Test if the value of a function f(i, j) depends more on the choice of i than on the choice of j Say I have a 24x7 matrix in which the rows represent hours in the day and the columns represent the days of the week. The values in this matrix represent the number of people present at a particular place.
Mathematically speaking the values are thus the result of a function: presence(hour, day)
I am looking for a way to statistically analyse this matrix to answer the following question: does the number of people present depend most on the day of the week, or most on the hour of the day? And to which extent (using some coefficient).
Preferably I would like to do this with Excel.
Note that the situation I describe here is just an example. In fact my data is about something also than days/hours/presence, however my problem is analogous: I have two-dimensional matrix and I want to know whether its values are more dependent on the rows than on the columns, or vice-versa.
Edit
Here is an Excel file with the example problem (hours/days matrix) and my actual data: http://brussense.be/temp/MatrixRowColDependency.xlsx
Edit 2
Please consider that I have a pretty light knowlegde of statistics and therefore may not be using the right wording to express my question. I'm not even sure whether I was right to tag this question with "correlation".
Edit 3
Maybe my hours/days example is causing more confusion than clarity. Hence, I'll explain what my actual data is about.
The data I have results from a calibration experiment for a sound level meter. The meter was exposed to pure tones (sounds consisting of a single frequency) of varying frequencies and varying sound levels. In total we produced tones at 27 frequencies and 7 sound levels, resulting in 189 sounds, each of which we measured with the sound level meter being calibrated as well as with a second trusted meter which we use as a reference. For each measurement we than computed the absolute measuring error. For example:


*

*say I have a tone of 1 kHz played at 70 dB (as measured by the reference), but the sound level meter being calibrated indicates 68 dB. Then the absolute error for that sound is 2 dB (w.r.t. reference).

*say I have a tone of 12.5 kHz played at 85 dB (again as measured by the reference), but the sound level meter indicates 89 dB. Then the absolute error for that sound is 4 dB (w.r.t. reference).


Doing this for all 189 sounds results in a 27x7 matrix, in which each value is an absolute measuring error occurring at a specific frequency and a specific sound level. I.e. the matrix rows represent frequencies, the columns represent (reference) sound levels.
The question I am trying to answer is whether the errors differ more between sound levels (columns) than between frequencies (rows), or vice-versa. Or in other words (although this may in fact be another question entirely), whether the error depends more on the level than on the frequency, or vice-versa.
The second sheet in the Excel file linked above holds this data.
 A: Unless you have more than one week of data, it will be impossible to come up with a statistical answer.
Let me elaborate a bit. You might be able to estimate a model with day of the week dummies and some sort of hours curve (which seems nonlinear in your db data), but in my experience with this sort of data, the effect of day of the week is different at different hours, so that sort specification above imposes a very strong assumption on the data. For example, the Friday effect at lunch is typically different than the Friday effect at dinner. With only one week of data, you cannot break out the two in a rigorous manner, and I think for the calculation you want to do, you need to show that there's no interaction between DOW and hour.
Maybe this is a reasonable assumption for your actual data, but it's hard to know without knowing more. 
Edit: I think you should transform your data so it looks like this:
freq dB error
50  60  13.3
50  65  17.8
50  70  19.8
...

Then you can run an ordinary regression of error on freq and dB. Based on a quick look at the data, you might want to enter dB coded as a dummy variable. I am not sure about freq. There may be other transformations, such as logs, that may be appropriate here, but this is not my area of expertise. In a model where freq enters linearly and dB is a dummy, it looks like dB is the main factor.    
Edit 2: Here's how you define a dummy variable. You create a variable that is 1 if dB=60, and 0 otherwise. You do that for other levels of dB. This gives you 7 binary variables. Then you do a regression with no constant or intercept (to make interpretation easier). Say I am interested in interpreting the coefficient on the dB=85 dummy. Let's say it's 5.25. This says the effect of dB at 85 is a measurement error of 5.25 units, all else being equal.
In the model I estimated, the effect of a one unit increase in freq was .0002097, so the estimated equation for dB=85 was error=.0002097*freq + 5.257775. For dB=90, it was error=.0002097*freq + 5.872589.
You could have also estimated the equation with everything entering linearly. This gives you error = 16.12038 + .0000352*freq -.1251323 *dB. The coefficients are the increase in error from a one unit increase in explanatory variables. The difference between the two models is that in the first, the intercept depends of dB, while in the second it is the same. The second model also imposes a restriction that effect of dB is linear.
Edit 3: Sadly, I don't think the answer is so simple in this case. Here's another way to look at the data that may give you some guidance. The error is represented by hue/temperature (red/hot means higher error, blue/cold indicates lower):

As you can see, the error gets worse towards the top of the range for freq, though the effect is somewhat offset by higher dB and it drops off. There are also some strange cycles too. Maybe this all makes sense if you understand the physics better, but I am not sure how much you can disentangle with this data.   
