Interpreting Data Like a Pro - Dawn of the Data Age Lecture Series
- 1. Dawn of the Data Age Lecture Series Interpreting Data Like a Pro
- 2. Hi. I’m Luciano Pesci… Co-Founder & CEO, EMPERITAS ● A Services as a Subscription team of economists and data scientists delivering bi-weekly Customer Lifetime Value intelligence so our clients can beat their competitors for the most profitable customers. Founder & Director, Utah Community Research Group, Univ. of Utah ● Teach microeconomics, statistics, applied research & data analytics, & American economic history. ● Teach data science for Westminster and developed their 3-class MBA emphasis in data science. 2
- 3. Today’s Lecture Outline ● Teach you how to identify data types & context. ● Explain the right way to select analysis methods. ● Show you the core data interpretation skills. 3
- 4. 4DATA TYPES & CONTEXT
- 5. Defining Data Differently ● There are many ways to define data, each requires a different approach when utilizing it: ○ Origin - How it was created. ○ Totality - If it’s a sample or a census. ○ Scope - Whether it’s been captured over time. ○ Measurement - How it was quantified. 5
- 6. What’s The Origin Story? ● Understanding the origin of your data is key to grasping its context: ○ Experiments produce data with strong causal patterns but it’s costly to collect & analyze. ○ Survey data is easy to get, but it shows intent or attitude, not necessarily actual outcomes. ○ Observational data is mostly captured by machines and shows actual outcomes, but it’s very rigid. 6
- 7. What’s the Totality? ● If you have data on every possible unit in a population of interest, then it’s Census data. ● In most cases you’ll only have a Sample which can be used to infer patterns about the larger (unknowable) population. 7
- 8. Scoping Time? ● If your data contains different variables, all measured at the same time, it’s Cross-Sectional. ○ Most data that you encounter will be cross-sectional. ● If your data contains multiple measurements of the same variable over time, it’s Time Series. 8
- 9. Data Measurement? 9 ● All data fits into 4 basic types based on how it was measured: ○ Nominal & Ordinal = CATEGORICAL. ○ Interval & Ratio = CONTINUOUS. ● Identifying data types is a critical skill to develop. ○ Analysis selection & interpretation depend on it.
- 10. 10SELECTING THE RIGHT ANALYSIS
- 11. Categorical or Continuous? ● The biggest difference when selecting analysis is based on whether the data is categorical (nominal, ordinal) or continuous (interval, ratio). ○ So much of what you can or can’t do is determined by the data’s measurement type. ○ Time Series vs Cross-Sectional is another important distinction that radically changes your approach. 11
- 12. Looking for Differences ● Tests of difference, like comparing medians or means, are a good way to find unique subgroups within the data. ○ This should only be done with continuous data, though you can use categorical variables for subgrouping when testing for differences. 12
- 13. Looking for Similarities ● Measures of association (like correlation) are a good way to find patterns that move together in the data. ● While correlation doesn’t equal causation, theory can help you understand when correlations are likely to be real or not.* 13 *Source: www.tylervigen.com/spurious-correlations
- 14. 14CORE DATA INTERPRETATION SKILLS
- 15. Looking At It Both Ways ● You should look at tables & visualizations. ○ Each tells a unique part of the data’s story. ● 3 very specific things to find (when possible based on your data type): ○ Shape of the distribution ○ Center of the distribution ○ Spread of the distribution 15
- 16. Understanding Shape ● The shape of any ordinal, interval or ratio data is important to its interpretation. ○ Can show multimodality and/or outliers. ● This is much easier to see through a visualization than from a table of numbers. 16
- 17. Understanding Center ● The central value of interval or ratio data tells you what to predictively expect (it's potentially the most frequent value). ○ You should calculate both the median & mean. ■ If they differ this is a sign you have skew in the data, possibly from outliers. 17
- 18. Understanding Spread ● The spread of interval and ratio data tells an important story about precision of your predictability. ○ Calculate the Interquartile Range (IQR). ■ This is found by subtracting the 1st quartile from the 3rd quartile, and shows 50% of the data. ○ You can also calculate the variance and standard deviation. 18
- 19. 5-Number Summary ● Between a visualization and the 5-Number Summary you get most of the information you need to interpret what’s going on with your variable. ○ This will show you the min value, quartiles, median/mean, and max value. ○ The only thing that’s missing is the number of observations (n-count). 19
- 20. 20WORKED EXAMPLE
- 21. The Example Data’s Origin ● The example data comes from a survey of festival goers (aka customers) and was linked to observational data about their lifetime ticket sales. ● It’s a cross-sectional sample (n=3,834) since we don’t have every festival customer’s feedback and the data was captured at a single moment in time. 21
- 22. Inspecting Your Data File ● Before you start summarizing and visualizing your data, open the raw file and look around. ● Make sure you can identify what the rows are, and what each column measures. ○ When in doubt, ask for a data map or data dictionary. 22
- 23. Ordinal Data: Years Attended 23 68% of festival customers have been attending for less than 10 years. 1 in 10 have been attending for more than 20 years.
