Upgrade to Pro — share decks privately, control downloads, hide ads and more …

We Love Speed: Understanding Cognitive Biases in Performance Measurement

We Love Speed: Understanding Cognitive Biases in Performance Measurement

When measuring web performance, we often try to get a single number that we can trend over time. This may be the median page load time, hero image time, page speed score, or core web vitals score. But is it really that simple?

Users seldom visit just a single page on a site, so how do we account for varying performance across multiple pages? How do we tell which page’s performance impacts the overall user experience? How do various cognitive biases affect the user’s perception of our site’s performance?

As developers and data analysts, we have our own biases that affect how we look at the data and which problems we end up trying to solve. Often our measurements themselves may be affected by our confirmation bias.

This talk is targeted at individuals who want to understand the business impact of their site’s performance, and how biases in data can affect that.

In this talk, we’ll go into different biases that may affect user perception as well as our ability to measure that perception, and ways in which to identify if our data exhibits these patterns.

References:
Ebbinghaus, Hermann (1913). On memory: A contribution to experimental psychology
Kahneman, Daniel (2000). "Evaluation by moments, past and future"
Baumeister, Roy F.; Finkenauer, Catrin; Vohs, Kathleen D. (2001). "Bad is stronger than good"
Staw, Barry M. (1997). "The escalation of commitment: An update and appraisal"
Arkes, Hal R.; Ayton, Peter (1999). "The sunk cost and Concorde effects"
The impact of network speed on emotional engagement
Ericsson ConsumerLab neuro research 2015
Wikipedia Paper on User Satisfaction v/s Performance
Toward a more civilized design: studying the effects of computers that apologize
The fastest way to pinpoint frustrating user experiences

Philip Tellis

May 11, 2023
Tweet

More Decks by Philip Tellis

Other Decks in Technology

Transcript

  1. Understanding Cognitive Biases in Performance Measurement Finding the factors that

    lead to abandonment https://speakerdeck.com/bluesmoon/we-love-speed-understanding-cognitive-biases-in-performance-measurement
  2. Philip Tellis Principal RUM Distiller @ Akamai • Analyses real

    user performance data from mPulse • Author of the OpenSource boomerang RUM library twitter:@bluesmoon ⦿ github:@bluesmoon speakerdeck:@bluesmoon
  3. BIAS is an Expectation Our Journey Today... ★ Understanding Cognitive

    Biases ★ Signs of cognitive biases in browsing data ★ What can we do?
  4. Bias stems from experience – It’s Normal • Helps us

    learn Perceptual/Sensory Dissonance • Keeps us safe Safety Bias, Loss Aversion, Negativity Bias • Find our people Similarity Bias, Proximity Bias Boston Shipyard Artist’s Community
  5. Cognitive Biases • Similarity Bias • Expedience Bias • Experience

    Bias • Proximity Bias • Safety Bias • Serial-position effect • False memory • Duration neglect • Peak–end rule • Negativity bias • Escalation of commitment • Loss aversion • Zero-risk bias • Next-in-line effect • Misattribution of memory • Sunk cost • Levels-of-processing • Spacing effect
  6. Cognitive Biases - Related to Performance on the Web •

    Similarity Bias • Expedience Bias • Experience Bias • Proximity Bias • Safety Bias • Serial-position effect • False memory • Duration neglect • Peak–end rule • Negativity bias • Escalation of commitment • Loss aversion • Zero-risk bias • Next-in-line effect • Misattribution of memory • Sunk cost • Levels-of-processing • Spacing effect
  7. Cognitive Biases - This Talk • Similarity Bias • Expedience

