No association between metformin initiation and incident dementia in older adults newly diagnosed with diabetes

Metformin has been suggested to reduce dementia risk; however, most epidemiologic studies have been limited by immortal time bias or confounding due to disease severity.

Many cohort studies defined metformin exposure based on metformin use at any time during entire follow-up after diabetes diagnosis [14][15][16][17][18][19], and these studies overall identified a lower dementia risk in metformin users; however, the benefit might have been overestimated by immortal time bias [25].Metformin users under that exposure definition had to be dementia-free until metformin initiation, and as a consequence, dementia-free time before metformin use might have been excluded or misclassified as exposed time.The time to event in metformin users might have been lengthened erroneously by misclassified periods.Two studies compared prevalent metformin users to nonusers in a general population of type 2 diabetes [20,21]; nonetheless, the comparator consisted of a mix of individuals untreated or treated with other glucose-lowering medications from different stages of diabetes, making it difficult to infer the treatment effect and generalize the findings to clinical decision making [26].Furthermore, although additional studies observed an association of metformin versus other glucose-lowering medications with reduced dementia risk [22,23], confounding due to diabetes severity (i.e., time-lag bias) remains a major concern because users of other glucoselowering medications tended to be more severe cases [25].
Several other cohort studies showed that metformin versus sulfonylureas was associated with a 10%-20% lower dementia risk [9][10][11][12][13], but sulfonylurea use might increase dementia risk according to previous real-world evidence [27,28].Hence, it remains unclear whether metformin is associated with reduced dementia risk.Moreover, sul-fonylureas are no longer a recommended first-line option, and current guidelines recommend lifestyle interventions without any pharmacological interventions if glycemic control can be achieved [4].
To infer whether metformin may reduce dementia risk, the present study investigates the association of early metformin monotherapy initiation versus delayed or no initiation of glucose-lowering medications with incident dementia in older adults newly diagnosed with diabetes.This study applied various bias mitigation strategies to provide robust evidence.The result from this study will suggest whether early metformin initiation should be considered to mitigate the risk of developing dementia in older adults with diabetes.The result will also imply the feasibility of repurposing metformin as a treatment to slow cognitive decline in people with diabetes or possibly in a broader context.

Data sources
This population-based retrospective cohort study used linkable administrative databases housed at ICES (https://www.ices.on.ca/DAS) in Ontario, Canada.ICES is an independent, nonprofit research institute funded by an annual grant from the Ontario Ministry of Health.As a prescribed entity under Ontario's privacy legislation, ICES is authorized to collect and use health care data for the purposes of health system analysis, evaluation, and decision support.Secure access to these data is governed by policies and procedures that are approved by the Information and Privacy Commissioner of Ontario.Variables were ascertained using the following databases: Ontario Drug Benefit (ODB) database, Ontario Laboratories Information System, Registered Persons Database, Ontario Health Insurance Plan (OHIP) Physician Billing database, National Ambulatory Care Reporting System, Discharge Abstract Database, and ICES Physician Database.More information on these databases is summarized in Table S1

Population
To ensure a 1-year lookback window for medications, the study only included individuals ≥66-year old.Ontario residents newly diagnosed with diabetes at the age of 66 or greater from January 1, 2008 to December 31, 2017 entered the cohort.Individuals newly diagnosed with diabetes were identified using a validated algorithm for diabetes that exhibited a high specificity (99.1%) and a high positive predictive value (91.4%) [29]: "three OHIP physician claims with diabetes diagnosis in one year."The date of cohort entry (i.e., incident diabetes diagnosis) was the date of the first claim.As incident diabetes diagnosis was defined based on physician claims, the cohort entry was aligned with the moment of a treatment decision.
An overview of the study design is illustrated in Fig. S1.People using any glucose-lowering medication in the year before diagnosis were excluded, to exclude prevalent users and define drug initiation.People using insulin as the first glucoselowering drug were also excluded, as they could have type 1 diabetes or severe type 2 diabetes.We further excluded people with baseline HbA1c (hemoglobin A1c) <6.5%, as this is the cut-off for prediabetes.Those with baseline HbA1c ≥8.0% or with baseline eGFR (estimated glomerular filtration rate) <45 mL/min/1.73m 2 were also excluded, to select individuals who could initiate either lifestyle interventions alone or a standard dose of metformin monotherapy.The CKD-EPI equation without the race coefficient was used to derive eGFR from creatinine levels [30].Baseline HbA1c or eGFR measure was the latest laboratory measure during a 1-year lookback window.

