W e publish our ATOM (Automated Talent Ontology Machine) rate benchmarks at the metro area level. Why? Because this makes good economic sense.
A metropolitan area is a “region consisting of a densely populated urban core and its less-populated surrounding territories, sharing industry, infrastructure and housing.” Critically, these areas function as economically integrated units. All the cities in a metro area are socioeconomically linked with the urban core.
This means that within a metro area we expect temporary labor rates to equilibrate. Contractors within a metro area can theoretically choose to work for any buyer organizations in that area, while buyers can (also theoretically) choose from the entire pool of labor living and working there. Within such individual markets, supply and demand work together to set labor rates.
We incorporate metro area features such as population, unemployment rate, educational attainment as measured by percentage of bachelor degree holders, and industry mix into our predictive benchmarks. The benchmarks also take into account full-time hourly wages by occupation in each metro area. To this, we mix in actual temporary labor bill rates from our application, from clients that have agreed to participate in our benchmarking. Combining these diverse sources of data gives us more complete and more accurate results compared to simply calculating averages off of our VMS data or applying transformations to full-time wage figures.
How do we know? Because we’ve tested it. A fundamental practice of predictive analytics is using hold-out data sets to see how well the approach generalizes to unseen data. A hold-out data set is created by setting aside an independent portion of your data on which you calculate evaluation statistics. We’ve found that our approach achieves better predictive performance on unseen data than you can achieve using transformed full-time rates or simply calculating medians off of raw VMS data.
Rolling up metro areas
We have almost 300 metro areas in our data set, but that may be too fine a granularity for some analytical and program management needs. Many of our clients like to roll up their U.S. rates to broad geographic areas, a handful at most.
How can you create meaningful groupings of metro areas? Some programs choose to do this using cost-of-living indexes alone, but our analyses have shown that other features of metro areas also influence contractor bill rates. So we’ve turned ATOM loose on our metro area data set to figure out how best to group metro areas into higher level rollups. We used something called “k-means clustering” to find distinct clusters of metro areas based on multiple dimensions including population, unemployment rate, industry mix, and more.
K-means clustering is what data scientists call an “unsupervised learning” algorithm. That means it finds classes without being given any examples of already-classified elements. This algorithm calculates the distance between pairs of cities plotted in a multidimensional space defined by the different variables we’ve measured about each city. Then it finds groups of cities that are close together. We found that six clusters worked well with our metro area data.
Here’s a visualization that shows our six-cluster solution. I’ve plotted metro areas by their cost-of-living index and unemployment rate, with the size of the bubble scaled according to the population as of 2014. Then I colored each metro area according to the cluster that ATOM assigned it to. I named the clusters according to common characteristics of the metro areas assigned to each.
You can see that ATOM assigned all large metro areas to one cluster and then found a second cluster for secondary metro areas such as IQN’s headquarters of Denver. College towns with high percentages of bachelor’s degree holders, low unemployment, and moderate to moderately high cost of living formed a third cluster. The remaining small metro areas divided into three clusters: one for high-cost-of-living, high unemployment central California cities; one for high-unemployment, low cost-of-living, low educational attainment “working class” small cities; and a final cluster of small cities with low unemployment, low cost of living, and high educational attainment.
Using geographic rollups in visualizations
We can use clusters to make exploratory visualizations more meaningful. I’ve plotted ATOM benchmark time to fill versus ATOM benchmark rates by job category, below, and colored each bubble with its cluster. Each bubble represents a metro area/occupation pair.
Coloring the bubbles using the cluster surfaces patterns, we wouldn’t otherwise be able to detect:
- Rates (shown on the x axis) are generally lowest for the Small Working Class metro areas, and highest for the Primary Metro areas, as we’d expect. An exception to this pattern is Engineering/Operations. We see that such jobs in Primary Metro areas have generally lower bill rates than such jobs in other clusters.
- In Manufacturing, jobs fill faster in Primary Metro areas versus in other metro areas, and slowest in the College Towns. You can see this by looking at time to fill, displayed on the y axes. The higher you go, the longer jobs take to fill.
- In Finance/Accounting, jobs fill most slowly in the Small Working Class metro area cluster. Perhaps in these areas there is an undersupply of such talent, due to relatively low educational attainment.
There is plenty more detail to dig into here. ATOM understands more than 100 different occupations and will tag your assignments with the occupation that best fits its job description. With an ATOM subscription, you can explore patterns for exactly the jobs that are most important to your company.