    Bias • Experience Bias • Proximity Bias • Safety Bias • Serial-position effect • False memory • Duration neglect • Peak–end rule • Negativity bias • Escalation of commitment • Loss aversion • Zero-risk bias • Next-in-line effect • Misattribution of memory • Sunk cost • Levels-of-processing • Spacing effect
  8. Pause Statistique A 500ms connection speed delay resulted in up

    to a 26% increase in peak frustration and up to an 8% decrease in engagement. Tammy Everts – The impact of network speed on emotional engagement
  9. some definitions Bounce Rate: Percentage of users on the site

    who leave after viewing one page. Retention Rate: Percentage of users on a particular page who remain on the site for at least one more page view. Conversion Rate: Percentage of users on the site who complete a goal or particular task. Goal: A task like a conversion, purchase, visiting a particular page, or viewing a certain number of pages. Frustration Index: A metric derived from multiple timers on a page that correlates with user frustration during page load.
  10. Serial-Position Effect …is the tendency of a person to recall

    the first and last items in a series best, and the middle items worst. Ebbinghaus, Hermann (1913). On memory: A contribution to experimental psychology
  11. Serial-Position Effect …is the tendency of a person to recall

    the first and last items in a series best, and the middle items worst. • Retention rate might be a function of the first and latest pages • The recency effect suggests that the latest page has a higher weight Ebbinghaus, Hermann (1913). On memory: A contribution to experimental psychology
  12. Peak-End Rule People judge an experience largely based on how

    they felt at its peak & at its end, rather than the sum or average of every moment of the experience. Kahneman, Daniel (2000). "Evaluation by moments, past and future"
  13. People judge an experience largely based on how they felt

    at its peak & at its end, rather than the sum or average of every moment of the experience. • Retention rate depends on the best/worst and latest performance • Conversion rate depends on the best/worst performance and that of the page just before the conversion Peak-End Rule Kahneman, Daniel (2000). "Evaluation by moments, past and future"
  14. Negativity Bias Even when of equal intensity, things of a

    more negative nature have a greater effect on one's psychological state and processes than neutral or positive things. Baumeister, Roy F .; Finkenauer, Catrin; Vohs, Kathleen D. (2001). "Bad is stronger than good"
  15. Negativity Bias Even when of equal intensity, things of a

    more negative nature have a greater effect on one's psychological state and processes than neutral or positive things. • Conversion rate should correlate with the ratio, or average of worst experience to best experience. • Active Listening can confound the results Baumeister, Roy F .; Finkenauer, Catrin; Vohs, Kathleen D. (2001). "Bad is stronger than good"
  16. Escalation of Commitment / Sunk Cost An individual or group

    facing increasingly negative outcomes continue the behavior instead of altering course. A greater tendency to continue an endeavor once an investment in money, effort, or time has been made. Staw, Barry M. (1997). "The escalation of commitment: An update and appraisal" Arkes, Hal R.; Ayton, Peter (1999). "The sunk cost and Concorde effects"
  17. Escalation of Commitment / Sunk Cost An individual or group

    facing increasingly negative outcomes continue the behavior instead of altering course. A greater tendency to continue an endeavor once an investment in money, effort, or time has been made. • High session length for really bad performing sessions • Retention/conversion rate increases as session length increases Staw, Barry M. (1997). "The escalation of commitment: An update and appraisal" Arkes, Hal R.; Ayton, Peter (1999). "The sunk cost and Concorde effects"
  18. Hypotheses… • The most recent experience is very impactful. •

    The best and/or worst experiences are impactful. • The first experience may be impactful. • The amount of time someone stays on the site is impactful. Pacific Islander Navigation Map, Museum of Fine Arts, Boston https://www.flickr.com/photos/bluesmoon/1266590108/
  19. Pause Statistique Wikipedia found that a 4% temporary improvement to

    page load time resulted in an equally temporary 1% increase in user satisfaction. Wiki Research: Analyzing Wikipedia Users’ Perceived Quality Of Experience
  20. • Collection: Real user performance data collected with boomerang •

