adequately addressed? • Are reports of the study free of suggestion of selective outcome reporting? When assessing the risk of bias, we focused on the main outcome of interest, CIN, an outcome that is objectively measured by laboratory testing. Data Synthesis We reviewed primary studies, as defined by our inclusion criteria, and we performed de novo meta-analyses. The de novo meta-analyses included all studies that met our inclusion criteria. Prior to conducting meta-analyses, clinicians discussed differences in the study design and reporting to identify characteristics that would limit the clinical meaningfulness of pooled results, such as the variability in outcome definitions, type of contrast media used, and route of contrast media administration. Differences in these items either prevented the statistical pooling with meta-analysis or were used to stratify the meta-analysis estimates. Pooled risks of large comparison groups (with 18 or more studies) were calculated using a random effects model using the method of DerSimonian and Laird.24 Because the DerSimonian and Laird method often underestimates confidence interval (CI) when there is a small number of studies (less than 18), for comparisons with less than 18 studies, the pooled risks were calculated using the Knapp-Hartung small sample estimator approach. This method allows for small sample adjustments to the variance estimates and forms CIs based on the t distribution with k - 1 degrees of freedom. 25 Statistical heterogeneity was assessed using the I-squared statistic. When the Isquared value was greater than or equal to 50%, or the p-value was 0.2 or less, the clinicians were asked to re-evaluate the studies for clinical heterogeneity and decide if the meta-analysis should be reported despite statistical heterogeneity. After reviewing the available evidence on all of the comparisons of interventions for preventing CIN, we felt that the heterogeneity across comparisons and the differences between reference groups were too great to support a network meta-analysis. In many of the studies, the intervention group or the comparison group received more than one intervention. Therefore, we stratified the analyses according to the comparisons that were 10 made, taking into consideration whether the intervention group or comparison group received more than one intervention. For example, we performed separate analyses for the following comparisons: N-acetylcysteine with IV saline versus IV saline with or without placebo; Nacetylcysteine with IV saline versus IV sodium bicarbonate; and N-acetylcysteine with IV sodium bicarbonate versus other interventions. The most common co-intervention was administration of fluids. We specified what fluid type was given whenever that was part of the intervention. For the analyses of N-acetylcysteine, all of the studies included IV fluids as a cointervention with N-acetylcysteine, so we could not do a network meta-analysis or metaregression to assess the effect of the co-intervention. We used Harbord’s modified test for small study effects to determine whether there was asymmetry in effect estimates when plotted against the standard error of the estimates, which can occur when publication bias exists. Minimally Important Difference To assess the clinical importance of differences in the incidence of CIN, a binary outcome, we followed guidance for selecting a minimally important difference based on the overall observed event rate in the studies.26 Taking into consideration the potential effect of CIN on a patient’s overall health and well-being, the clinical experts on our team decided that a relative risk reduction of 25% would be clinically important, which is consistent with the guidance suggesting a relative risk reduction of 20% to 30% in determining optimal information size. Strength of the Body of Evidence The team graded the strength of evidence on comparisons of interest for the key outcomes. We used the grading scheme recommended in the Methods Guide, and considered all domains: study limitations, directness, consistency, precision, reporting bias, and magnitude of effect.27 Study limitations were determined for each comparison group for CIN and other reported outcomes. Study limitations were determined using the following algorithm for a body of evidence: A body of evidence was assessed as having high study limitations if greater than 50 percent of the studies scored negative in one or more of the criteria. A body of evidence was assessed as having low study limitations if most (51% or greater) of the studies scored positive in all five domains. Bodies of evidence not meeting one of the above criteria were assessed as having medium study limitations. Following the guidance of the GRADE Working Group,26 we rated evidence as precise if the total number of patients exceeded an optimum information size, and the 95% confidence interval (CI) excluded a risk ratio of 1.0. If the total number of patients exceeded the optimum information size, and the 95% confidence interval did not exclude the possibility of no difference (i.e., risk ratio of 1.0), we only rated the evidence as precise if the 95% confidence interval excluded the possibility of a clinically important benefit or harm (i.e., risk ratio less than 0.75 or greater than 1.25). For the main outcome of interest, CIN, we used an optimum information size of 2000 based on an expected 0.1 probability