Is it appropriate to use t-tests to find the most different outcomes? I have a large CSV file of spending by individuals on different drugs. I want to find the drugs with the most different spending patterns by gender. 
The file looks a bit like this, but with 7.5k rows and 2k columns:
id,     gender, spending_drug_a,  spending_drug_b, spending_drug_c ... 
345089, m,      304.9,            405.2,          117.0
609235, f,      285.9,            0.45,           89.9,
345089, m,      304.9,            405.2,          112.1
609235, f,      285.9,            8.43,           89.9

About 6.5k of the rows are male and around 1k are female. 
My question is: what approach can I sensibly use to find the drugs with the most distinct spending patterns? My immediate thought is


*

*for each column, create the null hypothesis that the two means are equal, then perform a t-test on each column

*look at the value of p for each column, and rank the lowest values first


Does this sound like a reasonable way to proceed?
I don't know if the spending distributions are typically normal, but let me know if that would be useful information. 
 A: If you're primarily interested in detecting changes in typical expenditure, and drugs tend to cost different amounts; I'd imagine something like ratios might be more relevant than differences (suggesting looking for differences in a model with log link), but any zeros would have to be treated separately; this suggests considering a zero-inflated model, perhaps something like a zero-inflated gamma or zero-inflated lognormal.
--
If you want to find which particular ones are different:
One possibility is to construct some measure of difference and perform a permutation test or randomization test. Your first interest in in the mean, so one can readily compute the sample mean for groups of interest and look at differences or ratios of those. Being nonparametric, as long as the distributions would be the same under the null, there's no issue about what the distribution shape actually is.
One could as easily use a more general measure for the difference in distribution, like a Kolmogorov-Smirnov, Cramer-von Mises or Anderson-Darling-type two-sample statistic and again use a permutation-type test.

If you really want to identify the ones that are most different, that would not involve hypothesis testing, but some measure of effect-size. Confidence intervals for (say) the mean differences might be useful supplementary information. One might use bootstrap intervals for those.
