New York Statewide Planning and Research Cooperative System (SPARCS)

The Dartmouth Committee for Protection of Human Subjects has approved this analysis as this is de-identified database. All patients with ruptured cerebral aneurysms who were registered in the SPARCS (New York State Department of Health, Albany, NY)[8] database between 2009 and 2013 were included in the analysis. For these years, SPARCS contains patient-level details for every hospital discharge, ambulatory surgery, and emergency department admission in New York State as coded from admission and billing records. This is an all payer database. SPARCS is different from other administrative databases by providing information on the specific physician, and hospital where the patient received treatment. Additionally, the ZIP code of the patient’s residence is available to the researchers. In our analysis, we used the limited version (not the public use file) of the dataset. More information about SPARCS is available at https://www.health.ny.gov/statistics/sparcs/.

Cohort Definition

In order to establish the cohort of patients, we used International Classification of Disease-9-Clinical Modification (ICD-9-CM) codes to identify patients in the registry who underwent surgical clipping (ICD-9-CM procedure code 39.51) or endovascular coiling [patient with one of the following ICD-9-CM procedure codes: 39.52 (should also have a code 88.41 and no 39.51 during the same hospitalization), 39.72, 39.75, 39.76; all these codes correspond to endovascular treatments of cerebral aneurysms and have been used before in the literature][2] for ruptured (ICD-9-CM diagnosis code 430) cerebral aneurysms between 2009 and 2013. To avoid bias arising from including patients in the cohort at multiple time points, we only considered the first intervention each patient received and measured the outcomes associated with that hospitalization. If patients underwent a different intervention at a later date for a separate aneurysm we did not reconsider them for the analysis, since they were already censored.

Outcome variables

The primary outcome variable was mortality during the initial hospitalization after treatment of a ruptured cerebral aneurysm. Secondary outcomes were length-of-stay (LOS) during the initial hospitalization, the rate of discharge to rehabilitation facility (any facility other than the patient’s home), and 30-day post-discharge readmission to any hospital.

Exposure variables

The primary exposure variable was the treatment method (surgical clipping versus endovascular coiling).

Covariates (S1 Table) used for risk-adjustment were age, gender, race (African-American, Hispanic, Asian, Caucasian, other), and insurance (private, Medicare, Medicaid, uninsured, other). The comorbidities used for risk adjustment were diabetes mellitus (DM), smoking, chronic lung disease, hypertension, hypercholesterolemia, peripheral vascular disease (PVD), congestive heart failure (CHF), coronary artery disease (CAD), history of ischemic stroke, transient ischemic attack (TIA), alcohol abuse, obesity, chronic renal failure (CRF), and coagulopathy. Only variables that were defined as “present on admission” were considered part of the patient’s preadmission comorbidity profile.

Statistical analysis

The association of treatment technique with our outcome measures was examined in a multivariable setting. Patients undergoing surgical clipping, or endovascular coiling in our cohort were selected for either procedure based on provider and patient preferences in a non-random way, before our study was executed. In order to account for this unmeasured confounding, and to simulate the effect of randomization, we used an instrumental variable analysis, an econometric technique.[9] This analysis controls for unmeasured confounders (such as SAH grade, or aneurysm location) by creating randomization on the treatment method. It is expected that this approach will balance the comparison groups in terms of variables we cannot measure. Such comparisons with instrumental variable analysis have been published before for different anesthesia techniques,[10] partial versus radical nephrectomy,[11] and stroke center versus no stroke center care.[12] In all these cases, but particularly in the last example (which used the same database as our study) the authors did not have data on the baseline NIH stroke scale or neurologic status of the patients. This was the reason they used an instrumental variable analysis.

The regional ratio of coiling (county level coiling ratio—defined as the number of coiled patients divided by the total number of interventions for cerebral aneurysms in a county) was used as an instrument. The regional ratio of a procedure has been used before to create pseudorandomization on the treatment method, using an instrumental variable analysis.[13] A good instrument is not associated with the outcome other than through the exposure variable of interest (a requirement known as the exclusion restriction criterion).[13] In our case it is unlikely that the regional ratios of coiling would be associated with treatment mortality in any way other than the choice of treatment. This has been accepted for similar instruments (regional ratio of a treatment) in prior literature.[10–12] A two stage least squares (2SLS) method was used for the calculation of the coefficients. The value of the F statistic in the first stage of the 2SLS approach was 30, which is consistent with a strong instrument (F statistic>10), based on a practical rule, published before in the literature.[9]

A probit regression was used for the categorical outcomes (mortality, discharge to rehabilitation, 30-day readmission),[14] and a linear regression for the linear outcomes (LOS). The covariates used for risk adjustment in these models were: age, gender, race, insurance, hospital ID, and all the comorbidities mentioned previously. Since the coefficients produced by the probit function are not interpretable, we used the marginal effects of our independent variables instead. The marginal effects are the partial derivatives of the coefficients, and reflect the change in the probability of the dependent variable, for 1 unit change in the independent variable, at the average value of all other covariates.

