The Five Models of People Analytics

Over the last few weeks, I have been working up a series of articles on Lean People Analytics, which I hope will eventually become a book. In these articles, I frequently refer to "Models". I realize now I have a bit a problem. The word model means different things to different people and it can mean different things in different contexts. Without defining what I mean by models I am afraid that we might not be on the same page.

Whether you are learning about Lean People Analytics or just people analytics or analytics more broadly... you either know or you will eventually know that models are central to your success. Any method of people analytics I propose is going to come back in some way to models.

So what do I mean when I say model?

What You Need to Know About Models (For Now)

A model is an abstract representation of an object to help people understand or simulate reality. (By object I can mean a physical object or an abstract object like a system, theory or concept.)

Some models are physical objects. For example, a model of a building. In this way, a model can be used by an architect to convey the ideas and test them before they are applied at scale. Everyone can look at the model and make a decision if they like it or not. If they don't like it the architect can ask why and change it. (By “at scale” I mean proportionally increased from where it is to its intended real-world size). The model removes non-essential detail and material. Those can be added back if the decision is made to proceed.

Not to get too metaphysical on you but the example of an architect is also a model of sorts. It is a metaphor. The architect is an abstract imperfect depiction of the job that is to be performed by people analytics. That provides the architecture for great companies. Architects don't hammer the nails but they have an important job.

In people analytics, we don't construct physical objects. We work in abstract concepts and mathematics. However, architects also don't just draw pictures of buildings. Architects need to understand abstract engineering concepts and mathematics. Mathematics becomes increasingly important to architecture as the scope of the project increases: tall buildings, bridges and other objects that people depend on for their lives. People depend on their jobs for their lives and the things people do at work impact the lives of others. The returns of the companies they work for fuel economic growth, which produces societal benefits. If the companies are not successful it has the opposite effect and we all bear those costs. All of this is why mathematics is important for Human Resources too.

Conceptual models are abstractions that connect or organize ideas serving the same purpose as a physical model but for things, you can't hold it in your hand. Most of what we work within people analytics is conceptual so we have to get comfortable with conceptual models.

A model’s primary objective is to convey the fundamental relationship and functions of the system of elements that it represents without unnecessary detail. When implemented properly a model should satisfy four primary objectives:

?        enhance understanding of the system of parts through their organization

?        facilitate efficient communication between stakeholders

?        provide a reference for analysts to make predictions, test ideas, extract meaning, while simultaneously offering grist for practitioners to formulate ideas on how to solve problems.

?        document the system of objects for future reference and provide a means for collaboration

It is important to understand the different types of models and how they fit together.

Different Models For Different Things

No I'm not talking about runway models. Come on!

The word model is vague. There are a lot of different models. There are mathematical models, scientific models, business models, business process models, data models and other types of models. I could be talking about any of these or all of these. 

In any given context I'm usually talking about one type of model at a time, but to form a complete blueprint for people analytics I am talking about combining five different types of models.

It is important to understand what they are, how they are used, how they work together and how to get them out onto the runway in the right order.

Once we have defined each relevant type of model then I will compare and contrast how a Lean People Analytics method uses the five models and how a traditional method of People Analytics uses the five models.

Five Models of People Analytics:

Business Models

Business models are frameworks that describe how a business creates value -- or, as management theorist Peter Drucker has said, business models are simply “theory of a business.”

A business model is a conceptual model that describes and represents the elemental structure of how a business will earn profit. These conceptual models describe the elements of a business that include: problem focus, target customer focus (market), unique value proposition, channels, methods of generating revenue, total addressable market (projected target customer market estimates), projected costs, projected revenues, and any believed or real defendable business differentiation advantages.

Business models have changed over the years through innovation. Companies such as Ford (mass production), McDonald’s (fast food), Amazon (e-commerce) and Netflix (digital streaming) have all helped introduce new models for business. 

Some modern theorists separate all business models into two large categories: Pipes (firms that create goods or services then sell them to consumers) and Platforms (creators of scalable digital networks to facilitate exchanges between groups).

People are in these business models somewhere, the question is where are they?