Exposure definition
The exposure was defined using the landmark method, in which a fixed period from cohort entry (i.e., landmark period) is selected for exposure ascertainment and follow-up begins after this period [31].A 180-day landmark period from incident diabetes diagnosis was selected because HbA1c measurement for monitoring glycemic target is usually done every 3-6 months [32].In the primary analysis, the exposure was classified into "metformin monotherapy initiation within 180 days after new diabetes diagnosis" or "no glucoselowering medications within 180 days."Thus, individuals with censoring or events within 180 days were excluded, and individuals initiating other glucose-lowering medications during the landmark period were also excluded.

Outcome definition
Censoring was defined as death, end of OHIP eligibility, or end of the study period.The date of incident dementia as the outcome was ascertained using a validated algorithm for Alzheimer's disease and related dementias [33]: "three OHIP physician claims (first claim) for dementia at least 30 days apart in a two-year period, one hospitalization with a dementia record, or dispensing of a cholinesterase inhibitor."This algorithm showed a sensitivity of 79.3%, a specificity of 99.1%, a positive predictive value of 80.4%, and a negative predictive value of 99.0% [33].
People with incident dementia at any time before cohort entry were excluded.To address disease latency and mitigate protopathic bias (reverse causality) that might result from differential prescription preference in people presenting early dementia symptoms [34], a 1-year lag time from the end of the landmark period was applied [35].Therefore, individuals with censoring or events within 1.5 years (545 days) after cohort entry were excluded.

Statistical analysis
Analyses were conducted in SAS 9.4 (SAS Institute Inc.).Propensity scores were calculated in a logistic regression with predictors listed in Table S2, inverse probability treatment weighting for average treatment effects among the treated was used to address confounding by indication, and a standardized mean difference (SMD) <0.1 was considered a negligible between-group difference [36].A 5-year lookback window from cohort entry was applied to baseline comorbidities, and a 120-day lookback window from cohort entry was applied to baseline exposure to medications for other indication.
Adjusted hazard ratios (aHRs) and 95% confidence intervals (CIs) were obtained using a weighted Cox proportional hazard model with a robust sandwich estimator.Weighted Kaplan-Meier curves were used to visualize cumulative incident dementia over time in each exposure group and assess potential non-proportional hazards.

Secondary analysis
The secondary analysis compared continuous metformin monotherapy use to no exposure to any glucose-medications.Data were censored after initiation of other glucose-lowering medications, metformin initiation in those with delayed metformin initiation, or metformin discontinuation in metformin initiators (whichever happened first).However, as changes in treatment status might be caused by early dementia symptoms, we extended the observation window up to 1 year from the moment of treatment status change.In this analysis, individuals discontinuing metformin monotherapy within 180 days were excluded, such that both exposure groups had no censoring occurring during the landmark period.
Baseline propensity score weights for treatment were recalculated for individuals included in the secondary analysis.Inverse probability of censoring weighting was used to further correct for potential informative censoring related to treatment status change.The stabilized weights for every 30-day interval were calculated from a marginal structural model, which incorporated probabilities of censoring estimated from two pooled logistic regression models.The numerator was the probability of censoring predicted from all baseline covariates listed in Table S2, and the denominator was the probability predicted from all baseline covariates plus follow-up time as a quadratic function.The analysis used a time-dependent Cox regression model with a robust sandwich variance estimator to incorporate time-varying weights derived from products of censoring weights at each time point and baseline treatment weights.

Negative control outcome and quantitative bias analysis
Metformin use was suggested to be an indicator of healthy user effects [37], a source of confounding bias stemming from differential healthseeking behaviors.To assess healthy user effects, the rate of fecal occult blood testing (FOBT) prior to censoring or events was selected as a negative outcome [38], as metformin should have no effect on cancer screening rates, but individuals with poorer health-seeking behaviors are less likely to have cancer screening.FOBT was ascertained using the following OHIP fee codes: G004, L179, L181, Q043, Q133, and Q152.A propensity-score weighted Poisson regression was used to determine the between-group incidence rate ratio (IRR).
To quantify the influence of an unmeasured confounder, we performed a quantitative bias analysis using the array approach [39] in R 4.0.5 [40,41].We used the observed hazard ratio to approximate the observed relative risk as a parameter, and the prevalence of a confounder in the unexposed group was assumed to be 0.25.

Sensitivity analysis
A Fine-Gray model with all-cause mortality as a competing risk was used to explore the robustness to mortality-related informative censoring.In another sensitivity analysis, we implemented a 2-year lag time from the end of the landmark period as a more conservative approach to mitigate protopathic bias.A sensitivity analysis with a 1-year landmark period was performed to examine whether a similar estimate was seen with a longer landmark period.For sensitivity analyses with reduced population size, the propensity scores were recalculated.
Although referring to the earliest claim could more accurately ascertain dates of incident dementia among individuals who first satisfied the criterion for three physician claims, this approach was subject to at-risk time misclassification in true dementia cases with censoring before the third physician claim.Thus, as a sensitivity analysis attempting to adjust the risk sets, the last of three claims that first satisfied the criterion was considered the event date derived from physician claims.To explore whether inadequate sensitivity in event ascertainment was an important source of misclassification bias for the hazard ratio in this study, cholinesterase inhibitor dispensing was removed from the original event definition, as this step was expected to preserve a high specificity but decrease the sensitivity [33].