In order to demonstrate the robustness of our data in a sensitivity analysis, we used standard techniques to account for measured confounding while accounting for clustering at the hospital level. For categorical outcomes we used a logistic regression model with hospital name as a random effects variable, while controlling for all the covariates mentioned previously. In an alternative way to control for confounding, we used a propensity adjusted (with deciles of propensity score) logistic regression model. We calculated the propensity score of coiling, using a separate logistic regression model, adjusting for all the covariates mentioned previously. For continuous outcomes, similar analyses were conducted using linear models. Logarithmic transformation of the values of LOS yielded identical results and is therefore not reported further.

Regression diagnostics were used for all models. The number needed to treat (NNT) was calculated when appropriate. All results are based on two sided tests, and the level of statistical significance was set at 0.05. This study, based on 4,098 patients, has sufficient power (90%) at a 5% type I error rate to detect differences in mortality, as small as 10.4%. Statistical analyses were performed using Stata version 13 (StataCorp, College Station, TX), the 64-bit version of R.3.1.0 (R Foundation for Statistical Computing), and SPSS version 22 (IBM, Armonk, NY).

Patient characteristics

In the selected study period there were 4,098 patients undergoing treatment for ruptured cerebral aneurysms (mean age was 53.8 years, with 59.6% females) who were registered in SPARCS, 2,585 (63.1%) underwent endovascular coiling, and 1,513 (36.9%) underwent surgical clipping. The characteristic of the two cohorts at baseline can be seen in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Patient characteristics.

https://doi.org/10.1371/journal.pone.0137946.t001

Inpatient mortality

Overall, 117 (7.7%) inpatient deaths were recorded after clipping and 264 (10.2%) after coiling (Table 2). Clipping was associated with decreased mortality in comparison to coiling (OR, 0.73; 95% CI, 0.58 to 0.91) in unadjusted analysis. Likewise, there was no association of treatment technique with mortality (ME, -0.56; 95% CI, -1.03 to 0.02) after using a probit regression with instrumental variable analysis (Table 3). This persisted in a mixed effects logistic regression model (OR, 0.88; 95% CI, 0.69 to 1.14) and a propensity score adjusted model (OR, 0.83; 95% CI, 0.65 to 1.04).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Outcomes.

https://doi.org/10.1371/journal.pone.0137946.t002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Multivariable models examining the association of surgical clipping with outcomes.

https://doi.org/10.1371/journal.pone.0137946.t003

Discharge to rehabilitation

Overall, 948 (62.7%) were discharged to rehabilitation after clipping and 1,324 (51.2%) after coiling (Table 2). Clipping was associated with an increased rate of discharge to rehabilitation in comparison to coiling (OR, 1.54; 95% CI, 1.34 to 1.78) in the unadjusted analysis. This persisted (ME, 0.63; 95% CI, 0.24 to 1.01) after using a probit regression with instrumental variable analysis (Table 3). We found similar results in a mixed effects logistic regression model (OR, 1.65; 95% CI, 1.39 to 1.95) and a propensity score adjusted model (OR, 1.69; 95% CI, 1.45 to 1.96). This corresponded to 9 patients needed to be treated with coiling to prevent one discharge to rehabilitation.

30-day readmission

Overall, 108 (7.1%) were readmitted within 30-days after clipping and 179 (6.9%) after coiling (Table 2). Clipping was not associated with increased rate of 30-day readmission in comparison to coiling (OR, 1.03; 95% CI, 0.80 to 1.32) in the unadjusted analysis. Similarly, there was no association (ME, -0.30; 95% CI, -0.82 to 0.22) after using a probit regression with instrumental variable analysis (Table 3). We found similar results in a mixed effects logistic regression model (OR, 1.05; 95% CI, 0.81 to 1.38) and a propensity score adjusted model (OR, 1.05; 95% CI, 0.82 to 1.35).

Length-of-stay

The average LOS was 18 days (SD 13) after clipping, and 16 days (SD 13) after coiling (Table 2). Clipping was not associated with increased LOS in comparison to coiling (beta, 1.00; 95% CI, -1.14 to 1.16) in the unadjusted analysis. This persisted (ME, 1.72; 95% CI, -3.39 to 6.84) after using a linear regression with instrumental variable analysis (Table 3). We found similar results in a mixed effects linear regression model (beta, 1.26; 95% CI, -0.10 to 2.42) and a propensity score adjusted linear regression model (beta, 1.16; 95% CI, -0.02 to 2.33).