Regarding “accuracy or precision” – the wikipedia article makes a big issue of lot of terminologies; accuracy is simply % of accurately classified obs (the tabulation they give is for binary classifiction). So total predicted in the diagonal of the “confusion matrix” / total sample.

The question of trade off between accuracy and precision is legendry; however the graphical example seems to give the impression all problems are continuous targets; if it is a question of classification, where the decision is whether one hit the target or not, then the graphics does not do justice to the problem.

I like to kindle thoughts; this is not a retort! Great group…keep going.

I have some comments on attribution analysis; comes later.

Often I hear people say this can’t be done because of media silos etc. but there must be *someone* in charge of all Marketing efforts who can make this happen. Just takes some substantial guts.

Same old problem, not of measurement vs practicality: Turf protection and the budget $$$ associated with silo’d channels. Multi-channel campaign measurement & allocation has traditionally been a touchy issue in any environment (advertiser or vendor space) where competing spend tactics all desire to lay claim to the revenue recognition. So, instead of a realistic value proposition to attempt to accurately associate comparative value, it usually results in a zero sum game to trump the other guy’s revenue momentum.

Yesterday, it was TV & direct mail. Now it’s display & paid search. Same arguments in both cases.

The real business question you adeptly imply in your post, “why spend limited capital & scarce resources attempting to optimize an illusive & deceptively futile measurement problem?”

Doesn’t it make far better practical sense, to put that same level of energy into simply improving each of those separate silos to their maximum potential – especially for large branded advertisers who have fragmented channel responsibility themselves?

From my experience in the industry for the last 11yrs Allocation is the Model and attribution is the real-time application credit of the action to the session or series of events during a period time defined by the allocation model.

I apply a global allocation model to my system and attribute the credit value as it happens to the event. Allocaiton Most recent /Expires 30 days/Attribute credit to the event within my 30 day period

Just my two cents — but the assumption is correct in that not many people get this and not many use the terms correctly.

It seems to me that so many people are asking how we should model attribution effectively, few – other than your good self – are asking whether we should try.

As you so rightly say, just because a particular ad shows up in a user’s path to conversion, that doesn’t mean that ad had a positive effect, negative effect or no effect whatsoever – particularly given the cookie-flooding practices so common and the general issue of, for example, banner blindness.

However, with all that said, your last comment – about the ability to follow just one user’s path to conversion – points us to what makes online tracking technologies (like our own brilliant TagMan) so interesting. Online we do not need to plan against a defined (and fictitious) ‘audience’ – we can take an individual conversion and work backwards. Do this enough times and common sense tells us what’s happening.

Then we can test our theories and THEN we can make decisions of allocation and – if you so choose – attribution, though in my view allocation is just the ultimate form of attribution since, if I don’t allocate you any spend, you stand no chance of being attributed with any sales.

Like I say, great post.

Best,

Philip Buxton
Marketing director
TagMan

When you have actual customer behavior data, and are acting on individuals and not “the masses”, the need for mix modeling is substantially reduced, IMHO. Goes hand in hand with whether you are a big spender on untrackable media. If so, then have at it!

The problem with mix modeling now is that most of it is in CPG, but many other businesses can’t do it, and don’t get it. When they do, look out!

In the last 10 years many Fortune 500 companies such as P&G, Kraft, Coca-Cola and Pepsi have made Marketing Mix Modeling an integral part of their marketing planning.