Featured Product
This Week in Quality Digest Live
Management Features
Henning Piezunka
Businesses and leaders influence the kinds of ideas they receive without even realizing it
Chris Caldwell
Significant breakthroughs are required, but fully automated facilities are in the future
Dawn Bailey
Helping communities nurture the skilled workforce of the next generation
Brent Simpson
Even if it works in your favor
Mike Figliuolo
Stay cool. It all works out.

More Features

Management News
A tool to help detect sinister email
Developing tools to measure and improve trustworthiness
Manufacturers embrace quality management to improve operations, minimize risk
How well are women supported after landing technical positions?
Adds increased focus on governance
Survey shows 85% of top performers rely on it to achieve business objectives
Key takeaways from Marcum’s 2023 National Manufacturing Survey

More News

Davis Balestracci


Two More Lurking Mess-Ups for Any Experiment, Designed or Not

The ‘human variation’ factor is always present

Published: Monday, September 19, 2016 - 14:14

Referring back to June’s column, I hope you’ve found C. M. Hendrix’s “ways to mess up an experiment” helpful in putting your design of experiments training into a much better perspective. Today, I’m going to add two common mess-ups from my consulting experience. If you’re not careful, it’s all too easy to end up with data that’s worthless.

Balestracci’s Mess-Up No. 1

Underestimating the unintended ingenuity of human psychology will mess up your experiment—and this includes the study planners! 

Trust me, there is no way you could make up the things busy people will do to (unintentionally) mess up your design and its data. “But I just want to run a simple 2 × 2 × 2 factorial,” someone might say.

To refresh your memory: Suppose you’re interested in examining three components of a weight-loss intervention:
• Keeping a food diary (yes or no)
• Increasing activity (yes or no)
• Home visit (yes or no)

It’s easy to set up the 2 × 2 × 2 factorial design matrix, and you might think the only remaining issue is how many people should participate.

Remember from last column, when a client’s initial request for “just a design template” turned into three consults totaling more than two hours? Using this weight-loss experiment scenario, let’s see why.

Item 1: Some things need to be clarified

“Weight loss” is pretty vague. What exactly do you mean by that (i.e., the operational definition)? Pounds? Percent of original weight? Meeting a pre-established goal? Something else?
• Is a weight-loss goal of any kind going to be involved? For everyone?
• Imagine you have the data in hand. What are you going to do with it? Will it allow you to take the action you desire?
• Are counting tallies of any kind going to be recorded?
• Is the threshold between any “nonevent” (0) and “event” (1) clear?
• If two people were evaluating a participant, would they get the exact same number(s)?

What time period is going to be studied? One month? Two months? Three months? Six months? A year? And what exactly do you mean by “keeping a food diary,” “increasing activity,” and “home visit?” Do you mean daily? Weekly? Do you want them to formally record the activity? Does a phone call count as a “visit?”

What characterizes the people you would like to study? 
• How will you decide whom is admitted into your study? 
• How will you sample to obtain these people?

Are you going to make it absolutely clear to the study participants what is expected?
• How will you know that they understand? 
• Should assessing this understanding be an ongoing part of the home visit?
• How will you get this information from the people not having home visits? Would the knowledge of a scheduled check-in phone call bias their results?

Don’t underestimate a very serious “headache factor” that I’ve often encountered: Is this study going to collect data from several sites or departments? If so, how are you going to make sure everyone agrees on the answers to the questions above? How would you know? Would a simple designed data-collection sheet (with definitions on it) help reduce this variation? (Hint: Yes.)

Should you PLAN a brief initial study whose only objectives are to reduce the human variation in perceptions of 1) the execution of the protocol, 2) how to define and collect the data, and 3) the ease of using the designed data collection sheet? (Hint: Yes! But these data won’t be used in the study itself.) How will you choose people for such a study?
• DO the brief study.
• STUDY the results with everyone's input—both planners and participants. What mess-ups occurred during both the study’s execution and data collection? How could they be avoided? Should the data collection sheet be redesigned or simplified for easier recording?
• ACT on these conclusions and begin your next PLAN—another process/data study or begin the experiment? 
• Should you PLAN some type of ongoing data-collection assessment during the study to prevent data contamination from human variation? (Probably.)

