Comparison of Trials Using Ivermectin for COVID-19 Between Regions With High and Low Prevalence of Strongyloidiasis

Key Points Question Does prevalence of strongyloidiasis interact with the relative risk (RR) of mortality in ivermectin trials for the treatment of COVID-19? Findings In this meta-analysis of 12 randomized clinical trials involving 3901 patients, favorable mortality results were limited to trials in high-prevalence regions, with no evidence that ivermectin had a mortality benefit in low-prevalence regions. Meta-regression found an association between the regional prevalence of strongyloidiasis and risk of mortality, with a decrease in RR of 39% for each 5% increase in strongyloidiasis prevalence. Meaning Evidence supports that strongyloidiasis prevalence interacts with the RR of mortality in ivermectin trial results; no evidence was found to suggest ivermectin has any role in preventing mortality in patients with COVID-19 in regions where strongyloidiasis is not endemic.


eMethods. Subgroup and Sensitivity Analyses and Database Search Details
Given the nature of the trials, their protocols, locations, and reported results, it was decided to have the model of random effects for the subgroup analysis.
The assumptions of the meta regression regressing natural log relative risk of all-cause mortality on strongyloides prevalence are assessed here. From the knowledge of the studies, it was assumed a priori that the studies had the same τ2 value, were independent from each other, and that there was little to no measurement error within the individual studies. The approximate linear relationship between the log relative risk and the strongyloides prevalence was assessed via a scatter plot. (eFigure 1) The distribution of the σ2 (sampling error) was assessed via the fitted values vs standardized residuals plot. (eFigure 2). The distribution of the error terms were reasonably met, as the variance is relatively equal across the fitted values.
As a sensitivity analysis, a permutation analysis (useful for assessing whether there is overfit by reassessing the model on resampled data) was performed on our main meta regression model. The permutation analysis on the model returned an adjusted p-value of 0.0136 for the linear coefficient of -0.0983 of strongyloides prevalence on the natural log relative risk, suggesting that the model is not overfit and is robust to the uncertainty for our estimate of τ.
As another sensitivity analysis, a Knapp-Hartung adjustment was considered for both the main subgroup analysis and main meta regression models. However, given the estimation of τ2 being 0 and the issues associated with the modification when τ2 is 0[1], we decided to default to the results without the modification. However, we will provide the changes in results with the modification below: For the subgroup analysis (Knapp-Hartung), Ivermectin trials that took place in areas of low regional strongyloides prevalence were not associated with a significant decreased risk of mortality, Relative Risk (RR) = 0.8397 [0.6252; 1.1280] (P = 0.2). By contrast, ivermectin trials that took place in areas of high regional strongyloides prevalence were associated with a significant decreased risk of mortality, RR = 0.2507 [0.1225; 0.5129] (P < 0.01). Test for subgroup differences revealed a significant difference between the results of low and high strongyloides prevalence groups. Chi 2 = 22.09, (P < 0.0001).
In addition, the meta regression (Knapp-Hartung) analysis revealed a linear coefficient of -0.0983, (P=0.0186) for strongyloides prevalence (percent)'s effect on the natural log relative risk for all-cause mortality. The estimates and confidence intervals for τ2 and I2 were the same as the main model. The additional permutation analysis for the model with the Knapp-Hartung adjustment returned an adjusted p-value of 0.0145 for the linear coefficient of -0.0983 of strongyloides prevalence on the natural log relative risk, suggesting that the model's estimates and resulting inferences would not qualitatively change.
We decided to perform a sensitivity analyses that excluded trials with improper randomization protocols, which excluded three trials (Hashim, Gonzalez, and Okumus). The subgroup analysis and meta regression results are shown below: For the subgroup analysis, ivermectin trials that took place in areas of low regional strongyloides prevalence were not associated with a significant decreased risk of mortality, RR = 0.96 [0.6466; 1.4254] (P = 0.84). By contrast, ivermectin trials that took place in areas of high © 2022 Bitterman A et al. JAMA Network Open.
regional Strongyloides prevalence were associated with a significant decreased risk of mortality, RR = 0.25 [0.09-0.70] (P < 0.01). Test for subgroup differences revealed a significant difference between the results of low and high Strongyloides prevalence groups. Chi 2 = 5.7, (P = 0.017).
(eFigure 3) The estimate and confidence interval for τ2 (the variance of the study effect sizes) was 0 [0.0000; 2.3662] and the estimate and confidence interval for I2 (percentage of variability that is explained by between-study heterogeneity) was 0.0% [0.0%; 64.8%]. While the τ2 and I2 are smaller than assumed to be a priori, the confidence intervals indicate that τ2 and I2 are compatible with values up to 2.3662 and 64.8%, respectively.
The meta regression in the sensitivity analysis revealed a linear coefficient of -0.1044, (P=0.0350) for the Strongyloides prevalence effect on the natural log relative risk for all-cause mortality. From this, the estimated relative risk percent decrease for each 5% increase in strongyloides prevalence was calculated to be 40.68% [3.6% -63.49%]. (eFigure 4) The estimate and confidence interval for τ2 (the variance of the study effect sizes) was 0 [0.0000 -0.5341], and the estimate and confidence interval for I2 (percentage of variability that is explained by between-study heterogeneity) was 0.0% [0.0% -57.0%]. While the τ2 and I2 are smaller than assumed to be a priori, the confidence intervals indicate that τ2 and I2 are compatible with values up to 0.5341 and 57.0%, respectively. Since there was no estimated heterogeneity, the test for residual heterogeneity predictably returned a test statistic of QE(df = 10) = 2.94 (P=0.8905). The additional permutation analysis returned an adjusted p-value of 0.0132 for the linear coefficient of -0.0983 of strongyloides prevalence on the natural log relative risk, suggesting that the model estimates and resulting inferences would not qualitatively change.
The funnel plot analysis with the Harbord test did not show significant funnel plot asymmetry (p = 0.1568) (eFigure 5).