Causal inference is a simple idea, right? You do this and that happens. It winds up, though, that in science, making a causal inference about the world is tricky. Saying that when you do this, that happens is one thing. But understanding exactly why it happened and whether doing this is what caused that to happen is a difficult puzzle. In the organizational sciences such as management and industrial organizational psychology, the things we study are complex and due to multiple factors, so isolating a cause is extremely difficult. That doesn’t stop us from trying, as we apply complex statistical modeling in an attempt to isolate causal factors that allow us to determine what drives important outcomes, such as productivity or employee well-being. Unfortunately, the most complex statistics in the world aren’t able to provide more than a hint at what might be a cause of something. You must manipulate something to make a causal inference.
Causal Inference in Science
Science is a systematic means of discovering how the world works. We conduct scientific studies that enable us to better understand the world, and to figure out how to achieve our objectives, whether it is to cure disease, grow food, or keep ourselves entertained. Experimentation is the main tool scientists use to figure out how things work. Of course, it is possible to shed light on the world merely by observing it, but at the end of the day, if you want to know if a particular drug will cure a disease, or a fertilizer will increase crop yields, you have to try them out to see if that drug or that fertilizer is effective. In other words, we manipulate something to make a causal inference. We conclude that the drug cures a disease or fertilizer increases crop yields because we watched it happen.
Manipulate Something To Make a Causal Inference
The explanation of how manipulation allows causal conclusions comes from philosophy of science. As explained by Illari and Russo, the manipulationist point of view is that you “wiggle” the cause to see if the effect then wiggles too. This links the proposed cause to the effect by manipulating the cause first and observing the effect later. This is what we do in an experiment–we manipulate our independent variable first and measure the dependent variable afterwards. Depending upon the situation, we might measure the proposed effect before and after (a pretest-post test design) or we can conduct a randomized trial and compare those who get the proposed cause (e.g., a new drug) with the control group that did not.
Unfortunately, things are not as simple as this seems when it comes to drawing a causal conclusion. There are several limitations to keep in mind. The proposed cause we study in organizational sciences and many other fields are theoretical notions. If I want to know if leadership behavior causes engagement, for example, I have to figure out what the relevant behavior is, and then how to manipulate it. I might do it with a training intervention where I ask all managers to attend a training program (I provide a different example here). I then have to meet three conditions.
- The change in engagement is due to the change in leadership behavior. In other words, there is nothing else going on that might have been the real cause of the change in engagement. For example, at the same time the training was occurring, the company announced a new bonus system that is the real cause.
- The training course only affects leadership and not engagement itself. It is possible that the training has a direct effect on employees that has little to do with their supervisor’s leadership. Perhaps upon learning that the company is investing in leadership training, employees conclude that the company cares about them, and this increases their engagement.
- The training course does not affect anything else that might affect engagement. The training course might have an impact on things other than the intended leadership behavior. Perhaps it increases supervisor engagement rather than their leadership behavior, and that engagement is infectious, leading to employee engagement independent of leadership behavior.
Research Is a Complex Puzzle
The randomized experiment where you manipulate potential causes is considered the gold standard. They are hard to do in field settings such as organizations where what we want to study occurs in a complex environment. However, it is often possible to conduct research on what is occurring naturally. For example, data can be collected before and after some planned intervention such as leadership training, to assess the impact on outcomes of interest. For academic researchers, this means developing collaborative relationships with companies that will provide data in return for helping them figure out what does and does not work.
Most organizational research mainly provides insights about what is related to what. We have a large body of research showing hundreds of correlations among variables of interest to both practitioners and researchers. Little of this research allows us to draw the causal conclusions we would like, no matter what statistics are used or claims are made. If you want to be able to make claims about how to change a desired outcome, you need to manipulate to make a causal inference.
Image by Chat-GPT 4.0. Prompt: “image of dominos set up with a finger just touching the first one before it falls on the second”, “do in a cubist style”, “brighter colors”
SUBSCRIBE TO PAUL’S BLOG: Enter your e-mail and click SUBSCRIBE