Don’t be Fooled: Predictive or Not?

How do you know if you are truly conducting predictive analytics? If you’re not sure, you’re not alone. Few organizations fully harness the potential of predictive analytics due to a number of misconceptions about predictive analysis methods. It is not uncommon for an organization to invest in a predictive analytics program that is actually not predictive.

Walk yourself through the questions below. If you answer “yes” to any of them, we have bad news for you: your approach isn’t actually predictive in nature. The good news: we can steer you in the right direction. If you vehemently say “Yes; I’m doing it correctly,” we’d still like to challenge you to this test because you may be surprised.

Are you using trend lines to project the future?

Anyone that uses score-based trends to claim they’re providing predictive analytics is wrong (sorry if this is you!). Examining a trend line calculated from tracking changes in scores over time tells you nothing about why any ups and downs have occurred or where that line could go in the future. Unfortunately, HR has gravitated toward data visualization tools. Although these tools create pretty trending pictures, they are not predictive and do little to actually move the needle on the HR data being tracked (or “visualized”)—ignoring the other fact that these HR data haven’t been connected or shown impact on an actual business outcome. In short, trend lines simply track progress; these are not predictive analyses.

Are you simply using correlations?

Although a correlation is a statistical measure of a relationship, it is not predictive. Just because a relationship exists between two variables does not mean that one causes the other. It’s the classic example of ice cream sales and shark attacks being correlated. No one would argue that shark attacks cause ice cream sales to go up nor that ice cream sales cause more shark attacks—they both just happen to increase during the summertime. Similarly, having high engagement and high business outcomes doesn’t mean that engagement causes the other. Because these methodologies do not tell if a true relationship actually exists, they are not considered strong enough for prediction. If investments are going to be made based on analysis, then correlation is certainly not a strong enough basis for driving and securing a significant financial investment.  Correlations should only be used as a starting point to determine where to continue additional analyses. They are not predictive.

Do you believe that regression is the way to go?

You may be amongst the many who believe regression equals prediction. While regression does identify variables that predict (i.e., are antecedents to an outcome), they are not necessarily causal.  For example, the winner of the Super Bowl is a strong predictor of stock market performance (an NFC champion predicts a bull market) – however this is obviously not causal in nature. As a result, there are limitations to the usefulness and interpretation of simple regression analyses. We agree that multiple regression is closer to modeling real-world relationships because multiple factors can be tested as predictors of an outcome. This approach allows for the examination of each factor’s unique effects on the outcome. Although a step in the right direction, this method is still not the strongest method to determine cause and effect relationships because measurement error cannot be modeled; only one dependent variable can be included, and, most importantly, causation cannot be determined with much, if any, certainty. If you’re here, you’re getting close, but you still have not reached true predictive analytics.

Predictive Analytics Done Right

If you said “yes” to any of the questions above, don’t be discouraged just don’t stay your course. The method you need to employ for making predictions is an advanced statistical modeling method called Structural Equations Modeling (SEM). SEM has four large advantages over the other analytic methods:

  • Multiple inputs or “causes” can be tested along with multiple outcomes concurrently.
  • It provides the ability to correct for measurement error.
  • An accurate assessment of ROI can be calculated.
  • Causation can be determined.

You can get a quick glimpse of SEM’s advantages versus other methods below.

Comparison of Analytic Methods for HR Prediction

Simply put, SEM is the best approach to predictive analytics for HR. The caveat to utilizing SEM is that it requires specialized statistical software and a highly trained statistician to be correctly implemented. However, this should not be a deterrent for any HR practitioner hoping to leverage this type of analytics. Universities often have professors or graduate students with the skills to conduct this type of analysis, or the analysis can be outsourced to a consulting firm with expertise in predictive analytics.

Whatever the organizational goal (e.g., reduce turnover, increase hiring success, increase ROI of your HR initiatives), true predictive analytics, when done correctly, will not only allow for the identification of key drivers, but will also allow for their prioritization. The approach will make planning and resource allocation straightforward and give your company a massive competitive advantage.

To learn more, read our white paper.

Predictive Analytics Myths Debunked

We recently presented at the Predictive Analytics World Workforce conference where one of the industry’s hot topics – predictive analytics – was sliced and diced in a hundred different ways. As PhDs, we geek out when talking about big data, modeling and analyses. That also means we can’t help but point out some fallacies in some of the statements we heard while at the conference. Our takeaway? Companies are making strides toward understanding and embracing predictive analytics but there are still many that need to take one, two or maybe three more steps to achieve true predictive analytics. We don’t fault companies for not being 100 percent quite there yet – we spent years in graduate school to feel comfortable with data science. So, to help move us all in the right direction, we’ve debunked a few predictive analytics myths below.

Myth: As long as you have a “model,” it’s predictive. It doesn’t need to be reliable, validated or even tested to see if it’s working.

Gasp! A model on a PowerPoint slide does not make predictive analytics. A model must meet two key criteria: (1) Testable: The model should be able to be tested –plain and simple, yet vital. In order to know if your model works or, if it accurately and reliably predicts the outcome(s) of interest (e.g., turnover, sales, profit, etc.), you have to test it. Without testing your model, all you have is a theory. (2) Directional: It should be clear how the variables are interrelated and connected. We saw many ‘models’ presented that more or less consisted of a group or cluster of variables with no information on how these variables led to each other or the outcomes they were trying to predict. It is important to remember that when modeling, it is not a catch-all exercise. The goal is not to list every single variable that you can think of as a predictor of your outcomes. Instead, modeling is about examining the interrelations between your input variables to find the strongest combination of predictors of your outcomes.

Myth: Predicting engagement is a good business practice.

We’d like to give you a little background on why employee engagement became such a fad. Back in the 1990s, a book was published called “First, Break All the Rules” from Gallup that introduced the concept of employee engagement. Unfortunately, it was based on flawed research and purported to show that 12 magical things were all that any company – of any size and industry – needed to be successful. The research was flawed because companies that were already successful were examined and common themes emerged. Discovering that companies that make a lot of money and are well-run also have happy employees is not surprising. The problem is that HR leaders (and Gallup) jumped to the conclusion that making people happy or engaged is what caused those companies to be successful. Wrong! Actual cause-effect research/analytics that we conducted showed that the employees were engaged because those companies were successful. Bottom line: employee engagement is not a business outcome and it is not a driver of business outcomes; it is a by-product of a well-run organization. As such, it is definitely not a good business practice so no need to expend time predicting it. Think of the movie Field of Dreams and its “If you build it, they will come” mantra. If you work on the key drivers of business success, engagement will naturally happen.

Myth: Predictive analytics can’t link to ROI.

With sound predictive analytics approaches, such as structural equation modeling, you can certainly ascertain return on investment (ROI). For example, SMD’s patented technology, SMD Link, does all of the work to quantify the impact of employees on business outcomes; calculate an expected ROI for investments in employees, and define the relationship between HR processes and business outcomes. The ROI is calculated based on the complex algorithms from the statistical models built in SMD Link. That means you can actually focus on the drivers that have the largest ROI on your business and watch associates in action! Our stance is, if you can’t track whether your prediction is valid or not, why do it?

We could talk about predictive analytics all day! If you’re interested in the hour-long version, join us at our “Big Data and Predictive Analytics the Right Way” webinar on April 16 at 1:00 p.m. ET. Register here.