    Sessions: Anonymous session ID attached to continuous sessions; discarded after 30 minutes of inactivity. Limited to sessions of 30 pages or fewer. • Samples: Analysis was done across multiple websites with millions of data points each. • Timers: We looked at Page Load Time (PLT), Time to Interactive (TTI), Largest Contentful Paint (LCP) and Frustration Index for Full Page as well as Single Page Apps. Notes about the Data
  21. First, Last, Fastest, Slowest • There is a strong negative

    correlation between conversion rate and the performance of the first page. 3.5% @ 1.8s 0.8% @ 18s 1.6% @ 9s
  22. First, Last, Fastest, Slowest • There is a strong negative

    correlation between conversion rate and the performance of the first page – this is related to bounce rate. 40% @ 1.1s 60% @ 18s 50% @ 9s
  23. First, Last, Fastest, Slowest • There is a strong negative

    correlation between conversion rate and the performance of the first page – this is related to bounce rate. • The last page distribution has strong drop after an initial peak. The two peaks are for XHR & Full Page. 13.5% @ 300ms 0.4% @ 18s 1% @ 5.5s
  24. First, Last, Fastest, Slowest • There is a strong negative

    correlation between conversion rate and the performance of the first page – this is related to bounce rate.. • The last page distribution has strong drop after an initial peak. The two peaks are for XHR & Full Page. • The fastest page has to be really fast. 10.5% @ 500ms 0.4% @ 9s 1% @ 5s
  25. Conversions x First, Last, Fastest, Slowest • There is a

    strong negative correlation between conversion rate and the performance of the first page – this is related to bounce rate.. • The last page distribution has strong drop after an initial peak. The two peaks are for XHR & Full Page. • The fastest page has to be really fast. Too slow, and users bounce. • Correlation with the slowest page is a little weird…
  26. 0.5% @ 1s 3.4% @ 19s 3% @ 5s •

    It seems that conversions increase as performance gets worse • It turns out that a slow experience is part of the conversion flow. • The low conversion rate on the left is a result of bounces. Very fast pages are typically caused by JavaScript errors resulting in a mostly blank page. (we see the same when the fastest page is under 100ms) Is Slower Better?
  27. looking at the 2nd Slowest Instead… 0.5% @ 1s 1%

    @ 19s 4.8% @ 2s 1.9% @ 6s 1.3% @ 12s
  28. Conversions x First, Last, Fastest, Slowest • There is a

    strong negative correlation between conversion rate and the performance of the first page. • The last page distribution has strong drop after an initial peak. The two peaks are for XHR & Full Page. • The fastest page has to be really fast. Too slow, and users bounce. • The slowest page doesn’t matter, but you cannot have too many slow pages.
  29. Retention Rate x First, Last, Fastest, Slowest • Retention Rate

    of a page varies based on the page. • For Homepages and other Landing pages, the performance of the first page appears to be the biggest indicator of retention. • For Product Detail, Category, and Search Results Pages, it’s a combination of the fastest & latest, and sometimes the first page. • The worst and second worst performing pages do not have an impact on retention.
  30. Negativity Bias • To determine if negativity bias is in

    play, we look at combinations of the best and 2nd worst performing pages. • The ratio (worst/best) has a strong negative correlation with conversions. • The geometric mean has a high, narrow peak. • A heatmap shows low tolerance for deviations in the fastest load time, and inverse dependence between the fastest and slowest times. Ratio of Slowest to Fastest Geometric MEAN of Slowest & Fastest Fastest → ← Slowest 1 10 20 30 40 50 60 0 10 20 30 40 50 0 0.4 0.7 1.0 1.3 1.6 1.9 2.2 2.5
  31. Negativity Bias 1 10 20 30 40 50 60 0

    0.4 0.7 1.0 1.3 1.6 1.9 2.2 2.5 • We have a practical lower bound on the fastest page • We have a tolerable upper bound on the fastest page • Slow pages are tolerated only when paired with a fast page that’s at least 15x faster. • This results in an upper bound on the slowest page. Fastest → 0 10 20 30 40 50 ← Slowest
  32. A greater tendency to continue an endeavor once an investment

    in money, effort, or time has been made. Escalation of Commitment / Sunk Cost
  33. A greater tendency to continue an endeavor once an investment

    in money, effort, or time has been made. 29% after 30 0.6% after 5 pages 7.1% after 10 21% after 20 Pages -> Conversion Rate -> Escalation of Commitment / Sunk Cost
  34. Looking across Load Times… 1 10 20 30 40 50

