Tag Archives: Bayesian

The Dudes Do ASHA 2011: Day 4. On Bondage, Booze, and Bayesian Statistics

It’s a cheap trick I know, but the “bondage” referred to in the title refers to the binding of a laptop with a smartphone, otherwise known as tethering. Casual visitors who have landed here via a Google search hoping for some images of people tied up and engaging in fetishistic practices should leave now because there are none. Tethering is a little trick you can use if you find yourself without an easily accessible WiFi connection – such as at the San Diego Convention Center. Interestingly, the brochure in my hotel room touts the city as being one of the most “wired” places on the world but it conveniently leaves out the phrase “at a price.” There is a WiFi connection to be had inside the Convention Center but it is a fee-for-use option that, as always, seems steep for 10 minutes, which is about the time I needed to upload the latest blog posts.

Tethering an android phone

Computer bondage

However, with a cable and a piece of free software, it’s possible to tether a computer to a smartphone and use the phone’s Internet connectivity as a router. Specifically, I’m using a Dell Latitude, a Droid 3, a USB cable to connect the two, and a piece of software called PdaNet – one version running on the phone and its partner-in-crime running on the laptop. [1]

I say “in crime” because the phone company don’t really want you to do this. What they want is for you to pay an extra $20 per month on top of the $30 per month you are currently using in order to download through your phone. This strikes me as odd because whether I download data to my phone or my laptop via the phone, it’s the same data use. And seeing as I also have an unlimited data plan, what’s the difference? So, the outcome of this is that tethering is frowned upon and can, if the phone company finds out, can lead to a stiff warning followed by being kicked off the system.

In truth, I never tether if I can possibly avoid it, simply because it’s actually much, much easier (and more stable) to use WiFi or an ethernet connection. I reserve tethering to emergency situations. Such as the SDCC.

The tethering took place immediately following my last ASHA session for 2011; a short course entitled Evidence-based Statistics for Measuring Strength of Evidence, a title that sounds benign enough but turned out to be a wolf in sheep’s clothing. The presenters, David Maxwell and Eiki Satake, are both Professors at Emerson College and clearly know their stuff.

Evidence based practice pyramid

EBP Pyramid

I became an SLP partly because the part of the brain that deals with numbers is clearly non-functioning in my case. I know enough math to keep me away from the casinos [2] and overpaying at the bar, but when equations and formulae come onto the scene, I suffer from math blindness. Which is why this particular course had me cranking up my brain volume to eleven in order to follow some of the arguments.

And that’s where the Bayesian [3] statistics came in and I learned the difference between a Bayesian  and a Frequentist approach to statistical analysis. Let me try to summarize what the presenters were suggesting and how it relates to EBP. Essentially, they say that;

(a) Current research practice in Speech and Language Pathology uses a Frequentist model that relies on creating a null hypothesis and testing it against a P-value.

(b) This model inherent works on a “pass” or “fail” basis but tells you little about the strength or power of a treatment.

(c) Such p-values are not the best way to deal with clinical outcomes.

(d) A Bayesian model uses a modified P-value – the “precise P” as opposed to the “imprecise P” – which is more meaningful for clinical data.

(e) Using a Bayesian precise P-values tells us more about the strength of experimental observations and is, therefore better for evidence-based practice.

(f) There is a difference between statistical significance and clinical significance, and the latter is better evaluated using a Bayesian method.

This summary is probably hugely unfair because condensing a 3-hour short course into five sentences is hardly “best practice,” and there was a 150+ slides handout to accompany it, and that’s tough to condense! Nevertheless, if there’s a single take-away idea, it’s that we should be considering Bayesian methods for EBP. [4]

Fortunately, there is a book available that explains Bayesian statistics as applied to EBP; The Handbook of Statistical Methods: Single Subject Design (2008) [5]. I have the sneaky feeling that this is unlikely to be light, bedtime reading but I’ll see if I can track a copy down in the library before splashing out on a copy for the bookshelf.

Handbook of statistics

Satake et al. 2008

Which brings us to the booze and perhaps the second most useful piece of information to come out of the short course; beer and statistics do mix!

At some point in their training, SLP’s will have come across the “Student’s t-test,” which is not, as it suggests, a “t-test” for a “Student” to use but one originated by someone called “Student.” It turns out that “Student” was, in fact, the pseudonym of William Gosset, a statistician who worked on a project with folks at the Guinness corporation to test the quality of the beer. The t-test provided a way for beer testers to take a few small samples, measure them, and then use the test to get a quality score. Now there’s a job! Unfortunately for Gosset, as a Guinness employee, he couldn’t publish what was proprietary information. Nor did he want to face any ribbing from his colleagues because he was working for a private company – a beer maker at that. However, he was eventually allowed to write papers but under the nickname of “Student,” hence the ultimate name of “Student’s t-test.”



At this point, I recommend you’ve had enough statistics for the day and head for a local bar to do your own research into how effective the t-test is at quality control. The sample size is up to you.

[1] Setting up your phone and laptop for tethering is not rocket science and Google is your friend for using the PdaNet 3.0 software. As I said, it’s not illegal but if the Cell Network Mafia come a-knocking, I take no responsibility if you’re found at the bottom of a river with concrete overshoes. If you’re using a Mac and an iPhone, I can’t help you, but undoubtedly “there’s an app for that” and it’s so intuitive you can give it to a 6-month-old to set it up for you.

[2] Gambling (or as the industry prefers to call it in a miserable effort to make it sound more friendly, gaming) is an example, par excellence, of applied probability. All you really need to know in order to avoid living your life under bridges and drinking rubbing alcohol hidden inside a brown bag are two things; (a) the gambling industry only exists because the odds are always in their favor, and (b) your odds of becoming extremely wealthy by gambling are so small as to be, in most cases, negligible to zero. Casinos always tell you about the big winners but don’t provide you with a list of losers. By all means go out with $100 in your pocket for a night of fun, at the end of which you expect to have lost it all, but when you start thinking, “OK, maybe just another hundred and then I’ll call it quits,” start practicing how to make the sentence “Honey, I lost the house” acceptable.

[3] It’s pronounced /’bəɪziˌjʌn /. I add this because I was saying it wrong for years!

[4] I found a short summary of Satake’s critique of Frequentist statistics in an article he wrote in 2010. Entitled Moving Forward to Evidence-Based Statistics: What Really Prevents Us?, he explains why the traditional P-value is misleading and why. http://www.pluralpublishing.com/web_flyer/web_flyer_community_august/web_flyer_community_august.htm 

[5] Satake, E., Jagaroo, V., and Maxwell, D.L. (2008). Handbook of Statistical Methods: Single Subject Design. San Diego, CA: Plural Publishing.