Effect of Long-Acting Insulin Analogs on the Risk of Cancer: A Systematic Review of Observational Studies
Abstract
OBJECTIVE Observational studies examining the association between long-acting insulin analogs and cancer incidence have produced inconsistent results. We conducted a systematic review of these studies, focusing on their methodological strengths and weaknesses.
RESEARCH DESIGN AND METHODS We systematically searched MEDLINE and EMBASE from 2000 to 2014 to identify all observational studies evaluating the relationship between the long-acting insulin analogs and the risk of any and site-specific cancers (breast, colorectal, prostate). We included cohort and case-control studies published in English on insulin glargine and detemir and any cancer incidence among patients with type 1 or 2 diabetes. The methodological assessment involved the inclusion of prevalent users, inclusion of lag periods, time-related biases, and duration of follow-up between insulin initiation and cancer incidence.
RESULTS A total of 16 cohort and 3 case-control studies met our inclusion criteria. All studies evaluated insulin glargine, and four studies also examined insulin detemir. Follow-up ranged from 0.9 to 7.0 years. Thirteen of 15 studies reported no association between insulin glargine and detemir and any cancer. Four of 13 studies reported an increased risk of breast cancer with insulin glargine. In the quality assessment, 7 studies included prevalent users, 11 did not consider a lag period, 6 had time-related biases, and 16 had short (<5 years) follow-up.
CONCLUSIONS The observational studies examining the risk of cancer associated with long-acting insulin analogs have important methodological shortcomings that limit the conclusions that can be drawn. Thus, uncertainty remains, particularly for breast cancer risk.
Introduction
NPH insulin has been the mainstay treatment for type 1 diabetes and advanced type 2 diabetes since the 1950s. However, this insulin is associated with an increased risk of nocturnal hypoglycemia, and its relatively short half-life requires frequent administration (1,2). Consequently, structurally modified insulins, known as long-acting insulin analogs (glargine and detemir), were developed in the 1990s to circumvent these limitations. However, there are concerns that long-acting insulin analogs may be associated with an increased risk of cancer. Indeed, some laboratory studies showed long-acting insulin analogs were associated with cancer cell proliferation and protected against apoptosis via their higher binding affinity to IGF-I receptors (3,4).
In 2009, four observational studies associated the use of insulin glargine with an increased risk of cancer (5–8). These studies raised important concerns but were also criticized for important methodological shortcomings (9–13). Since then, several observational studies assessing the association between long-acting insulin analogs and cancer have been published but yielded inconsistent findings (14–28). Such discrepancies may be due to methodological limitations, including inadequate durations of follow-up between insulin initiation and cancer incidence, protopathic bias, detection bias, the inclusion of prevalent users, and time-related biases such as immortal time bias, time-window bias, and time-lag bias (29).
Randomized controlled trials (RCTs) have reported the effects of long-acting insulin analogs on the risk of any cancers (30–32), but most of these RCTs were designed to study efficacy (e.g., fasting plasma glucose level) and not designed to assess cancer. The most notable RCT, the Outcomes Reduction with Insulin Glargine Intervention (ORIGIN) trial, did not observe an effect of insulin glargine on the composite outcome of any cancer (33). Although the ORIGIN trial had several strengths, including the power to detect a clinically important effect of insulin glargine on any cancer and adjudication of cancer outcomes, it was not powered to detect site-specific cancers, and follow-up was relatively short (<7 years) given the long latency of cancer.
Several meta-analyses of observational studies have investigated the association between insulin glargine and cancer risk (34–37). These meta-analyses assessed the quality of included studies, but the methodological issues particular to pharmacoepidemiologic research were not fully considered. In addition, given the presence of important heterogeneity in this literature, the appropriateness of pooling the results of these studies remains unclear. We therefore conducted a systematic review of observational studies examining the association between long-acting insulin analogs and cancer incidence, with a particular focus on methodological strengths and weaknesses of these studies.
Research Design and Methods
This systematic review was conducted following a prespecified protocol and reported following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines (38).
Search Strategy
We systematically searched MEDLINE and EMBASE via Ovid from 1 January 2000 to 8 October 2014 for observational studies examining the association between long-acting insulin analogs and cancer incidence. The detailed search strategy is reported in Supplementary Table 1. Briefly, the search included MeSH terms, Emtree terms, and keywords for diabetes, long-acting insulin analogs, neoplasm, and observational studies. The publication type search terms used in this search strategy were adopted from the Scottish Intercollegiate Guidelines Network (SIGN) group (39). The search was limited to articles published from 2000 onwards because long-acting insulin analogs were not available globally until after 2000. Our search was also limited to studies published in English. We hand-searched relevant systematic reviews and meta-analyses to identify additional articles that were not identified in our electronic literature search.
Inclusion and Exclusion Criteria
Cohort, case-control, and case-cohort studies evaluating the association of long-acting insulin analogs (glargine and/or detemir) and cancer incidence among patients with type 1 or 2 diabetes were eligible for inclusion. Inclusion was restricted to studies reporting any incident cancer or site-specific cancers as primary or secondary outcomes. Studies that did not exclude prevalent cancer cases were eligible for inclusion. We excluded studies that did not meet these inclusion criteria.
The literature search was conducted independently by two reviewers (J.W.W. and M.K.D.), who assessed the titles and/or abstracts of identified publications. The full text of any publication deemed potentially relevant by either reviewer at this stage was retrieved for detailed review. Discrepancies in determining whether the study met our inclusion criteria during the full-text review were resolved by consensus or, when necessary, a third reviewer (K.B.F.).
Data Extraction and Quality Assessment
We developed a data extraction form, which was pilot tested on six included studies. Two independent reviewers (J.W.W. and M.K.D.) extracted data, with disagreements resolved by consensus or a third reviewer (K.B.F., L.A., and S.S.). Disagreements could have occurred when extracting individual data points (e.g., study characteristics and measures of association) or when evaluating the quality of the studies.
