Tag Archives: McNemar’s test

Choosing the right statistical test

I cheat.  Well, I use “cheat” sheets.  I’ve had plenty of statistics training, but honestly, my brain just doesn’t want to hold onto the assumptions associated with this or that test.  So I create little charts or tables, which I reference often.  This frees up brain space to continually re-assess the pros and cons of buying a used Jeep Wrangler versus a Subaru WRX (three years strong and I still haven’t decided).

Anyway…here’s a quick reference guide for choosing a statistical test:

1. What is my variable type?
2. Is the comparison group data paired or unpaired? (i.e., can you link data from individual respondents in the two comparison groups or not)
3. What is the sample size?

Once you’ve answered these questions, use the following to identify which statistical test of significance to use:

 Variable type Paired data Sample size (in each group) Test of significance Categorical No >5 Chi-square <5 Fisher’s exact test Yes N/A NcNemar’s Test Continuous No >30 Independent samples t-test <30 Mann-Whitney U test Yes >30 Paired t-test <30 Wilcoxon signed-rank test Ordinal No N/A Mann-Whitney U test Yes N/A Wilcoxon signed-rank test

Part IV: What? More about effective size?

Over the previous three posts, I introduced effect size, discussed its calculation and interpretation, and even provided an example of how you can use effect size to demonstrate the effectiveness of your overall CME program.  My intention was to present a method for CME assessment that is both practical and powerful.

For those a bit more statistically savvy, you likely noticed that my previous effect size example focused on paired, ordinal data.  That is, I used a pre- vs. post-activity survey (i.e., paired) comprised of rating-scale (i.e., ordinal) questions.  I chose this path because it’s fairly common in CME outcome assessments and it’s the most straightforward calculation of Cohen’s d (which was the effect size measure of interest).

Here are some other scenarios:

1. If you’re using pre- vs. post-activity case-based surveys, you’re now working with paired, nominal (or categorical) data that has most likely been dichotomized (e.g., transformed into correct/evidence-based preferred answer = 1, all other responses = 0).  In this case, the road to effect size is a bit more complex (i.e., use McNemar’s to test for statistical significance, calculate an odds ratio[OR], and convert the odds ratio to Cohen’s d).  Of note, an OR is itself an effect size measure, and converting this to Cohen’s d is optional.  The formula for this conversion is d = ln(OR)/1.81 (Chinn S: A simple method for converting an odds ration to effect size for use in meta-analysis. Statistics in Medicine 2000, 19:3127-3131).
2. If you’re using post-activity case-based surveys administered to CME participants and a representative control group, you’re now working with unpaired, nominal data (that is typically dichotomized into correct answer vs. incorrect answer).  In this case, you’ll use a chi-square test (if the sample is large) or Fisher’s exact test (if the sample is small) and also calculate a Cramer’s V.  You’ll then need to convert Cramer’s V to Cohen’s d (which you can do here).

If you’ve been doing this, or any other analysis incorrectly (as I have in the past, often do in the present, and bet on in the future).  Don’t fret.  Statisticians are constantly pointing out examples of faulty use of statistics in the peer-reviewed literature (even in such prestigious journals as JAMA and NEJM).  Keep making mistakes, it means you’re moving forward.