- 24. Interval Data: Likelihood to Recommend 24 Min 0 1st Quartile 9 Median 10 Mean 9.2 2nd Quartile 10 Max 10 5-Number Summary ~80% of festival customers are likely to recommend (9’s & 10’s).
- 25. Making It Ordinal: Net Promoter Groups* ● It’s always possible to transform data from continuous to categorical, but not the other way around. ○ Likelihood to recommend can be transformed into categorical groups to create a simpler metric: ■ Net Promoter Score (NPS). 25 *Source: https://hbr.org/2003/12/the-one-number-you-need-to-grow
- 26. From Interval to Ordinal Data: NPS 26 ● The Festival’s NPS is: 75% ● We could use these groups for testing differences, like in their Customer Lifetime Value. ○ This is often why you want to create categorical data from continuous data.
- 27. Ratio Data: # Tickets Purchased Per Visit 27 Min 0 1st Quartile 2 Median 4 Mean 5.8 2nd Quartile 6 Max 400 5-Number Summary The presence of outliers hides an important pattern in this data. To see it, we will drop outliers who purchase more than 13 tickets per visit.* *You should ALWAYS note when you drop outliers from analysis.
- 28. Ratio Data: # Tickets Purchased Per Visit 28 Min 0 1st Quartile 2 Median 4 Mean 4.4 2nd Quartile 6 Max 12 5-Number Summary With outliers removed the mean falls to ~4 tickets, and we can see multimodality for even-numbered purchases. People don’t like to go to the festival alone.
- 29. Ratio Data: Customer Lifetime Value 29 Min 0 1st Quartile 124 Median 336 Mean 1510 2nd Quartile 1125 Max 479878 5-Number SummaryAs with tickets purchased, the presence of outliers is obscuring any detail in the visualization. The maximum value of $479,878 is suspiciously high (though it turns out to be an accurate value, despite being 55 standard deviations above the mean*). *Values more than 3 Standard Deviations from the mean are considered outliers.
- 30. Ratio Data: Customer Lifetime Value 30 Min 0 1st Quartile 112 Median 249 Mean 486 2nd Quartile 642 Max 2624 5-Number Summary Dropping outliers above $2,624 causes the mean to fall significantly from its previous level. This shows EXTREME leverage in the data.
- 31. Ratio Data: Customer Lifetime Value 31 Like most data, the festival’s Customer Lifetime Value exhibits the Pareto Principle (aka the 80/20 rule). This means 80% of all CLV comes from 20% of customers.
- 32. What We Learned About CLV? ● Most festival customer have been attending for less than 10 years, but there’s a small group that’s been coming for more than 20. ● Festival customers are unlikely to come alone, they’ll buy 4 tickets, and virtually all are likely to recommend the festival. ● The average CLV is $486 and 80% of all CLV comes from just 20% of festival customers. 32
- 33. Next Step: Analytics & Predictive Modeling ● The next step for this data would be multivariate analytics. ○ Tests of difference & measures of association. ○ Present discounted value of future ticket sales. ● After that, we could use all of the data to build a predictive model for CLV. 33
- 34. JOIN US FOR THE NEXT LECTURE Turning Analytics into Actionable Insights, Thursday October 19th 2017 emperitas.com/lecture