Issue and
Opportunity… Traditional improvement technology is heavily analytical.
Complexity and chaos sciences offer new analytical techniques for
teasing out more information about patterns in time series data. These
patterns in the data can lead to a deeper understanding of the dynamics
of the underlying complex adaptive system. These better understandings
might lead to better ideas for improvement. This could be a major
extension of analytical improvement techniques; potentially as powerful
and paradigm altering as Shewhart's control charts were in the 1930s.
New Understandings
from Time Series Data... "Magic" of time series is that even though
it represents only a single variable, the causal influence of other
variables are embedded in its dynamics (Taken's embedding theorem)...
We can look at time series data and see whether there are many or
few casual variables and get some indication as to what they are doing...
Example: study of teen pregnancy data in Texas indicates "pink noise"
allowing model that says that teen's decision is based on local peer
and family attributes.
Aggregating data
(e.g., over a quarter) loses some of the information in the time series...
Can never recapture dynamics that are quicker than your sampling (reporting)
rate.
An emerging
data analysis path might be…¨
- Identify
"special causes" in our data with control charts, or correlations
with traditional statistical tools. Understand and act upon.
- Now we are
left with what we have traditionally called "random variation."
(Common cause variation in control chart lingo). But complexity/chaos
studies show that even further distinction can be made... there
may be periodic and chaotic patterns... different "colors" of "noise"
(randomness) arise from different underlying generating mechanisms.
- Analyze "random"
variation by looking at power law structure and Hurst exponent.
When multiple causes exist and are related to one another in a multiplicative
(dependent) sense, a log-normal (power law) distribution arises.
When multiple causes exists and are related to one another in an
additive (independent) sense, a Normal distribution arises. Classic
statistics and QI often assume additive, independent multiple causes
(the Normal distribution).
- Depending
on the Hurst exponent found in analysis of the data, we can have
four colors of noise (what we have traditionally lumped all together
as "randomness"). (Note: Dooley has posted more detailed descriptions
of these different colors of noise.)
- White
noise is pure randomness, no memory in the system, truly no
telling what will happen next.
- Brown
noise is accumulation of white noise over time, classic Brownian
motion... What happens next is dependent on what happened before,
but no telling how far the next step will jump you from where
you are now.
- A time
series with Black noise shows a long term persistence in apparent
trends and cycles, but these patterns are non-linear in nature...
But there is a pattern that we can identify.
- A time
series with Pink noise tends to revert back to the mean more
often than a purely random pattern would... The sand pile model
of self-organizing criticality gives rise to pink noise.
Open question…
How do we make the leap from abstract concepts like "colored noise"
into practical insights into the real systems of health care that
we can play a part in influencing?
Can we predict
complex system behavior? and Can we ever be prescriptive about events
in a CAS?… We can predict within a range, but this might be good
enough to let us experiment, establish good enough vision, and state
min specs. We might also know enough to describe "basins of attraction;"
and again, this may be all we need to know to be practical.
Of course, we
can force a CAS to behave the way we want through extraordinary means
(example: prison camp), but practically speaking the best we can hope
to do is to influence behavior by changing the very context of the
CAS itself through 15% actions and butterfly effects. The metaphors
of managers as gardeners, and managers as the wizard behind the curtain
in Oz, were introduced to describe the approach.
Because we cannot
predict and be prescriptive... "observation may be the keenest sense
for managers to develop, the ability to postulate associations --
their greatest skill, and their ability to take risk in facilitating
the association -- their greatest attribute." We agreed that learning
to see pattern is key, but how do we do this practically?
Pattern is
not the same as "path"… We can describe areas of performance and
outcome that are possible through the pattern, but we cannot say what
precise sequence of outcomes will unfold. A potential area for additional
exploration involves using the information that we might have about
a complex pattern to say where the system will not ever go in it's
performance. We still need to work out the potential usefulness of
this knowledge.
Tales From
Past Work and Literature… We have Kevin Dooley's past work on
teenage births in Texas. Ary Goldberger's work also illustrates this
for a pure clinical issue. There is also literature from Priestmeyer
et. al. on the use of phase-plane plots for things like nursing staffing.
New Demonstration
Projects… Several avenues of analysis potentially exist…
- Medication
Errors. VHA-East has sent some data on medication errors to Kevin
Dooley. Recall that the idea was to illustrate power laws and self-organizing
criticality. Can small, no-harm errors be used to roughly predict
the future rate of occurrence of larger med errors (even with the
problem that Jacquie Byers raised about underreporting of med errors)?
Initial analysis suggests that power law distribution is confirmed
by the data. This leads to the conclusion that tracking near-miss
med errors could greatly enhance the power of data collection on
med errors.
- Patterns
in Census Data JB Collins has daily census data going back five
years at his place! Let's see if Kevin Dooley can find any interesting
complex patterns in that. At this point, it is unclear how this
will relate to QI, but let's first see what patterns and new insights
emerge from the analysis.
- Data Mining
Another possible demonstration project comes from Perry Pepper at
Chester County Hospital. Perry is pursuing complex data mining using
software called Clementine from a company called Integral Solutions
Ltd. Perry also mentioned contacting SMS to see if they were interested
in working on this with him.
- Gregg Bennett
may have some HBSI data.
- Kevin Dooley
has suggested that we might test the hypothesis that what has been
called "small area variation" might actually be nothing more than
fractal noise. This could be a major contribution to improvement
science in healthcare.
- We might
learn something about common healthcare CAS's by constructing a
simulation. This remains an open possibility. (Note: Kevin Dooley
is currently building a CAS simulation to study the effects of neighbor
influences on the adoption and success of TQM.)
It would be great
to get some waiting time data and some outcomes data to analyze. These
are common in QI projects and would, therefore, make powerful demonstrations.
Ultimate Goal
of This Line of Thinking… Extending traditional QI thinking beyond
the "lump" category of common cause random variation... i.e., learning
to see more patterns than we have with existing methods. The next
step in analytical QI comparable to Shewhart's introduction of the
control chart.
