Employment Application Arrival Model for Talent Acquisition Simulation and Management (2021-US-30MP-816)

[#1]

Level: Intermediate

Thor Osborn, Principal Systems Research Analyst, Sandia National Laboratories

Talent acquisition is a critical element of the talent management cycle. Employment application arrival is a stochastic process that poses limitations and delays on the subsequent vetting, decision making, and negotiation steps necessary to hire and onboard talent. Analytical comprehension of this process may be useful for guiding the content of job postings and the expectations of hiring managers, as well as for simulating flows and timing from the issuance of job requisitions to onboarding. This presentation demonstrates that application arrival behavior may be effectively modeled using an ensemble of Gamma-Poisson distributions with mean rate and overdispersion parameter distributions correlated to work site, field of practice, and career stage, using application data from a large research organization collected over several years. 

Auto-generated transcript...

Speaker	Transcript
Christy Spain	Good afternoon.
	hello, how are you today.
	I'm Okay, how are you.
Christy Spain	Good thanks.
	Where are you located Thor.
	yep Christy Spain.
	And albuquerque new Mexico.
Christy Spain	that's what I was thinking.
	got the right.
Christy Spain	triple digits out there.
	Not right now.
	Okay we've got a.
	Hurricane coming in from the.
	ball hogs gonna knock the temperatures down.
Christy Spain	Okay what's the name of that one.
	I don't know normally doesn't happen much and I've kind of not been paying much attention.
	well.
Christy Spain	I am I kind of caught me off guard, but you know that's hit the golf they're pretty hard yesterday so everybody here is filling up on gas, because the pipeline is now shut down.
	So where are you.
Christy Spain	I'm in cary North Carolina headquarters yeah so.
	um so let's see have you done, did you do any recordings for us last year or.
	I made my own last year.
	Possibly one of the reasons why we're doing this.
	This time.
	Oh.
Christy Spain	that's funny.
	um let me see, I have a little checklist let me open that up like actually close too many things down when.
	I was trying to clear everything out, so I didn't get any pop ups, have you closed out all your Apps and.
	I haven't yet, and I also had a question.
	Sure um.
	It said that I'm supposed to upload things before my my recording but I can't because I don't have a link.
Christy Spain	I'm upload is and what what things.
	And presentation materials.
Christy Spain	Okay i'll find out what that link is for you.
	And so I went back and looked through my old emails I couldn't find it I'm thinking well.
	I'm.
	Not a big deal but.
Christy Spain	We should definitely be able to just go ahead and record, though.
	Okay yeah.
Christy Spain	Let me find a little checklist but part of that checklist is closing at your Apps.
	i'll go do that.
	You don't need Microsoft teams.
Christy Spain	You don't have anything any logos on your shirt so that's good.
	No logos.
	Probably don't need no right now, either.
Christy Spain	I couldn't figure out how to do not disturb teams.
	The teams well.
	One of the things about modern.
	Technology is that people aren't giving you instructions that's anymore they're giving you this thing and you're supposed to figure it out.
Christy Spain	Well, I typed in and the help section D amp D and it was acted like I was looking for any messages or anything of that sort, and my Okay, or it was fine I closed it out so.
	Okay, so.
	I have a few applications open but.
	Look at that little parasitic teams thing.
Christy Spain	we're good so.
	Well, they put it in this little hidden menu so that it's still there, along with Skype because they want to be able to contact you, no matter what.
Christy Spain	And you know I think for maybe we should have you took one of the background.
	It says no copyright infringement, but you have you know a lot of books back there so maybe.
	Oh well, I do have books back here, I thought that wasn't a big deal because you can't really read them, but all right.
Christy Spain	I don't know i'd hate to.
	have to call you back and say I'm going to let it go.
	yeah I don't want to do it twice.
	No.
	I mean I'm sure this will be great fun, but I don't want to do it again.
	me just get rid of teams.
	they're all the teams is gone.
	Okay, so these backgrounds, I got a thing on that.
	I have to go back to outlook.
Christy Spain	Or you could just maybe we'll.
	See, I never used them.
	A truly never used them.
	yeah because they analysis to blur.
	yeah.