Scientific Models

A scientific model is the conceptual model that describes and represents the component structure, relationship, behavior, and other views of a scientific theory for a physical object or process. A scientific model is a simplified abstract view of a complex reality. Sometimes these are represented in mathematical expression and other times they remain strictly conceptual diagrams for more accessible expression of theory.

E=MC^2 is a mathematical expression of a fundamental principle of how our universe works from the very large to the very small. It was developed through theory, refined with mathematics and tested through experiments. It abstract but it has helped us do everything from go to the moon, to creating large explosions, to harnessing the energy of the universe. Of course, we don't always like the side effects of those things but never-the-less the model works.

The quality of a scientific field can be assessed by how well the mathematical models developed on the theoretical side agree with the mathematical results of repeatable experiments. Lack of agreement between theoretical mathematical models and experimental measurements often leads to important advances as better theories are developed based on the nuances of the findings.

Mathematical / Statistical Models

A mathematical model is a conceptual model that describes and represents the mathematical structure, relationships, behaviors, and other views of real world situations, represented as equations, diagrams, graphs, scatter plots, tree diagrams, etc.. A mathematical model is a simplified abstract view of a more complex real-world phenomena. Mathematical models can take many forms, including dynamical systems, statistical models, differential equations, or game-theoretic models.

Simple Example: When a population doubles each year, the function P(n) = I x 2^n represents the population P after n years, where I is the initial population.

The National Council of Teachers of Mathematics (NCTM) says:

“Modeling involves identifying and selecting relevant features of a real-world situation, representing those features symbolically, analyzing and reasoning about the model and the characteristics of the situation, and considering the accuracy and limitations of the model.” (Principles and Standards for School Mathematics, p. 302)

Mathematical models are useful for a variety of reasons.

·      Models allow condensed communication of the numerical expression of a real-life situation without non-essential detail.

·      Mathematical models allow for logical testing of ideas.

·      A model may help to explain a system and to study the effects of different components, and to make predictions about behavior.

For example - a rocket manufacturer should begin by designing a mathematical model and conduct simulations on a computer, rather than incur the costs of building million-dollar rockets and blowing them up for testing purposes. That might get expensive.

Eventually, you have to test your rockets in the real world, but only after you believe you have worked out the mathematical model. As you launch real rockets, you collect data to see if the rocket performs as predicted and to adapt your model when things go off track.

By the way this is exactly what Elon Musk did with SpaceX, as well as some of his other companies.

A mathematical model of a complex phenomenon in physical science usually includes the following elements:

·      Governing equations

·      Supplementary sub-models

·      Defining equations

·      Constitutive equations

·      Assumptions and constraints

·      Initial and boundary conditions

·      Classical constraints and kinematic equations

Data Models

A data model is a conceptual model that describes and represents the component structure, relationships, behaviors, and other views of data elements generated in information systems that represent objects or processes of the real world. It is a simplified abstract view of simple or complex data relationships. Unlike statistical models, data models are not generally represented in a mathematical expression.

Data models assist software engineers, testers, technical writers, IT, analysts, business users, and other stakeholders to understand and use a common data definition of the concepts represented by data and their relationships with one another:

o  to facilitate the design of systems,

o  to facilitate efficient data management in databases/data warehouses (aka data repositories) and reporting applications,

o  to facilitate multiple stakeholders in analyzing data in a consistent way,

o  and to facilitate integration of multiple information systems that contain common elements.

Data Models can be broken into three parts and/or phases

The data requirements are initially recorded as a conceptual data model which is essentially a set of technology independent specifications about the data stemming from or used to discuss initial requirements with stakeholders. This conceptual data model formally describes the abstract object semantics of a domain. This consists of entity classes, representing kinds of things of significance in the domain, and relationship assertions about associations between pairs of entity classes. For instance, a data model may specify that the data element representing an employee be composed of a number of other elements which, in turn, represent the job, location, start date, and manager of an employee in an HR system or for example the customers, products, and orders found in a sales system.

The conceptual model is then translated into a logical data model, which documents structures of the data that can be implemented in databases: for example concepts such as entities, attributes, relations, or tables. The logical data model describes the semantics, as represented by a particular data manipulation technology be they tables and columns, object-oriented classes, XML tags, or other things. The implementation of one logical model may require multiple sub-models.

