Can a Netflixian Approach Work For Compensation?

Hello, username, we have recommendations for you!

You've seen the recommendations. Books, movies, clothing, dining options, and even friends - everything has a recommendation for you. Predictive modeling helps us weed through the hundreds of choices of books / movies / restaurants / friends and find what we want. Or, rather, what the algorithms think we want.

Companies like Netflix and Amazon have gotten pretty good at this. You can thank Big Data and predictive modeling. From books and movies, clothing and accessories, household furnishings and even food items, Big Data is bellowing at us and predictive modeling is the Oz behind the curtain. But what role - if any - does predictive modeling have in compensation?

My Cafe colleague Laura Schroeder talked about the issue last week in the context of employee retention. Develop a predictive model, and identify how much impact individual levers such as compensation have on retention to inform decision-making.

I agree with Laura that predictive modeling can - and should - be used to inform decision-making. But the key word is inform, not replace. Predictive models place emphasis on the statistical context. They cannot incorporate personal judgment and experience.

Some think that this is a good thing. Michael Schrage says you should ignore your gut - using personal judgment and experience will only lead to making the same mistakes over and over again.

Personally, I think a healthy mix of empirical information and judgment is the way to go. Relying solely on "gut feelings" and hunches makes it difficult to demonstrate the business justification. As Laura stated, "you need more precise data to get more money from the people whose job it is not to give you more money."On the other hand, emailing output from a predictive model with the note "model says Bob needs more money or he will leave" is also not likely to get more money from the people whose job it is not to give you more money. 

Human judgment comes into play in interpreting results. The predictive model says Bob and Sam are going to leave if they don't get more money. It would be foolish to simply obey the algorithm and give Bob and Sam more money without first asking, "do we care if Bob and Sam leave?" Sam may be essential to the completion of a mission-critical project; the wise choice would be to give him what he needs to stay put. Bob may be important, but the cost of retaining him may be too high vis-a-vis recruiting and training his replacement. 

Theoretically, we could build a predictive model complex enough to consider all available information so that these cost/benefit decisions could be made by the model. But if you were able to do that, you wouldn't be in the business you're in now - you'd be in the business of building predictive models.

From my perspective, the bottom line is that predictive modeling can be very useful in informing the decision-making process. I don't think that, at least right now, predictive modeling can be relied upon to actually make the decisions. 

But that may change in the future...

Stephanie R. Thomas is an economic and statistical consultant specializing in EEO issues and employment litigation risk management. Since 1999, she's been working with businesses and government agencies providing expert quantitative analysis. Stephanie's articles on examining compensation systems for internal equity have appeared in professional journals and she has appeared on NPR to discuss the gender wage gap. Stephanie is the founder of Thomas Econometrics Inc., the host of The Proactive Employer radio show, and author of the upcoming book Compensating Your Employees Fairly: A Guide to Internal Pay Equity. Follow her on Twitter at proactivemployr.

Comments

[Dan Walter - Performensation]

A colleague once wrote an great article about data. He discussed that in a world where all cars are either red of white, survey data would declare all cars pink.

While the survey data would be useful at making it clear that there were NO green, blue or Yellow cars, it would also be completely wrong.

I love predictive analysis. I also think that every decision should be based on data. Lastly, I think that the experience, insight and general human-ness of the people involved are the best way to evaluate and interpret data.

So, let's move toward MUCH better predictive modeling and MUCH MUCH better analysis and understanding.

[E. James (Jim) Brennan]

Living in the world of informatics, I agree with all above. More and better information can be produced from the rich and robust streams of data now available. Modeling predicts what will occur with precise reliability statistics... which simply means that we can measure the amount of predictive error or standard deviation from the norm. None of that, however, implies that the norm is best or always appropriate, or that the outlier practices might not be superior in a particular case. Or that we should actually be interested in what is being modeled.

Knowledge can be improved with better information, but it still requires thought for proper application. The human brain that informs "the gut" is still the most versatile analytical machine known, because it can connect gaps in data and perceive patterns better than any alternative system.