Quasi-Experimental based on Cross-Section Historical Data: An Initial Alternative Framework for Analyzing Causality in Manufacturing Process

Abstract

The classic design of experiment (DoE) requires randomization within the treatments to ensure statistical independence and fulfill mathematical assumptions for causality investigation purposes. For some cases with difficulties in randomizing the treatment, a quasi-DoE becomes an alternative with all its weaknesses. Meanwhile, a non-random cross-section historical data from such a smart manufacturing should be considered as one of the available resources, and there are big challenges to retrieving hidden information within. This paper proposes an alternative framework to select observations from historical data and treat it as a quasi-experimental that meets a type of classic DoE, followed by performing statistical analysis to build an evaluation and interpretation.  As an initial validation of this framework, three factors historical data from a CNC milling process were recorded for a case study. The statistical analysis is successfully conducted by selecting an observational subset that matched a DoE design with satisfying its properties. It has been concluded that selecting the desired observations subset gives a similar interpretation to a classic DoE.