	You know, whenever you're doing stuff to move I see people use them and stuff blurs and they see all the junk that's in their room.
	See so from.
	tanya.
	tanya so new symptom.
	Here we go so.
	Those are social tiles you know what social tiles.
	Virtual backgrounds so blue scatter plot.
	i'll just do that.
	Come on.
	So that's that now the question would be how do I.
	Use them.
	let's see.
Christy Spain	I thought it would be under the little three dots but it's not as.
	Okay I'm not clear on where the ducks are.
	Where do you find dots.
Christy Spain	Well, you know at the bottom, where you have share screen and chat.
	And all of that there's like a little more.
	That but it's I don't see any option for change in the background, there.
	don't either.
Christy Spain	We go back and forth between so many different.
	tools here it's hard.
	it's let's do a quick search.
	Virtual background under slick camera.
	Oh.
	So.
	And then you have to know where the virtual background is because they didn't say.
	Which one, you should use or anything like that, no, no, no.
Christy Spain	Oh, that looks Nice.
	Oh good because I wasn't sure I was going to be able to do much else OK, so now we have the background and if I move a lot that's working pretty well actually only a little bit of screw up okay.
Christy Spain	and
	Did you auto had your taskbar.
	let's see how do I do that.
Christy Spain	You go down to your taskbar and right click.
	Yes, on a place that doesn't have anything on it.
	No.
	problem is my taskbar is fully subscribed.
Christy Spain	mine's not hiding either I don't know why.
	Is it yeah.
	taskbar settings here we go.
	it's Monday automatically hide the test test bar in desktop mode.
	On.
	OK, so now it appears to be auto hidden except it's not hidden.
Christy Spain	Well, if you scroll up it goes away.
	So the only shows up if.
	I have something activity goes away good.
Christy Spain	And then your sound is great.
	and
Christy Spain	i'll do you want to practice sharing, so we can check your resolution, and all of that.
	Well let's see.
	For sure a specific thing that's going to be a problem.
	For share the screen I'm going to be showing what you look like.
Christy Spain	Well, you can go ahead and share your content.
	yeah, but I want to flip back and forth between slides and the the JMP application so I want to be able to show.
	My screen.
	So somehow I ended up.
	With the screen and I've got you in a small window here.
Christy Spain	At the top, you can do view and change it to side by side or that's what I have.
	To turn my camera off because nobody wants to see me anyway.
	it's all about you today.
	Show small active speaker video thumbnail video.
	Okay, that shows you.
	It shows both of us this shows a grid.
	And that makes the whole thing really small.
	And I can't see myself anymore.
	But I can move this thing.
Christy Spain	Well, I can see you sad your PowerPoint.
	Well, so here's a question I got my PowerPoint gets in the regular writing mode.
	But if I go into presentation view.
	It gets bigger, but then I have to keep popping back and forth between that and the other view.
	When you do the video.
	Are you going to scope in on the parts that are relevant or you're just going to show the screen whatever is coming through.
Christy Spain	yeah we're just going to show whatever.
	So, most people do presentation mode and then, if you need to get out of that and pull up your JMP screen, you can do that.
	and go back does that make sense, and if we end up seeing you know some of your.
	yeah that looks great.
	Okay, so this I guess is what we'll do and i'll just escape out and go to JMP where necessary.
Christy Spain	Do you want to try that real quick, so we can make sure it's the right size whatnot.
	yeah.
Christy Spain	looks great.
	Okay.
	So it's visible.
Christy Spain	OK, so now, I need to.
	confirm your name company and abstract title, please.
	Name is Thor Osborn. Company is Sandia National Laboratories.
	The abstract title should be Employment Application Arrival Model for Talent Acquisition Simulation and Management.
Christy Spain	Perfect. And then you understand this is being recorded for use in Discovery Summit.
	JMP conference and will be available publicly in the JMP User Community. Do you give permission for that use and recording?
	Thank you.
Christy Spain	And I think we're ready, Thor, so I'm going to mute myself and
	You can go whenever you're ready. The goal is to get go straight through. I'm not gonna say anything or stop you unless something catastrophic happens. How about that?
	Okay, the goal is to go straight through and how long, am I supposed to have?
Christy Spain	30 minutes.
