Help needed to model the impact of a long term promotional tactic on sales

[user552]

Hi,

I am working on a study where the objective is to calculate the impact of a promotional tactic (coupon redemption data) on the sales that took place some time back, for this I have the sales information of the subjects (around 800) who used the coupons along with their coupon redemption data and number of calls given to these subjects as a part of the regular promotional work all a month level for 12 months.

I generally use a Test-Control approach to estimate the impact of promotional activities but not experienced modeling relationships for panel data.

Can some one provide a direction on how to approach this, based on the literature review I've done till now I feel that best approach is to use proc panel, but I am not sure which model to use given a wide variety of options available, below is a snapshot of the partial data.

Sorry if this sounds very naive, but it would be really helpful if some one could refer to any documentation that can help me here.

Thanks!

Subject Number	Month	Sales	Call	Coupon redemption
1	1	4000	1	300
1	2	5000	1	300
1	3	5000	2	500
1	4	5000	3	300
1	5	4000	0	300
1	6	3000	0	200
1	7	3000	1	200
1	8	3000	1	200
1	9	4000	1	200
1	10	4000	2	200
1	11	5000	2	500
1	12	3000	2	500
2	1	2000	4	100
2	2	2000	2	100
2	10	5000	2	500
2	11	4000	2	500
2	12	4000	2	500

[ets_kps]

A very good question and thanks for bringing it to us.

You indeed have panel data and it is likely that PROC PANEL contains some estimators that will help you find the impact of the campaign (both calls and coupon size). Since it appears you only have information on customers using coupons, you will likely only be able to estimate the impact of changing the coupon size. But that should be ok.

Now for the hard part.  If you are willing to assume away any autoregressive process and are willing to assume the impact of the program is say, 3 periods, then create three lagged values of coupon, include those as regressors and estimate a two-way fixed effect model. The general syntax could be found here. SAS/ETS(R) 13.2 User's Guide

However, if you are unwilling to make that assumption (no serial correlation), then you have to move to a more complicated estimator to get consistent estimates.  An example of that can be found here. SAS/ETS(R) 13.2 User's Guide

As a first, pass I would try my first recommendation.  Best of luck.

[user552]

Hi Ken,

Thanks a lot for the reply! I will get started on these.

I have couple of questions here

• In the two way fixed effect model, how would I calculate the total impact of coupons (we get 4 coefficients, three for lag terms and one for current coupon)?
  • are the coefficients additive? (α1+α2+α3+α4) or should we use some decaying exponential form for the coefficients corresponding to the lags 
    
• How can I calculate the total impact (annual) of the program, from the regression model I would get the impact on average per subject for a period of 4 months, can I project it to 12 months? 

Thanks!