DOE: An Oft-ignored Continuous Improvement Tool

Design of Experiments, or DOE, spun out from industrial production where process improvements can only be performed without operational stoppage. DOE allows operators to identify the combined effects of process controls and their interaction. This analytical tool found use in the service industry because of these features. To demonstrate the use of this tool, we revisit Midwest Social Services. This fictional agency applied DOE to optimize the computer model of its cash assistance unit.

Darcy Harkin, income maintenance services director, began her presentation to the administrator and other area managers. "I was intimidated by this analytical tool when I received training for my first process improvement project. DOE was surprisingly user friendly. The team not only reduced process variables to an essential few but also measured the effects of their interaction. The latter information would not have been available if we had focused on one variable at a time. In our current project, staff were similarly apprehensive but decided to use DOE after an hour of training.

The Results

The team downloaded an Excel waiting-line computer model from the internet. We initially modified the parameters so that the model mirrored our process. The client waits before the Comprehensive Intake and Assessment section and before the Cash Assistance section. The total time a client stays in this system is our performance criteria.

In summary, the project team inferred that by improving the Cash Assistance Service cycle time by 6 minutes, we could reduce the time the client stays in the system by an hour and 22 minutes. Close to the end of the month, a client drops-in into the service center every hour. The client stays in the system for about a day while the application for cash assistance is being processed.

The team was satisfied with their results. DOE gave staff confidence in the improvement of the actual intake and cash assistance processes.

The Design

The Excel waiting-line computer model modified to resemble our processes keeps track of the time it takes all virtual clients to go through the service system. With an average arrival rate of one per hour, the walk-in client waits for a turn for an interview and assessment of the need for cash assistance as well as any other human services need. Once the intake process has been completed, the client waits for a turn in the evaluation of employment opportunities and on to the cashier’s window. Our intake process cycle time, or variable A, averages 0.80 hours per client which we coded as 1. We assumed a realistic improved scenario where the intake process operated at 0.60 hours per client. This was coded as -1. Currently, our cash assistance process cycle time, or variable B, averages 0.85 hours per client and was coded as 1. The improved scenario was set at an attainable level of 0.65 hours per client and was coded as -1.

Well-designed experiments with two variables require four combinations: (1) ab, where the two variable were at their low levels variables, (2) aB, where variable A was low and variable B was high, (3) Ab, where variable A was high and variable B was low, and (4) AB, both variables were at their high levels. The interaction variable was coded as the product of the codes of the two variables.

For example, while referring to the first line of the results table above, the experimental run on combination ab, gave an average time to service a client of 3.8 hours. The comprehensive intake process was coded as -1 and so was the cash assistance process coded as -1. The interaction variable was at 1 the product of -1 and -1. The design was trice replicated.

The Analysis

The contents of the table above were entered in an Excel worksheet and selected as input to Excel’s regression analysis add-on. The coefficients of the resulting regression equation indicate the significant effects of the two operating variables and the insignificant effect of the interaction term.

The p values of the intercept coefficient, as well as the coefficients of intake and cash assistance variables, were all below the threshold of p = 0.05, while the coefficient of the interaction term was insignificant with p = 0.46. Substituting the data at the first row of the table above into the equation calculates the time the client is in the system at 3.48 hours, or (5.73 + (0.88 x -1) + (1.37 x -1). This number derived from the equation was closer to the experimental value of 3.8 hours in comparison to the value of 3.27 hours when the interaction term was included in the calculation.

By Mgmtlaboratory.com staff and Noel Jagolino, management consultant. 2021