Big Data Big Pharma

Steve Gaskin explains how using big data at the physician level saved a flagship pharma company.

Video Transcript

Steven P. Gaskin: I’m going to talk to you today about using data at the physician level and how it helps save or turnaround one of the flagship products of a pharmaceutical company that I work with. And I’m going to talk about how some of the challenges that Big Data present and a remedy and show what’s possible when we use some cutting edge models and integrate them into operations and how to get there. I say that’s important because I’ve been doing mathematical models for about 30 years now which is sort of a chastening experience at times and I feel like success for mathematical modeling and business is about 25% technology and 75% diplomacy and they didn’t teach us the diplomacy part at MIT, unfortunately. So I’ve tried a little bit. I’m sure my wife would say I have a long, long way to go but, you know, we try and every now and then we find a client that is very clever at working these things into the organization.

So what I’m talking about here is a US pharmaceutical company, whom I shouldn’t name for confidentiality reasons, they had a great product, they had a big market share. And over the years, competitors came out and that market share declined. I don’t know if you’ve ever seen this sort of situation, but it’s sort of a sad situation really. So my client, they–it was going in such a direction that they said, what the heck? Try a mathematical model out. What have we got to lose, really? You know. So we gave it a shot and as you see there is–we launched our model in about 2011 or so. And we turned around the trend or at least so my client thinks and I think so too. And in fact, it–that little upward slope on the line represents about $100 million, so it was a pretty good payoff really. It’s not bad, really. You know, I’ll take it. Considering they paid about, you know, a few 100,000 to do it. That’s pretty good. So at least that’s what we’re going to talk about but we’re going to talk about how we use this model for better targeting, increased investment where it would pay and improved operations and really how that was done. So with Big Data, you have to do a lot of integration. That’s the field of IT people. I think my chief donation or contribution to this subject is putting these colors on the little cylinders there. Does anyone know what blue, red, black, yellow, green signifies? Or any of you cyclists. It’s the world championship stripes if you’re a bicycle racer. I thought there might be many of you here. So–but maybe not. Okay. But that’s why I did it. And I don’t want–by the way, I don’t want to belittle that. That was really the hardest part of the whole exercise technically. Was getting all that data together and then me putting those colors on the cylinders. Okay. So I have a saying, Big Data, big opportunities, big headaches. I’m kind of old. I’m a curmudgeon, I suppose, and not only am I tired of standing in big conference rooms with no windows late in the afternoon but I also think Big Data, haven’t I heard of that before? Wasn’t that data mining a while ago? My–the professors I work with at MIT go, isn’t it the same as small data, don’t you get the same insights? I don’t know. But despite what you hear and the challenges it presents, I think it is possible to use it in your decisions and one way that Big Data has become better if it is big data is that we have a more complete picture of what is going on. In the olden days, we had maybe prescriptions by physician but we didn’t really know what we did to them, you know, in terms of marketing and sales effort physician by physician or not all of it. And now we still don’t know what other companies are doing with their–our company’s competitors.

We want to know what eventually but we’re getting there. Now, do any of you work with data? Yes. Yes. I should shoot one of those arrow things at you. Did you see those before lunch? Those are the best. Okay. When I get data from some manager, this is what I get. The cloud. Where there is no relationship, whatever, and I think oh I’m so depressed. This is horrible but it turns out sometimes aggregation can cause some problems. If you look at the average response across the number of physicians, it might be pretty weak but if you disaggregate then you might see one doctor as a certain response with a certain pattern, another has another, and so on. And so if you get down to a good enough level, you know, you can tell what’s going on. That’s good, right? Now, I’m going to give you a technical description of our model. I like to use a sausage machine that I actually bought my client one of these on eBay and he has it in his office. So what it does is it takes all these inputs, you know, physician calls, direct mail programs, marketing, e-mail, whatever, and it runs through this big cranking and grinding process called Linear Algebra and it finds what drives sales for particular physician and you might think, okay great, sausage machine, I don’t care. Well, let me tell you, we were talking about sales people before lunch and why they think they’re so great and to me, sometimes a sales person, they have a good understanding of their client but they don’t know the other stuff that’s going on in all of the marketing world. They might be like a person–have you ever watched a football game and to make the team win, like the guy is rubbing the bald head of the guy next to him? And that’s what causing the win, that’s what your sales people can be like sometimes. They think I am so awesome I am causing all of these sales, when in fact it might be something else entirely another part of the marketing mix. That’s what this model can help tell for each physician, okay? Which is helpful.

