An Approach for Using Data Mining to Support Theory Development

Abstract

The rapid and constant change in information technologies (IT), organizational forms, and social structures is challenging our existing theories of the impact IT on organizations and society. A basic problem for researchers is how to generate testable hypotheses about the given area of research. However, new IT offer opportunities for information processing and problem solving that could extend the capacity of researchers to generate hypotheses and systematically explore the limitations of any theory. The idea of using IT to support IS research is not new. In this chapter, we explore and illustrate how data mining techniques could be applied to assist researchers in systematic theory testing and development.

Appendices

Appendix A

1.1 A Procedure for Implementing the Methodology

• Step 1: Preparation (Researcher)

  • 1. Identify Dependent variables: Identify dependent variables (e.g., Performance).

  • 2. Identify Possible Mediator variables: Identify possible mediator variables (e.g., Utilization).

  • 3. Identify Independent variables: Identify independent variables (e.g., System Quality).

  • 4. Specify the Discretization Method: An example of such a method is presented in Table 1.

  • 5. Identify the Target Events for the Dependent and Mediator Variables: For each dependent variable, identify target events (e.g., Performance is High ≡ the value of Performance is in the [5.5–7] interval) that may be of interest. Do similarly for the mediator variable(s).

  • 6. Discretize All Ordinal Variables: For each ordinal variable, discretize the given variable using the specified discretization method.

  • 7. Specify DT Generation Parameters: Specify relevant values for data partitioning parameters (e.g., distribution of cases in training, validation, and test datasets; stratification variables), DT induction parameters (e.g., splitting method options, minimum observations per leaf).

  • 8. Specify Thresholds: Specify α, the significance level for statistical testing of the hypotheses, and τ ₀, the threshold for p ₀. A Strong Single Rule Hypothesis will not be generated for any rule for which the highest p ₀ that is supported by the data is below τ ₀.

  • 9. Specify the Set of Decision Rules for Abducting and Evaluating Sibling Rules Hypotheses: A candidate Sibling Rules Hypothesis is formulated based on the relative frequency distribution of the target event for the set of sibling rules that are associated with the predictor variable. Appendix B provides an example of a Set of Decision Rules for Formulating and Evaluating Candidate Sibling Rules Hypotheses.

• Step 2: Hypotheses Variables (Automatic)

  • For each dependent variable:

    • Substep 2a: Generate DTs for the given Dependent Variable: Generate a set of DTs using the discretized dataset and the combination of DT generation parameter values that were specified in Step 1.

    • Substep 2b: Abduct and Evaluate Sibling Rules Hypotheses for the given Dependent Variable: In this substep, for each DT, a Sibling Rules Hypothesis will be abducted for each set of sibling rules if the relationship between the associated relative frequencies is included in the Set of Decision Rules for Formulating and Evaluating Candidate Sibling Rules Hypotheses. Each such abducted Sibling Rules Hypothesis is indirectly evaluated by subjecting the associated set of surrogate hypotheses to statistical testing. If they are all accepted, then there is good reason to believe that the Sibling Rules Hypothesis will not be rejected, and so it is not rejected.

    • Substep 2c: Abduct Strong Single Rule Hypotheses for the given Dependent Variable: For each sibling rule that is statistically different from each of its other sibling rules at significance level α, determine whether it is a strong rule by testing the surrogate hypothesis: p ₀ ≥ τ ₀. For each such surrogate hypothesis that is accepted, use the given rule to abduct a corresponding Strong Single Rule Hypothesis.

• Step 3: Identify Mediator Variables (Automatic)

  • Examine the set of supported hypotheses from Step 2 to determine whether any potential mediator variable is included at least one of the abducted hypotheses for one of the dependent variables.

• Step 4: Generate Hypotheses for Mediator Variables (Automatic)

  • Step 4 is executed for each mediator variable that was included in a condition event of a hypothesis for one of the dependent variables. If there is no such mediator variable, then this step is bypassed.

  • For each mediator variable:

    • Substep 4a: Generate DTs for Predicting the given Mediator Variable. Similar to Substep 2a.

    • Substep 4b: Abduct Sibling Rules Hypotheses for the given Mediator Variable. Similar to Substep 2b.

    • Substep 4c: Abduct Strong Single Rule Hypotheses for the given Mediator Variable. Similar to Substep 2c.

• Step 5: Step 5: Abduction of Theoretical Model (Automatic):

  • Generate a theoretical model by integrating the set of causal links between predictor and target variables that are associated with the abducted directional and Single Rule Hypotheses.

Appendix B

2.1 Examples of Decision Rules for Formulating and Evaluating

2.1.1 Candidate Sibling Rules Hypotheses

Assuming a discretized target variable with 3 bins, Table 9 could be used to formulate and test a candidate Sibling Rules Hypothesis of the form: Predictor variable X has { Impact Type } on Target variable Y.

Table 9 Example of decision rules for formulating candidate Sibling Rules Hypotheses