T-test and the number of observations I have a sample of the number of bank loans issued each year from the period 2004-2012. I have split the sample in two (2004-2008 and 2009-2012) and want to test if the number of bank loans issued in the second period (2009-2012) is significantly lower than the first period. 
Question:
When using the T-test, would N be the total number of bank loans (640 in this case) or the total number of years observed (9)?   
 A: Taking a step back, why did you combine years? Unless you have some strong reason to suspect that something dramatic happened exactly at the division of the two periods, this seems to waste information.
My first step would be to make a line graph of bank loans per year (year on the x axis, number of loans on the y axis). 
The dependence question that others raised is also interesting. I am not sure if the data are dependent or not, there isn't enough information given. I wrote about dependent vs. independent data on my blog
A: Very generally, N is typically used to denote the number of observations, not the sum of all values that a variable takes in a sample.
If I understand you correctly, what you have is a number of bank loans for each year, so you can look at it as one observation (the count) for each year, 9 in total. The number of bank loans in a year is the value your variable takes.
That said, the question is much more complex that it seems and using a t-test looks like a bad idea for a number of reasons (measures from different years cannot be assumed to be independent from each other, t-test is not recommended for count data). If you want useful advice, you should probably give more details on your problem and what you are trying to achieve, perhaps in a follow-up question.
A: If I understand your question correctly, you wish to know if the number of loans per year differs between the two periods. This would mean that you have nine observations of what you want to compare: the number of loans for 2004, for 2005, ..., 2012.
Given this small number of samples and the difficulties in assessing the requirements for using the t-test (normality and whether or not the variances are equal), I might advise looking into alternatives for the t-test, like rank tests (Wilcoxon, in your case) instead.
A: You want to test, if the mean is significantly different, this is a difference of means test (t-test) and its test statitics depends also on the number of observations in the subsamples. These are the number of observations, what in your case are the number of bank loans and not the number of years.
With this test your are just testing the difference of the mean in two subsamples. The test statistic does not 'recognize' from which year the subsamples are. So you basically could create a subsample of ten observations of e.g. the years 1988,2000,2000,2000,2000,2000,2000,2000,2000,2000 and e.g. the period 1900-1909. So the number of observations would be for each subsample 10, since you have 10 observations. From which year they are, is not relevant.
attachment: test statistic:

You see, $n_1$ is the number of observations of the first subsample and $n_2$ is the number of observations of the second subsample.
