The evening before the space shuttle Challenger explosion, scientists at NASA caught what they thought was a potentially catastrophic risk with the o-rings considering the unusually cold temperature expected for the morning’s launch. They brought the issue to management attention but failed to influence the final decision enough to stop the launch. Your failure to influence may not cost lives but it could be “catastrophic” for your business. This talk presents my top tips for using data to influence others toward better decisions.
Learning outcomes:
- The do's and don'ts of visualization
- How others lie with data
- What makes an effective dashboard
- How to communicate uncertainty
16. @LMaccherone @TheAgileCraft
Truths about cognitive bias
1. Very few people are immune to it.
2. We all think that we are part of that
small group.
3. You can be trained to get much, much better.
Douglass Hubbard – How to Measure Anything
4. We do a first-fit pattern match. Not a best-fit pattern
match. And we only use about 5% of the information
to do the matching.
5. We evolved to be this way (survival trait).
19. @LMaccherone @TheAgileCraft
Trained/Calibrated
Untrained/Uncalibrated
Statistical Error
“Ideal” Confidence
30%
40%
50%
60%
70%
80%
90%
100%
50% 60% 80% 90% 100%
25
75 71 65 58
21
17
68 152
65
45
21
70%
Assessed Chance Of Being Correct
PercentCorrect
99 # of Responses
We are overconfident when assessing our own uncertainty
But, training can “calibrate” people so that of all the times they
say they are X% confident, they will be right X% of the time
Copyright HDR 2007
dwhubbard@hubbardresearch.com
20. @LMaccherone @TheAgileCraft
Equivalent Bet calibration
What year did Newton published the Universal Laws of
Gravitation?
Pick year range that you are 90% certain it would fall within.
Win $1,000:
1. It is within your range; or
2. You spin this wheel and it lands green
Adjust your range until 1 and 2 seem equal.
Even pretending to bet money works.
90%
10%
22. @LMaccherone @TheAgileCraft
How to avoid cognitive bias in decision making
Don't focus on consensus.
Ritual dissent is a much more successful approach.
“But that doesn’t explain _______”.
An FBI agent knew that some folks were being trained to fly
but not take off and land.
Assign someone the role of devil’s advocate.
Israel’s 10th man.
In other words… Really consider the other ALTERNATIVES
25. @LMaccherone @TheAgileCraft
2. Shows causality, or is at least informed by it
The primary chart used by the
NASA scientists showed O-ring
failure indicators by launch date.
Tufte's alternative shows the same
data by the critical factor,
temperature.
The fateful shuttle launch occurred
at 31 degree. Tufte's visualization
makes it obvious that there is great
risk for any launch at temperatures
below 66 degrees.
27. @LMaccherone @TheAgileCraft
4. Is credible
Calculations explained
Sources
Assumptions
Who (name drop?)
Drill-down
How?
Etc.
28. @LMaccherone @TheAgileCraft
5. Has business value
(or value in it’s social context)
The ODIM framework
O U T C O M E
D E C I S I O N
I N S I G H T
M E A S U R E
THINK
EFFECT
like Vic Basili’s
Goal-Question-Metric (GQM)
but without
ISO/IEC 15939 baggage
29. @LMaccherone @TheAgileCraft
6. Shows differences easily (1)
aka:
Save the “pie” for dessert Credit:
• Stephen Few (Perceptual Edge)
• http://www.perceptualedge.com/ar
ticles/08-21-07.pdf
30. @LMaccherone @TheAgileCraft
6. Shows differences easily (2)
Can you compare the market share from one year to the next?
Quickly: Which two companies are growing share the fastest?
One pie chart is bad. Multiple pie charts are worse!!!
37. @LMaccherone @TheAgileCraft
Top 10 criteria for great visualization
1. Answers the question,
"Compared with what?”
(SO What?)
2. Shows causality, or is at least
informed by it.
(NOW WHAT?)
3. Tells a story with whatever it
takes.
4. Is credible.
5. Has business value or impact in
its social context.
6. Shows
differences
easily.
7. Allows you to see the forest
AND the trees.
8. Informs along multiple
dimensions.
9. Leaves in the numbers where
possible.
10. Leaves out glitter.
Credits:
• Edward Tufte
• Stephen Few
• Gestalt
(School of Psychology)
11. Uses good visual grammar
38. @LMaccherone @TheAgileCraft
11. Good visual grammar (1)
When to use a straight line series?
1. Cumulative data
(later data points include
items from earlier data points)
example: Burn or Scope
series
2. Benchmark for column series
3. Connected by “ordinal inertia”
(ordinal usually = temporal)
However, spline is usually
what you want in this
instance.
39. @LMaccherone @TheAgileCraft
11. Good visual grammar (2)
When to use a dashed line?
1. “Ideal” or reference
2. Forecast
3. Regression or fit
line
40. @LMaccherone @TheAgileCraft
11. Good visual grammar (3)
When to use a column (vertical bar)
series?
1. Counts/sums independent of
neighbors
example: Throughput or
velocity chart
41. @LMaccherone @TheAgileCraft
11. Good visual grammar (4)
When to use a bar (horizontal)
series?
(Rarely)
1. The x-axis variable is causal?
2. In a table
3. Stylistic reasons
• Length of labels
• Two-variable display
42. @LMaccherone @TheAgileCraft
11. Good visual grammar (5)
When to use stacked area?
