Association of Bariatric Surgery With Cardiovascular Outcomes in Adults With Severe Obesity and Nonalcoholic Fatty Liver Disease

This cohort study compares the risk of experiencing cardiovascular events after bariatric surgery vs nonsurgical treatment in adults across the full spectrum of fatty liver disease.


Missing Values
The analytic sample had complete data on all included demographics (age, gender, type of insurance, region of residence etc.). Nevertheless, under-reporting could result in missing data related to smoking status and medical history variables. For example, a participant without a diagnostic code for hypertension will be classified as without hypertension. However, the absence of diagnostic code could be due to either a true negative or false negative (patients had hypertension but was not coded as such). Consistent with prior studies, we limited our analysis to participants with at least one year of continuous enrollment prior to the first date of nonalcoholic fatty liver disease (NAFLD) diagnosis to ensure accurate reporting of patients' medical history in both the exposed and unexposed groups (1)(2)(3). As such, the presence of any missing data will result in non-differential misclassification which would only bias our findings towards the null value of no association between bariatric surgery and reductions in the risk of cardiovascular disease.

Severe Obesity Sample Restriction
We restricted the study sample to severely obese patients to reflect on both the clinical guidelines implemented during the study period and the nature of the administrative data used in the analysis. Both the 2008 and 2013 clinical guidelines by the American Association of Clinical Endocrinologists, the Obesity Society, and the American Society for Metabolic & Bariatric Surgery (ASMBS) suggest that bariatric surgery should be offered to 1) individuals with BMI ≥40 kg/m 2 with or without coexisting comorbidities, and 2) patients with both BMI ≥35 kg/m 2 and one or more severe obesity-related co-morbidities (4,5). Several health insurance providers do not include NAFLD/NASH as a qualifying condition for surgery coverage for those with BMI ≥35 kg/m 2 . In those NAFLD patients with BMI ≥35 kg/m 2 , the probability of receiving surgery is conditioned on the provider's policy. We conducted a sensitivity analysis to examine the impacts of extending the study sample to NAFLD patients with BMI ≥35 kg/m 2 .

eMethods 2. Sensitivity Analyses
We conducted multiple sensitivity analyses to assess the robustness of our main findings. First, we redefined the incident diagnoses of all eight cardiovascular disease outcomes as the presence of at least two separate inpatient or outpatient claims for each outcome made ≥90 days following the first date of NAFLD diagnosis (eTable 5). Second, we expanded our cohort to include all NAFLD patients with BMI ≥ 30.0 Kg/M 2 . We then estimated the inverse probability of treatment weights (IPTW) using inverse probability (IP) weights drawn from the expanded cohort (n=126,341). Next, we utilized the IPTW drawn for the expanded cohort to examine the association between the risk of cardiovascular disease outcomes and bariatric Surgery (eTable 6).
Third, we conduct a sensitivity analysis using the inverse probability of censoring weighting (IPTCW) to examine the effects of any potential selection bias due to informative censoring (6). As such, we used logistic regression models to estimate the probability of not being censored using patients' medical history and demographics. This model included all variables listed in table one with the addition of attained age at the end of follow-up, follow-up duration, surgery status, socioeconomic status, in-hospital mortality, pregnancy status, and employment status. Next, we further corrected the initially calculated IP weights of receiving surgery by multiplied them by the probability of not being censored to obtain the IPTCW (7). Subsequently, we used the IPTCW in Cox proportional hazard regression analyses to estimate the adjusted relationship between bariatric surgery and the risk of cardiovascular disease (eTable 7). Finally, we limited bariatric surgeries to Roux-en-Y gastric bypass and sleeve gastrectomy (eTable 8

eMethods 3. The E-value and Bias Factors Sensitivity Analyses for Unmeasured Confounding
While we controlled for known confounders in our main and sensitivity analyses, the nature of the claims data used in the study might leave room for unmeasured confounding. In response, we examined the potential impacts of any unmeasured confounders on our main findings by computing the unmeasured confounders bias factors and the E-values (8,9). The bias factor is the maximum relative amount by which an unmeasured confounder could reduce an observed risk or hazard ratio (8). To estimate the bias factor, the maximum risk ratio for the unmeasured confounders-outcome association (RR UD ) and the risk ratio for the maximum exposureconfounders relationship (RR EU ) need to be specified (Equation S1).
To obtain conservative reference values for the confounders-outcome association, we fitted bivariate Cox proportional hazard models for the associations between smoking status, diabetes, hypertension, and dyslipidemia and each of the study outcomes (i.e., any, primary, and secondary cardiovascular outcomes, as well as, for each of the eight endpoints). We also fitted bivariate logistic regression models with surgery status as the outcome and smoking status, diabetes, hypertension, and dyslipidemia each as exposures to ascertain the reference values for the exposure-confounder relationship (eTable 9). Next, we used the highest hazard and odds ratio point estimates obtained from the fitted models as inputs in equation S1 to quantify the bias factor associated with an unmeasured confounder that has the same strength of association as the highest crude impacts of known cardiovascular disease risk factors and predictors of undergoing bariatric surgery. We then divided the observed point estimates and the limits of the confidence intervals by the estimates bias factor to obtain the maximum values by which the unmeasured confounder could move the point estimates and confidence intervals towards the nullcorrected estimates (eTable 9).
The E-value is a validated measure of the observed associations' robustness to potential unmeasured confounders (10). The E-value is the minimum value of the association on the risk ratio scale that an unmeasured confounder would need to have with both the outcome and exposure conditional on the measured covariates to fully explain away a specific exposureoutcome association (Equation S2-1 & S2-2) (8). The higher the estimated E-value for a particular association, the stronger the unmeasured confounder needs to be to explain away the observed effects (point estimate becomes the null). As such, unmeasured confounders with weaker association magnitude than the E-value cannot explain the observed association. eTable 9 shows the corrected estimates relative to observed results for all the study outcomes.
After assuming an unmeasured confounder with the same magnitude of association as the strongest known risk factor for cardiovascular disease, the corrected estimates show that bariatric surgery remained significantly associated with a lower risk of any, primary, and secondary cardiovascular outcomes. Furthermore, surgery remained significantly associated with lower risks for heart failure, secondary ischemic heart events, secondary cerebrovascular events, and atherosclerosis.
In our study, the observed associations between bariatric surgery and the risks of any, primary, and secondary cardiovascular outcomes were hazard ratios (HRs) 0.51 (95% confidence interval [CI] 0.48 to 0.54), 0.53 (95% CI: 0.48 to 0.59), and 0.50 (95% CI: 0.46 to 0.53) respectively. Based on the estimated E-value for the risk of cardiovascular disease, the HR of 0.51 could be fully explained (i.e., HR becomes 1) by an unmeasured confounder that has an HR of 2.56 associated with both bariatric surgery and cardiovascular disease, in addition to the confounders we adjusted for in the analysis (eTable 10). Moreover, an unmeasured confounder needs at least an HR 2.43fold association with bariatric surgery and cardiovascular disease for the observed 95% CI to contain the null value of one. To put these E-values in perspective, the crude association of any cardiovascular disease and primary cardiovascular outcomes in our sample were HR 1.36 (95% CI: 1.26 to 1.47) and HR 1.65 (1.45 to 1.85) for those with vs. without diabetes (eTable 10).