5 Steps Every Data Analyst Must Know


Many people make it through statistics class without a clear understanding of the common process used for all statistical hypothesis tests. In fact, it's not unusual to complete the class without realizing there is a common process. That's not the students' fault -- most professors don't emphasize it.

Executives need accurate, relevant information that is presented simply, quickly, and clearly. They need answers to their questions, and they need to have confidence, not just in the data's message, but also in the messenger.

What happens to analysts who don't know the process and don't use it to clarify their thinking? Most often, executives just don't find those analysts persuasive. Their presentations sound too esoteric, and their answers to questions do not satisfy.

Here's a confession: I love data analysts who are full of hot air. It is so easy to disarm these windbags. All I have to do is smile sweetly, look them straight in the eye, and ask a little question -- What were your assumptions here? -- or make a simple point -- The method you've used was designed for ratio measures, but this metric is ordinal -- and it's all over. Everybody in the room understands. Don't let that happen to you. The five-step process common to every statistical hypothesis test will make your work bulletproof.

  1. State a null and alternative hypothesis. If this doesn't seem familiar, or you are not sure of how it's done, read any introductory statistics textbook. It's in all of them.

  2. State the assumptions that are reasonable for the application. These are conditions we believe to be true: simple random sampling, normally distributed population, and so on.

  3. Select a test type. The type of test to use depends on what you are measuring and the assumptions you can reasonably make. Each test has its own assumptions, and you may have to comb the books carefully to find them. Make sure the assumptions fit your situation. If they don't, you've selected the wrong test.

  4. Calculate a p value. This can be done for any type of statistical hypothesis test. The p value is interpreted the same way for all tests. Very low p values imply strong evidence against a null hypothesis. High p values represent weaker evidence against the null hypothesis -- or no such evidence at all.

  5. State your conclusion in a simple, plain-language sentence. This one is the downfall of many a mathematical genius. State the results in terms any business person can understand, such as "Our tests show no evidence that the test coupon will outperform the coupon we are using today." Avoid sentences like "The p value was 0.12, so we failed to reject the null hypothesis."

Do you use these five steps in your work? Please share your experience.

Meta S. Brown, Business Analytics Consultant

Meta S. Brown is a consultant, speaker, and writer who promotes the use of business analytics. A hands-on analyst who has tackled projects with up to $900 million at stake, she is a recognized expert in cutting-edge business analytics. She has conducted more than 4,000 hours of presentations about business analytics, and written guides on neural networks, quality improvement, statistical process control, and many other statistical methods. Meta's seminars have attracted thousands of attendees from across the US and Canada, from novices to professors.

Tell Me a Story: Why Data Analysts Must Be Storytellers, Too

Data alone won't make an analyst's work memorable or actionable in the eyes of a business executive. A story puts it into perspective.

It's the Data, Stupid

When it comes to acquiring the data that will feed your analytics initiative, "free" isn't always the best approach.


Re: Communication
  • 9/28/2013 6:03:34 PM
NO RATINGS

The points were well sumarized. I'm thinking the last one might be a neglected one though. Being able to summarize the results so the executives can easily see what's going on and the recommendation can turn into a overblown recitation that goes over the heads of many and leaves confusion if not stated simply and with as few words as possible.

Re: Good advice
  • 9/25/2013 10:22:49 AM
NO RATINGS

@Pierre,

"Professional discussions around the topic could inspire more discussion of about displaying technical results to support data visualization, I imagine."

That's a valid point. Data analytics in the enterprise should not be a "one man show". It is a decision making process that normally involve many persons/skills in the company.

Communication
  • 9/25/2013 10:11:16 AM
NO RATINGS

"Avoid sentences like "The p value was 0.12, so we failed to reject the null hypothesis.""

Agreed! Most data scientists will "likely" not communicate the results of their analysis to the executive in those terms. But it is good to know.

Re: Good advice
  • 9/25/2013 8:09:31 AM
NO RATINGS

Michaeljackson, very true on eliminating technojargon - I do think to employ simplicity in practice takes time to digest the results and think about what expression best represents what needs to be conveyed. I wonder about the degree of difficulty analysts struggle with delivering information simply while meeting deadlines.  Simplicity and refinement do not mix with urgency at times.

Re: Good advice
  • 9/25/2013 1:06:24 AM
NO RATINGS

Nice points, Meta, about how basic stat concepts can be understandable, and how the last point, describing uncertainty, can send a misunderstood meassage about the opportunity possible from the result. Professional discussions around the topic could inspire more discussion of about displaying technical results to support data visualization, I imagine.

Re: Bad advice
  • 9/25/2013 12:43:57 AM
NO RATINGS

Sounds like the real problem is Stupid (and probably overpaid) Executives. Executives should at minimum understand the business they are in. Data analysis is usually done in order to answer a question (solve a specific problem.) Your procedural is ineffective when "here's a bunch of data we got off of Facebook, tell us what it means" scenario.

"Garbage in, Garbage out". Wisconsin has the highest per capita consumption of cheese and also the highest per capita rate of rectal cancer. The numbers show high statistical correlation and the numbers pass your test. Still wrong though.

 

Re: Good advice
  • 9/24/2013 6:50:57 PM
NO RATINGS

very true! it is really from transforming business problem to analyzing data to the final business answer of that original business problem!

Re: Good advice
  • 9/24/2013 3:43:41 PM
NO RATINGS

"State your conclusion in a simple, plain-language sentence. This one is the downfall of many a mathematical genius. State the results in terms any business person can understand, such as "Our tests show no evidence that the test coupon will outperform the coupon we are using today." Avoid sentences like "The p value was 0.12, so we failed to reject the null hypothesis."

This point, in my opinion is the most valid.  If you can communicate your point across to me without all of the techno jargon, then I can better understand you and feel more comfortable in doing business with you.

Re: Good advice
  • 9/24/2013 8:45:57 AM
NO RATINGS

Vincent,

I did not, and have never, recommended stating a p value in a presentation to an executive. I recommended calculating the p value as a part of a process that leads to a plain English statement of the conclusion. However, it is important that the analyst know and understand the p value, and be prepared to discuss it if asked, as some executives, and certainly some staff members, are aware of the concept and will ask.


The use of confidence intervals in place of hypothesis tests is a variant arising out of the same statistical theory, not a revolutionary alternative. As a presentation technique, it's a good idea. However, this should be regarded as a supplement, not an alternative, to hypothesis testing.

 

Re: Good advice
  • 9/24/2013 1:26:32 AM
NO RATINGS

I never provide p-values. This concept is too esoteric and most executives i work with don't know what it is and they  like when i speak their language, rather than mine. indeed, i never perform statistical tests, but instead I provide confidence intervals based on model-free inference. Google "Analyticbridge first theorem" for details.

Page 1 / 2   >   >>
INFORMATION RESOURCES
ANALYTICS IN ACTION
CARTERTOONS
VIEW ALL +
QUICK POLL
VIEW ALL +