We have received exceedingly positive feedback from our prospective clients on a critical component relevant in feature engineering ( for attrition modeling) i.e. “attrition network analysis”.
Taking clue we present an exclusive write-up of “attrition network analysis”, broadly covering the concept, challenges, misconceptions, and its application scope in HR analytics. This is centric around our expertise, recent professional work and research we have executed in ANA.
The attrition data to generate ANA maps primarily includes data from banking and pharma. Some maps use data from employees who recently quit and some use dual data ( quit as well as current employees )
A little background about the perennial problem of attrition in HR
A stratified sample survey conducted in 2017 observed that approximately 50% of the global working population isn’t fully engaged or satisfied at their workplace
27% of these employees mentioned inadequate compensation as the primary reason, followed by 22% as a mismatch between their designation and their actual work profile. 15% stated poor understanding and relationship with their reporting manager.14% as overall communication and organization culture & work timings!
These employees remain susceptible to attrition. A survey conducted in the United States reported that an average of 4.9 million monthly separations (employee turnover) occurred between August 2016 and December 2016 alone, which represent a monthly rate of 3.45% of total employment.
Implicit & explicit costs associated with attrition
The costs associated with attrition are humongous and therefore it remains a core problem area for CHROs, particularly for high performing employees. There are Implicit costs including headhunter fees, overworked remaining staff, loss of specialist knowledge, interviewing and training costs, etc. Even when a replacement is found It takes a year or longer for a new employee to reach full productivity level, moreover some types of skills/ competencies can never be truly replaced.
Traditional approaches to estimating attrition and its limitations
Classification modeling on attrition data has unique limitations. This includes severely imbalanced data, censored cases, constraint due to confidentiality/ compliance or incomplete data itself.
Most of us are aware that attrition models may perform well on test data sets, however real-world cross-tabulated accuracy (unseen data) is usually poor, particularly for the critical positive class recall.
This severely impacts confidence in HR decision making and new initiatives to reduce attrition, particularly for high-performance employees. Our endeavor is to inject innovation in this area via “attrition network analysis”.
We disperse information via a mock Q&A Session below. The questions have been generated in course of client interactions and engagements
Question 1: We are all familiar with ONA ( Organization Network Analysis ) and it has gained significant ground, particularly over the last two years, however, we haven’t heard of ANA ( Attrition Network Analysis ). What is it exactly?
Response: ONA is well documented, discussed and understood within global HR. On the other hand, ANA is indeed a new novel application of “network analysis on attrition data”. Summing up in brief ANA is similar in some aspects to ONA but also very different in many other key aspects.
Question 2: So how is ANA it different from ONA and more importantly how does it augment the prevailing industry approaches to attrition modeling/ management in HR?
Response: ANA can help attrition management in a multitude of ways.
- Implicitly by providing insights for generating feature engineered attributes for attrition modeling itself ( a topic that we will cover shortly )
- .Explicitly by providing deep visual insights on strategic/ tactical decision making for mapping, understanding and managing the problem of attrition.
Question 3: If ANA is that useful why hasn’t it been explored as yet in the case with ONA. Why hasn’t there been an industry experiment and adoption of it already?
Response: Good question. The primary reason is while gathering/ generating data for ONA analysis is a fairly straightforward, well-documented process, it’s rather complex for ANA.
Typical data for ONA analysis constitutes of email headers from mail exchange servers or from other employee communication mediums of businesses.
Most of us are aware that the crux of ONA is to observe “magnitude” of connections between a group of employees ( clusters ) based on their communication and collaboration patterns ( typically via email exchanges ). Informal networks identify informal groups and individuals of influence, including progressively changing influence.
Technically ONA does this by mapping various statistical centrality and dispersion measures on communication/ connection data and then portraying this in the form of nodes and edges.
Nodes = Employees. Edges = closeness of the connection between employees
On the other hand, ANA or attrition network maps depend on derived or implicit connectedness between employees via dimensionality reduction (PCA ) on generic employees data.
Typical data could be a comprehensive profile, performance, engagement or employee survey data, in essence. This data is usually available with operational HR.
