red tree

Identifying Root Causes

During the measure phase we use control charts to identify the variability in the output and to fix assignable causes.

The challenge for the analyse phase is to discover which of the inputs to the process are the dominant causes of the overall process variation and are therefore targets for potential improvements to the process.

It is helpful here to remember the meaning of the “root” in “root cause”. Roots are found under the ground, which means that you can expect to have to dig deep to seek out root causes. The obvious and easy answers are unlikely to be root causes, which means that fixes are likely to be superficial and disappoint.

Root causes are specific, actionable, and have a significant impact.

Cause and Effect Diagram

The classical method for generating ideas about root causes is to create a cause and effect diagram. This is merely a systematic way of surfacing what the participants believe to be candidates for the reasons for the wiggliness - variation - in the process factors, and can be mapped onto the same set of generic inputs as shown in the process schematic.

The cause and effect diagram is thus an extension of the generic process concept, which allows us to represent potential sources of variation for team discussion and evaluation.


Normal brainstorm rules apply here - we are interested in generating a visible record of ideas, not in producing an elegant and refined model. With a team of 4 - 6 people, one would be looking to generate 80 - 100 ideas at this point - it is important not just to settle on the immediate and obvious ideas. Ensure that all factors are considered. Especially beware of cause and effect diagrams that are loaded towards the “people” category. 

Testing our Theories

Karl Popper’s view of knowledge is that we start with a theory, which can come in any shape or form, and that we test the theory according to what predictions it makes or implies. (If the idea does not make testable predications, then it is not a very useful theory.)

If the test does not support the theory, then we need to improve our theory (or the way we are testing it). If the test does not disprove the theory, we may use the theory - until something happens to disprove, or until a better theory comes along. We can never be sure; we can only have a best working model with what we currently believe.

So if for example one of the ideas for sources of variation is “lack of adequate training for front line staff” the prediction would be that better training would lead to measurably better outcomes, or that there is some measurable difference between staff performance that maps on to the amount and quality of training they have received. If neither of these proves to be the case, then the theory is not an adequate explanation. (If I select very short individuals for my basketball team, no amount of training will make them competitive in the big league!)

Because testing of this nature tends to be time consuming and expensive, it is not feasible to test all of the ideas generated. Generally, follow up discussions will surface a few plausible candidates for further investigation and analysis, focussing on the criteria of being specific, actionable, and thought to have a significant impact.

Stratification and Correlation

As discussed in Data Collection, stratification means looking for specific differences between input conditions that correspond to differences in outputs. Typical dimensions for stratification include, e.g. between locations, teams, product types, or time of day.


  • If we are tracking defects, the most common tool to use with stratification is the Pareto chart, which shows the number of defects associated with each value of the stratification dimension, sorted into descending order by frequency and / or by impact. The reasons can come from anywhere on the cause and effect diagram, e.g. we could have both "truck breakdown" and "poor weather" as reasons, although one would be equipment and the other environment on the C&E.
  • If we are tracking variables, e.g. time to process, the most common tool to use with stratification is the Box Plot. The types will usually be different categories for a single source of variation on the cause and effect diagram, e.g. type of query, day of week, location of office. We are looking for the boxes to have no overlap as an indicator of difference. In this case, there appears to be no difference between P and Q, P and R, and Q and S, but there may be differences between P and S, Q and R, and R and S that are worth following up.
  • If we want to show the relationship between a variable input and a variable output, the appropriate tool is the Scatter plot. Note that if the points fall in a perfectly straight line, this implies that there was no other source of variation in the outcomes, which is implausible.

Each of these visual representations have small print and caveats, and may be further evaluated using appropriate formal statistical tests to determine the likelihood that they arose by chance, or conversely our confidence that the variations in the outputs are genuinely related to the differences in the inputs.

The conventional six sigma approach is to use hypothesis testing to ascertain the probability that the difference between the scenarios could arise by chance alone.