Program Satisfaction Knowledge Item 10 Extended: Results from HRSA/HAB's SPNS Innovative Models of Care

Extended Results
Knowledge Item: CA-Program Satisfaction-10

Model 1: Totally Empirical Analysis

Exhaustive CHAID (CHi-square Automatic Interaction Detector) modeling was used to study the relationship of the need and vulnerability factors discussed in Knowledge Item CA-Initiative Impact-07 to a total satisfaction score formed from the responses to nine items. The resulting pattern of significant relationships is shown in the figure below. As a context for this model, note that the satisfaction scores are heavily skewed toward the highest possible ratings. Note that an alternate model for these same data is shown in the Additional Statistics section of this Knowledge Item.

While the sample can be split on several factors to yield groups of patients who are relatively more or less satisfied, it should be recognized that the vast majority of all patients are highly satisfied with their services.

Note that in the following CHAID diagrams, the green bars represent the distribution among clients/patients of Total Satisfaction Score.

Click graphic to expand. (IE 6 users may also have to click the graphic expansion button in the new window.)

More Information: CHAID and CHAID Diagram

Click here for a list of the possible predictor variables that would have been entered, if statistically significant, into this CHAID model, and a discussion of several important methodological issues about these indicators.

Note on CHAID models: CHAID is a useful method of summarizing data, and can show major natural divisions of the sample by various defining variables. It must be recognized, however, that CHAID is analogous to a "forward" stepwise regression analysis and has all of the possible inferential difficulties of such stepwise regression methods. The statistical significance tests are sequential ones dependent on prior splits of the sample. In many cases, the models presented should be considered as suggestive, but not absolutely definitive as there may be alternate models that also fit these data in a statistically or theoretically acceptable manner. This model may have been manually altered very slightly from the automated CHAID modeling trees; specifically, categories for "missing values" may have been separated (or re-split) from categories for actual values with which they were statistically merged if the authors judged this would give a more clear interpretation of the data; the separation may result in a "missing category" with only a few cases that could be statistically merged with one of the other categories. [In those cases where the "missing value" category is combined with actual values, it was judged that the automated split was a better representation of these data.] The use of Bonferroni confidence intervals to correct for the potentially large number of statistical tests in this model building method and the use of more stringent alpha levels results in relatively conservative data representations. All patient-client model analyses were conducted by the senior author [GH] so that consistent data fitting techniques and judgments would be employed in the different areas studied. In many cases, alternate models are presented so that the viewer can judge the appropriateness of one or more ways of looking at the same data.

Simple Effect Models

The following charts are simple predictive CHAID (CHi-square Automatic Interaction Detector) models. Note that the green bars represent Total Satisfaction Score. Each CHAID in this next series of diagrams is split by a specific service need, vulnerability or demographic factor.

The difference between the two alternate simple effects models shown is as follows. The top alternative includes cases with a missing value on the splitter variable. Additionally, the split is made by forcing each category of the splitter variable to be separate. Since this decision was made a priori, the probability shown has not been adjusted for the number of possible ways of grouping the categories of the splitter variables (that is, there is no Bonferroni adjustment of probabilities). For the second alternative, only cases which have values on the splitter variable are used. Additionally, the program optimally splits the single predictor variable in a hierarchical way using statistical fitting. For this reason, in the second alternative, the probabilities are adjusted for the number of possible statistical tests using the Bonferroni method. In some cases, the two alternatives produce identical results; in such cases, only one tree is shown.

Race-Ethnicity