The models described so far have all been based upon the assumption that QoL scales can be represented as in Figure 5.1, with observed variables that reflect the value of the latent variable. For example, in the HADS, presence of anxiety is expected to be manifested by high levels of Qi, Q3, Q5, g7, gn and g13. However, many QoL instruments include a large number of items covering diverse aspects of QoL. For example, the RSCL includes 30 items relating to general QoL, symptoms and side-effects; it also incorporates an activity scale and a global question about overall QoL. For simplicity, we restrict consideration to 17 items. Adopting a conventional EFA model, a path diagram such as that of Figure 5.5 might be considered.
This model assumes that a poor QoL is likely to be manifested by psychological distress, and that "psychological distress" is a latent variable that tends to result in anxiety, depression, despair, irritability and similar signs of distress. This much does seem a plausible model. However, a patient with poor QoL need not necessarily have high levels of all treatment-related symptoms. For example, a patient receiving chemotherapy may well suffer from hair loss, nausea, vomiting, and other treatment-related side-effects that cause deterioration in QoL. However, other cancer patients receiving non-chemotherapy treatments could be suffering from a completely different set of symptoms and side-effects that cause poor QoL for other reasons.
Thus a poor QoL does not necessarily imply that, say, a patient is probably experiencing nausea; this is in contrast to the psychological distress indicators, all of which may well be affected if the patient experiences distress because of their condition. On the other hand, if a patient does have severe nausea, that is likely to result in—or cause—a diminished QoL. Hence a more realistic model is as in Figure 5.6, where symptoms and side-effects are shown as causal indicators with the directional arrows pointing from the observed variables towards the "symptoms and side-effects" factor, which in turn causes changes in QoL. The observed items reflecting psychological distress are called effect indicators, to distinguish them from the causal indicators. Thus effect indicators can provide a measure of the QoL experienced by patients, whilst the causal indicators affect or influence patients' QoL.
Fayers and Hand (1997a) analysed RSCL data from an MRC trial of chemotherapy with or without interferon for patients with advanced colorectal cancer. There appeared to be four factors, representing psychological distress, symptoms, nausea and vomiting, and pains and aches. At first sight the second factor, labelled "symptoms", contained a strange combination of items: lack of appetite, decreased sexual interest, dry mouth, tiredness and lack of energy. However, these five symptoms were precisely the items that the study team had reported as the main treatment differences in the randomised trial. In other words, the second factor is an interferon-related cluster of symptoms, and the item correlations arise from treatment differences and not through any sense of this necessarily being a single meaningful QoL construct.
Exploratory factor analysis is ill-equipped to deal with causal variables. Instead, a more general approach has to be considered, with models that can represent structures such as those of Figure 5.6 and can estimate the coefficients and parameters describing the various paths. This approach is known as structural equation modelling (SEM). At present, SEM requires the use of specialised software, and is not included as a standard part of most statistical packages. Programs for SEM models include AMOS (Arbuckle, 1997), EQS (Bentler, 1995) or LISREL (Joreskog and Sorbom, 1996).
One major difference between EFA and SEM is the emphasis that the latter places upon prior specification of the postulated structure. Thus one form of SEM is also known as confirmatory factor analysis, since a factor-analytic structure is
Was this article helpful?