The Psychology of Quality and More

A Toolbook for Quality Improvement and Problem Solving (contents)

Design of Experiments (DOE): How to do it

The Quality Toolbook > Design of Experiments (DOE) > How to do it

When to use it | How to understand it | Example | How to do it | Practical variations

<-- Previous | Next -->

How to do it

1. Identify the objective of using experiments. Typically, this may be one of:

• Finding true causes of problems.
• Finding how causes interact.
• Finding the best solution to a problem.
• Testing a solution to ensure it has no undesirable side-effects.

2. Define what is to be measured to show the results of the experiment. It will make the experiment much easier if this is a single measurement that can be easily and accurately performed. Be clear about other factors, such as who, when and how the measurement will be made.

3. Identify the factors that are to be controlled during the experiment. Consider all things that may affect the measured result, then select those that are to be varied and those that are to be held steady or otherwise monitored. As with step 2, any measurements should be clearly defined. Examples of factors include price, dimensions, temperature, errors, brands of fertilizer, age ranges.

When there are many factors, reduce the list of those that are to be varied by selecting those known to affect the result and those whose effects are uncertain. It might also be appropriate to perform a series of experiments, starting with a smaller subset of 'high-probability' factors.

 

4. For each factor selected in step 3, identify the set of levels that the experiment must consider. This will typically be a small set of possible values such as: 20, 24 and 28° C; 'GroFast' and 'EasyGrow'; present and absent.

There will be fewer trials to perform and the subsequent analysis will be easier if very few levels of each factor are selected. Two levels are sufficient for many cases, as this will show whether changing the factor level changes the experimental result. Three or more levels can be used to check for a non-linear response.

Select the levels to be representative of the range of normal values. Thus they should be sufficiently separated to be able to identify changes over the normal operating range, but not so spread as to meet any boundary effects, such as where a liquid boils.

Ensure the factors can be controlled and measured at these levels. If they cannot be controlled, then it may be sufficient to measure them and sort them into ranges.

 



 

Fig. 1. Building an experiment

 

5. Select the actual trials to take place. There are a number of possible ways of doing this, depending on the type of experiment being performed. Some simple methods are described in the section on practical variations, below. Other more complex methods are described in the references at the end of the chapter..

The decision on how many trials to perform may include economic factors, such as time, effort and cost. For example, crash-testing of vehicles is expensive and time consuming and is impractical to do too often.

When trials are selected, check that they are balanced, with the different levels of each factor occurring the same number of times. Also check for orthogonality, with each pair of factors having each combination of levels occurring the same number of times (as illustrated above).

 

6. Perform the trials as planned. This may be a simple set of experiments or may require significant organization and funding. In any case, be careful in controlling factors at the selected levels and in measuring and recording results.

Consecutive trials should not have any chance of affecting one another; if this could happen, perform trials in random order.

Results may be recorded in a simple table, such as illustrated, which shows one trial per row, with levels and results on the same row. This will help analysis, as results may be visually correlated with the selected factor levels.

 

7. Analyze the results. A simple approach is to average and plot results for each factor, level and combination, as illustrated. More complex methods are given in the references.

Where there are more than two levels, this will result in lines through more than two points. If the lines are not straight, then this indicates a complex effect.

8. Act on the results. This will depend on the objectives from step 1, and thus may be one of:

• Eliminating what is now a known cause.
• Selecting the most effective solution to a problem.
• Acting to remove undesirable side-effects.

<-- Previous | Next -->