	Okay, now I think you have a little leeway because they like to leave time at the end of the playback for question and answer so if it goes a little bit over that's okay.
	Okay, now, before we go there's this little window that says talking and who's talking.
	Because they shrink the little video thing down.
	Okay does my screen show the presentation view, or does it show that, with this little insert in the lower right.
Christy Spain	All I see is the presentation view.
	Okay, so they're somehow floating this.
Christy Spain	That's got to be annoying.
	It's okay, as long as it doesn't get in the way I mean I don't put anything down there.
Christy Spain	that's okay.
	yeah.
	All right.
Christy Spain	Okay I'm gonna go on mute and then you'll be ready.
	Okay.
	So, my name is Thor Osborn. I work at Sandia National Laboratories as a principal systems research analyst.
	My talk today is Employment Application Arrival Model for Talent Acquisition Simulation and Management.
	And I hope you enjoy it.
	I've got three basic objectives here. One is to make the business case for why you'd want to model employment application arrival times. Another is to show a straightforward process for creating a concise, broadly applicable model using this kind of data.
	And then I'll demonstrate that process on data from a large research organization and show what you can do with that, briefly.
	Technically i'll go over understanding the data because I think this kind of data is a little unusual for most folks.
	The source models, essentially, this is a source modeling
	analysis, so i'll be talking about the Poisson and Gamma-Poisson models.
	i'll do some analysis and some transformation to make this easier to work with and then briefly touch on how you might go about making models.
	So, as far as motivations go there are three basic areas of motivation. One is to improve the talent acquisition process and business function.
	Another is to improve the understanding of executives leading the company, and then the third is to help frame expectations for hiring managers.
	i'll start with talent acquisition. The hire rate and lag depends on flows through vetting stages in the talent acquisition pipeline. But the critical thing is, you have to have sufficiency of employment applications before you can go anywhere with the process.
	And the rates and patterns vary quite a bit depending upon the field, the specificity of the postings that are put out, the competition, how much you advertised, and so forth.
	A common mathematical framework for application arrival could enable better understanding of the trade space for improving application capture rates,
	as well as other things. Some key relationships application capture rate and variance versus employment context, job site, career level, field of practice.
	Capture rate impacts of adjustable variables advertisement job posting specificity, the posting language--and by that I don't mean English versus Spanish, but rather the way things are framed.
	And targeted recruiting efforts. And then the capture rate impacts of external factorsthings like economic conditions, competition in the field of practice,
	how big is the professional population that you're really reaching within your recruiting area.
	And I think it's important to recognize that talent acquisition faces the typical constraint that you often see with project or program execution--time, quality, and cost.
	It takes time to collect applications. You want to get a high quality of applicants. That costs money, and the faster you try to go the more it's going to cost.
	So there are trade offs here.
	Executives often get involved in the process through workforce planning, which is a catch all term but, basically, you know say an annual plan.
	But absent relevant models and feedback on the cost of things, the consideration of the typical constraint and allocation of staffing budget or processes could be subjective or absent.
	And by that I mean at the executive level, HR is often viewed as simply a necessary cost, but not necessarily a lever for moving the company forward. Partly due to lack of information, so providing information could be a way of improving that situation.
	Also, hiring manager expectations.
	Basically we're dealing with small number statistics, small numbers of applications.
	Human beings are really subject to pattern bias, and most hiring managers don't hire people every week, so they're not used to looking at this from the perspective of a lot of different cases over a long period of time.
	And that can lead to anchoring in the last situation they faced, that can lead to pattern bias where they look at the number of applicants heading down from one week to the next and figure that must mean that they've accessed everyone who's interested.
	And the problem with these intuitive responses to biases that can lead to overreactions because, as always with small number statistics, you get misinformation if you don't take it in that context.
	So a bit about the data. Arrival data are tied to specific job requisitions. So, you have a job requisition, there's a posting for that requisition, and the applications arriving in consequence of the posting.
	They're characterized in several wayssite location; career phase;
	visibility, how broadly it's made visible;
	field of practice, what you're doing;
	and specific requirements that can be very broad or very narrow within those fields.
	And the applications can be submitted during the window of time when a job posting is accessible, so there's a defined time frame usually.
