Disheartening error bars, or, when you realize your data is not significant

Let’s find the subgroup analysis that looks somewhat significant! Hmm this one inhibitor in one situation looks really promising… To think of it, we didn’t cover the ethics or practice of leaving out data and focusing only on the positive results?

There are two conflicting forces at play whenever I begin to make figures: the need to explicitly and accurately convey technical information, and the desire to have visual cohesion and minimalism. This wasn’t a problem for the Mod1 report; although we performed various analyses using existing programs, they already looked visually acceptable and clear.

And then we learned about statistics. The guest lecture by Shannon was quite illuminating; I had been using t-tests and ANOVA p-values without precisely knowing how each value was calculated, and I learned a lot in this lecture. However, now there is no pretending like we don’t know how to calculate 95% significance from our data. The inclusion of these confidence intervals led to many of us experiencing the discomfort of having error bars that are larger than the sample space (Ie: 5 +/- 10).

This was of course an element in my figures that I didn’t find necessarily beautiful, so I tried several ways of pooling the data together such that it would be optimal in both conflict realms.  It was good in some respects, in that it did generate results acceptable of reporting, but it also led to me agonizing over what were the best ways to represent data over multiple variables. In the end, I’ve learned that aesthetics and comprehensive accuracy in figures are not a dichotomy, and you can have both with experience, and of course, more samples.

Significance is a funny thing; it is both definitively objective (p < 0.05) and entirely relative (I’ve learned a significant amount from this class!).  

Should I represent it all on one graph? Or four different sub graphs? Or maybe just compare xrs6 to…