Dear Editor,

We have read with great interest the recent paper of Alkabbani et al. [1] on the risk of myopathy related to sodium glucose co-transporter 2 (SGLT-2) inhibitors and statin use. Alkabbani et al. used 2202 myopathy cases and 686,186 non-cases from the US FDA Adverse Events Reporting System (FAERS). For their statistical analyses, they referred to the methods published by Thakrar et al. [2], who proposed two different models (an additive model and a multiplicative model) to assess the association between a disease and two drugs. Thakrar et al. especially emphasized to assess interactions between both drugs, where these interactions can be given on either an additive or multiplicative scale. Both of these models are generalized linear models with a binomial error and an identity/log link in the additive/multiplicative case. Binary indicators for the two drugs and their interaction are included as covariates. Thakrar et al. gave advice as to how these models can be fit with the GENMOD procedure in SAS (SAS Institute Inc., Cary, NC, USA). In the Appendix, we report the SAS code for the additive model based on the data given in Table 2 of the study by Alkabbani et al. [1].

Conveniently, the parameters from the additive model can be interpreted as risk differences. Thus, taking statins is associated with a larger absolute risk for myopathy of 0.818 percentage points compared with the reference group that took none of the two drugs. In contrast, taking SGLT-2 inhibitors is associated with a reduced absolute risk of myopathy of 0.045 percentage points. If there were no interaction between statins and SGLT-2 inhibitors, the two risk differences would sum up, yielding (with some minimal rounding error) a risk difference of 0.818 + (− 0.045) = 0.772 percentage points when taking both drugs compared with the reference group. However, when directly calculating this risk difference for taking both drugs, we found 60 / (60 + 13,293) – 735 / (735 + 481,097) = 0.449 % − 0.152% = 0.297 percentage points. This value is substantially smaller than the value that is expected under the assumption of no interaction (0.772 percentage points). The interaction parameter from Table 1 is, by definition, equal to the difference of the two figures: 0.297 – 0.772 = − 0.475 percentage points = − 0.00475 with a 95% confidence interval [CI] of − 0.00603 to − 0.00347.

Table 1 Results from fitting the additive model of Thakrar et al. [2] to the data set for SGLT-2 inhibitors and statins on the risk of myopathy, as given in Table 2 of the study by Alkabbani et al. [1]
Full size table

Therefore, there is a clear clinically relevant and statistically significant additive interaction between statins and SGLT-2 inhibitors. If we would assume the observed results to be causal, this interaction would be highly protective, and additionally taking SGLT-2 inhibitors to statins would reduce the detrimental effect of statins on myopathy by 0.00475 / 0.00818 = 59% [95% CI 44%–73%]. The respective public health impact would be considerable. Focusing on the German situation, we refer to Ihle et al. [3], who obtained population-based nationwide data. They reported a 1-year incidence of myopathy in German incident statin users of 1.93% (= 10,250 / 531,672). Thus, one could prevent 0.00475 / 0.00818 * 10,250 = 5960 [95% CI 4469–7450] myopathy cases each year only in Germany, if new statin users started simultaneously with SGLT-2 inhibitor therapy.

In line with the findings from the additive model, the interaction from the multiplicative model is also negative and statistically significant (data not shown). Thus, SGLT-2 inhibitors also lower the statin effect on myopathy on the multiplicative scale. This is in contrast to the assertion of Alkabbani et al., that “both estimates of the interaction were not significant from the additive and multiplicative models”.

It should be noted that our computations were based on the crude data from Table 2, whereas Alkabbani et al. had additional access to potential confounders and have used them in their analyses. While there are some differences between adjusted and unadjusted analyses, we are confident that the negative interactions would also be found in adjusted analyses.

Interestingly, we recently saw a similar potentially protective drug–drug interaction for the use of dipeptidyl peptidase 4 (DPP-4) inhibitors and statins on myopathy risk [4], and we refer to this previous letter for an explanation of the multiplicative model of Thakrar et al. [2].

To conclude, the main aim of this letter is not to criticize Alkabbani et al. for the problems in their statistical analysis. Indeed, we have also seen the same problems with the Thakrar models in other publications [4]. Unfortunately, the original paper by Thakrar et al. is unnecessarily technical, whereas the actual software code for model fitting and the interpretation of parameters is rather simple. As demonstrated above, the central results of interest such as risks, risk differences, and sums of risk differences can even be calculated from the raw data using a pocket calculator (albeit without their CIs). We thus encourage future researchers who are assessing drug–drug interactions and are uneasy with the technical details of the Thakrar models to contact a biostatistician. The safety data from the large pharmacovigilance databases are just too precious to allow suboptimal statistical analyses to limit their value for pharmacoepidemiology and clinical research.