ANA has very little to with direct communication patterns between groups of employees and draws maps based on implicit connectedness
Technically speaking for establishing implicit connectedness ANA uses fine tuned PCAs ( principal component analyses ) or factor analysis .
Each principal component needs to capture a distinct aspect of employee character like background, competence, engagement levels etc.
Deriving meaningful principal components for attrition network maps necessities the application of deep statistics and machine learning and is an iterative, cohort and corroborative process with functional HR
Once computed this data is then portrayed on a multitude of networks maps (similar to ONA ). In essence, ONA is dependent on direct observations vs. and ANA dependent on derived observations.
Therein lies the core challenge.
Question 4: So in real terms give us a practical rundown as to how ANA maps help in managing the problem of attrition?
Answer: There are various types of ANA maps, each providing specialist insights.
Every map in this article has integrated analysis, you could go through them for detailed information
However we also enumerate the key uses of ANA maps
1. ANA identifies clusters of employee groups ( informal ). These clusters are assumed to have common patterns of behavior, core interest and response to stimuli ( they are ideal starting points for the design of experiments / tactical initiatives and deeper conjoint exploratory analysis especially for high performing employees )
2. Via iterative indexing and grouping measures, ANA identifies influencers among current employees / or employees who have just. Influencers have a different context and meaning in ANA vs ONA. Influencers are prime candidates for HR design of experiments
3. ANA maps can graphically bifurcate employees who quit vs. a vs. current employees to observe patterns and characteristics that exhibit points/ anomalies of interest
4. Do certain informal groups have unique patterns of attrition, which group have the least attrition. ANA maps can throw light on this
5. Clusters represent interesting factors/ potpourri of factors ( including sequentially developed over time ). For example, the circumstances leading to the quitting/ increased probability of quitting of star employees.
6. Node tagging can be based on departments, gender, performance, and involvement helping in acute insights for strategic analysis
Other insights available are Influencing employees, influencing groups, outlier employees, patterns, trends in attrition ( via stochastic actor-oriented models ) across departments, geographies, genders, age groups, performance ratings etc.
Outlay emphasizing on different points of interest in employee attrition
Same outlay emphasizing on different points of interest in employee attrition
Comprehensive conjoint tactical analysis of ANA maps assists in:
1. Defining the design of experiments that have the maximum probability of reducing the probability of attrition ( particularly for high-performance employees ).
2. Identifying employees who are best candidates for the design of experiments like innovative compensation components, flexible working etc.
3. Identifying star employees on the map with a higher probability of attrition and observing which informal group/ cluster they belong to. A deeper study of the unique characteristics of the informal group can then provide insights for tactical measures.
4. Co-factors or combined factors having an influence on attrition like gender – pay parity combo, compensation – qualification combo, a disparity in employee rating etc.
5 Identifying outlier candidates and factors responsible for them being outliers. Outliers mean they do not belong to any informal groups, or they should logically belong to one group but don’t.
Further, there is a lot of scope for creative analysis ( corroborated by expert attribute agreement analysis) compared to ONA maps. This is organization and requirement specific.
Question 5: we now understand the essence of “principal components” and their “connectedness” and that in the in the context of ANA is an implicit measure. As a point of interest is there scope for network analysis even beyond HR?
Absolutely absenteeism, succession planning ( via stochastic actor oriented networks ) even dynamic compensation management. We will cover all these areas subsequently.
Question 6: What are the typical cautions that we need to maintain in applying ANA maps to manage attrition?
Yes of course. The types of maps relevant and the possibility to generate various types of maps itself are depended upon the quantum, dimensionality, and quality of employee/ former employee data.
Further is important to keep in prospective the unique pain points and concerns regarding employee turnover for every specific organization in question before conceptualizing ANA maps.
All these aspects need to be fine tunes and generating ANA maps ( and determining the principal components ) needs to be an iterative, collaborative process between operational HR and data scientists within an integrated improvement project framework
No two cases can ever be the same!
About the author:
Raja Sengupta: Head of product development and Principal Consultant at Workplaceif. A Masters in applied statistics with 17 years of experience in statistics teaching, quality control, HR, product development and database programming. His expertise includes analytics, statistical quality monitoring, supply chain, andHR. This included multiple designs of experiments around blockchain too.
51,452 total views, 3 views today