Dear stack exchangers,
in the course of an internship at a microfinance bank in Tajikistan, I have been tasked with assessing the efficiency of different marketing expenses. I will try to expose the situation:
-I have monthly data (15 months)
-the individuals are cities (7 cities)
-as a regressand/dependent variable (called Y), I will be using the monthly variation in clients at the bank (three analyses: global, for deposit products, for loans. I may use the number of new clients, if it is avaible.
-my regressors/independent (called X) variables are the monthly expenses in one marketing tool (newspapers, billboard, leaflets, website ads, tv+radio), per city.
--> I'll have 105 observations to treat
My objective is to evaluate which media brings the most clients per Somoni spent. To this end, using STATA, I will conduct a panel analysis, using errors ajusted for clustering, and a fixed effect per city (the number of new customers per months may be affected, among other things, by the population of the sector).
The question that motivated this post is: how could I include spending that is national (e.g. TV/radio) into my equations, and obtain a meaningful coefficient for it?
If I include the value of national spending in each equation, then the estimated coefficient would give the average impact of national spending on Y at the city level. But national spending has an impact on all cities at the same time.
Because the coefficient is an average, I guess multiplying it by 7 (as there are 7 cities) will give me the global impact of national spending on the number of customers. What would be the problems with such a solution? What alternatives are there?
I also have two subsidiary questions:
-A lot of Xs take the 0 value. And in a lot of cases, an observation has only one X different from 0. Does this pose a particular problem, apart from those of statistical significance? (I don't think so)
-Finally, I have a few questions concerning the design of my study. Do you think that such a design will yield exploitable coefficients?
Is the exclusion of the interest rate variable a heavy problem? (I might not be able to obtain that data)
Should I worry about AR(1) problems? (maybe the number of new clients is affected by network effects, so Y(t) depends on Y(t-1))