    60 70 80 90 100 110 115 0 10 20 30 40 50 0.1s 2s 4s 6s 8s 10s 15s 20s 25s 30s Load Time -> <- Number 0f Pages Conversion Rate
  35. Pause Statistique The average rise in mobile users' heart rates

    caused by delayed web pages — equivalent to the anxiety of watching a horror movie alone. Ericsson ConsumerLab neuro research 2015 38%
  36. The slowest page in a session should be no more

    than 15x the latency of the fastest page.
  37. Acknowledging when you didn’t meet the user’s expectations can alleviate

    negative perceptions. Practice Active Listening https://affect.media.mit.edu/pdfs/02.klein-moon-picard.pdf https://uxdesign.cc/the-fastest-way-to-pinpoint-frustrating-user-experiences-1f8b95bc94aa https://doi.org/10.1016/j.ijhcs.2004.01.002 https://www.sciencedirect.com/science/article/abs/pii/S1071581904000060?via%3Dihub
  38. A fast page increase pages per session which in turn

    increase the likelihood of a conversion.
  39. Pause Statistique Users are most patient when using the web

    from the office and least patient when using their phones. Median Lethal Frustration Index study in mPulse data
  40. Cognitive Biases – Developer Edition • Amdahl's Law Assuming every

    millisecond is the same. • Outcome Bias Choosing data that confirms past outcomes. • Survivorship Bias Assuming what we’ve measured is all there is. • Selection Bias Choosing dimensions based on our instincts. • Pareidolia Preferring data that renders interesting shapes. • Insensitivity to Sample Size Forgetting that smaller samples have larger variance. • Clustering Illusion Seeing patterns in small samples where none exist. • Confirmation Bias Choosing data that confirms our pre-existing beliefs.
  41. Ignoring Amdahl’s Law You may have read reports that say

    something like: “every 100ms decrease in homepage load time worked out to a 1% increase in conversion” Citation redacted to protect the innocent
  42. Survivorship Bias • In 2012, Youtube made their site lighter

    but aggregate performance got worse. • It turns out that new users who previously could not access the site were now coming in at the long tail. • The site appeared slower in aggregate, but the number of users who could use it had gone up. Chris Zacharias: Page Weight Matters.
  43. Insensitivity To Sample Size We often get questions like: “Why

    is performance on tablets worse than performance on mobile devices?” It turns out that mobile generally has 50x the amount of traffic than tablets. That results in far less variance in the data. A customer recently asked me this question.
  44. Anscombe’s Quartet Anscombe's Quartet Frank Anscombe Plot of Anscombe's Quartet

    by Schutz & Avenue • 4 data sets with the same summary statistics: ◦ 𝜇 x = 9, 𝜇 y = 7.5 ◦ s x 2 = 11, s y 2 = 4.125 ◦ 𝜌 x,y = 0.816 ◦ Linear Regression Line: y=3 ◦ ℝ2 = 0.67 • Anscombe’s Quartet shows us why it’s important to visualize data and not just look at summary stats
  45. References • Ebbinghaus, Hermann (1913). On memory: A contribution to

    experimental psychology • Kahneman, Daniel (2000). "Evaluation by moments, past and future" • Baumeister, Roy F .; Finkenauer, Catrin; Vohs, Kathleen D. (2001). "Bad is stronger than good" • Staw, Barry M. (1997). "The escalation of commitment: An update and appraisal" • Arkes, Hal R.; Ayton, Peter (1999). "The sunk cost and Concorde effects" • The impact of network speed on emotional engagement • Ericsson ConsumerLab neuro research 2015 • Wikipedia Paper on User Satisfaction v/s Performance • Toward a more civilized design: studying the effects of computers that apologize • The fastest way to pinpoint frustrating user experiences • Serial-position effect • Peak–end rule • Negativity bias • Escalation of commitment / Sunk cost • Levels-of-processing • Amdahl's Law • Outcome Bias • Survivorship Bias • Selection Bias • Pareidolia • Insensitivity to Sample Size • Clustering Illusion • Confirmation Bias • Time Saving Bias