The last step in data modeling is transforming the logical data model to a physical data model that organizes the data onto physical assets that accommodates access, speed and other situational needs. The physical data model: describes the physical means by which data are stored. This is concerned with hardware, partitions, CPUs, tablespaces, and the like.

The significance of this three-part approach is that it allows the three perspectives to be managed independent of each other. The table/column structure can change without (necessarily) affecting the conceptual model. Storage technology can change without (necessarily) affecting either the logical or the conceptual models.

Entity Relationship Models

Entity-relationship modeling (ERM) is a conceptual modeling technique used to graphically represent the entities and relationships of entities in a database or information system. The entities can represent independent functions, objects, or events. The relationships are responsible for relating the entities to one another. Examples of diagramming conventions for ERM include IDEF1X, Bachman, and EXPRESS. I know nothing about these, :-) but someone does and we can all read if we had to.

Data Flow Models (DFM)

Data flow modeling (DFM) is a basic conceptual modeling technique used to graphically represents the data flow between the major functions of a system.

Systems Architecture Models

A system architecture model is a conceptual model that describes and represents the component structures, relationships, behaviors, and other views of a technology system or set of systems. It is a simplified abstract view of a complex technology landscape. 

What Are The Limitations of Models?

?        A model is not a perfect reality, and never will be.

?        A model can feel abstract to others if they are not yet familiar with the pieces.

?        A model may not apply well between industries, companies, locations and situations. So general models need to be validated and rebuilt for specific situations.

?        A model regarding dynamic systems (e.g. people) can and will change over time. You are never done.

What is The Benefit of Models?

Models:

?        enhance understanding of the system of parts through their organization

?        facilitate efficient communication between stakeholders

?        provide a reference for analysts to make predictions, test ideas, extract meaning, while simultaneously offering grist for practitioners to formulate ideas on how to solve problems.

?        document the system of objects for future reference and provide a means for collaboration

This brief interview with Michael Lewis regarding his book Moneyball describes what we are doing broadly in human resources with people analytics. There are numerous concepts in this brief interview that are spot on.  Most notably the importance of developing a model to identify the things that matter. Sometimes what we learn from models is intuitive and sometimes counterintuitive but what we learn is always useful to strategy.

The main distinction between baseball and corporations is that baseball is always played the same way and the rules are the same for everyone. In baseball teams may take alternate strategies but it is the same game with the same rules. In the case of people analytics each company operates by a unique model. The formulaic method of finding this model is embodied in a framework I call Lean People Analytics, which was designed from my experiences to be applied across employers, regardless of budgetary issues or starting point.

This is in contrast to a traditional approach to people analytics that by nature assumes a general model and dataset exists. In my article, Making a Business Case for People Analytics (with the three A's of Lean People Analytics) I suggest the whole of human resources can be boiled down to three principle problems: attraction, activation and attrition. This can be generalized and much of the data you would use to understand these can be made common across companies, however the data you need to define a solution cannot.

Case in point, if you have 5 or more pets at home there is a high probability you are a great hire for PetSmart. This same fact does not predict success for many other companies. At any other company collecting this information is wrong, wasteful and at the same time prevents you from collecting something else that would be more useful. You need a way to know what data to collect, which absent a theoretical model, approaches the entire known or possible universe of data. e.g. infinity. You have to create the right dataset for a unique business. There are some common data elements (roster lists, exit lists, hire lists, etc.), but these won't lead you to the answers without augmentation. We won't be able to derive a unique model from a common dataset. We have to have a methodology that leads us to get the right data. It is not impossible, it just requires a thoughtful methodology. This too can be represented in a model.

The priority of work shifts from a data and system orientation to the conversations and actions required to develop a thoughtful and unique people model for each business, which then informs the data and eventual analysis. Once you have the right data the analysis part is methodical and therefore easy. The emphasis in the field on technological or statistical complexity, miss attributes where the more important job is.

Maybe my point is totally lost, but Michael Lewis's interview is still a great interview that foretells success if we can make what we are doing in people analytics clear to more people.

How Do I Use Models?

What do I do with models? Everything.

The work of people analytics at any company is to develop a unique model of how people deliver value to customers and keep refining that model to create more happy customers before running out of time and resources. You work theories, you test them and if you learn something you improve your model. I use the other four models to represent how I will do this technique and to communicate the requirements to technical partners.

