Do people who complete three years of college actually stay on the job longer than people who graduate? Does someone who has held two jobs in the past five years really stay longer than someone who held one? This is the kind of information contained in every application blank, but it often languishes in dusty old files instead of being thoroughly examined for underlying patterns. Let's consider turnover, for example.

Why focus on retention data? One reason is that turnover is beautifully concrete. It is seldom subject to the "no one in our company is perfect," "everyone here is above average," "forced ranked," or "supervisory favorites" syndromes. The reasons for termination might vary, but there is no doubt about whether the employee is either here or gone.

Another reason is that turnover is exceedingly expensive. Figures range from a mere $10,000 per employee to several times each employee’s annual salary. Whatever the calculation, turnover can drain the bottom line. But savvy employers can reduce this expense by doing a little research into data they already have at hand.

Biographical Data
Biographical data (or "biodata," for short) represents the sum of our past experiences, job history, education, and a host of other factors. We all know intuitively that the information included in an application blank probably contains vital clues to the applicant’s potential for turnover and productivity. The secret is discovering what these clues are and how they combine.

Think of biodata as another form of an interview. A well-conducted interview includes: 1) pre-determined job standards; 2) carefully crafted questions that gather examples of past job-related skills; 3) a standardized scoring guide; and 4) multiple interviewers who arrive at a hiring consensus.

Research shows professional behavioral interviews predict as much as 10-20% of the variance in job performance. Biodata produces similar results. Like behavioral interviews, biodata uses examples of past information; but unlike interviews, biodata questionnaires take less time and actually add to the effectiveness of other methods. So, why aren't they used more?

I think it is because 1) it takes hundreds of people to provide enough data to identify meaningful patterns; and 2) it takes powerful statistics and expert knowledge to sort out relationships between biodata and turnover data.

Fancy Statistics
Imagine a large company with thousands of employees performing a similar job. They could be call center reps, delivery drivers, or coffee baristas. The products or services are unimportant; the key words are "big numbers," "similar jobs," and "unacceptable turnover."

Now suppose we have biographical data on current and past employees hidden away in an HR database that contains gobs of data accumulated from the Internet, scanned application blanks, performance reports, and retention records. This is the mother lode!

All we have to do is find the relationships. Where do you start? First, we call in a professional, someone who:

• Has a sound background in exploratory statistics.
• Knows how to operate highly sophisticated exploratory analytical software.
• Can recognize the clues between trustworthy models and bogus ones.
• Is well-versed in the intricacies of psychometrics.

Pick Your Targets
Now, we begin by defining the kind of retention we want to control. For example, there is early-career retention (which often has to do with the ability to learn job requirements or discovering what the recruiter never told you); middle-career retention (usually associated with having skills to perform the position); and late-career retention (generally associated with job dissatisfaction). As you can imagine, each category of turnover will have different causes.

We next pick the biodata that we think are most likely to be related to the cause. It could be frequency of jobs held, level of formal or informal education, driving distance from work, or even prior experience with a similar job. Like turnover, we have to carefully sort through each piece of information looking for clues. We might even run a few trials to narrow down our choices.

Build a Model
This gives us a clear target and a starter set of causal factors (or at least we think so). What's next? Power up the software, analyze and examine, and repeat until a robust model evolves. What does it look like? Well, it's seldom as simple as looking at one factor. In fact, it is often a combination of several.

Here's one example of a turnover-prediction rule we built for a client recently (there were six such rules for retention, and eight rules for termination):

• If the employee had prior experience using his car as an office;
• And he had more than five years' prior work experience;
• And he scored 65 or above on generating new ideas;
• And his interpersonal skills were maximum;
• And his competitive drive on our hiring test was 25 or lower;
• Then, there is a 95% chance he will still be employed at six months.

Of course, a smart HR manager can immediately grasp how this data could be used: 1) on a special application form; 2) as part of a web-based screening questionnaire; or, 3) to predict an applicant's probability of early termination. Assuming the analysis was done right, employers can have a significant impact on turnover by hiring only people with a high probability of retention.

Danger, Will Robinson, Danger!
As you can imagine, this kind of analysis is not for the faint of heart. Several warning labels are in order.

Warning one: Backward-looking data analysis must always be scrutinized to eliminate information that does not cause turnover. For example, if we included weight among our turnover predictors, we might conclude that an employee's weight affected retention rates. Perhaps fat people didn't stay as long. That could have been true, but may not have been causal – it may have been coincidence. Bottom line? The data in the analysis must be thought through and not just dumped into the statistical blender.

The second warning has to do with numbers. Making assumptions based on small numbers can lead to big mistakes. Take for example "terminated" or "employed." Those are two categories. Now add "education," "prior experience," "job tenure," and "driving distance from work." These are four categories, making a total of eight cells. Trustworthy analysis mandates having at least 25 people in each cell. 8 x 25 = 200 people as a bare minimum. Real-world examples typically require many more.

The third warning concerns the wrong-headed idea of only looking at one outcome (e.g., including only people who are terminated). A one-sided analysis only tells us about one group. That's nice, but, essentially worthless unless we know that the people who did not terminate were different from those who did. Trustworthy analysis requires knowing about both groups.

The last warning is personal. It takes a courageous HR person to mount the charge to do things differently. There are plenty of risks. But, how often to HR departments get to take credit for making a true and measurable bottom-line improvement? And you might even get the files cleaned up while you’re at it.

The Biggest Interview Mistake?
It may not be the kind you’re thinking of. A recent Human Resources Executive magazine article quotes Dr. Williams extensively, to the effect that some of the dumbest interview behavior actually happens on the employer’s side of the table.

The column offers some highlights from CareerBuilder.com’s survey of the all-time worst interview mistakes. While taking personal calls during the interview or confessing to a history of violence are definitely deal-breakers, the author recognizes that the interview process itself can be just as self-defeating as the most boneheaded candidate behavior.

The “everyday interview,” detached from any proven job relevance and devoid of any standardization, is really no better than chance at predicting success. In the long run, executive management will come to understand this. And recruiters who have persisted in believing otherwise may find they’ve made a career-limiting blunder of their own.