Response by Rumalla and Burkhardt to Letter Regarding Article, "Recent Nationwide Impact of Mechanical Thrombectomy on Decompressive Hemicraniectomy for Acute Ischemic Stroke".

In Response: We read with great interest a meticulous letter written by Neugebauer et al on account of our recent article entitled “Recent Nationwide Impact of Mechanical Thrombectomy on Decompressive Hemicraniectomy for Acute Ischemic Stroke”. The authors have requested that we comment further on several aspects of our statistical methodology. We agree with Neuegerbauer et al that our study design and data source are not sufficient to establish causality between the 2 procedures. This limitation is inherent to all cross-sectional analyses using the National Inpatient Sample. Multivariable analyses of retrospective, observation data are limited to the identification of independent associations between variables. Thus, we do not claim to have identified a temporal or causal relationship. The programming statements for applying the appropriate stratum, cluster, and discharge weights are discussed at https://www.hcupus.ahrq.gov/reports/methods/2015_09.jsp. The website provides example instructions for the SAS statistical package. In our study, we similarly applied the variables (DISCWT, HOSP_NIS, NIS_ STRATUM) in the complex samples menu. The authors also raise concern for confounding by indication and propose alternative statistical approaches. In our study, we employed statistical methodology that has been frequently used to analyze the National Inpatient Sample and similar data sets to attempt to limit (but not eliminate) confounding variables. This involved utilization of binary logistic regression modeling to adjust for demographics, hospital characteristics, comorbidities, and severity of illness. Univariate analysis for preselection followed by stepwise backward variable selection is a commonly cited practice in nationwide database studies. The authors expressed concern that our selection process may have eliminated clinically significant covariates. In our study, the only variable excluded during preselection and stepwise backward selection was sex. The statistical methodology used in our article is previously documented in the stroke literature. Vahidy FS et al studied predictors of 30-day readmission after ischemic stroke using the Nationwide Readmissions Database. In their methodology, variables were entered in multivariable analysis based on univariate statistical significance and clinical importance. Qureshi AI et al studied factors associated with prolonged length of stay during hospitalization for transient ischemic attack. The authors used univariate analysis and included only significant variables into a stepwise logistic regression model. The variables were retained if P<0.01. In conclusion, many of the limitations described in the letter by Neugebauer et al are inherent to all cross-sectional analyses of nationwide databases, including our study. Thus, it is important for limitations to be clearly discussed in all associated article. These limitations include but are not limited to confounding by indication and inability to establish causal relationships.