patients with dentine hypersensitivity. This study contributed two new mechanisms to the original response shift model, ‘acceptance’ and ‘habits’.
Acceptance involved accepting symptoms of dentine hypersensitivity and recognizing that the condition is incurable. Habits were manifest as routinized changes in oral hygiene routines, eating and drinking to cope with dentine hypersensitivity, that became almost sub-conscious.
Qualitative assessments allow the incorporation of individual concepts of QoL and importantly, to test variations within those meanings. Such data obtained must be interpreted carefully. Due to the intensive effort and time applied in the interviews, the sample size is small. Further, the analysis may be subjective and depend on the skills of the researcher (Schwartz and Sprangers, 1999).
Nonetheless, qualitative assessments are essential to illuminate QoL measures and to incorporate concepts of HRQoL into the different theories and models.
2.2.5.2.5 Statistical Approaches
With the development of technology and computational sciences, several new methods can be used in sophisticated data analysis. Statistical methods applied to the study of RS can be utilized in both primary and secondary data sets.
Thus, large data sets are much more manageable for response detection and the analysis of different aspects of RS is more reliable (Schwartz et al., 2013).
Several statistical methods have successfully identified response shift in such different disease populations as hypertension with coronary artery disease (Gandhi et al., 2013), stroke (Ahmed et al., 2005a) multiple sclerosis (Mayo et al.,
2009, Ahmed et al., 2011, Li and Schwartz, 2011, King-Kallimanis et al., 2011) cancer (Oort et al., 2005), obstructive pulmonary disease (Ahmed et al., 2009), and HIV/AIDS (Li and Rapkin, 2009). Those include techniques such as Structural Equation Modelling, Recursive Partitioning and Regression Trees Method (RPRT) and Trajectory Analysis with subject- centred residuals. Moreover, an increasing body of evidence supports the convergence among statistical and other methods of detection of RS (Mayo et al., 2008, Visser et al., 2005, Ahmed et al., 2005b), empathizing their inclusion in any study on RS.
Structural Equation Modelling
Structural Equation Modelling (SEM) uses different types of models to illustrate relationships among observed variables to test a theoretical model quantitatively. The aim of SEM is to determine the extent to which the theoretical model is supported by the sample data. It represents an extension of general linear modelling procedures such as ANOVA, multiple regression and confirmatory factor analysis (Bollen, 1995).
SEM defines two types of variables: observed and latent. The observed variables are measured whereas latent variables are indirectly inferred from the observed variables. For instance, a latent variable of socio-economic status could be considered by combining data on education, income and occupation. Thus, SEM tests the overall fit of a model and assesses direct and indirect links between observed and latent variables.
To test response shift, a common factor model is used to describe the
response shift, the difference between common factor means is used as a measure of true change (Visser et al., 2005).
After this analysis, response shift components are operationalized as follows (Oort et al., 2005):
- Recalibration is inferred from residual change in responses as a function of time or change in intercepts.
- Reprioritization is inferred from the change in variance in factors loading values over time.
- Reconceptualization is seen as zero versus nonzero factor loading pattern changes over time.
SEM has been useful in detecting response shifts in patients with cancer (Oort et al., 2005, Visser et al., 2005), stroke (Barclay-Goddard et al., 2009b), hypertension with coronary artery disease (Gandhi et al., 2013) and multiple sclerosis (King-Kallimanis et al., 2011). The main limitation of this method is that in the absence of external criteria, response shift cannot be detected if it affects most of the results in the same way (Schwartz et al., 2011). Furthermore, this method requires large samples (n>200).
Classification and Regression Trees (CRT)
Classification and Regression Trees (CRT) is a non-parametric statistical method developed by Breiman and colleagues (1984) commonly used in data mining to create predictive models. Different abbreviations found in the literature such as Classification and Regression Trees (CART, CRT, C&RT), Recursive Partitioning and Regression Trees (RPART) or Regression Trees Analysis (RTA),
are referred to the same method depending on the software employed, but this review will use the term CRT throughout.
CRT creates a regression tree as a representation of the data. Each of the terminal nodes or leaves of the tree represents a cell of the partition, and has attached a simple model that applies to that cell only. The members of the studied population are classified based on several dichotomous dependant variables (Li and Schwartz, 2011). CRT is non-model based, thus it enables intuitive predictions without predefinition of possible interactions among factors and allows exploration of non-linear relationships among variables in a graphical representation (Hastie et al., 2013).
This method has been used in data mining to detect different patterns and trajectories of response shift. (Li and Rapkin, 2009, Li and Schwartz, 2011). The different forms of response shift have been operationalized as follows (Schwartz et al., 2011):
- Recalibration is inferred by using trees indicating relationships between predictors and outcomes scores using different group-specific thresholds or cut points for selected predictors variables. The interaction terms are used to identify homogeneous groups over time.
- Reprioritization is inferred from changes in the order of domains in tree pathways over time.
- Reconceptualization is inferred from changes in the content and/or number of domains by group in a pruned tree over time.
The limitations of this method are that involves substantial qualitative interpretation of the results and there are no specific codes to detect different aspects of response shift.
Trajectory Analysis with subject- centered residuals
Latent Trajectory Analysis with subject-centered residuals (Mayo et al., 2009) consists of developing a predictive General Health model to examine patterns in discrepancies between expected and observed scores. A longitudinal model with a random intercept is created to predict General Health using only significant predictors, excluding predictors if their association with the outcome varies over time. This method detects reprioritization and reconceptualization as fluctuations in differences between observed and predicted scores or residuals over time (Ahmed et al., 2011).
Due to random error, there will be always some random variation in the data with over and underestimation, so masking the response shift detection.
Likewise, to correctly interpret residuals in terms of response shift, an external criterion such an appraisal process is required.
Disadvantages of statistical methods to detect RS
Several of the statistical approaches used to explore RS are not based on the theories of RS, but lies in the study design, sample size or variable distributions (Sawatzky et al., 2017). Additionally, some methods are parametric, assuming a normal distribution and homogeneity of data. Such assumptions are not always met in QoL data, which are often skewed and show substantial variability across groups (Beaumont et al., 2006). Thus, the replicability of the results may be limited and individual effects may be masked when observing group level data (Barclay-Goddard et al., 2009a).
However, as each major approach to study response shift (design-based, individualized) relies on a different operationalization, the use of a statistical approach in any study of RS is strongly encouraged (Ahmed and Ring, 2008).