Logit regression to evaluate real price test

#1
Dear statistically inclined people,

I'm asking for your help in a real world price test evaluation. I've had two years of stats and econometrics back at uni a couple of years ago and am trying to scramble all my knowledge together - would be awesome if you could help me a bit :).

[Here a bit of background info for anybody interested]
I am working for a startup in the Netherlands and we just ran a price test on an online webstore for the last couple of months.

The setup is as follows:
People visiting the online store are randomly put into either of 5 price groups

A - cheapest price offered by any of our competitors for any given product
B - + 7,5%
C - + 7,5%
D - + 7,5%
E - + 7,5%

The data currently consists of around 2000 unique user ID's and looks somehow like this:

User ID ; Price group(A/B/C/D/E); Conversion (0/1)

I'd now like to run a logit regression on the data so that I can say: A consumer paying 7,5% more than the cheapest price on the market (group B) is X% less likely to buy than a consumer seeing the lowest price (A).

From what I remember I would let the regression be:

ln(conversion) = ß0 + ß1(B) + ß2(C) + ß3(D) + ß4(E) + e

where the the independent variables are dummies and taking the lowest group A as the base group.

My questions are:

- does that make sense from a practical perspective?
- anything else I have to look out for/ test? (autocorrelation, homoscedasticity etc)
- anything else you would do with this set of data that would be interesting from a business perspective?

Really grateful for any help! :)

Andre
 

terzi

TS Contributor
#2
Hi Andul,

The logistic model makes sense and I see no problems with it, yet, since you have only those two discrete variables, you may obtain similar results with some simpler procedures. Remember that a model requires you to check assumptions such as linearity in the parameters, an appropriate distribution, etc. Assuming you want to block according user's ID, you should probably read about the Cochran's test, which I think is quiet appropriate for your problem.

Regards,