How to isolate impact of event in a product's lifecycle? I'm trying to figure out how a single event affects sales numbers of a song. For example, see what the effect of being featured in iTunes store compared to songs with comparable previous download numbers.
How should I go about modeling this question?
Thanks!
 A: One approach is to identify a pool of songs that are similar to the focal song, where all the songs in the pool have similar sales characteristics, and thus similar pattern of sales -- i.e. sales are highly correlated. The validity of this hypothesis can be tested rigorously with some variant of generalized linear model (GLM) to predict the sales of the focal song based on the sales of the pool songs.
You'd identify the pool by some clustering or classification procedure. Might be best to start with something simple and then only make it more complex if your GLM estimation doesn't yield good results.
Effectively, the pool becomes a control group for your event model.  Let's say you have 100 songs that have experienced the event (being featured in iTunes store).  You'd create 100 pools of similar songs, or maybe fewer if some of the featured songs are in the same pool.  You'd then estimate a GLM that includes a dummy variable for "iTunes_featured".  It would also probably be important to include a variable for seasonality, since seasonal variation could be very high.
EDIT -- Here's an academic article that describes this method, provides much more detail, and provides support for it's validity:
The Central Role of the Propensity Score in Observational Studies for Causal Effects

This approach is fairly generic and would work OK for many sales-event studies.  But there's a snag -- music can be a "hits" business, where a small number of songs/performers can "explode" in popularity, perhaps triggered by events like getting featured in iTunes store.  The GLM model above won't be adequate to model the cascading, positive-feedback processes involved in "hits".  You may not have any "hits" in your 100 song sample, or you might have one "mega-hit", or some "moderate hits" (Like all power law phenomena, the cascading mechanism of "hits" can produce sales of all magnitudes, not just the largest "mega-hits".)
For a good discussion of these complexities, see the book How Hits Happen (the author is a colleague of mine at George Mason U.).  Here's a slide summary of the book.  Also, here's a recent Wired article on predicting hits in the music business.  A key quote:

"...random choices of early users combined with social cues makes the market highly unpredictable."

What to do? I'd try to estimate a GLM model for "normal iTunes bump", excluding any and all "hits" from your sample set.  Then, I'd have a separate sample set of songs that did became "hits" (even minor hits).  Some will have been featured in iTunes, and some will not.  Here, your goal is to estimate the probability of a song becoming a "hit" if it is featured in iTunes, compared to not being featured in iTunes.  But this goal may not be feasible, given lack of data and confounding factors. To get more than a lower bound, you'd probably have to write a stochastic simulation and then fit the parameters of the simulation so the simulation results mimic the empirical data (especially the structure of cascades). This approach is discussed in the book How Hits Happen.
