Ghent University Academic Bibliography

Advanced

On quantifying quality of care

Machteld Varewyck (2015)
abstract
In this thesis, we examine how the analysis of quality outcomes, such as 30-day mortality for patients with acute stroke, can help compare the quality of care between hospitals. As in the neighboring countries, the demand for quality control in hospitals is growing but also, for example for residential care centers and schools, both by government and patients as well as centers themselves. Given the potentially large impact of reported results, this requires a careful statistical analysis of the available data, as discussed in Chapter 1. To estimate the causal effect of the quality of care on the outcome of interest, we have to control for differences between patients on admission, such as age and initial disease severity. This is necessary because they may influence the outcome and they are possibly distributed differently across centers. Otherwise, hospitals treating mostly elderly patients may show higher mortality risks, even though the given care is excellent. The research questions in this thesis were mostly inspired by the analysis of the Swedish register for acute stroke care, Riksstroke (http://www.riksstroke.org/eng/), but the discussed methods are more generally applicable. To account for measured patient characteristics we will use, depending on the research question, directly or indirectly ized risks as performance measure. It has been proven that when standardized risks are estimated based on the popular normal mixed effects model, the estimated quality of care may be shrunken towards the average, often masking outlying performance of hospitals (Normand et al., 1997; Ash et al., 2012). In Chapter 2 we therefore investigated the use of a Firth corrected fixed effects model and found little shrinkage of the center effects towards the overall mean. This approach is thus particularly valuable when some centers have a small number of registered patients since the convergence of this estimation strategy is better than for fixed effects models and a better detection of outlying performance is obtained than for normal mixed effects models. Secondly, we investigate undue model extrapolation when estimating for example, directly standardized risks, especially if patient mix differs substantially between hospitals. Extrapolation in combination with the use of misspecified statistical models can yield biased results with an underestimated uncertainty. Therefore, we examined a method that weights observations by the inverse of the so-called propensity score, i.e. the probability to be treated in the observed center (Shahian and Normand, 2008). The investigated doubly robust method is protected against model misspecification (Robins et al., 2007) and, if the propensity score is very small, the user will be warned for extrapolation via inflated variance estimates. Although promising, the obtained results suggested to use the Firth corrected fixed effects method. Common adjustments for differences in patientmix generally assume that the effect of the given care level on the outcome is constant across patient groups (Ohlssen et al., 2007b; Shahian and Normand, 2008). In practice, however, this may be violated when some centers are for example specialized in care for the elderly (Nicholl et al., 2013;Mohammed et al., 2009). If then no center-patient interactions are included in the outcome regression model, we found in Chapter 3 that the directly and indirectly standardized risks will only be biased if the distribution of that patient characteristic differs substantially across centers, otherwise bias is negligible. Being able to justify common practice is especially important in settings where it is simply impossible to estimate these interactions in the model, because insufficient information is available in small hospitals, for example see Ash et al. (2012). In Chapter 4 we also examined how the number of (expensive, genetic) measurements - and thus the cost per patient - can be reduced when predicting individual patient outcomes or estimating standardized risks for hospital quality evaluation. Stochastic search algorithms allow for a relatively quick and costefficient variable selection and they can easily handle multiple imputed datasets when some measurements are missing. We have also illustrated how the search time can be further reduced by a priori performing a cost-efficient generalized LASSO search. Because we believe in the broad applicability of the statistical methods in this thesis, we havemade them available via the R-package RiskStandard (www.cvstat.ugent.be), as documented in Chapter 5.
Please use this url to cite or link to this publication:
author
promoter
UGent and UGent
organization
year
type
dissertation
publication status
published
subject
pages
IX, 191 pages
publisher
Ghent University. Faculty of Sciences
place of publication
Ghent, Belgium
defense location
Gent : Campus Sterre (S9, aud. A2)
defense date
2015-12-18 17:00
language
English
UGent publication?
yes
classification
D1
copyright statement
I have transferred the copyright for this publication to the publisher
id
7026411
handle
http://hdl.handle.net/1854/LU-7026411
date created
2016-01-06 12:00:48
date last changed
2017-04-03 12:29:03
