Forecasting in the pharmaceutical industry is an essential task for companies. While firms have long used Excel-based models to conduct forecasts, those models are outdated and unsuited for the sheer amount of data currently available, maintain industry experts who point out that multiple Web-based options exist. In order to give an accurate picture that’s useful for everyone from executives to sales teams, forecasters should create a model that offers detailed segmentation of the market. Doing this is easier said than done, but strategies exist to help evaluate the different options.
“Segmentation is really an interesting topic,” maintained David Wolter, vice president of consulting services at IQVIA, who noted that in the pharma industry, “many companies are about to begin probably the busiest part of their forecasting season and going into budgeting.” And as complexity of drugs changes, the space has changed as well, he said during a recent webinar hosted by Pharmaceutical Executive.
“When we talk about ‘segmentation,’ we define it as splitting up the market that a forecaster is addressing into smaller buckets than the whole piece, and that could be done in many ways,” he told listeners. Data can be split up based on the type of patient, for example by disease severity — mild, moderate or severe — as well as by line of therapy or patient age group. “The segmentation impacts the level of insight that a forecaster can provide at the end of the day to their stakeholders. We all know it’s forecasters…that drive business planning, so this segmenting directly impacts what kind of insights that can come out of the forecast at the end of the day.”
“We’ve seen the segmentation strategy kind of evolve over the years,” said Rick Johnson, Ph.D., senior software solutions lead at IQVIA. “How do you build, re-segment, allow aggregation within segments? This is a real problem for pharmaceutical forecasting. In other industries, the concept of segmentation is much less important than in pharma, where you have very complex patient journeys and very complex modalities of treatment.”
“There are more options for segmenting today as data becomes more available,” explained Wolter. “Whether it’s primary data or secondary data, it’s a much richer space, and that allows the forecaster the option to segment more in their forecasting.” In addition, “the drugs and the markets are changing in such a way…with more subgroups for patients, more personalized medicine, more competition in each market,” that segmentation when forecasting is a needed step.
But it’s not as easy as just having as many segments as possible, he asserted. “Like many things, it’s a trade-off.…The forecaster really needs to strike a balance between getting the insights from breaking their forecast into many segments and the additional burden of and challenges of having multiple segments. And so one of the key responsibilities of a forecaster when they set up and frame their segments is to think about what insights am I trying to get at? What data do I have available? What modeling tools do I have available? How much time is available to do this forecast? Is it an assessment that we need to have an answer for in the next week, or is it for a longer time frame? What bandwidth is available in the organization? Is there an appetite for looking into this?”
Companies must evaluate the epidemiological data they have to determine potential segments, he explained, and consider how patients’ treatment journeys can be mapped not only against their products but also those of their competitors. “As you know, some of the labels for new drugs are even stating in the label that this is only for patients who have relapsed or have previously taken these other drugs. So it’s necessary to think through in the forecast these slim parts of the market or particular segments. Forecasters are also asking themselves, ‘When do patients transition between lines of therapy?’ So when did they switch from, especially in oncology, from a first-line to a second-line [agent]? When did they switch to a maintenance in some areas? And all of those are individual segments and could be in the forecast. And then there are, of course, all the regular inputs to consider in segments: prevalence, compliance, persistence, etc. And then finally there’s the issue of adapting the model in case new information becomes available and you need to re-segment your market. So there’s a lot of complexity here that each forecaster needs to consider.”
Wolter explained that the most common segmentation falls into four areas:
(1) Patient-based segmentation: In addition to age, line of therapy and disease severity, biomarker status has become “very common now.”
(2) Geographic area-based segmentation: While segmenting by country is very common, it could also be based on “regions within a country that behave differently than other regions. It could in some cases be subparts of a country — states or ZIP codes that are useful to break the forecast into.”