In Lean People Analytics we start by defining a business model, add to this a scientific model that incorporates people, and then create a statistical model, and finally, a data model to systematize this information. We proceed from one model to the next only as we receive a positive indication of success at the previous stage.

How does the Lean People Analytics way contrast with the traditional approach?

Here is how I order the models in the lean methodology...

You can read a lot more about this in my Introduction to Lean People Analytics.

In traditional people analytics, it is precisely reverse from what I just described. You implement a system model, which provides a given data model that you have to work with, from which you apply a statistical model, which may be informed by a scientific model depending on your methods, and after all this, you seek to apply whatever insights you can mine from this to a business model. (I add "depending on your methods" to this statement to indicate that if you apply a machine learning or artificial intelligence you may not be applying the scientific method.)

Following a traditional approach success is only achieved:

• if you have implemented systems,
• if you have joined the data the right way,
• if you have the right data in the first place,
• if you have clean data,
• if you have structured your data correctly for your analysis,
• if you can figure out if any of this connects to your unique business model, situation AND,
• if you can find someone who cares.

How many times have you heard, "If only we had all our data in one place then we could do people analytics" or "If only we had better data governance, then our people analytics would work" or "If only we could hire someone who can tell a story with our data in a compelling way then people would use it" or "If only executives wouldn't rush ahead to make decisions without data."

I'd like to point out: the traditional model of people analytics contains a lot of if statements.

Each if contains a contingency that can and will cause failure. Calculate the estimated probability of failure at each step of your overall model and combine them. This suggests a very low probability of success for a traditional implementation of people analytics.

For purposes of illustration, using the 7 primary if statements (above the paragraph) and assuming you have 75% of getting each right and assuming they are independent then your overall probability of success would be about 13%. (.75 x .75 x .75 x .75 x .75 x .75 x .75) This means on average you could only expect to be successful on traditional path approximately 1 out of 10 times. Finger to the wind I think that is about right. If you don't like this estimate you can input your own probability estimate assumptions and run it again.

To estimate the probable return, (multiply (the overall probability of success given the contingencies) times (the cost) times (some ROI factor)) minus (the cost).

The traditional method has a much lower expected value than the lean approach because of the number of contingencies and upfront costs, even if we assumed the same ROI. The reason Lean People Analytics has a higher overall expected value is that the approach preserves the investment in time and resources until the risk of the previous stage has been reduced or eliminated.

The order matters. Ideally, you want to resolve your greatest risks first. We know that we don't have a straight .75 probability of success with each contingency. I can assign a near perfect probability of getting data models and system models right provided I know I have a problem worth solving to someone. If I don't work on the right problem then it doesn't matter I was successful in the preceding steps because the entire effort was wasted. Worse, the probability of working on a problem worth solving to someone is much lower than the odds of technical failure. So we need to vet that first before we make big investments. Absent a theoretical model we also have fairly low odds of identifying the right data. We need to get this right early too. We increase our odds of success by addressing these two problem areas first with models and we leave our more certain tasks (as well as costly) until the end.

The entire premise of Lean People Analytics is that by re-ordering and risk gating we remove waste, improving the odds of overall success, consequently increasing the expected value of the overall system of work substantially.

Here is a quick overview:

If you are on a traditional path then at all steps the option that costs the least money creates the higher overall expected value. This is counterintuitive. A traditional path is like going down a treacherous road in a luxury car when an economy car would do. The economy car doesn't guarantee success, but you are much less likely to go bankrupt buying new luxury cars when the cars repeatedly get destroyed, which is fairly certain. I had a friend at work once who drove to work in his sneakers and put on his leather shoes in the parking lot. Maybe shoes are a trivial example of this problem, but that's the idea in a nutshell. We are not saying don't have nice shoes. We are saying, save your expensive shoes for the low-risk carpet. Use your sneakers to get you there.

A Texas No-Limit Hold'em Example Just For Fun

If you know poker then the level of probability of a traditional approach to people analytics is something like the type a starting hand in Texas No Limit Hold'em where everything you know suggests you should either fold this starting hand or "limp in" to see what three common cards the dealer puts out. (Limp-in means to pay the minimum bet to see what happens next. Hope for luck.) Let say, for example, you start out with an 8 and a 10. In theory, it is playable but the poker pros will say you should stay out of these types of hands unless you are skilled in the difficult conditions that follow. e.g. Low probability hands can pay off but you should only play low probability hands when you have already played a lot of poker in your lifetime. Also, you should only play them when there will be a high reward.