@phdthesis{7026411,
  abstract     = {In this thesis, we examine how the analysis of quality outcomes, such as 30-day mortality for patients with acute stroke, can help compare the quality of care between hospitals. As in the neighboring countries, the demand for quality control in hospitals is growing but also, for example for residential care centers and schools, both by government and patients as well as centers themselves.
Given the potentially large impact of reported results, this requires a careful statistical analysis of the available data, as discussed in Chapter 1. To estimate the causal effect of the quality of care on the outcome of interest, we have to control for differences between patients on admission, such as age and initial disease severity. This is necessary because they may influence the outcome and they are possibly distributed differently across centers. Otherwise, hospitals treating mostly elderly patients may show higher mortality risks, even though the given care is excellent. The research questions in this thesis were mostly inspired by the analysis of the Swedish register for acute stroke care,  Riksstroke (http://www.riksstroke.org/eng/), but the discussed methods are more generally applicable. To account for measured patient characteristics we will use, depending on the research question, directly or indirectly ized risks as performance measure.
It has been proven that when standardized risks are estimated based on the popular normal mixed effects model, the estimated quality of care may be shrunken towards the average, often masking outlying performance of hospitals (Normand et al., 1997; Ash et al., 2012). In Chapter 2 we therefore investigated the use of a Firth corrected fixed effects model and found little shrinkage of the center effects towards the overall mean. This approach is thus particularly valuable when some centers have a small number of registered patients since the convergence of this estimation strategy is better than for fixed effects models and a better detection of outlying performance is obtained than for normal mixed effects models. Secondly, we investigate undue model extrapolation when estimating for example, directly standardized risks, especially if patient mix differs substantially between hospitals. Extrapolation in combination with the use of misspecified statistical models can yield biased results with an underestimated uncertainty. Therefore, we examined a method that weights observations by the inverse of the so-called propensity score, i.e. the probability to be treated in the observed center (Shahian and Normand, 2008). The investigated doubly robust method is protected against model misspecification (Robins et al., 2007) and, if the propensity score is very small, the user will be warned for extrapolation via inflated variance estimates. Although promising, the obtained results suggested to use the Firth corrected fixed effects method.
Common adjustments for differences in patientmix generally assume that the effect of the given care level on the outcome is constant across patient groups (Ohlssen et al., 2007b; Shahian and Normand, 2008). In practice, however, this may be violated when some centers are for example specialized in care for the elderly (Nicholl et al., 2013;Mohammed et al., 2009). If then no center-patient interactions are included in the outcome regression model, we found in Chapter 3 that the directly and indirectly standardized risks will only be biased if the distribution of that patient characteristic differs substantially across centers, otherwise bias is negligible. Being able to justify common practice is especially important in settings where it is simply impossible to estimate these interactions in the model, because insufficient information is available in small hospitals, for example see Ash et al. (2012).
In Chapter 4 we also examined how the number of (expensive, genetic) measurements - and thus the cost per patient - can be reduced when predicting individual patient outcomes or estimating standardized risks for hospital quality evaluation. Stochastic search algorithms allow for a relatively quick and costefficient variable selection and they can easily handle multiple imputed datasets when some measurements are missing. We have also illustrated how the search time can be further reduced by a priori performing a cost-efficient generalized LASSO search.
Because we believe in the broad applicability of the statistical methods in this thesis, we havemade them available via the R-package RiskStandard (www.cvstat.ugent.be), as documented in Chapter 5.},
  author       = {Varewyck, Machteld},
  language     = {eng},
  pages        = {IX, 191},
  publisher    = {Ghent University. Faculty of Sciences},
  school       = {Ghent University},
  title        = {On quantifying quality of care},
  year         = {2015},
}

Chicago
Varewyck, Machteld. 2015. “On Quantifying Quality of Care”. Ghent, Belgium: Ghent University. Faculty of Sciences.
APA
Varewyck, Machteld. (2015). On quantifying quality of care. Ghent University. Faculty of Sciences, Ghent, Belgium.
Vancouver
1.
Varewyck M. On quantifying quality of care. [Ghent, Belgium]: Ghent University. Faculty of Sciences; 2015.
MLA
Varewyck, Machteld. “On Quantifying Quality of Care.” 2015 : n. pag. Print.