Population characteristics
Within 180 days after new diabetes diagnosis, there were 12,331 people initiating metformin monotherapy and 22,369 people not initiating a glucose-lowering medication (Fig. 1).Highlighted characteristics by exposure groups are summarized in Table 1, and full characteristics are summarized in Table S3.All covariates had an SMD <0.1 after propensity score weighting.

Dementia risk
Over a mean follow-up of 6.77 years from cohort entry (i.e., new diabetes diagnosis) and a total of 183,177 person years of at-risk time, there were in total 2746 observed incident dementia cases (Table 2).Initiation of metformin monotherapy within 180 days after new diabetes diagnosis was not associated with incident dementia, compared to delayed or no initiation of pharmacotherapy (aHR [95% CI] = 1.05 [0.96-1.15];Fig. 2).
The secondary analysis had 8880 metformin monotherapy users and 22,369 individuals without a glucose-lowering therapy over a total of 89,605 person years of at-risk time (Table 2).Continuous use of metformin monotherapy compared to no pharmacotherapy also showed no association with dementia risk (aHR [95% CI] = 1.08 [0.92-1.26];Fig. S2).The characteristics for the secondary analysis were summarized in Table S4, and all characteristics had an SMD <0.1 after weighting.

Negative control outcome and quantitative bias analysis
Metformin monotherapy initiation within 180 days after new diabetes diagnosis was not associated with FOBT rates compared to delayed or no initiation of pharmacotherapy (IRR [95% CI] = 0.97 [0.93-1.02]).
The quantitative bias analysis (Fig. S3) demonstrated that an unmeasured confounder that masked a clinically important effect size had to moderately increase dementia risk with a large prevalence difference between exposure groups, or be a strong risk factor with a moderate prevalence difference.
As the sensitivity analyses with 2-year lag time and a 1-year landmark period had reduced population size because of censoring or events prior to the start of observation window, propensity scores were recalculated for these analyses, all characteristics had an SMD <0.1 after weighting (Tables S5  and S6).An aHR of 1.05 (95% CI = 0.96-1.15)was seen when the last of the first three claims meeting the criterion was considered the event date derived from the OHIP database.The analysis defining events without cholinesterase inhibitor dispensing exhibited an aHR of 1.04 (95% CI = 0.94-1.

Post hoc analysis
There was crossing in the Kaplan-Meier curves, and there appeared to be a higher risk of dementia over the first 5 years of observation in metformin initiators (Fig. 2).Therefore, we ran a post hoc analysis restricting a maximum of 5-year observation window, to obtain the estimate over the first 5 years of observation.A slightly higher but nonsignificant risk of dementia was observed (aHR [95% CI] = 1.10 [0.98-1.22]).