(3) Provider- or facility-based segmentation: This is based on the entity that’s purchasing the drug. “If you have a drug that is prescribed by both general practitioners but also specialists or prescribed by both allergists and pulmonologists, it can be useful in a forecast to separate down into those different physician types. It may also be a situation where you want to separate down to certain facilities or hospitals. Sometimes we see forecasts that are broken down into different groups, like commercial payers or government payers.”
(4) Competitor-based segmentation: This is based on “splitting your model up to handle competitors.…It’s segmenting the market down by which part of the market am I going to get, and which part of the market are competitors going to get?”
According to Wolter, “the more you can segment your forecast, the more you can break out your forecast assumptions and your research into separate segments, and you can generate insights at these levels.” Forecasters could determine potential uptake of a drug based on types of physicians, which can help with decision making for companies: “Do we want to move more resources over to the physicians where we think we can get faster uptake, or do we want to move more resources in the direction where we want to put more emphasis and try to convert those physicians, who might not be our easiest segment? It can also inform what the sources of business will be, which is very helpful in business planning. As an example, will we get more patients from government and Medicare insurance plans, or will we get more patients from commercial insurance plans, and, therefore, how should we think about our pricing strategy? So in a way, segmentation is sort of the lens by which we see the commercial forecast.”
Common Segmentation Challenges Exist…
Forecasters will need to address some segmentation challenges. One of these is “a need for adjustment to a new situation,” he said. “Often in a forecast you will set your segmentation, but then as you gather market research, you learn things that may change the situation. You may learn in physician interviews, for example, that physicians really don’t treat all moderate patients the same, but rather there are two categories of moderate patients” such as biomarker-negative and biomarker-positive, resulting in the need for more segmentation.
Financial or time constraints in data gathering also may require a change in approach. Often these situations are caused by forecasters having “to stick with the existing available research and data purchases of your company, and those may be aligned to the older segmentation structure,” stated Wolter. “For your forecast models, in some cases, models can be unwieldy, especially custom-made Excel models where it’s just hard to add new segments or to collapse segments. And then there can be organization issues, where maybe if we segment out a forecast to different payer types, we have to talk to different groups within our company to get information, which may or may not be feasible. So there’s a whole challenge in this area of dealing with changing segmentation in the middle of a forecast.”
Another potential issue is when the data doesn’t sync up with the segmentation the forecaster would prefer to use. “Usually the forecast comes from two sources: One is primary research — actual phone calls or interviews or focus groups with physicians — and then secondary data: looking at epi [i.e., epidemiological] reports, preferably and usually in most cases looking at real-world data, claims data or electronic medical records data,” Wolter said. “Both of these often don’t fit segmentation in a perfect way. The market research may come from not the exact region you want or not the exact type of doctors you’re looking for, and the real-world data may come pre-segmented in a way that’s not the exact match of the detail you want.”
“A third challenge we see is there’s just a math and data management issue that comes up with breaking your forecast into segments,” he said. “Forecasters struggle with the mechanics of keeping data consistent between segments.” IQVIA envisions a standard funnel-shaped model of broad categories such as patients that can then be broken down into disease severity or eligibility for a therapy. So while some of these sub-elements such as new patients, percentage of eligible patients and percentage of patients treated might be the same across segments, others like dose timing and quantity may differ by segment.
…But So Do Solutions
According to Wolter, forecasters should take certain steps that can help alleviate these challenges. One is having “consistent funnel norms,” which can be challenging, he asserted. “There are two elements that can make a forecast complex. There is the how far you segment it and then within each segment, how complex you make the funnel. And what we think about in our forecasts and what we encourage clients to do is to think carefully about both of those because in some ways, if you want to go more complex in the amount of segments, you may want to be a little less complex in the amount of detail within your funnel.”