Item 2: What threshold of weight loss would result in proceeding with your investment in such a program? Or is it a matter of “met personal goal” (1) or “did not meet personal goal” (0)? The latter measure of “percent of patients who met goal” isn’t recommended. Read on to see why.

Item 3: How badly do you want to detect this difference?

Item 4: How many people do you need? 

Key question: What is the ratio of this desired result from item 2 above, relative to your standard deviation? Are any actual data available from similar weight-loss research studies (preferably using the same time period) to measure the weight losses as well as unintended gains? If so, what are some of the reported standard deviations for such a group of people? Are they consistent enough to come up with an approximate value?

Remember the client dialogue from last newsletter regarding the tar scenario?
• In wanting to detect an effect of 1, its resulting ratio relative to the tar-process standard deviation of 4 was (1/4) = 0.25, which required 500 to 680 experiments. If this was your desired ratio as well, then, like them, you would need 500 to 680 people.
• Similarly, when wanting to detect an effect of 2, (2/4) = 0.5. If this was your desired ratio, then you would need 130 to 170 people.

If you obtained sample sizes like these, might you consider the possibility of studying two additional variables’ effects on your results? Perhaps consider: “Person does not set a goal” (0) or “Person sets goal” (1)?

Experimental logistics aside, where do these sample size numbers come from?

They come from answering three questions. The first is answered pretty much by default:
1. What risk are you willing to take for declaring an effect significant when it isn’t? Usually, 5 percent (p < 0.05).

The other two are:
2. What size effect must I detect?
3. How badly do I want to detect it?

Finding the answers to the three questions above automatically answers the implicit fourth question: What size sample do I need?

If you have a specific sample size in mind and plan on using p = 0.05 for significance (that’s two questions answered), you can work backward to calculate various combinations of the other two, either:
• What effect you can reasonably detect (2), given your specific answer to (3)—i.e., desired probability to detect it
• The probability of detecting a desired effect (3), given your specific answer to item (2)—i.e., desired effect to detect

Important point: If all you do is default to p = 0.05 to detect effects, your design will, by default, answer questions to items 2, 3, and 4 (as in the calculated sample sizes above). This could possibly waste your hard work unless you reconsider your objectives.

For those of you who are wedded to the “rapid cycle PDSA” methodology, have I brought up some things that might need consideration when you PLAN? (Hint: Yes.)

Here’s an additional mess-up, which, unlike the first, isn’t necessarily guaranteed—if you carefully PLAN:

Balestracci’s Mess-Up No. 2

Vague planning on a proposed vague solution to a vague problem usually results in vague data, on which vague analyses are performed—yielding vague results.

Many times, I see data collection addressed as an afterthought, usually ad hoc. Collecting poorly designed data (or not even collecting data at all) virtually guarantees non-trivial human variation seeping in—an open door for introducing the toxic and very human “Constant Repetition of Anecdotal Perceptions.” (CRA...).

More discussion of sample size next time.


About The Author

Davis Balestracci’s picture

Davis Balestracci

Davis Balestracci is a past chair of ASQ’s statistics division. He has synthesized W. Edwards Deming’s philosophy as Deming intended—as an approach to leadership—in the second edition of Data Sanity (Medical Group Management Association, 2015), with a foreword by Donald Berwick, M.D. Shipped free or as an ebook, Data Sanity offers a new way of thinking using a common organizational language based in process and understanding variation (data sanity), applied to everyday data and management. It also integrates Balestracci’s 20 years of studying organizational psychology into an “improvement as built in” approach as opposed to most current “quality as bolt-on” programs. Balestracci would love to wake up your conferences with his dynamic style and entertaining insights into the places where process, statistics, organizational culture, and quality meet.