Here is an example–it’s about–it’s not the best example but it’s a pretty good example, not the worst example–of the model fit for one physician. What we’ve got here is those top 2 lines that are sort of close together, I think, you know, I’m sort of nearsighted but I think the blue one is actual and the green one is what our model predicted for that physician and the red one is what would happen if you did a standard ordinary lease squares regression on the whole populace and that’s what it would predict and you’ll admit that the green one predicts better than the red one, I think. One doesn’t have to have statistical measures to see that. And the reason it can do that is that it’s looking at that physician’s particular data and then it’s doing a process to make sure the regression is reasonable. It’s using a version of–a hybrid version of hierarchical base regression. Does anyone here use hierarchical base regression? I’m just curious. Someone? Did someone say yes? Come on, don’t be shy. No? Okay. Well, what it does is it basically does a regression for each doctor and then to the degree that that regression is noisy or ill-behaved in terms of bad coefficients or something like that. It averages that doctor’s coefficients based on that doctor’s work alone with the global average to make it more well behaved and the less well behaved that doctor is, the more we make them look like the global average because that’s really our best guess. We don’t have really good enough data on that doctor. But for a lot of doctors you have a pretty good idea of what’s going on. And so what this regression gives you is a coefficient for each of your element of your marketing and sales mix and then you can use that to make calculations about how much of your business came from each of those things and how much you could increase it within reason if you in–or how much it would change or increase or decrease say the amount of calls to that doctor.

So the forecasting accuracy is better with these individual level models. On the left there, you’ll see what we would get and we calculated this with the traditional regression and on the right with the individual physician approach. And again, I’m sure some of you are data experts and you’re going, come on, Steve, how can a graph be that good? You’re lying to me. You just want my money. Shut up, you know. And that’s what I would say I think. But the reason that it does that well is that, well the regressions are pretty good but also you have sort of the base prescribing level of each doctor as the intercept for each regression. So you’re pretty close to each doctor just for starters, and then the regression coefficients make it that much closer. So it works pretty well. Now, what do these models deliver? I know you don’t want to hear how it works. I don’t think I do. Well, I sort of do but it’s sort of painful. What you get is this response coefficients for various things and what you’re going to look at is which doctors are actually responding to these different elements of the sales and marketing mix. Are you calling on someone and thinking they’re doing great and really responding to you and they’re not? Then maybe you shouldn’t be calling on that person so much. If they are respond–you want to call in the people that do respond. You want to send your marketing programs to people that do respond to it. You’re getting some bang for your buck. I think that’s pretty simple and so what you can do is you can see the relative influence or importance of each type of program and also how the different doctors stack up on each of these programs and eliminate the ones that don’t respond to the degree you can and then increase the effort to the ones that do. So what you can do is you can run this once a month, you can keep it up-to-date, you can keep your finger on the pulse of the market basically and you can aggregate it because it’s in the individual level in any way you want. This Lego doctor head is my own particular contribution to the slide. So you could look at different channels or types of promotion. You can aggregate by geography or by function, physician group, you know, decile. All those things you’ve been cutting and slicing and dicing for, for years. So what? Well, I’ll give you an example. It always takes a bit of convincing at a company that this model is working, that the sales people should listen to it. And what we did was we took two–one territory and another one had really good response from the physicians on average. Much better on average than the one–than the other territory and then we started talking to the sales force and we had them talk together and discuss what was going well in the territory that it really worked out well for and what was going not so well in the other territory. And they actually found some differences in their approaches and they tried those better approaches in that other territory and it actually did raise the sales. So it was helpful in that way.