I am not making this up. Here is what the pros say about "limping in":

However, below is an example of two people who have played a lot of hands. In this situation, Negreanu started with the better hand but ended with the worse hand, but loses this hand the right way. Why is losing the right way important? Because he preserves capital to play future hands. In this situation winning or losing was out of his control, however he controls his bet sizes to maximize the information received so as to either maximize his return if he wins and preserve capital for future hands if he loses. This is a perfect analogy of what we are doing with risk in Lean People Analytics.

In poker, counterintuitively, when you proceed into low probability hands you should bet more money than your hand is worth so that if in the rare instance your hand improves with the next deal you will get a bigger payoff for it to make it worth all the times you will lose with the exact same hand. It is a high-risk high reward scenario. In the Negreanu example that is what Hennigan's call of Negreanu, pre-flop raise represented. He paid more than he should have for that hand. (The flop is when the dealer puts the three shared cards in the middle). He knew he was probably behind Negreanu and others at the start (before the three cards came out), but he knew that when he sees the three cards he will either make his straight or not, which makes his subsequent decision almost binary. If he makes his straight (or gets close) there could be a high reward because that is a relatively strong hand and his position holding this hand is obscured. If he doesn't make his straight (or get close) he can just fold then and not lose any more money. On the other hand, if you take every low-risk hand you will lose all your money on the antes (the cost to see the deal).

As the scenario unfolds in the video above you can see that it can get complicated fast, but in this situation, both players did the right things. Of course, you can't control what cards come out or some of the crazy things your opponents might do. Sometimes you will do the right things and you will not win. That is why you play a lot of hands. Over a longer period of time, the poker players that make better decisions will come out ahead.

Poker theory only works if you get to play a lot of hands. In the world of people analytics project management, you don't get a lot of hands. So you want to gate risk and de-risk before proceeding to pour resources on. You spend more time and money commensurate with increasing certainty. Like an advanced poker player, you need to spend the right amount of resources to figure out if you are up or down before you shove your chips in the pot, which could bankrupt you if you are wrong. You need to preserve your ability to play hands that are in your favor of paying off. You do that by getting out of hands that are not. That's lean in a nutshell.

If poker intrigues you as much as it does me you might enjoy this interesting piece of published research: Evolving Opponent Models for Texas Hold ’Em. The researchers conclude that having a system of classifying opponent behavior leads to a consistently higher return over a large number of hands. Guys like Mike Caro created strategies built on classification many years ago - combining the first-hand experience, theory and science - consistently improving his theories and the range possible of strategies through testing. All of this without the help of artificial intelligence. In fact, even when you apply machine learning or artificial intelligence, as the Evolving Opponent Model article above does, it is important to have the right variables and parameters that you want to apply artificial intelligence to. Advanced methods simply increase the importance of models and expand the range of ways we can test them.

You do you and I'll do me, however my point is that the traditional approach to people analytics contains much higher upfront cost and risks so you better have clean models. Don't limp in and hope for the best. You also better be very experienced or have a team of experienced professionals to support you.

Lean People Analytics Series

“Introducing Lean People Analytics”, July 6, 2018, LinkedIn.

“The Ten Types of Waste in People Analytics”, June 19, 2018, LinkedIn.

“The Five Models of People Analytics”, July 7, 2018, LinkedIn.

“Making a Business Case for People Analytics (with the three A's of Lean People Analytics)”, July 2, 2018, LinkedIn.

“Getting Results Faster with Lean People Analytics”, June 15, 2018, LinkedIn.

More

• Find more of my writing here: Index of my writing on people analytics at PeopleAnalyst
• Connect with me on LinkedIn here: https://www.dhirubhai.net/in/michaelcwest
• Check out the People Analytics Community here: https://www.dhirubhai.net/groups/6663060
• Buy my book on Amazon here: People Analytics For Dummies , directly from the publisher (Wiley) here: People Analytics for Dummies , or from other places where books are sold.