included in the MA,

| Baseline glycaemic status (HbA1c)
A total of 62 effect sizes from 22 studies were included in the MA of HbA1c. Two studies were not included because of insufficient baseline data needed for OR calculation. 50

| Kidney-related problems
A total of 21 studies examined the impact of kidneyrelated problems in terms of chronic renal disease (CRD), nephropathy or based on the estimated glomerular filtration rate (eGFR) of <60 mL/min/1.73 m 2 . Only one study 50 was excluded due to insufficient data necessary for OR calculation, thus 20 studies were included in the MA, related to within-study variance while between-study variance for all studied factors were < 75% (Appendix S6). The results of the log likelihood ratio test (Appendix S6) indicated that the three-level model had a better fit for variability in data and better estimation of the pooled estimate.

| Outliers/influential studies
A total of 15 out of 96 effect sizes of sex data, 27 out of 88 effect sizes of age data, 18 out of 66 effect sizes of BMI data and 12 out of 61 effect sizes of renal data were detected as outliers; moreover, about half of the effect sizes of HbA1c data were detected as outliers (31/62). Histogram plots of all factors (Appendix S7) reflect that the potential outliers are not uniformly distributed around the pooled estimate. However, the results of the sensitivity analysis (Table 6) revealed a close overall OR and narrower but overlapped 95%CI of the pooled estimate after excluding the outliers compared to the one including the outliers. Nevertheless, it could not be determined whether the outliers did, in fact, bias the pooled estimate.
Cook's-D was measured for all factors (scatterplots in Appendix S7). None of the effect sizes included in the sex MA had a Cook's value exceeding 0.04(4/96), indicating that none had an influential effect on the pooled estimate. In contrast, two effect sizes of age and HbA1c were considered as influential cases in the model as they have a distance value of >0.05(4/88) and >0.06(4/62), respectively. 59,60,63,68 For BMI MA, only one study presented a distance value larger than 0.06(4/66). 68 Lastly, three effect sizes included in the MA of kidney-related problems were considered to have influential effect in the model with a distance value of >0.07(4/61). 49,61

| Publication bias
The funnel plots (Appendix S8) of all factors showed that all studies cluster at the top part of the plots, suggesting a possible presence of publication bias. Extended Eggers' test showed a significant possibility of asymmetry in the funnel plots of age, BMI and kidney-related problems (p < .0001, .0013 and <.0001, respectively), while the test  was non-significant for sex and HbA1c (p = .101 and .329, respectively).

