Our umbrella review has evaluated data from 74 meta-analysis publications on the association of common medications with cancer risk. While 11 articles have reported on a total of 16 increased risk estimates, several of these claimed associations may not be supported by the overall analysis of the literature. The large majority of surveyed meta-analyses have found no significant associations, and the number of meta-analyses that have shown significantly increased risk is similar to those that have shown significantly decreased risk for the same medications. Few claims of increased risk are based on the data from randomized trials and/or large sample sizes. We also documented a very strong inverse relationship between the amount of evidence and effect sizes: when more data were available, estimates of increased risk were much smaller. Finally, for several meta-analyses that claimed increased cancer risk, we found others that were larger and/or included data from better controlled studies and in all these cases, the larger and/or better controlled meta-analyses indicated no increased burden of malignancy.
Our analysis suggests that most associations of commonly used medications with cancer risk, if present, are likely to have small or modest effects. Large estimates of risks in some meta-analyses were documented with limited evidence from small studies. These may correspond to either small or no effects, when large studies are carried out. This may be due to the winner's curse (a regression-to-the-mean phenomenon), where results selected on the basis of statistical significance are expected to have inflated effect sizes, even if some association is genuinely present. Documentation and validation of relative risks in the <1.20 range will require very large datasets, careful designs and protection from selective reporting and other biases to minimize noise. It is possible that several such small relative risks may still remain undetected based on the current evidence which remains underpowered, even when data are combined in meta-analyses. It is also possible that small relative risks can reflect heterogeneity of cancer risks across population subgroups defined on the basis of demographic, clinical and other biological factors such as genetic susceptibility. Conversely, several of the seemingly detected increased risks may reflect the impact of common limitations of pharmacoepidemiology studies, including the lack of control for drug dose and duration, recall bias from self-reported data, short follow-up times, confounding by indication and duration of disease, detection bias, as well as selective reporting and other biases in a setting where there is often substantial unaccounted multiplicity of comparisons that may cause false positives. False positives may manifest as either increased or decreased risk and we observed an equal number of claims of increased and decreased risk in meta-analyses. Feinstein had seen the same phenomenon in an evaluation of observational studies 25 years ago and had suggested that such observational risks may simply reflect false positives. We also suspect that the number of null results in meta-analyses of cancer risk may be substantially underestimated due to reporting bias.
Of the five types of medications that have been associated with significantly increased risk of all malignancies combined, ARB and TNF inhibitors seem unlikely to confer substantially increased overall cancer risk. Very large meta-analyses have found no significant associations and the 95% confidence intervals are very tight, excluding relative risks >1.05. The combination of ACE inhibitors and ARB also has weak evidence with a very small relative risk (1.14) and very modest statistical significance. At least one other meta-analysis shows no risks or decreased risks with these antihypertensive agents. Insulin therapy seemingly has the strongest evidence for an association with increased overall cancer risk, and this is also reflected in significant associations with specific cancer types, in particular colorectal and pancreatic cancers which have each been documented in two separate meta-analyses. However, a recent large randomized, trial on over 12 000 randomized participants (ORIGIN) found no increased risk of cancer with insulin versus standard care over a median follow-up of 6.2 years. Immunosuppressive therapy for AAV also seems to have evidence for an association with increased risk, although the number of cancer cases is very small and thus there should be some reservation on the exact magnitude of the effect.
The sporadic associations of medications with specific cancer types at one site should be seen with even greater caution. Given that there are several dozens of cancer sites and types, and perhaps additional subgroup considerations, such associations have an additional layer of multiplicity and a traditional P value of <0.05 is likely to be a very weak discriminating tool for identifying genuine associations. We will focus here on the two site-specific associations that had P < 0.001, and which thus account for a Bonferroni correction of 50-fold for the multiplicity of comparisons.
