It seems obvious that Slovenia Email List 1 with its 50% success rate is better than Commercial 2 which has a 40% success rate. You give him a bigger bonus to reward him for his performance. Only there you have it, one day Commercial 2 comes to you and tells you that he feels you are treating him unfairly. In addition, he claims to have a better success rate than Commercial 1, which you well rewarded, but not him. You retort that you are, on the contrary, very fair. Besides, you show him the painting to justify yourself. In reality, we see that Commercial 2 has a higher success rate than Commercial 1:
In fact, if you are looking to win your next move, you have every interest in mobilizing Commercial 2 rather than Commercial 1. The numbers are there, and you can verify that it fits with the 2 tables. It is in cases like this that we find the Simpson paradox . It can happen when you try to generalize too much and miss confusing factors . It is a factor which influences both the explanatory variables and the result that one seeks to explain. In the example I gave, we tried to measure the performance of salespeople with only 1 success rate. This rate hides the reality of disproportionate samples.
What really happened
The confounding factor is the type of commercial proposal: response to a call for tenders or response to a request from the network. I imagined a few frequent cases of confounding factors. Do not hesitate to leave a comment with yours in your field. In our example, the type of answer strongly influences the chances of winning. You can benefit from segmenting your indicators. It is the business sense that counts here: proposals with a low success rate vs. proposals with a high success rate. You don’t always have to do this, especially if you want to encourage a good overall success rate. But you have to make the decision consciously and not suffer it and make it suffer.
Knowing about this paradox has made me more wary of the numbers I see and makes me question the ones I build. As part of your business, your dashboards should help you manage your business or help you manage. But the need for synthesis forces your BI tools to aggregate a large amount of data . It may be relevant, at the time of construction or now that you have read this article, to question the choice of indicators displayed. In the course of your life as a citizen, you will perhaps be subjected to abusive generalizations (“the Xs are more Y than the Zs. It’s not me saying it, it’s the numbers”). But they can be supported by very real data.
Frequent cases in a dashboard
I had to insist like crazy to explain to him that feedback with the national federation was necessary and that I wanted to participate (and that it would be useful to use the log of all maintenance tickets or changes that had been reported) -> I quickly understood that he had not recorded any requests for correction or changes reported and had instead strived to do the maximum (which is very well if he has an infinite capacity to do them which is obviously not the case). The idea for me was to get off to a good start and redo a pilot in 2022 by organizing myself properly to be in a dynamic of improvement before deployment.
It is up to you to require an expert analysis, to pay attention to confounding factors, to do a relevant segmentation or to choose the right explanatory variables. It may prevent you from concluding the opposite of what is actually happening! With all that, it could not go well, and out of the sample of pilot associations that agreed to wipe the plasters, only one really wiped them, the others having withdrawn very quickly. But guess what … I naively thought that after this mini catastrophe, our digital genius would have understood. And there, my computer genius colleague but nevertheless friend definitively sawed me by leaving me: