Analyzing Complex Data: Robust Regression, Fixed Effects, and Two-Stage Least Squares

2.50
Hdl Handle:
http://hdl.handle.net/10755/157913
Type:
Presentation
Title:
Analyzing Complex Data: Robust Regression, Fixed Effects, and Two-Stage Least Squares
Abstract:
Analyzing Complex Data: Robust Regression, Fixed Effects, and Two-Stage Least Squares
Conference Sponsor:Western Institute of Nursing
Conference Year:2009
Author:Blegen, Mary A., RN, PhD, FAAN
P.I. Institution Name:University of California San Francisco, Community Health Systems
Title:Professor
Contact Address:2 Koret #0608, San Francisco, CA, 94143, USA
Contact Telephone:415-476-2599
Co-Authors:Shin Hye Park, Research Assistant; Joanne Spetz, Associate Professor
Purpose: The purpose of this project was to explore ways to control statistically for complex reality and imperfect data in estimating a causal model. Background: Cross-sectional data are often the only data available to evaluate causal models used to describe a complex reality that exists in an environment in which experimental studies cannot be conducted. Several approaches have been suggested to improve the rigor of these analyses. Methods: This research project used two data sets created by the University Health System Consortium. The Clinical Data Set contained patient discharge data and the Operational Data Set contained nurse staffing (direct nursing care hours and patient days of care) at the level of the patient care unit. Risk adjusted outcomes were calculated using the AHRQ algorithms for Inpatient Quality Indicators and Patient Safety Indicators. Staffing data were calculated per patient day adjusted for short stay and observation patients. In addition, data describing the environment in which the hospital operated were also collected and linked: RN supply and general unemployment rate in the surrounding area. Ordinary Least Squares (OLS) regression results were contrasted with OLS with robust Standard Errors, Fixed Effect Regression, and Instrumental Variable analyses (2 stage least squares). Robust Standard errors were used to correct for nesting of quarterly values for staffing and outcomes within hospitals. Fixed effect regression considered all the hospital characteristics fixed and evaluated the variance across quarters in staffing and outcomes. Instrumental variable analysis was used to estimate nurse staffing levels with exogenous predictors to control the likely endogeniety of staffing in predicting outcomes. Results: OLS regression coefficients with and without robust standard errors were useful in estimating the model, although inferences about statistical significance differed. Fixed effects regression found few significant coefficients using only the variance across four quarters to estimate the model. The selected instrumental variables were good predictors of nurse staffing in 2 stage least squares but the STATA diagnostics suggested that staffing was not endogenous in the model. Implications: Advanced statistical analyses are necessary in estimating complex causal models with cross-sectional data. However, with the data available to this project, the simplest of those techniques was sufficient.
Repository Posting Date:
26-Oct-2011
Date of Publication:
17-Oct-2011
Sponsors:
Western Institute of Nursing

Full metadata record

DC FieldValue Language
dc.typePresentationen_GB
dc.titleAnalyzing Complex Data: Robust Regression, Fixed Effects, and Two-Stage Least Squaresen_GB
dc.identifier.urihttp://hdl.handle.net/10755/157913-
dc.description.abstract<table><tr><td colspan="2" class="item-title">Analyzing Complex Data: Robust Regression, Fixed Effects, and Two-Stage Least Squares</td></tr><tr class="item-sponsor"><td class="label">Conference Sponsor:</td><td class="value">Western Institute of Nursing</td></tr><tr class="item-year"><td class="label">Conference Year:</td><td class="value">2009</td></tr><tr class="item-author"><td class="label">Author:</td><td class="value">Blegen, Mary A., RN, PhD, FAAN</td></tr><tr class="item-institute"><td class="label">P.I. Institution Name:</td><td class="value">University of California San Francisco, Community Health Systems</td></tr><tr class="item-author-title"><td class="label">Title:</td><td class="value">Professor</td></tr><tr class="item-address"><td class="label">Contact Address:</td><td class="value">2 Koret #0608, San Francisco, CA, 94143, USA</td></tr><tr class="item-phone"><td class="label">Contact Telephone:</td><td class="value">415-476-2599</td></tr><tr class="item-email"><td class="label">Email:</td><td class="value">Mary.Blegen@nursing.ucsf.edu</td></tr><tr class="item-co-authors"><td class="label">Co-Authors:</td><td class="value">Shin Hye Park, Research Assistant; Joanne Spetz, Associate Professor</td></tr><tr><td colspan="2" class="item-abstract">Purpose: The purpose of this project was to explore ways to control statistically for complex reality and imperfect data in estimating a causal model. Background: Cross-sectional data are often the only data available to evaluate causal models used to describe a complex reality that exists in an environment in which experimental studies cannot be conducted. Several approaches have been suggested to improve the rigor of these analyses. Methods: This research project used two data sets created by the University Health System Consortium. The Clinical Data Set contained patient discharge data and the Operational Data Set contained nurse staffing (direct nursing care hours and patient days of care) at the level of the patient care unit. Risk adjusted outcomes were calculated using the AHRQ algorithms for Inpatient Quality Indicators and Patient Safety Indicators. Staffing data were calculated per patient day adjusted for short stay and observation patients. In addition, data describing the environment in which the hospital operated were also collected and linked: RN supply and general unemployment rate in the surrounding area. Ordinary Least Squares (OLS) regression results were contrasted with OLS with robust Standard Errors, Fixed Effect Regression, and Instrumental Variable analyses (2 stage least squares). Robust Standard errors were used to correct for nesting of quarterly values for staffing and outcomes within hospitals. Fixed effect regression considered all the hospital characteristics fixed and evaluated the variance across quarters in staffing and outcomes. Instrumental variable analysis was used to estimate nurse staffing levels with exogenous predictors to control the likely endogeniety of staffing in predicting outcomes. Results: OLS regression coefficients with and without robust standard errors were useful in estimating the model, although inferences about statistical significance differed. Fixed effects regression found few significant coefficients using only the variance across four quarters to estimate the model. The selected instrumental variables were good predictors of nurse staffing in 2 stage least squares but the STATA diagnostics suggested that staffing was not endogenous in the model. Implications: Advanced statistical analyses are necessary in estimating complex causal models with cross-sectional data. However, with the data available to this project, the simplest of those techniques was sufficient.</td></tr></table>en_GB
dc.date.available2011-10-26T20:19:32Z-
dc.date.issued2011-10-17en_GB
dc.date.accessioned2011-10-26T20:19:32Z-
dc.description.sponsorshipWestern Institute of Nursingen_GB
All Items in this repository are protected by copyright, with all rights reserved, unless otherwise indicated.