need help with inference

hi all,
I am trying to remember stuff i learned long ago ... in order to argue my position in the following dispute: I said that there was no significant drop in gas consumption in the wake of hurricane Katrina in Texas, some other people said it was. Regardless of the relevance of this discussion, please help me figure this one out.

I will receive data relating to gas consumption on a daily basis in the city of Dallas, for a good period of time, i believe it goes back about 10 years. With this data at hand, how exactly am i supposed to test my hypothesis?

I will gladly do my homework if anybody puts me on the right track. For now, I only know I will have to get the consumption data and probably look at consumption on the same days at the end of August, beginning of September from 2005 and 2004, and as far back as the data goes. I am not sure though how would I build my samples ... I guess my question would be can you really make an inference about the gas consumption based on data for just a few days. What if the consumption data were weekly and I would have to make an inference about one week only, so I would look at one measurement only?

Its quite a bit of homework =)~

This is a classic time series problem in a linear regression format.

The text I used presented it thusly:
Concord California around 1979 to to 1981 wants to know if its campaign for water usage reduction is having an effect. The problem is that there are natural cycle to these things. You code for the months that are after campaign (after katrina) as 0 and 1 (or something more elaborate..)

But if you do that you'll discover the residuals are "screwed". and youll need to do all sorts of work on that.

I did a project with some oil problems it turned out they were whats called "ARIMA (1,1)" So if yours is any similiar it is a nontrivial time series you are working.

Same sort of idea. Theres a quite a bit of background though to being able to do this stuff though =/ Not much I can tell you in the way of shortcuts.


New Member
If you just want to make a point and not do something fancy, you could just just line up the 365 days (leave out Feb 29), of the last, say 5 years, and make a repeated measures ANOVA. That way you just get the significance and avoid the regression models.
If I am trying to make a point for the null hypothesis, as you are, I would pairwise t-test each of the previous years, and possibly a few after, and see if any has a <0.05, if this is so, it strongly supports your assertion, since the critical p value should be even lower, due to multiple testing.
But consider this more of an exploratory analysis, just for in house consumption. If you have to report this as part of your job, you will have to go the way Round suggests.