Extracted information included the following:
1) study characteristics (source population, country, sample size, study design, type of database used to ascertain information about exposure and outcome);
2) patient characteristics (age);
3) exposure and comparator definitions (ever vs. never use, duration of use, dose, use of time-independent or -dependent approach);
4) incidence of any and/or site-specific cancers;
5) odds ratios, risk ratios, rate ratios, or hazard ratios (HRs) with corresponding 95% CIs;
6) methods of adjustment for confounders (matching, regression-based adjustments, propensity scores, disease risk scores) and list of potential confounders; and
7) quality of the studies.
We extracted any site-specific cancers but did not report on relative risks (RRs) for sites that were not commonly reported among the included studies.
No available quality assessment tool adequately captures the methodological issues and biases that are particular to pharmacoepidemiology. Therefore, we assessed the quality of studies for key components, including time-related biases (immortal time, time-lag, and time-window), inclusion of prevalent users, inclusion of lag periods, and length of follow-up between insulin initiation and cancer incidence.
Immortal time bias is defined by a period of unexposed person-time that is misclassified as exposed person-time or excluded, resulting in the exposure of interest appearing more favorable (40,41). Time-lag bias occurs when treatments used later in the disease management process are compared with those used earlier for less advanced stages of the disease. Such comparisons can result in confounding by disease duration or severity of disease if duration and severity of disease are not adequately considered in the design or analysis of the study (29). This is particularly true for chronic disease with dynamic treatment processes such as type 2 diabetes. Currently, American and European clinical guidelines suggest using basal insulin (e.g., NPH, glargine, and detemir) as a last line of treatment if HbA1c targets are not achieved with other antidiabetic medications (42). Therefore, studies that compare long-acting insulin analogs to nonbasal insulin may introduce confounding by disease duration. Time-window bias occurs when the opportunity for exposure differs between case subjects and control subjects (29,43).
The importance of considering a lag period is necessary for latency considerations (i.e., a minimum time between treatment initiation and the development of cancer) and to minimize protopathic and detection bias. Protopathic bias, or reverse causation, is present when a medication (exposure) is prescribed for early symptoms related to the outcome of interest, which can lead to an overestimation of the association. Lagging the exposure by a predefined time window in cohort studies or excluding exposures in a predefined time window before the event in case-control studies is a means of minimizing this bias (44). Detection bias is present when the exposure leads to higher detection of the outcome of interest due to the increased frequency of clinic visits (e.g., newly diagnosed patients with type 2 diabetes or new users of another antidiabetic medication), which also results in an overestimation of risk (45). Thus, including a lag period, such as starting follow-up after 1 year of the initiation of a drug, simultaneously considers a latency period while also minimizing protopathic and detection bias.
We also assessed the studies for traditional epidemiological biases such as selection bias, information bias, and confounding. For confounding, we considered three potential sources:
1) imbalances between measured baseline covariates that were not addressed analytically;
2) residual confounding due to unmeasured confounders; and
3) lack of adjustment for time-dependent confounders.
This assessment focused on the discussion of key components of design and analysis rather than on the creation of an aggregate score, as has been suggested elsewhere (46). We used the primary analysis of each included study for the qualitative assessment, but if the issue or bias was addressed in an appropriate sensitivity analysis, we considered this in the qualitative assessment.
Data Analysis
Given the methodological focus of this review and heterogeneity among published studies, we conducted a systematic review without a meta-analysis. Nonetheless, forest plots were constructed with Stata 13 software (StataCorp LP, College Station, TX) to graphically present the available data.
Results
Study Selection
Our search of MEDLINE and EMBASE yielded 4,417 potentially relevant articles (Supplementary Fig. 1). Following our inclusion criteria, 16 cohort and 3 case-control studies were included in this systematic review (5–8,14–28). All studies evaluated insulin glargine, with four studies also investigating insulin detemir (15,17,25,28).
Study Characteristics and Effect Estimates
The study populations ranged from 1,340 to 275,164 patients (Table 1). The mean or median durations of follow-up and age ranged from 0.9 to 7.0 years and from 52.3 to 77.4 years, respectively. Thirteen studies examined ever use of long-acting insulin analogs, which was defined as at least one prescription, compared with nonuse, other, human, or NPH insulin (5–8,14,16,18,19,21,23,25–27). One study examined duration of time since starting long-acting insulin analogs, and one examined mean daily dose (22,28). Four studies used time-dependent exposure definitions (15,17,20,24). All included studies evaluated cancer incidence as a primary outcome.
Characteristics of observational studies examining the association between long-acting insulin analogs and cancer incidence
Of the 16 studies that evaluated the relationship between long-acting insulin analogs and any, colorectal, and/or prostate cancer, 13 reported no associations (Fig. 1 and Supplementary Fig. 2) (5,8,14–17,19–21,23,25,26,28). Four of 13 studies reported an association of insulin glargine and breast cancer (8,19,21,24).
Forest plots of RRs (solid squares) and 95% CIs (solid horizontal lines) from studies on insulin glargine and any (A), breast (B), colorectal (C), and prostate (D) cancers. For exposure and comparator definitions in each study, please refer to Table 1.
Quality Assessment
The different key components of the quality assessment are summarized in Table 2 and discussed in detail below.
Pharmacoepidemiology biases in studies examining the association between long-acting insulin analogs and cancer incidence
Immortal Time Bias
Of the 19 studies in this review, immortal time bias may have been introduced in one study based on the time-independent exposure and cohort entry definitions that were used in this cohort study (14). For the exclusive user definition, patients needed to have insulin glargine or human insulin only between the first and last pr