1. Multiple cumulative series
(the ordinal sum means
something)
example: a cumulative
flow diagram (CFD)
44. @LMaccherone @TheAgileCraft
11. Good visual grammar (7)
When to use a single spline series?
1. Filling in for missing granularity
2. Rolling average
48. @LMaccherone @TheAgileCraft
Other ways to lie with statistics
Sampling bias
Self selection bias
Leading question bias
Social desirability bias
Median vs Mean
The big “zoom-in”
All together now…
Correlation does not necessarily mean causation
51. @LMaccherone @TheAgileCraft
Quality of decision depends upon:
1. alternatives considered, and
2. models used to forecast the
outcome of those alternatives.
53. @LMaccherone @TheAgileCraft
1. Different Models
2. Different Values
3. Different Risk Tolerance
Why do people disagree?
favor different
alternatives
Fear-based decision
making
54. @LMaccherone @TheAgileCraft
Models and Values
Models calculate probability in terms of proxy variables
Values translate those probabilities into money
Different models example:
Joe forecasts that alternative A will make the most money
Sally forecasts that alternative B will make the most money
Different values example:
Betty favors the alternative with higher quality
George favors the alternative that will get to market faster
61. @LMaccherone @TheAgileCraft
Example model and “values” system
Team performance feedback
Duplicate great
performance
Bend the curve early on
declining performance
Decide what process
improvements to
implement next
Dimensions of performance
1. Productivity
2. Predictability
3. Time-to-market
4. Responsiveness
5. Defect-free-ness
6. Customer satisfaction
7. Employee engagement
8. Build-the-right-thing
9. Code quality
62. @LMaccherone @TheAgileCraft
Economic value weighting of team performance
Proxy variables translated to something unit-less
Weight (adding to 100%) applied to chosen dimensions
Examples:
Medical device manufacturer values QUALITY
60% for Quality
10% each for Productivity, Predictability, Time-to-market, and
Responsiveness
Mobile game team values TIME-TO-MARKET
40% for Time-to-market
20% each for Productivity and Responsiveness
10% each Quality and Predictability
65. @LMaccherone @TheAgileCraft
The rider and the elephant
Direct the rider
Motivate the
elephant
Shape the path
Jonathan Haidt
The Happiness Hypothesis
(also mentioned in Switch)
67. @LMaccherone @TheAgileCraft
… but for those brave enough to journey
into the dangerous world of
agile measurement
there are great riches to be had.
The trick is to slay or avoid the dragons.
68. @LMaccherone @TheAgileCraft
The Dragons of Agile Measurement
If you do metrics wrong, you will harm your agile transformation
1. Dragon: Measurement as a lever
Slayer: Measurement as feedback
2. Dragon: Unbalanced metrics
Slayer: 1 each for Do it
fast/right/on-time, and Keep doing it
3. Dragon: Metrics can replace
thinking
Slayer: Metrics compliment
thinking
4. Dragon: Expensive metrics
Slayer: 1st work with the data you
are already passively gathering
5. Dragon: Using a convenient
metric
Slayer: Outcomes
Decisions Insight Metric
(ODIM)
6. Dragon: Bad analysis
Slayer: Simple stats and
simulation
7. Dragon: Single outcome
forecasts
Slayer: Forecasts w/
probability
69. @LMaccherone @TheAgileCraft
Now what?
Enter feedback in Sched App
Come to the AgileCraft booth:
• Questions answered
• Demo of how AgileCraft gets you
from “What?” to “So what?” and
then “NOW WHAT?”
70. @LMaccherone @TheAgileCraft
When you come to a
fork in the road…
take it!
~Yogi Berra
What?
the metrics and analysis
So what?
how it compares/trends
what it means
NOW WHAT?
every decision is a forecast
71. @LMaccherone @TheAgileCraft
Now what?
Enter feedback in Sched App
Come to the AgileCraft booth:
• Questions answered
• Demo of how AgileCraft gets you
from “What?” to “So what?” and
then “NOW WHAT?”
Notas del editor
The bottom illustration is a re-visioned version of the Tufte chart, which itself is just black and white and not quite as clear as this one.
The bottom illustration is a re-visioned version of the Tufte chart, which itself is just black and white and not quite as clear as this one.
READ THIS
1687
READ THIS
1687
READ THIS
1687
Tell the story about the creation of Insights and “This is crap!”
READ THIS
1687
The NYT used Bellman Equation. We’ll use the expression at the core of that equation…
Let’s say that getting a 1st down is worth 2 “points contributed” and the chance of getting a 1st down is 50%. The weighted value of this alternative is 1 point contributed.
Now, let’s say that the value of putting the team back at their 20 yard line vs giving them the ball right here on say our own 40 yard line is 1 point contributed and the probability of that happening is 80% (could have a blocked punt or big run back). It’s weighed value is then 0.8 points contributed.
For any given roll of the dice, the punt could come out better, but statistically, over the long haul, going for it in this situation is the better alternative.
The difference between 50% and 80% is significant but doesn’t explain why coaches rarely choose to go for it on fourth down. The difference is that when they don’t make it, they are shown to have been wrong.
Now a real example that shows two different risk profiles in a business scenario
Every decision is a forecast. You are now armed with ways to avoid cognitive bias, creating models that produce probabilities, translating proxy variables into money using “values”. Make better decisions.