Simulation Uncovers the “Hawthorne Effect” in Distribution Center Operations Planning

Bruce Pic

Bruce Gladwin – Vice President, Consulting

The past 30 years has seen some dramatic changes in the ways we consumers purchase everything from books to batteries, cameras to clothing and TV’s to trampolines. The changes are not only in the products themselves, but in the channels of delivery. For example, when I was growing up in the 70’s and 80’s the only chain businesses that I can remember were Sears and McDonald’s. Even then, you had to live in a fairly good sized town to have either of these stores. Today, any town of any size in America has not only a McDonald’s, but a Wal-Mart, Home Depot, Starbucks, Olive Garden, etc., etc. Yes, the chain stores have certainly replaced the mom and pop shops that were so prevalent just a generation ago. This change has led to an explosion in the number and size of distribution centers around the country and the folks running these DC’s are constantly looking for ways to get more product through their systems in less time and for lower cost.

I recently had the chance to work with one of our customers who was modeling a proposed layout change to their DC floor plan. The change was already being piloted in a particular DC, and a model was being developed to test the merits of the new layout under various product demand and mix scenarios. The early results from the real-life pilot were promising, showing that relieving forklift congestion could increase the throughput of the system, even if the forklifts had to travel longer distances to avoid congestion. As such, a change was about to be implemented in each of this customer’s DC’s across the US, at the cost of over $2 million.

My contact had the opportunity to use simulation to determine the expected improvement of the new floor layout in other DC’s throughout their network. To his surprise, the model predicted exactly the opposite results! Not only would the new layout actually reduce throughput, it would increase the operating cost of the system by requiring more forklifts and, hence, more drivers to maintain the previous level of productivity. How could this be since the real-life pilot was showing moderate (but not stellar) improvements? Enter something called the Hawthorne effect…

Stated simply, the Hawthorne effect is a temporary improvement in productivity that results when management pays greater attention to an established process. In other words, what gets measured gets done better and faster than it was before. In this case, the fact that the pilot area was the focus of management attention at this DC meant that the workers were unconsciously improving their performance. The simulation, on the other hand, was completely unbiased and used the same assumptions regarding forklift travel speeds, put and pick times, and operator work habits across both scenarios.

In the end, the existing system proved to be more efficient because the frequency and duration of forklift congestion events was less detrimental than the additional travel distance in the new layout that was required to eliminate those events.

In summary, whenever human operations are a significant part of a production process, consideration must be given to the methods in which physical simulations, i.e. pilot tests are performed on the floor. Otherwise, the Hawthorne effect just may result in a wrong (and costly) change to procedures.

4 thoughts on “Simulation Uncovers the “Hawthorne Effect” in Distribution Center Operations Planning

  1. Logistics modeling isn’t easy. What one thinks isn’t biased is more than likely highly biased. If the delays, etc. that were built into the model are the result of adding values from subordinate events/actions that have been measured separately, the resultant value or values may be too inclusive, i.e., too pessimistic. Alternatively, if the model values were derived by removing delay, i.e., by exclusion, then the model values may be too optimistic. A good approach is to make runs using values in both ways and take the average, assuming that it isn’t too costly to derive both. As wargaming has demonstrated, it is very difficult to capture the vagaries of human actions; and consequently, a model is never exact. And because of this, wargamers will warn modelers to not read to much into the results of their models. Rather, as this article suggests, use the model and the results to prompt questions and gain insight.


  2. There is one other thing that has to be considered when trying to determine the “average” time necessary to complete a human-performed operation. If you ask the operator for the average time required to perform a task, their response will usually be a “most likely time”, not a mathematical average of many observations. This is important in simulation because it points out the relevance of using a Triangular distribution over the more familiar Normal distribution. A Triangular distribution use three parameters which include the minimum, maximum and most likely. A Normal distribution requires several observations to determine the mathematical average and standard deviation.


    • You are quite right about human responses and using the triangular distribution. I’m curious as to how the flow in DC was modeled. My assumption is that a discrete event simulation approach was taken and that it was modeled as a network of nodes with each node representing an intersection of nominal paths through and around the DC (somewhat like the runway and taxiways of a modern airport) and each node having a queue and service time associated with it, e.g, the delay associated with waiting for a forklift to pass through the intersection or the amount of time to collect items from stock. From such a model, one might be able to deduce that non-stock related nodes with a non-zero queue size implies congestion. Non-zero stock related nodes might suggest spreading out and duplicating high frequency items to reduce the congestion. But I’m curious how how such a two-dimensional (possibly three-dimensional) problem was approached since typically queuing problems are dimensionless in terms of space.

      Likewise, from the data collected was the floor design, specifically the placement of stock, refined by the use of a Linear Programming Model? I would think that the application of multiple objective functions (assuming that conflicting objectives exist) could potentially offer an improvement of the floor design, but that’s just conjecture.


Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s