Discussion
Early metformin monotherapy initiation versus delayed or no initiation of pharmacotherapy exhibited no association with incident dementia in older adults newly diagnosed with diabetes, consistent with a previous landmark analysis [24].The result was robust to all sensitivity analyses.The negative control outcome analysis reflected that confounding due to health-seeking behaviors was likely minimal, and the quantitative bias analysis implied that unmeasured confounders impacting the interpretation of the result were unlikely in most scenarios.The secondary analysis exploring continuous metformin monotherapy versus no pharmacotherapy yielded a comparable estimate.These analyses did not support metformin as a drug to prevent dementia.
Associations between metformin and a lower dementia risk were mostly observed in cohort studies that did not address immortal time bias [14][15][16][17][18][19] and time-lag bias [22,23], which may have overestimated the benefit of metformin [25].Two studies compared dementia risk between prevalent metformin users and metformin nonusers in a general population of type 2 diabetes [20,21]; nevertheless, substantial residual confounding was possible, because the nonuser group was not sufficiently comparable to the metformin group in terms of characteristics, and postexposure covariates rather than pre-exposure covariates were adjusted.Other studies that found associations with cognitive assessment scores were also susceptible to these concerns by virtue of that exposure definition [20,21,[42][43][44][45][46].In contrast, the present study employed the landmark method to minimize immortal time bias [25,31], and included people newly diagnosed with diabetes at an earlier disease stage to mitigate time-lag bias [25].To adequately and properly reduce confounding by indication, early metformin monotherapy initiation was compared to no or delayed pharmacotherapy for diabetes in people who had an opportunity to receive either treatment strategy [4,47], and the propensity scores were calculated based only on pre-exposure covariates.
An active-comparator new-user cohort design can also mitigate immortal time bias, confounding by indication, time-lag bias, and overadjustment bias due to postexposure covariates [48].Several active-comparator new-user cohorts demonstrated that metformin compared to sulfonylureas was associated with a 10%-20% lower dementia risk [9][10][11][12][13].However, sulfonylurea use may not be an ideal comparator for this question because sulfonylureas might elevate dementia risk [27,28], possibly as a result of sulfonylurea-induced hypoglycemia.Severe hypoglycemia is proposed as a strong risk factor for dementia [49,50].To overcome this limitation, the landmark method was chosen to investigate the timing of metformin initiation, rather than an active-comparator newuser design.In addition, as sulfonylureas are no longer a recommended first-line option, and lifestyle interventions without any pharmacotherapy is a therapeutic option for an earlier stage of diabetes [4], the present study better reflects current clinical practice than cohorts comparing metformin to sulfonylureas.
We implemented additional bias mitigation strategies to strengthen the evidence from a previous landmark analysis [24].Additional strategies here included (1) defining cohort entry based on physician visits rather than HbA1c measurement to ensure the landmark period began from a clinical treatment decision, (2) selecting a population with a narrower range of baseline HbA1c to facilitate balanced characteristics after propensity score weighting, (3) using lag time to address disease latency and possible protopathic bias, and (4) performing a negative control outcome analysis to evaluate the degree of healthy user effects.These approaches increase confidence that there was no association between early metformin initiation and incident dementia.
Current guidelines for older adults with type 2 diabetes were unable to recommend interventions to prevent dementia because of limited clear evidence [2,3].The findings from the present study and a previous landmark analysis [24] highlighted that earlier or immediate initiation of metformin monotherapy in newly diagnosed type 2 diabetes might not reduce dementia risk.The present study extends that finding by showing that continuous use of metformin monotherapy compared to no pharmacotherapy for diabetes was not associated with a lower dementia risk.Thus, metformin was unlikely to be effective for dementia prevention in people with diabetes.Previous studies that used metformin as a comparator indicated that first-line sulfonylurea use might elevate dementia risk [9][10][11][12].In toto, the evidence suggests that neither first-line treatment protected against dementia.Despite metformin being a longstanding first-line therapy, sodium-glucose cotransporter-2 (SGLT2) inhibitors are of interest as possible future first-line therapies [51].New use of an SGLT2 inhibitor versus a dipeptidyl peptidase-4 inhibitor was associated with a lower dementia risk in a previous cohort [52].Future cohorts with adequate sample size, or randomized controlled trials, might investigate cognitive outcomes associated with newer glucose-lowering agents.
Limitations include an inability to account for unmeasured confounding due to body mass index and vitamin B12 deficiency.Although the algorithms to identify diabetes and dementia showed low false positive rates in prior validation studies, their performance for incident dates was not quantified [29,33].Moreover, non-differential misclassifications in outcome ascertainment might cause bias toward the null, but this study balanced health care utilization variables to potentially reduce detection bias resulting from differential frequencies of health care contact.Furthermore, the findings should be considered to generalize only to older adults who were dementia-free and newly diagnosed with diabetes; no inference can be made as to whether metformin might slow cognitive decline in people with dementia, or in people without diabetes.As dementia development is considered a chronic process, a relatively short landmark period and follow-up time might not well represent dementia pathogenesis; however, a priori selection of landmark period length according to diabetes management guidelines attempted to avoid bias, and the estimates were comparable in the secondary analysis and in the sensitivity analyses with a longer landmark period or lag time.Lastly, the OHIP claims did not differentiate dementia subtypes, and therefore time to different dementia subtypes could not be studied.
In summary, using different bias mitigation strategies, our analyses found that earlier initiation of metformin was not associated with prevention of dementia in older adults.Further evidence that appropriately deals with important sources of bias is needed to help balance the benefits and harms related to metformin initiation.

Fig. 2
Fig. 2 Weighted Kaplan-Meier curves for the primary analysis.The figure illustrates the cumulative proportions (failure probability) of incident dementia from the beginning of the observation window (i.e., 1-year lag-time after the end of the 180-day landmark period).
. All Ontario residents at any age are eligible for OHIP coverage, but only residents ≥65-year old are eligible for the ODB program.The study was approved by the research ethics board at Sunnybrook Health Sciences Centre (REB Number: 2021-4839).

Table 1 .
Highlighted demographics for individuals initiating metformin monotherapy versus not initiating of any glucoselowering drug within 180 days after new diabetes diagnosis.

Table 1 .
(Continued) Note: Proportions (counts) were reported for categorical variables, and mean (SD) was reported for continuous variables.Full demographics are reported in TableS3.Abbreviation: SMD, Standardized mean difference.

Table 2 .
Summary of results.