	They're tracked by date.
	Now, in this case, using real data for the analysis later I have to use the date of last submitted application as the posting closure date, but it's not necessarily true. But, as often happens in real data, we don't have all the data we'd like to have.
	And i'll say finally that
	you're counting the number of applications per day. When you see zero applications for a day that's a count of zero.
	So thinking of a national pool of potential applicants for a job posting,
	employer puts out information into the world.
	Some of the people in the national pool of potential applicants make applications.
	it's generally assumed that that will be few relative to the total who could.
	And so you have this vast applicant pool. It's a source and any applications or minor perturbation on that source.
	And, just to be very explicit, which this could be very boring for people who are used to dealing with Poisson mathematics, but
	if you see one application arrival on day one, that's one instance of a count of one. Day two there's no application arrival. That's an instance of a count of zero. And so on the far right, you see three instances of zero, three instances of one, a two and a three.
	And that can be fit, although it's not very much data, it can be fit with the Poisson model, and you see the blue line there in the lower right.
	So an average rate of one application arrival per day, but the actual count per day is going to vary across the distribution. And that's it.
	So, the Poisson distribution can be used to describe probability of a count produced in a unit of time by randomly emitting source of discrete items that have some constant mean rate of emission Lambda per unit time.
	And you get this equation. And, so in the case of a count of four, which showed a very small probability in the previous slide,
	you essentially have one to the fourth.
	times e to the minus Lambda, so basically 37%
	divided by four factorial, or 24, and so something in the neighborhood of one and a half percent, which jives with a small but but nonzero probability.
	So, as a first hypothesis
	we'd assume that members of the nationally distributed pool of potential applicants for this broadly advertised job act in a uncoordinated manner
	regarding employment opportunities, and so a Poisson model could make sense.
	And so it's a reasonable initial hypothesis.
	Now, what if they do interact.
	Then, that can be conceived in terms of the Gamma-Poisson source model, which is essentially a Poisson components using a Gamma distribution as a mixing distribution.
	Or you can think of it as a blurred out Poisson because the rate, the average rate, is not a constant average rate, but it changes from one moment to the next.
	And this can happen under an alternative hypothesis that applicants behave in a coordinated manner,
	which could happen if there were networks of people talking to each other about job opportunities and so forth.
	Now here is
	a slide on how to generate the Gamma-Poisson. This is basic stuff, but I find looking at these distributions can be helpful.
	So I'm going to break out and do a little DEMO in JMP.
	Now what you see here, random Gamma four-one, just means random Gamma with a parameter of four and a scaling parameter of one, so average is four.
	You look at the distribution. There's 100,000 rows here; you get a curve like you saw on the slide.
	Average is about four, which happens when you do something lots of times.
	it's almost exactly what it's supposed to be.
	Now, take that distribution and use it
	as the feed for a random Poisson.
	And you get this.
	Which doesn't look like the Poisson from before because it's been smeared out.
	If you look at a Poisson fit, you'll find that the best fit
	for four as an average, which is all it can really do,
	has a narrower dispersion than the Gamma-Poisson and just to verify that yes, it's a Gamma-Poisson,
	look at the red line there.
	And it fits really, really well.
	Almost perfectly.
	And so, essentially, this is, this is a case where having gone through that
	communication process, if you will, the Poisson is no longer the best model.
	The case I just showed is shown here in the upper right of this quad panel.
	Sigma equals two, the over dispersion parameter of two.
	If you have an over dispersion parameter of one, Sigma equals one, on the upper left, then that's essentially the same as the Poisson distribution all over again.
	But as the over dispersion parameter increases, goes from a hump followed by a tail, instead to something that looks more like a decaying exponential. You'll see that in the lower right in the blue,
	the colors are switched here opposite what they were before, so if there is a confusion on that, sorry.
	Now here's with real data.
	So the case of broadly accessible, early career mechanical engineering posting.
	140 applications to create 131 counts, many of which are zero, so this is spread out over time--one hundred and thirty-one days.
	The Gamma-Poisson fits best, and you see that in the upper left.
	I'm using Akaike's criterion, the AIC
	metric.
	In the lower left, you will see
	an experienced position. Now experienced professionals don't have the the networks of, say, university placement folks and so forth.