He cautioned against using different funnels for each segment, an approach that can lead to “chaos” with too many funnels. “A useful concept that we’ve found is for each segment to think about the common output that you want to make sure comes from out of that segment. So the funnel might be different across each segment, but you might want to make sure that for every segment, you have a certain common output.” This, he said, “really makes portfolio management and looking across the portfolio much more rich and accurate. You’re setting up your framework in advance in a way that allows you to compare apples to apples at the end of the day.”
Companies should be agile when it comes to segmentation. For example, perhaps a forecast started with two age groups but then there’s a need to split it into three age groups. “One of the concepts that’s useful here is linking,” which is the ability to turn off and turn on segments and to relabel or regroup them in order to adjust forecasting, explained Wolter.
According to Johnson, people don’t want to look at data in segments, so forecasts should “plan for segment aggregations and how segments will be ‘rolled up’ together to see combined outputs,” which can be done in many different ways but can be a challenge with Excel files. Companies may want to look into how many patients with mild severity exist across different geographic markets, for instance. “Different metrics are going to be different. So, for example, if we’re looking at market share, maybe we’re doing a weighted average or something like that to get at the overall number in the U.S.…So this idea of aggregation is actually super important.”
A final solution is real-world data, said Johnson during the webinar. “Real-world data has exploded over the last 10 years.…You can obtain data from lots of sources — lots of patient-based data, lots of different types of data [feeds] that can support data requests across the funnel. I think the really cool thing here is taking this data and figuring out, for example, if we have longitudinal data around how a patient has progressed through a therapy, how can we use that information to try and get better segments and better data within the segments? And so this typically means that we’re moving to a segmentation strategy that has hundreds of segments.”
Patient groups, for example, can be broken down by ZIP codes or Nielsen ratings, and providers and hospital networks can be broken down by their performance. In the latter situation, once a pharma company has data on how a particular hospital network is performing, “we can give that data to the accounting [department] in that hospital and say, ‘You are performing lower than this other equivalent hospital network in a different area. Why is that?’ It’s very actionable data if you can get this increased and granularity of segmentation.”
Integrating real-world data into this model can help provide much more detailed forecasting. “Automating this part of the process where you’re pulling historical data and grouping it or aggregating in certain ways, getting it into the right segmentation strategy that you have a kind of…pre-processing step is now something you can do with essentially one click of the button,” explained Johnson.
He recounted one large pharmaceutical client that that wanted to know about biosimilars’ impact on its innovator product’s market share over the previous three years. Some hospitals were continuing to prescribe the innovator drug, while others were moving quickly to biosimilars. “And of course, this is something that sales teams want to know, account teams want to know — how can we preserve our share within that hospital network as long as possible and what key messages are resonating? — which if you have this data, you can get,” he explained.
The company’s forecasting was done through an Excel spreadsheet model that had only 40 market segments, “very ungranular segmentation,” said Johnson. “That’s really not enough to tease out patterns from individual hospital networks and across individual payers.”
The firm worked with IQVIA to “set up a model that dynamically segmented the market based on individual providers and payer segments. Now of course there are a ton of providers, there are a ton of hospitals around the world, so you can’t look at all of them, so there’s sort of a pre-step where you’re picking certain sizes and certain similar types of institutions and grouping them together.” The model ended up with around 2,000 segments, which were then aggregated, producing “really nice top-level data while also allowing the account teams to drill down. And of course this was driven by a large data lake presence, which is a complex endeavor to set up, but once you have it set up, every time you get new data, the forecast can automatically update.”
“We’ve seen a tremendous growth in the number of segments that pharma companies are trying to model,” noted Johnson. But “creating new Excel worksheets and then kind of trying to manually link them back to the global summary just doesn’t work.” Web-based tools, he maintained, allow for quick re-segmenting of the market, “allowing forecasters to essentially focus the minimum amount of time on a single segment to get really good insights.” They provide “the ability to move more quickly from getting an insight from a forecast to actual action, to giving teams a way to take the data that they receive from a forecast and actually act on it. That’s something that really makes the forecast a very practical tool and something that’s very helpful.”