First, the claimed association of TNF inhibitors with increased risk of non-melanoma skin cancer was seen in a meta-analysis with four observational studies and 1258 cases. Despite the strong statistical significance, documentation of non-melanoma skin cancer is likely to be more susceptible to poor data collection for this type of typically non-aggressive cancer. Moreover, the increased risk estimate is seen in the same article where the estimate and 95% CI for overall cancer exclude relative risks >1.05. One cannot also exclude the possibility that some sort of diagnosis bias may exist also in these data, i.e., patients treated with a biologic agent that is considered to potentially increase cancer risk may be more likely to have more thorough screening for suspicious cancerous lesions, referrals for them and diagnosis of non-melanoma skin cancer. Of note, the earliest meta-analysis on TNF inhibitors and cancer had suggested a very large risk with a 3.3-fold increase in overall malignancy incidence, but it was based on the data from only 32 cancer cases. This highlights the major danger of drawing conclusions from early, limited data that may be subject to substantial subsequent regression to the mean.
Second, the evidence for an increased risk of renal cell cancer with antihypertensives seems to have strong statistical support. The statistical significance is more prominent for non-diuretics, but practically the risk estimate is the same also for diuretics, so it is unclear whether there is any discernible drug specificity for this risk. The evidence for increased risk comes from observational studies, while data on diverse antihypertensives from randomized trials and observational data show mostly no significant risk, and occasionally even decreased risks of cancer. It is unclear whether diagnosis bias may also exist for observational data on renal cell cancer, e.g., some patients with hypertension may be more likely to be subjected to evaluation of their kidneys, and this may result in more renal cell cancers being diagnosed.
Our umbrella review has limitations. Many of them reflect from the limitations of the primary data included in the 75 meta-analyses that we surveyed. Cancer risk may not be possible to detect in studies of short follow-up, and this caveat is very common in randomized trials in particular. Loss to follow-up may compound this problem. Underpowered studies and meta-analyses may fail to detect small or even modest cancer risks. Non-differential misclassification with under-diagnosis of cancer may dilute the estimates of association. Moreover, we did not evaluate the quality or the accuracy of the data and calculations of the 74 meta-analyses, a task that would have been too arduous or even impossible to carry out, given only what was published in the respective article reports. The meta-analysis authors did not have access to original data either and they did not adjust their meta-analysis for multiple testing. Publication and other selective reporting biases are likely to affect several of these meta-analyses, requiring caution in making inferences.
Furthermore, a larger meta-analysis may not necessarily be better than a smaller one with less data. Observational data are likely to be even more biased than randomized data and have a poorer replication record. It is well-documented that observational data may often find stronger effects than randomized trials, although this is not so clear for harms. All evidence should be examined comparatively, and we tried to obtain as wide a view as possible given the accumulated published information. Finally, it is possible that for some medication classes, only meta-analyses of observational data exist with no respective randomized evidence and these classes would not have been captured by our searches. This would have required scrutinizing tens of thousands of meta-analyses of observational data with low yield. Nevertheless, our approach has probably captured all the medication classes that have attracted major attention in the literature. It is unlikely that a medication has attracted substantial attention for its association with increased cancer risk in observational data and no one has been tempted to see what the respective data would suggest in its randomized trials.
Allowing for these caveats, our evaluation maps systematically the current landscape of the pharmacoepidemiology of claims of increased cancer risk. Many of the proposed associations with increased risk of malignancy may not be real or may be very modest in magnitude. The available evidence often cannot exclude small risks for many medications, or even modestly large risk in circumscribed population subsets (for example, based on genetic susceptibility). One would have to decide on a case-by-case basis whether small risks are clinically or otherwise important to document robustly. Documentation of relative risks <1.20 will require large studies, long-term follow-up and complete data in collaborative teams carrying out individual-level meta-analyses, where selective reporting of analyses is minimized and ideally eliminated. Bias, multiple testing and multiple modeling are all potential problems and would need to be properly accounted for. Evaluation of population subsets would require strong biologic rationale and careful protection from issues of data dredging and subgroup analyses. Analysis plans should also be transparent and ideally registered upfront, and should be clearly stated whether associations with specific cancer types or other forms of secondary analyses are pre-conceived or exploratory ones. The burden of multiplicity of end points and analyses should be carefully considered in making or not making strong conclusions on cancer risks from medications.