PDA

View Full Version : Advice Needed w/statistics



stanl
03-10-2007, 08:20 AM
I have been asked to use Excel to graph a 'trend' with 3 sets of values measured at different times from 1982 to 2006.

1. dated measures of wholesale % increases of Product
2. dated measures of retail % increases of Product
3. yearly measures of CPI % increases (Consumer Price Index)

the cumulative values of 1-3 graph nicely (and trend upward in general:rotlaugh: ). However, to avoid an apples/oranges objection from any party viewing the graph, could I apply a 'goodness of fit' measure (I am soooo statistically out-of-it) with say a .05 confidence level.

If so, would I make whosale a function of CPI, then retail a function of wholesale... the hypothesis being that the retail increases were a function of wholesale increases and not due to 'whims' - hope this makes sense. Stan

mdmackillop
03-10-2007, 08:59 AM
Hi Stan,
Doubtful if I can help here, but if you could post sample data I'd be interested to see what can be done.
Regards
MD

tpoynton
03-10-2007, 09:07 AM
I'm more familiar with the stats than your examples...it sounds like you want to use regression, predicting retail increases from wholesale increases, and CPI predicting wholesale? The data analysis toolpak will do the regression, although I have to admit I typically dont use Excel for regression. I've tried and failed a few times.

an alternative (but perhaps more appropriate) regression model would be to use CPI and Wholesale as predictors of retail sales, which is what I think you are trying to do. The results of the regression will allow you to pull out the relative influence of CPI and Wholesale on retail sales.

the 'sticking point' here will be how your data are structured; depending on how they are set up (how many points in time are captured, in particular) will determine whether or not regression is an option, and if your 'sample size' will be large enough to determine anything with confidence...

EDIT - as MD asserts, having the sample data would be helpful

stanl
03-11-2007, 05:12 AM
Thank you both. The data is confidential so I need to set up the table and graph as a 'widgets' example, which I hope to get around to early this week. Stan

stanl
03-11-2007, 09:03 AM
I'm more familiar with the stats than your examples...

Hope you didn't interpret my post as "this guy is Stupid!" - by suggesting Regression, you are assuming parametric statistics [If I may defer a bit: the assumption that normally obviates all psychological studies IMHO:bug: ]

The question, perhaps: are individual price increases the product of a normalized market and therefore parametric - yeah, say like gasoline princes increasing .50 in the last month - :dunno

Of course, to quote Mark Twain " There are lies, damned lies, and statistics".

tpoynton
03-11-2007, 10:18 AM
No offense taken :)

Since most of my work with statistics is in the realm of psychological variables, my slant is that their 'underlying properties' are in fact parametric, which is certainly a debatable point. I'm no psychometrician, or statistician for that matter, but I do dabble.

Most parametric statistics are fairly robust to violations of the 'equal variances' assumption to which I think you are eluding, and software packages such as SPSS will generate statistics that assess the degree to which that particular assumption is violated.

Also, there are methods of calculating regression for non-linear relationships. For example, a 'traditional' linear regression would likely not pick up on the relationship between test anxiety and test performance. A little bit of anxiety is a good thing, but too much anxiety will lead to lower test performance. This sort of curvilinear relationship needs to be assessed differently.

You could skip all this, and (likely) apply a chi square goodness of fit test. chi square is a non-parametric statistic that has no assumptions of normality, which is the benefit. The drawback is that it is less powerful at finding differences that do exist. On a positive note, this means that if the chi square indicates good fit (is significant at p < .05), you can be pretty confident that it is not a pack of lies...HOWEVER, if memory serves, doing this sort of test does assume the data are uncorrelated, which does not seem to be the case for all the variables here.

regression is my swiss army knife procedure. I'd love to know of other approaches to helping you present the info in meaningful, honest ways, but I am certainly biased! You might consider sticking with things familiar to the audience...sometimes the logical argument is better than the statistical one.