	Or necessarily the constant immersion with other people that you would see in a university setting. This is not proof of anything but just a way of rationalizing that perhaps that's why you see a more classic Poisson behavior with experienced professional responses typically.
	And just to show here, it seems to correlate with the number of applications as well. Other factors that would matter seem to be the
	career stage.
	As I just mentioned FLSA Status (Non-Exempt) meaning, say, technologists as opposed to exempt staff.
	Those factors seem to matter, but there's definitely a bias as the application count increases, becomes more likely to be Gamma-Poisson distributed.
	Now, how this data is prepared for analysis. Application dates by job req.
	Tabulated setting zeros for
	when there is no count.
	And then fitting the Poisson and Gamma-Poisson for each req to find out which one gives you the better AIC value.
	In parallel, requisition summary table was made with one line, or one row, per requisition, and then those are fused together to give parameters for the Poisson and Gamma-Poisson, as well as the AIC parameter that tells you which one is more likely a better fit.
	And you get a summary for Poisson and a summary for Gamma-Poisson after subsetting, because these are difficult to deal with together.
	You can use the Gamma-Poisson with a sigma of one, but then you end up with a zero inflation, and it becomes complicated to deal with, so instead I'm just treating them separately.
	Whichever subset is used, the process will show fits linear model for the parameter versus the drivers.
	The parameter is normalized by the model.
	And then I look to see if a common distribution is plausible across all the driver conditions, which in this case is essentially the requisition context of site location, early or late career, and so forth.
	If it's possible that they fall in the same distribution, then it's possible to create a common distribution parameter set and use that for the entire data set.
	Otherwise, the approach fails and I have, at this point, no mechanism to deal with it, so fortuitously that didn't happen, but it is something to think about for the future.
	So here's some Poisson parameter distribution by context where context has to do with the site, whether it's an early or late career, or established you could say,
	and whether it's, well, which field of practice. Every one of these has a "B" meaning broad because analyzing internal only job requisitions is a difficult, different thing with much more constrained opportunity, and so I'm sticking with simply the broad case in this presentation.
	What you can see is the mean variance differ by requisition contexts pretty obviously, and no, they don't fall within the same distribution. You can demonstrate that easily.
	Looking at a smaller subset, just the initial chunk, the Site A early set, it becomes more obvious that the means and the standard deviations are all over the map.
	Now here I'm going to also do a little demo.
	So I have this data.
	Now, this is a really simple model. I'm going to take the Poisson parameter output, and I'm going to use requisition context, which is this multi-level
	subjective variable.
	And run that, and you see that there's an RSquare of about point four. That means that
	it's not insignificant. The model fit is actually pretty good.
	And the fit for this parameter is very good. It doesn't cover all the variance, but that's okay because we don't expect it to; it's just saying that this is a pretty good model for what it does.
	It could also be done purely as a tabulation, but I like this approach because it makes it easy to work with, and it gives me a sense of how much variability is being dealt with in this fit.
	It can go here, and look at the
	actual by predicted, and you see, yeah, it's not a great model, but it does serve a purpose.
	And another thing you can do is save the prediction formula.
	And so you see here.
	Because I've already done this, this is subscripted to automatically.
	You see that what this amounts to is essentially just
	a constant
	plus an offset by requisition context.
	There's no rocket surgery involved.
	But now, if I want to do a normalization,
	add a column.
	Call this
	Norm Poisson Again because I've already done it once for the purposes of making this
	table in the first place.
	Find Poisson. There's my Poisson parameter. Divide that by the
	model fit.
	That was actually supposed to work.
	Not sure what's going on here.
	Okay, never do demos live that's the rule, but this has always worked in the past, and I'm really not sure what's going on.
	There's my Poisson parameter.
	There's my prediction formula.
	Oh. You should never try to divide something by
	itself. That's why it didn't work.
	It's being stubborn now.
	Okay.
	Sometimes subtle things matter. So here's the normalized Poisson.
	And what's interesting about that is that
	you can fit this.
	And from the available bottles you see that the Johnson SL fits the best and does a pretty good job.
	And you can also
	do Fit Y by X.
	Take this output
	and fit by requisition context.