| DISCUSSION
To our knowledge, no previous review either quantified the impact of several factors related to patients' characteristics on ADP; or compared their impact among different classes of ADDs. Age, baseline BMI and baseline HbA1c had the greatest impact on the selection of ADDs while patients' sex had the least impact. The significant variability in the pooled estimate of sex by class of ADDs could be linked to the differences in the number of studies investigating each antidiabetic class, or to the differences in the pharmacological characteristics of ADDs (mainly their safety and tolerability profile). The observed higher prescriptions of GLP1-RA for female patients compared to male patients could be explained in part by previous findings that GLP1-RA was better tolerated and associated with a lower cardiovascular risk among female patients. 71 On the contrary, the significantly lower prescriptions of TZD for female patients could be explained by the findings that female patients have experienced more side effects from TZD including weight gain, fracture and oedema. 72,73 This suggests a possible consideration of the variability in the effectiveness and tolerability of ADDs between female and male patients when making a decision on the appropriate ADDs in clinical practice. However, because of the limited number of studies examined the majority of antidiabetic classes, more studies are required to have a better understanding regarding the impact of sex on the choice of ADDs.
Despite the risk of SU-related hypoglycaemia being higher among older people, the pooled estimate of SU showed that older people were significantly more likely to use SU. The low cost of SU and the current availability of short-acting second-generation SU (e.g. glipizide) with fewer side effects might be partially responsible for the observed impact of age on SU prescription. 74 This could also reflect the legacy availability of SU for T2DM management as none of the newer ADDs were available 10 years ago, and patients started on SU may have stayed on the same regimen unless they developed intolerable side effects or required additional drug therapy. The safety of newer ADDs (GLP1-RA, SGLT2-I) in older adults was less studied thus prescribers might be less confident to prescribe the newer ADDs for older patients because of the higher concern that elderly patients are more susceptible to adverse reactions. [74][75][76] Furthermore, the higher cost of newer drugs, the cost of the required monitoring and the familiarity of prescribers with the update in clinical guidelines could contribute to the lower prescription of GLP1-RA and SGLT2-I for older patients. Therefore, further studies investigating prescribing quantity of newer ADDs for older patients are still required since older patients are more likely to have cardiovascular and renal diseases, the situations where the newer ADDs are recommended. The negative significant association between metformin prescription and age could be related to the fact that metformin is not recommended to be prescribed for patients with gastrointestinal complaints, functional impairment or with renal insufficiency, conditions that are increasingly present with increasing age. 75,77,78 This might positively reflect clinical practice adherence to drug characteristics when prescribing metformin to older patients with T2DM.
GLP1-RA and SGLT2-I were reported to have weight loss effect and metformin was accepted to have weight neutral or slight weight loss effects, while SU is associated with weight gain. [79][80][81] Thereby, the weight effect of ADDs might be responsible for our findings that overweight/ obese people were more likely to get a medication with weight neutral/loss effect (GLP1-RA, SGLT2-I, and metformin) but less likely to be prescribed a medication with weight gain effect (SU). Overall, these findings might indirectly reflect a consistency of ADD selection in clinical practice considering patient weight against drug features.
Baseline HbA1c level had the strongest association with insulin prescription where patients with higher baseline HbA1C were more likely to receive insulin, whereas higher baseline HbA1c had negative weak significant associations with metformin, TZD and DPP4-I prescriptions.
All aforementioned associations were consistent with the known effectiveness of each antidiabetic class relevant to HbA1c reduction, which partially indicate clinicians' consideration of disease severity (indicated by HbA1c) when selecting the most appropriate ADDs for each patient. Insulin is known to have the greatest effect on the reduction of HbA1c and this might explain the greater likelihood of prescribing insulin for patients with higher baseline HbA1c. 82,83 Lastly, the management of diabetes in patients with kidney-related problems is challenging as the impairment in kidney function might affect glucose metabolism and alter drug clearance. 84 This further complicates the selection of an appropriate ADD, considering the need for more frequent adjustment of doses and monitoring for the risk of hypo/hyperglycaemia. 84 Insulin has been considered as the best choice for patients with T2DM and kidney problems, yet still requires close monitoring and dose adjustment. 84 Also, DPP4-I is among the most acceptable option for patients with kidney problems considering dose adjustment based on the agent and degree of impairment. 84 In contrast, metformin is not recommended for patients with kidney disease, and it is contraindicated when eGFR is <30 mL/min/1.73 m 2 , because of the higher risk of lactic acidosis. 84 Collectively, that could explain the observed associations of higher prescription of insulin and DPP4-I and lower prescription of metformin for patients with kidney-related problems.
Despite that the use of SGLT2-I and GLP1-RA has been recently encouraged by several guidelines especially for patients with established or high risk of cardiovascular or renal diseases because of their cardioprotective and renal protective effects, 9,[85][86][87][88] the pooled estimates of studies investigating the prescription of SGLT2-I and GLP1-RA for patients with kidney-related problems were not in line with the previous recommendations. Nevertheless, those recommendations are relatively recent while the majority of included studies were conducted early after the introduction of GLP1-RA and SGLT2-I. Therefore, more studies are still required to further investigate prescribing of newer classes for patients with kidney problems in clinical practice considering different stages and types of kidney disease.

| Strength and limitations
To the best of our knowledge, this is the first SR/MA integrating the results of observational studies assessing the association of several factor with ADP to draw an overall estimate. This review provides a wide range of data by investigating each factor on seven different antidiabetic classes. Additionally, applying a three-level MA approach to account for the presence of dependency among effect sizes gave an opportunity to answer the research question without losing valuable data and to directly compare different antidiabetic groups.
Nevertheless, all previous results should be interpreted cautiously because of several limitations of the study. First, limited number of studies examined certain classes of ADDs, especially the newer ones; thus, more studies are required to draw a more robust conclusion. Second, the possible presence of publication bias especially for age, BMI and kidney-related problems may have affected the reliability of findings; however, there is no agreed-upon method available to adjust for publication bias in the threelevel MA model. Third, bias could have been introduced by including all studies in the pooled estimate regardless the type of data presentation (categorical vs. continuous) and the type of categorisation scheme; yet, subgroup analyses were done and showed no significant impact, and the pooled estimate of each sub-group was reported separately. Lastly, other important factors, including socioeconomic and prescriber-related factors, were much less frequently studied and further investigations are needed.

| CONCLUSION
In conclusion, all identified factors are crucial to be considered when making a decision regarding the most appropriate ADDs for patients with T2DM. The magnitude, direction and significance of influence of the identified factors on ADP varied according to the type of antidiabetic group. Age, baseline BMI and baseline HbA1c had the greatest impact on the selection of ADDs in which they had statistically significant associations with prescribing of four out of the seven investigated antidiabetic classes. On the other hand, sex had the least impact on ADDs selection which had only a significant influence on GLP1-RA and TZD prescriptions. The findings of this SR&MA could be helpful in determining the need of improving prescribing practice of ADDs by reflecting the consistency of prescribing decision of ADDs with guidelines recommendations and specific drugs features.

AUTHOR CONTRIBUTIONS
The authors confirm their contribution to the paper as follows: Study concept and design: FM, AK, AM; Study screening, data collection and validation: FM, HY, NA; Data analysis: FM; interpretation of results: FM, AK, AM, TM, CS, GR; Draft manuscript preparation: FM, AK, TM, AM, CS, GR, HY, NA. All authors reviewed and revised the manuscript, and approved the final version.

ACKNO WLE DGE MENTS
This work is part of a PhD project, and the authors acknowledge Jordan University of Science and Technology for sponsoring this PhD project at the University of Strathclyde.