Talking Numbers: The Significance of Statistical Significance

Standard

Statistical significance XKCDStatistical Significance. It’s so basic a concept that many modelers and statisticians don’t look for it in simpler analyses such as marketing campaign comparisons, population distributions and so forth. Nevertheless, statistical significance can make all the difference (pun fully intended) in whether results are jaw-dropping or trash-worthy. And when speaking with the business-side of the house, you’d better know which is which.

Tip for Translation: Less Insignificant Info is More

First of all, check for significance. That could go without saying… but it doesn’t (see above). If the results are not significant then do not put them out there for the business to jump all over… which they will. If they are close to significant and you want to share them, add a caveat/footnote indicating that the results may be the result of random variation. Do not use technical terms here.

Two common times that analysts try to give flack about this:

“Yeah, but my tests always appear significant because I have enough data that even tiny differences get picked up due to the sample sizes.”
Try bootstrapping smaller samples for comparison and see what happens. If you are still getting significant results, good on ya’. If not, maybe the results weren’t as significant to start with.

“But the business impact of this minuscule difference is huge so therefore it doesn’t matter if the result is significant.”
This is actually even more reason to validate your results. Presumably, if the resulting teeny difference would cause a major upheaval for the business, so too would a minor variation due to natural fluctuation. Use a significance level that aligns with the importance of finding a difference. For a super-important analysis, go with an alpha of .01 or even .001 rather than the usual .05.

To Show or Not to Show Statistical Significance…

More often than not, business stakeholders only want to see results that are significant. They want to know how the analysis can be used to better effect. In general, that does not mean that you go around flaunting a p-value. Just state the results and how to use them and move on. On the other hand, it is usually informative and interesting to business users to see statistically insignificant results when it confirms or debunks a long-held hypothesis.

For example, let’s say that there is a “gut feel” that customers who buy diapers also buy beer. After doing some testing on purchase data, you find that there is no significant link between these two product categories. The business stakeholders (and holders of the gut feeling) would likely need to know that these two items are not correlated. It impacts store and display layouts for the future.

Got some great examples of significant results that really weren’t? Or times when instincts were proven right/wrong? I’d love to hear all about your adventures with statistical significance in the comments.

Talking Numbers: Presenting Analytics

Standard

Use FIRST to help you when presenting analyticsOne of the most difficult parts of any analyst’s job is packaging and presenting analytics work for a business audience. It’s a matter of showing the “So What” rather than just the “What” of the data. As someone who not only does a lot of presenting, but also coaches others on how to improve their deliverables, I found myself in need of a consistent way to structure results that minimizes the technical or mathematical description and instead focuses on the business implications.

Here’s my framework, along with a handy acronym: FIRST

Findings

You have to present the facts of the model or analysis. These are the numbers that came from all your hard work. Talk about the hypotheses tested, what fell out, what stayed in, results of tests, etc. Detail any rabbits you chased in the data (anomalies, unexpected results, iterations, etc). Include visualizations wherever able to succinctly illustrate what you saw. This section is of particular interest to other data scientists, model auditing teams, and the statistically-inclined.

Insights

Now read into the numbers or the model and weave the story. Describe what you learned from the analysis or model and, especially, what it means to the business. Explain what the findings mean in a broader context.

Recommendations

Document further analyses, follow-on projects, or deeper dives that you recommend pursuing. Also note any follow-up questions that your business stakeholders come up with based on the findings and insights. Build your own backlog of projects and then track them down. This is where you plan to tie up loose ends.

Suggestions

Describe what you think the business should do, or how it should change, to make use of the insights or models. If there are specific processes that would benefit from incorporating a model or API, indicate how this might be done (do not show code – just say how it would revamp the process). How does this help the business make better decisions about their initiatives?

Takeaways

This section details who is doing what coming out of this analysis. If there were specific expected actions to be taken based upon results, identify whom is needed to complete them. Any time frames necessary should also be outlined.

Presenting Analytics in Documents

Putting FIRST into a document is pretty straightforward. Set up the project at the top, laying out the key questions, expected actions, and the planned analysis steps. Then include the FIRST sections with the majority of the content. Document as you go along so that you capture the findings roughly as they occur. This helps to ensure that you present all of the permutations of analysis performed.

Presenting Analytics in Slides

Most of the time, when presenting analytics, we are called upon to use slides or a slide-like format for conveying information. Again, present the project overview and key questions. Then immediately put forward the key insights. Yes, this is out of order for FIRST. However, the next section is where FIRST comes into play. For each key insight, present the FIRST elements on a single slide.

An example might be that customers using discounts are more valuable over time. On a slide, show a graph comparing the spend patterns of customers with and without discounts. Provide a bullet point indicating that this is the key insight. Then outline additional steps for:

  • analyzing types of discounts or time periods of discount as a recommended follow-on,
  • using promotions to increase total basket size as a suggestion, and
  • planning an upcoming promotion as a takeaway.

Wrap up with a summary of next steps so that there is a clear list of actions to be done.

What methods do you use today for presenting analytics? If you give FIRST a try, please leave comments about how it goes with your key stakeholders.

Calculating Percentiles in PostgreSQL

Standard

Slicing data into manageable chunks for viewing is crucial when you start dealing with more records than will fit in something like Excel (without PowerPivot, of course). One of the most common ways to look at data in a more easily-digestible manner is to use percentiles, or some derivative thereof, to group records based on a ranking. This allows you to then compare equal-sized groups to one another in order to form conclusions as to relative behavior.

Most SQL variants have workarounds for how to accomplish this task that may or may not actually cover 100% of your data (may drop a few records here or there when trying to round a percentage to a whole number of records to pull). PostgreSQL, on the other hand, has a handy function built in for doing this sort of thing without having to worry about getting full coverage on your table. Continue reading

Data Really Is Sexy

Standard

Recently, in speaking with the President of Kobie regarding my team, I used the term “data munging” to describe a lot of the work that we do. He laughed, thinking I had said “data munching” (mmmm, tasty!) and asked if that was a technical term. The short answer is that yes, it is another term for data wrangling (which, incidentally, is one of my favorite terms in the industry). Continue reading

Randomizing Random Selection

Standard

Random Numbers SerifThis one had been bugging me for a while now. There are a lot of analyses where it is useful to select multiple random groups. Usually, this would involve picking a bunch of numbers out of your head and trying them as the seed values (I like using phone numbers without the area codes – then I can call the person and tell them they rocked my randomization).

But today, as I struggled with pulling a multitude of sample sets, I decided to come up with a more elegant solution for generating random number seed values. Behold the random loop: Continue reading