### The Concept of Causation

Causation in ROI, measurability, and accountability measurement discussions is almost always never mentioned. And if it were ever mentioned, it is in a manner that is simplistic and in a ‘by-the-way’ fashion. A number of media and ad companies claim that ROI is as simple as running models after the gathering of data and transformation of these data and ﬁtting them in models - most likely linear regressions using OLS or MLE, econometric modeling, or worse, how the numbers "fit" against norms. Fits are measured by R-squared and areas-under-curves.

But causation is far more than that.

And that is probably why causation is rarely mentioned: it is far more than that - and is far more complex.

Because causation is very closely related to correlation: a strong correlation between two variables, for example, would potentially imply a causative relationship. Two variables - X and Y - that are strongly

correlated could suggest a potential causative relationship: that X causes Y. But the reverse may also be true: that Y causes X.

Alternatively, there may be an even deeper explanation - a third (or even a fourth) variable that explains the movements across the two variables. Simply put, there might be a variable, say M and Z, that essentially a ‘catalyst’ between X and Y, such that X ‘increases’ the value of Y because there is also an upward change in the difference between M and Z.

#### The bottom line is simple, really: Correlation doesn’t necessarily suggest causation.

However, because correlation is far easier to explain and calculate than causation/cause-and-effect, we stop at correlation. And worse - assume that whatever we are looking at is "causative" rather than "correlative".

### Which is “Better”: Correlative or Causative Modeling?

Nothing is technically better than the other - it’s really dependent on what you want to achieve.

**If the end goal is simply to ﬁnd out the nature of the relationship between two variables - say, advertising spends and sales - then, correlative modeling may be sufﬁcient**. One can go as far as saying “A dollar spent in advertising appears to correspond with a 0.2 increase in sales revenues, assuming a linear model”.

But that is only as far as one can go in correlative research.

There are no further explanations that can be gathered from the these research - no matter how many data points and no matter how robust the sampling methodology and framework is.

If the end-goal of the project is to explain the impact of advertising on sales and revenues, then causative modeling is what we ought to conduct.

Note the subtle difference between the two objectives - so subtle that even analytics people seem to forget it:

- The correlative approach simply wants to know the presence or absence of a relationship - if and when it does exist -and whether such a correlation is linear or non-linear.
- The causative approach aims to explain relationships between two or more variables, and underlying, intervening variables: It presupposes that there is a relationship - a hypothetical, theoretical relationship - between the two variables. It seeks to explain - not merely to describe - the movements between the two variables, and ultimately arrive at an answer to the question “Why?”.

Neither is better than the other - **it all depends on what one wants to achieve in the ﬁrst place**.

One is simply different from the other because simply, correlated variables do not necessarily indicate causal relationships.

### The Perils of Correlative Research

Simply put, the perils of correlative research is when one starts to look at the data as if they were causative.

Let’s take the case of advertising spends and sales. Assume that a correlation of 0.70 has been found between 104-weekly data points. Can we conclude that advertising spending caused the improvement in sales? Not necessarily.

Whilst the correlation - and the R-squared - is close to 0.50, it is still is not sufficient to conclude that increasing advertising spends means increased sales.

Similarly, if another set of data points actually see a correlation of 0.30 - low by standards of correlative research - between sales and advertising spends, we cannot simply say “advertising efforts are wasted because they did not contribute to sales”.

Just to reiterate: **There is nothing wrong with correlative research; what is worrying is when we look at correlative research results as if they were indicative of causal relationships. It is when we start misusing these results that we get misleading results. **

Thanks to Rob Ireton on Flickr.Com for the photo.

### Causative Research - and Its Perils

Causative research is far more extensive and time-intensive than correlative research. Causative research indicates a deep-dive into theories, hypothesis, and testing.

It goes beyond pure equations and r-squared (or other indicators of goodness-of-ﬁt) into areas that such as model-testing (through structural equation modeling and latent modeling across time).

Causative research requires a deeper understanding of the variables - thereby making the equations far more complex.

**But the beauty of causative research is that it can explain - or at least attempt to explain - why and how things work. **

In the case of ad-spends versus sales, an inquisitive, curious mind might not just measure these two variables - but also intervening variables: recall, executional cut-through, correct-brand attribution, message-relevance and afﬁnity, likeability, and believability of the message.

This curious mind might also incorporate the characteristics of the respondents themselves into the model to determine if being female or male, or if being under-20 or above-20, could potentially ex-

plain the movements between ad-spends and sales. She would also be interested in short-term, long-term, and base-effects of previous campaigns - and explain why these remain to be so.

Moreover, an inquisitive, curious mind bent on ﬁnding out “why” and “how” will dive deep into academic literature, best practices, and other models that have been uncovered by others in the past to test if these models and hypotheses ﬁt her own set of data.

### Causative Research has more explanatory power...

Causative research is no panacea. In fact, there will be lots of discussions from causative research - but the main discussions will be centered, disciplined, and rigorous. Because causative research demands rigor:

- Causative research ﬁrst deﬁnes the problem. A clear definition of what the project is all about is the first step. Are we looking for indicators of correlations or are we seeking to explain why and how things move?

- Causative researchers then trawl through studies of past cases, in search of potential underlying, latent, and intervening variables that need to be taken into account - either in the research or in the explanations. These variables may or may not be included in the model itself.
- Causative researchers then come up with a potential model that could explain the “dependent” variable - together with the underlying, latent, and/or intervening variables.

(In the case of advertising: If all that matters is "you advertise, you get sales", then we would have discovered the answer to the age-old question, "how do we increase sales?" We would have found the answer to all marketeers and sales executives' problems: "Just advertise".

- They then come up with a research design and a data-gathering strategy that is appropriate for the hypothesis or hypotheses to be tested.

I stand on the belief that there is no single research design nor a single data-gathering strategy that answers every single marketing problem there is. Any research project or learning project needs to be customized - against the problem and against the hypotheses derived from the review of past cases, academic papers, and other sources. - They then verify the results - applying statistical techniques including correlative techniques (if the model demands such), linear and non-linear model-ﬁtting, response-sensitivity, structural equation/conﬁrmatory factor analysis, and others.
- Causative researchers then either accept, reject, or revise the initial hypotheses that they have come up with and come up with explanations and recommendations for further ﬁne-tuning of the research.

No one research can answer all the questions in the world. Research studies build on the results of previous research studies.

**Causative research takes time, effort, patience, and curiosity - a deep-seated belief that there are stories behind numbers. And that numbers are typically shy - and secretive of their stories. It takes some teasing to see the stories they tell. **

Your point on the rigor involved in causative research is right on.

Thanks for this. Excellent post.

Posted by: Ian | 27 February 2009 at 16:17

Hey Ian - thanks very much for the comment. I do believe that it's all about rigor - and well, knowing what we want to achieve. These days when information is so easily 'affordable', we fall into the trap of just simply running these numbers through stat procedures without really thinking through. Thanks and please do come back! :)

Posted by: Philip Tiongson | 01 March 2009 at 23:35