(if the slightest hint of math is intolerable to you, please don't read on.)

1. i'm doing a meta-analysis for my quals. you know this.

2. a meta-analysis is a quantitative compilation of studies within a topic area; this is superior to a qualitative compilation because (a) one avoids bias in study selection (everything goes in), and (b) the result is a definitive answer as to whether a given treatment has a significantly greater effect than 0 (as opposed to: "findings are in conflict, and more work needs to be done in this area".)

3. this is all very nice in theory.

4. in practice, looking in 4 textbooks gives you 6 different formulae for every step of the analysis you have to do.

4.1. there are many steps.

4.2. why can't statisticians agree with one another.

(4.3.) they do, says the housemate, but only after they get tenure.

5. the problems right now are numerous, but there are two big ones. first of all, there is no agreement whatsoever among anybody as to how to calculate sampling variance. the idea is this: each study is weighted by a coefficient that represents the error due to sampling from a random population -- the larger the study, the greater the weight. i strongly suspect that the differences between all the formulae i've accumulated are extremely subtle, and will make no difference to the final result, but having honed my anal-retentiveness to a razor-sharp point over the last 5 years, i shall endeavor to find out. big problem #2 is even more troublesome: i have a mix of between-subject and within-subject designs in my analysis (between = experimental group/control group; within = everyone does everything and each subject is his own control). apparently, and i only just found this out, there are different equations for the two types of studies, and conversions are necessary before everything can be combined. specific problem: calculating effect sizes for between subjects studies is easy, and in fact can be done with a few button presses right here. within-subjects designs, however, are trickier, because the error term is not clear. a bunch of people recommend that one uses the variance of the change score, which is fine in theory except that who on earth reports that in a manuscript, and i'd rather tapdance naked in perelman quad than email sixty billion labs to ask them to give me that data. so...how? use the pooled variance? use the post-treatment variance (ugh). i hate practical solutions when i just know there are pretty ones out there.

6. excel is serving me well, and minz, if you're still reading this, you will be pleased to know that pivot tables are too (finally).