	And get this Oneway plot.
	That looks a lot more regular than the previous one that I showed. Much more plausible that these have the same distribution. You will see a larger range when on the ones that have more data but that's not,
	that's not surprising, really.
	And if you look at the unequal variances test, you can see that there's no indication that they unequal variances.
	That's not proof that they don't, but it's a way of saying that it's reasonable to use a model where you have assumed that they have the same distribution.
	So.
	Back to this.
	Oh, I also did a Kolmogorov–Smirnov test of each context versus all of the rest of them, and of so 78 cases, only one of those showed a p-value of less than .05.
	P-values don't prove anything, but what it does give an indication of is that it's reasonable to treat these just coming from the same distribution.
	And here's just the blown up version for that one smaller subset just before, but you can see that this is plausible; it's not proof, but it's plausible.
	And with the goodness of fit test
	again, it seems reasonable that
	the Johnson SL is a reasonable model to use here to describe the Poisson parameter distribution, the normalized Poisson parameter distribution, for this dataset.
	This same thing can be done for the Gamma-Poisson and that subset of data, and that Lambda parameter, which is the equivalent parameter.
	And once again that works out.
	For doing the Sigma parameter I actually use a Sigma minus one because Sigma is on a baseline of one.
	And the model here is even lower quality, if you will, but it's necessary to use
	separate modeling step because Lambda correlates to Sigma, and so you want to be able to have that effect modeled within it.
	Again, you get a decent fit.
	In this case it's the Johnson Su.
	And so here's what's going on. We generate synthetic random distribution parameters.
	From a data subset, generate a linear model based on context.
	Normalize a parameter distribution and fit it to a common parametric continuous model.
	And then to generate the synthetic parameter obtain a random number from the normalized parameter distribution and multiply that by the appropriate linear model outcome, which is to say, to "de-normalize" it.
	Now evaluating synthetic random model parameters, I used again the KS test, and you can see that there's no way of really telling them apart, so at least it is a way of saying that it looks believable
	compared to the real cases.
	And I also made a composite model.
	Turned that into a function so it can be a callable function that would then, and this is in JSL, callable function in JMP Scripting Language
	for generating parameters for random job requisition with the context as the input, all the different context features.
	Now, not going to go into all that detail, but the point is that you can create a visualization using synthetic random parameter pairs that gives you a sense of not only where things typically are in this heat map but also
	kind of an idea of the outliers what's a plausible outlier and how far does it really go.
	And you can also see if you do a correlation that the correlation between Lambda and Sigma
	and between synthetic Lambda and synthetic Sigma are almost the same.
	It's modeling that relationship reasonably closely.
	And then, if you look at this with the real data thrown in,
	which is little black dots in the main graph.
	Ninety-eight percent of the synthetic density is in the reddish region indicated by the cross hatch, which is the middle inset with a green square around it.
	All of the data points, of which there are 43,
	fall there.
	And so.
	This again is the synthetic is modeling what really happens reasonably well. It's believable but it shows you what could happen in that other few percent of cases.
	So that's as far as I've taken the modeling, but what you can see is that you can use context; you can use other variables like how much
	specificity is put into the language for the job posting that's going to make fewer people qualified, is going to maybe scare off some people.
	You can look at the way that the language is crafted. Nowadays, people are using tools in HR to craft language that is more acceptable. You can see how much difference that makes in terms of what you get over time and multiple job requisitions.
	You could treat these as factors for the model, and then by that you could tune how you apply
	these features in creating requisitions depending upon what you need.
	But now I'm going to switch into, okay, what does this look like again. Again basic principle, so for a case of a Poisson with seven per week, 30% of the time, the count will be five or fewer.
	So you should expect one of those is that you should expect that the number of applications require
	time required to obtain a reasonably competitive selection for hire is going to vary, because even with an average of seven the number that you get is going to vary quite a bit.
	The variance of the Poisson is the same as the parameter, and so you get a broad variation.
	And so, if you're used to thinking in terms of long term averages
	or anchored by another case you could be thrown off by any specific instance.
	And the problem is this pattern recognition bias.
	The clustering illusion is this tendency to consider that the inevitable streaks or clusters arriving in small samples means that there's a nonrandom effect going on, there's some kind of intent to it.
