Modified Stepwise Regression Approach to Streamlining Predictive Analytics for Construction Engineering Applications

A literature review has identified the absence of a robust framework that guides the development of streamlined and valid multiple linear regression (MLR) predictive models for construction engineering applications. A reliable MLR model requires an appropriate set of input variables that can satisfy the underlying assumptions of best linear unbiased estimators (BLUE). In this research, an analytical framework is proposed for developing MLR-based predictive models by (1) selecting input variables based on a modified stepwise approach, (2) verifying the BLUE assumptions, and (3) validating the prediction performance of the regression model. The resulting MLR model only contains the most-relevant input variables while also fulfilling the BLUE assumptions. By utilizing statistical inference techniques, the MLR model also produces reliable range estimates around its point-value prediction according to a particular confidence level. To illustrate the application procedure of the proposed framework, a data set intended for workability control of ready-mixed concrete from the University of California, Irvine (UCI) machine learning repository is used. A practical case study based on a real-world bridge construction project is provided to further demonstrate the application of the proposed methodology in modeling the precast span installation cycle-time.