The data are from a prospective study whose purpose was to characterize the spontaneous recovery of language functioning in patients during the first 3 to 4 months following a stroke. Details of the design and results of this study may be found in Holland et al. (1985) and Holland, Greenhouse, Fromm, and Swindell (1989). Briefly, patients admitted to a hospital with the diagnosis of stroke and who met the study inclusion criteria were seen daily (6 days/week) for 15 minutes beginning 48 to 72 hours poststroke and continuing until hospital discharge. The daily visits consisted of semistructured conversational interactions and systemic scoring of the patient's communication behavior. At hospital discharge, patients were formally tested using the Western Aphasia Battery (WAB). If the WAB indicated aphasic dysfunction (AQ < 93.8), the patient was seen and assessed at 1 month postdischarge. If this second WAB demonstrated continued difficulty the patient was assessed again at 2 months postdischarge. The data set to be discussed here consists of 61 patients who had unilateral (right- or left-hemisphere) stroke, ischemic or hemorrhagic etiology, and impaired language functioning at time of hospital discharge. We will use these data to investigate the correlates of language recovery during the two month period post-hospital discharge. Definitions and the codes for the variables of interest are presented in Table 2.
The multiple logistic regression model will allow us to investigate
simultaneously the effect of a number of different variables on the
likelihood of spontaneous language recovery. A feature of the
logistic regression model is that it allows us to use explanatory
variables that are either quantitative or discrete. Using the
variables defined and coded in Table 2 for the stroke
recovery study,
the specification of the deterministic component of the multiple
linear logistic regression model on the logit scale is
. (1)
Note that if we were to dichotomize length of stay, age and WAB1 then
a multiple cross-classification table of these data would consist of
cells or subgroups and that with 61 patients such a table
would have numerous sparse or empty cells which would be a problem for
a chi-square analysis. An advantage of a statistical model such as
the one given in equation (1), however, is that the model provides a
structure that smooths the fit of the probability of a response to all
the cells in such a table.
Table 1: Example of what the data file 'stroke.dat' looks like
Table 2: Variables and Codes from the Stroke Recovery Study