Omitted Variable Bias: The reason why models fail to deliver

In standard econometric experiments, there is a test for omitted variable bias; if an independent variable is missed, which if included in the regression would have changed the outcome, such a variable better be included. The examples of this abound from selling to setting up of business models to mundane things like deciding on which courses to opt for getting jobs. Let me start with the more mundane ones.

 

Imagine a student trying to figure out which courses raise his chances of getting a job. If the job he is looking for involves technical work, the one to one relation could be easily established and the obvious focus would be on the degree of absorption of a technical course that is related to the job. If the student enrolls for more than one such course the odds of getting the job changes for the better. Suppose the job actually needs apart from the command over the technical aspects also exposure on analytical skills, perhaps the student needs to sharpen his skills in quantitative techniques. Most jobs also require managerial skills as well, so the simplicity of the model where a dependent variable was linked to an independent one through a linear relationship gets into the multivariate function. Most students suffer from an omitted variable bias, which is all the more reason why some are better prepared for a job than the other.

 

Think about sales promotions and the related sales. There seems to be a direct correlation and this is a simple Y= mX + C, if X= Sales promotion and Y= Sales. But imagine that we have a situation that after repeated promotions, the sales does not go up at the same pace as it did in the first few rounds (in fact repeated promotions could have a debilitating effect); when to withdraw away from promotion could become important in such an event. There could be another independent variable Z= price discounting which impacts sales and the interplay of promotion and discounting could well be acting in concert. The challenge is that the simplicity of the regression with Y= mX + C, could be quite different with a multivariate non-linear environment and the estimates of Y on the basis of the different independent variables (if some of them had stronger correlation) could vary very differently. Salesmen suffer from an omitted variable bias when they ignore some variables while constructing their sales function.

 

In the world of business we have examples a plenty. Business models are built around a large numbers of unknown; some of the estimates could be regressed over experiments. When experiments are done and the test results are validated, the confidence is lifted to embark on a larger canvas. But there is always this possibility that we could be trapped in the tyranny of an omitted variable bias.

 

The best example of an omitted variable bias is the shale gas experiment in U.S. When the structural shift of natural gas prices had already happened in U.S. the shale gas experiment came with an even greater promise. If one follows the post 2008 crisis, the natural gas prices moved on a trajectory that moved sideways and lower, many actually attributed this to a structural shift in natural gas prices in U.S. The hydraulic fracking techniques improved the chances of taking out entrapped gas, in dry or wet form, even further, so the supply side had an even bigger boost. It was assumed that this supply would bring in a large shift in the economic balances in manufacturing in U.S. and a plethora of investments in the associated industries would take the overall GDP to a higher trajectory. The rise in natural gas output growth (together with the impact of shale gas) was 33%, which was seen as a new industrial revolution in U.S.

 

The actual reality was found to be different, the economic impact in GDP terms attributed to shale gas boom has been found to be rather muted in the region of 0.04% incremental GDP growth over natural gas (0.84% against 0.8%); the recent oil shock and its impact on shale gas comes on top. So where did it go wrong?

 

It was a classic case of the omitted variable bias. The natural gas supply was regressed with economic output growth through higher consumption by industry as a direct linkage. A few things were missed (possible omitted variable bias):

 

–          Per capita energy use in U.S. was falling

–          Gas prices for household was falling even before shale gas boom

–          Household electricity prices were rising

–          Industrial electricity prices were rising

 

So while the import of oil came down steadily from 11 Million barrels/ day to 8 million barrels per day, no one actually saw that the consumption variable was changing as well as so was the basket.

 

 

Omitted variable bias in the case of logistics is very common as most of the cost function misses to take logistics as an independent variable. No wonder India’s logistics cost is the highest in the whole world at 13% of GDP, which is as much because of lack of infrastructure as due to missing to identify that logistics could be costly and that wastes could be very high.

 

Today’s E-commerce is also the case in point. Here the model assumes that customer’s choice automatically factors in the shipping costs, but more often than not the customers after pushing items to the cart realize that the total shipping costs could become quite different as the cart grows in size. The cost of logistics for e-commerce as compared to retailing on a like to like basis would be higher, but there could be ways of bringing that down through strategies that allow flexibility to the customer vis-à-vis shipping costs. The customer can actually be willing to pick up an item from a place instead of home delivery if the costs make an allowance for it. The integrated technology is a clear winner on this to be able to connect customer preferences on an on-going basis. The reverse logistics is one more killer if the process is not clearly laid out such that the customer understanding matches with the ability to execute it.