Wheeler & Lyday vs R&R
The key points connecting and separating the conventional R&R
approach and the Wheeler and Lyday approach can be summarised as follows:
Both approaches use the same data
collection plans: data gathered by one can be readily used for the other
(this has obvious implications for a software package like Mesa).
The key difference between the
two is that the R&R study attempts to break up the measurement
error into two, and sometimes more, components.
The main two are equipment variation (EV) and appraiser variation
contrast, in the W-L approach, measurement error variation (MEV) is idealised as being due just to test-retest error.
All other sources are identifiable and removable, at least in
R&R studies, the MEV is referred to as the R&R variation.
Furthermore, the R&R variation is broken up into equipment
variation (EV), and appraiser variation (AV) using the sums of squares rule.
See the R&R geometry description for an illustration of this.
The main problem with this is
that it greatly complicates the analysis and interpretation stages,
especially when, as may be common, the variation does not actually split in
a simple additive way over these components because of interaction effects,
e.g. between operator and equipment.
problem of interaction effects leads proponents of the R&R approach to
include the Analysis of Variance (ANOVA) in their options. ANOVA
is an excellent and interesting way to analyse data but is probably best
left to the statistician or statistically advanced engineer.
We do not need it for MPE.
If important interaction effects are present, we can expect to see
evidence for them in the plots we use - for example by departures from
parallelism between plots of parts means found by each operator.
The viewpoint which has guided
the development of MESA is not against R&R studies per se - they have
the same goals in mind as the Wheeler & Lyday/MPE approach, but rather that there is no
obvious benefit from going to the trouble involved.
Wheeler (1992) also points to
technical deficiencies in the R&R approach.
Sources of variation are commonly expressed as a percentage of either
tolerance bands or of total variation, but this is done without squaring the
values involved, and hence these “percentages” can be very misleading
and will not, of course, add up to 100%.
The percentages used in R%R studies can be even more
misleading when expressed as a percentage of a tolerance spread.
The key point to note about the
Wheeler & Lyday approach is the pivotal role of data displays involving the test-retest
variation. Since other sources of variation are more readily avoidable and
are always bad news, the approach basically says “let us assume that the
only measurement error is test-retest and see what stands out”.
If in fact there are important amounts of other error variation
sources present, such as AV, then they will be detected by violation of the
limits on one or more of the charts produced to test for consistency and
* Acknowledgments to John Shade of Good
Decision Ltd for the basis of the above text.