So you can use it for a lot of different things and you can optimize also as we’ll discuss. But–so how did they do this? We took it in phases because I had worked with this client about 10 years ago in another category and he said, this would be great for my product and so he convinced his boss to try it and we educated the team, we looked at the impact of the different promotions, we found some opportunities, we did some pilots and we showed that it really did kind of work. And in fact, they invested about half a million and it paid off by their measurement about $100 million. So they said, okay. I almost believe it. That’s almost enough payoff for me to like a mathematical model which is sort of a high bar, you know. I wish it were easier. But at any rate, they tried it and then they said, okay, for the next year, we’re not just going to invest in this model, we’re going to invest in some programs. We’re going to put more money in the programs that really seemed to be working. We’re going to eliminate some others, we’re going to track some more categories with the model and we’re going to build some tools to track the performance and there they think they invested about $30 million and got about $125 million out of it. So still again a nice payoff and I think the general consensus even among the sales force and the rest of marketing is that it’s working. Again, you can look at the–these–across the bottom in that left hand graph. It’s just the different kinds of effort that are shown in that bubble chart. You can look at the relative payoff of these things. You can eliminate channels that aren’t paying off. And you can increase the ones that are–and we have an optimization component that we build in that lets us say, okay, there are sort of four situations. One is what if we just did random targeting of this? What would we get and it shows that on the left hand side. How about our current methods, how much does that pay in terms of monthly units? And then if you optimize using your current budget, how much would you get? That’s the third one. And then if you just sort of let the budget ride and do as much as really pays, that’s the right hand column. And so they have two major products. On one product it paid–it going from the second to third column when you optimize, doubled the response and on the other it went up by about 70%. So which is also not bad.

Okay. So how has it been used? Well, all over the place. I’m hoping you will find your function on here and come up to me and go, Steve, one of those is my function. You’ve got to do this for me. But, you know, we’ll see what happens. At any rate, it’s useful for sales in terms of making the sales department as effective as possible. Doing structure call planning things like that. Rapid uptake for launching a product marketing mix optimization, you name it. So I think the basic message though is that you really want to be able to look at the individual physician level, aggregate it as you wish and make your decisions on the different parts of your marketing and sales force promotion mix, and just really go where it’s effective, where it’s paying off and you can see that with this. So that’s really how we modelled at the individual level to take some of the headaches out of Big Data and this shows you what’s possible with some of these analytics. And I really thank you for your courage at this time of the afternoon. I’m going to go get some light box therapy after this but, you know, at any rate, thank you very much and if you have any questions, I’ll take them.

[applause]

Male Speaker 1: So, questions for Steve?

Male Speaker 2: Yeah. I’m in the marketing mix optimization box.

Steven P. Gaskin: Yes.

Male Speaker 2: So I do similar models of this except using HP regression, we do mix models.

Steven P. Gaskin: Could you say that more clearly?

Male Speaker 2: Instead of using HP regression, we do mix models instead.

Steven P. Gaskin: Yeah.

Male Speaker 2: I guess, and so but before I was doing secondary data analysis, I did primary market research and a lot of times we were doing a lot of conjoint measurement and one of the big outlets of that was utility scores and we actually, that became an input into segmentation. So you actually had benefits segments.

Steven P. Gaskin: Yeah.

Male Speaker 2: Have you ever taken your outputs here and used that as an input for segmentation so you literally have responsive in the segments as opposed to…

Steven P. Gaskin: You know, we’ve grouped them by responsiveness…

Male Speaker 2: You have?

Steven P. Gaskin: But we haven’t used as part of a larger sort of segmentation scheme like you might with conjoint part worth which is a good thing to do, I think. It sounds like what you’re doing is good. You could add it in with attitudes about sales, effort or things like that. There’s other variables you could combine. You could also combine it with their product, the amount of prescriptions they write and things like that. So there’s a lot of ways to combine this sort of thing and I think that might be profitable. You’re right.

Male Speaker 2: Have you tried using it for instance for looking at other factors which you do not actually model that might explain why some people are more responsive than others?

Steven P. Gaskin: Yes. We did–we took their response coefficients and used those as a dependent variable and we used a logit model to try to–to predict those and we also tried discriminant analysis and it worked reasonably well. I think we could’ve used more variables maybe but it did have some effect and we were able to learn something from it.

Male Speaker 2: Thanks.

Steven P. Gaskin: Okay. You’ve asked the obligatory one question. I congratulate that gentleman there. Here I’ll shoot an arrow to him and then, here, we’ll see if you can catch it. Oops. Oh. Okay. So having avoided a lawsuit, I will conclude but thank you very much for watching my talk.