4

I am working on something to test advertisements. I have 3 independent variables I want to test (with mixed numbers of variations of each variable), and I would like to find the best combination of the three by looking at their effect on a single dependent variable (the percentage who purchase).

TITLE     IMAGE  DESCRIPTION  (Percentage who purchased)
Title1    Image1 Decription1  .05%
Title2    Image2 Description1 .02%
Title3    Image? Description1 .08%
Title4    Image? Description1 .02%

There can be any number of variations of each variable, sometimes there will be 10 images, other times there will only be one, etc. My goal is to find out what Titles/Images/Descriptions affect the purchase percentage.

  • How can I test and analyze something like this?
  • Is it possible to not have to test every single variation and still find the best combination?

(I've read about chi-square tests but I don't know if they are the proper fit)

Jeromy Anglim
  • 42,044
  • 23
  • 146
  • 250
  • There might be a bit of text analytics involved. Are "Title" and "Descriptions" true independent variables, or do they just supply information about image? – Ralph Winters Mar 25 '11 at 13:28
  • @ralph-winters If I understand correctly they are all independent. Some will work better with others, but they all have an affect on the Percent who purchased. Image and title usually have the biggest impact though. –  Mar 25 '11 at 16:56
  • Can you supply some real-type examples of the columns? – Ralph Winters Mar 28 '11 at 20:21

1 Answers1

4

In marketing-oriented statistics the analysis you need is called conjoint analysis. You construct a number of product "scenarios" which are various "mixes" of attributes such your 3 attributes, each being allowed to vary across some "levels". The Conjoint will tell you what's the best scenario and what are "utility" coefficients for each of your attributes - these allow to predict popularity of scenarios which you didn't test. Algorithmically conjoint analysis could be based on regressions or multidimensional scaling or other technique. Look for software performing conjoint analysis.

ttnphns
  • 51,648
  • 40
  • 253
  • 462
  • (+1): Did not know about conjoint analysis before (just read this article:http://en.wikipedia.org/wiki/Conjoint_analysis). Question: So basically one collects preference/purchase data for all or a well chosen subset of combinations and then performs an analyis afterwards to determine the variable importance. Is this correct ? – mlwida Mar 25 '11 at 12:37
  • Yes it is. There are many variations of the analysis. Some, so-called computer-based (or adaptive, or on-the-fly) conjoint exposes a respondent with not all or random scenarios but with scenarios that will yield most valuable results, based on his/her previous responses. – ttnphns Mar 26 '11 at 11:05