	And it's clearly irrelevant for Poisson distributed data as R.D. Clarke found in 1946 with his analysis of German V-bomb.
	sites falling on the London area.
	People would see the groupings occurred, and they would think that that meant something, but he was able to demonstrate that it was pretty well satisfied by looking at the process as Poisson
	distributed, meaning that you're going to get clusters anyway sometimes, and certainly about as often as you saw.
	So that meant really what was happening was not that they were aiming for anything in particular, but that they were aiming in general for London and they kind of generally hit somewhere in London.
	In this case here, looking at jobs, the likelihood of getting a sequence over a span of three weeks, where you get a decreasing or increasing count is about 12%.
	In this example, the likelihood of getting a declining two week count, which is to say one's bigger than the other, is 45%.
	These are really short patterns; they shouldn't be thought of as meaning anything, but they can be thought of that way by people who aren't aware of the nature of the statistics. That's really the point here. And that can lead to overreactions.
	Now the variability by field of practice is about an order of magnitude.
	Some fields are harder to source than others, so people's expectations can be skewed if they don't understand those differences.
	And if you look at requisitions in general, not just by field of practice, you can see about a factor of 25 difference in the 95%
	range from 2.5 percent up to 97.5 percent. It's still a huge variability
	in the rates.
	And then if you understood how this was being affected by how the language of the posting was crafted, how narrowly, how generally,
	what the language is attracting, advertising all these factors,
	then you could decide how much
	you want to put into making those those rates higher and getting the process done faster.
	So here's a case for a specific job posting using that simulation model that I talked about earlier.
	Established professional at Site A and discipline field of practice 9, again this is all proprietary stuff, so it's not going to be disclosed as to what that is; that's just a particular kind of professional.
	You get this very broad difference from
	across the 95% confidence, somewhere between zero and nine with a median of two.
	This can throw people, this small numbers statistics can throw people, because they expect things to be more regular that.
	There's limitations of the model and the approach.
	Always the data, the quality of the data, matters. Infrequently hired fields with rare skill sets are going to be essentially missing from the data set in all likelihood if data collected over a short time frame.
	But on the other hand, a lot of the external variables that you don't have control over
	may be
	changing during the timeframe of data collection if you use a long timeframe as the basis for these analyses.
	And the modeling approach does not consider self cannibalization. If you have two requisitions open within a field, and do applicants apply to both or do they pick one. Maybe in cases they pick the one that they feel is the best chance.
	This model doesn't do anything about that but would represent the real world outcome just by capturing what happens.
	Doesn't represent the scope of opportunity missed, though, because if you'd probably like them to apply to all of them that they could be qualified for just to see where they might fit.
	So conclusions. Some employment application response to a job posting tends to be distributed as Poisson or Gamma-Poisson.
	And this was checked over a large data set consisting of about 2,500 total requisitions.
	The distribution parameters for the application response varies substantially.
	If normalized
	they can be fit to common profiles.
	And those concise models that result from the common profiles facilitate generation of synthetic random requisition models.
	This would allow a person to do some kind of scoping or analysis on
	likelihoods
	over different circumstances.
	I think it's important to mention that application arrival models can fill an important gap for understanding the complete employee lifecycle because they'll provide perspective for hiring managers and staffing professionals regarding those pattern biases.
	They can be used in discrete event or agent based models as a way of generating applicants if you want to do that in a way that's realistic.
	And
	perhaps most importantly, from an economic perspective,
	it's a way of framing cost per applicant versus characteristics of the job and various other adjustable and out of one's control variables external factors.
	That, then, would give a better understanding of what to expect in terms of how quickly things could be filled, how quickly openings could be filled, and how much it would cost to do that
	depending upon relationships with advertising and so forth.
	That concludes what I meant to present today and
	i'll just say that if you have any questions you should have contact information from this
	video, or where it's placed, and feel free to reach out.
Christy Spain	Okay, that was a great.
	yeah except for the part where my thing didn't work because I screwed up the DEMO.
	But that was a good, safe, it was a good SAVE I finally realized what I was doing.
	yeah JMP will not do stupid things, no matter how much you want it to.