Cover Story
Number Crunchers
By Howard Bell
Biostatisticians are more essential than ever to medical innovation—and yes, they do have a sense of humor.
Here’s one: What’s the difference between a biostatistician and a physician? A physician makes an analysis of a complex illness whereas a biostatistician makes you ill with a complex analysis.
Biostatisticians swap jokes like this about themselves on websites and in hallways, perhaps as psychological balm for the behind-the-scenes role they play in medicine. Like the Wizard of Oz, biostatisticians work behind the curtain, even though their hands are on important levers. In peer-reviewed papers, their names typically appear last in a long string of authors. But biostatisticians’ work is essential to the validity of the research in nearly every clinical trial and observational study done nowadays. And although they don’t treat patients, they play a huge role in ensuring that the evidence for protocols and practices is accurate.
What do these number crunchers bring to medicine? “We’re trained to think differently than M.D.s are,” says Thomas Flottemesch, Ph.D., a biostatistician with HealthPartners who spends half his time at Regions Hospital in St. Paul and half at the HealthPartners Research Foundation in Bloomington. The foundation employs three biostatisticians and plans to hire more. “M.D.s,” he says, “are trained to think in terms of procedures—tests and treatments to relieve symptoms. Biostatisticians think in terms of process. We try to model and understand mathematically how data is generated and determine mathematically how well an experimental drug, device, or procedure improves patient outcomes.”
Biostatisticians don’t just analyze results. They also help shape the design of studies and even prepare grant proposals. For example, Flottemesch helps principal investigators (PIs) clarify what they need to think about: What should the sample size be? How long should the study run? What design will produce statistically significant results that other researchers can replicate in future studies?
From Farming to Pharmaceuticals
Biostatistics is not a new field. Its story goes back to the early 1900s, when Britain’s Ronald Fisher, Ph.D., developed many of the statistical methods used today. Back then, they were applied to agricultural experiments with seed varieties and growing conditions. Biostatistics became widely used in medicine in the early 1940s, when statistical analysis became part of clinical trials. Computers catapulted biostatistics to new lofty heights. Today, many biostatisticians work on clinical trials—either at drug companies, medical device makers, or at big medical research centers such as Mayo Clinic and the University of Minnesota. But job opportunities are growing fastest at pharmaceutical companies.
Two things are powering the growth in demand for biostatisticians—the ballooning volume of complex data that must be managed and analyzed and the mandates attached to new NIH money.
The National Institutes of Health is granting Clinical and Translational Science Awards (CTSAs)—$4 million to $6 million per year—to big medical centers for personnel and lab facilities that enhance a center’s overall research capability. The goal is to increase the amount of research that is translated into practical, everyday use by making other medical research centers more like Duke University’s Clinical Research Institute, which has been extremely successful at carrying out important clinical research, according to John Connett, Ph.D., who directs the Division of Biostatistics at the University of Minnesota’s School of Public Health. “One reason for Duke’s success,” Connett says, “is their big biostatistical and data management support staff, which has helped minimize proposals that are weak on design and data analysis and likely to produce results that are inconclusive, unsound, or inadequately powered.”
Fifty to 60 CTSAs will be given nationwide. Twenty have been awarded so far. Mayo has one. The University of Minnesota hopes to get one in 2008. “CTSAs will increase demand for biostatisticians and job opportunities nationwide,” Connett says.
Bull Market for Biostatisticians
Connett has seen the growing demand for biostatisticians firsthand. “Our graduates have no trouble getting jobs,” he says. “A lot of our students are from other countries. No way could we satisfy the demand for biostatisticians from just domestic students.”
The University of Minnesota is the only school in Minnesota that offers degrees in biostatistics, according to Connett. During the 1980s, the university’s biostatistics program graduated one Ph.D. per year. Now it graduates eight, along with 25 from the master’s program. Those numbers are likely to increase.
Biostatisticians who hold Ph.D. degrees are usually considered scientific peers with others working on a research project. They help frame questions and decide what direction the research should take. They are more likely than those with master’s or bachelor’s degrees to work on developing new statistical methods and applying those methods to clinical studies. And they’re more likely to write parts of research papers and grant applications. Those with master’s and bachelor’s degrees are more likely to apply prepackaged or already-developed statistical methods, put together tables and graphs, and do background analyses for grant applications.
Although some biostatisticians have degrees in mathematics, rather than biostatistics, most employers look less for people with theoretical inclinations and more for those interested and experienced in applied biostatistics. They especially want those who’ve collaborated with physicians or other principal investigators. “We look for people who are excited about solving clinical problems, not just working on theories,” says Karla Ballman, Ph.D., a biostatistician at Mayo Clinic, who chairs Mayo’s division of biostatistics. Mayo employs 155 biostatisticians: 30 with Ph.D.s, 50 with master’s degrees, and 75 with bachelor’s degrees. All specialize in a particular medical field—cancer for example, which is Ballman’s.
Drinking from a Fire Hose
The sheer volume of data medical research cranks out these days, likened to drinking from a fire hose, has dramatically increased medicine’s need for people who can help manage and analyze it in gulp-sized portions.
Certainly the most prolific source of data is coming from genomewide association studies, which attempt to find the genes that cause or increase risk for diseases so that treatments can then be tailored to individuals. Researchers analyze DNA from blood samples, looking for points along chromosomes called single nucleotide polymorphisms (SNPs). Some SNPs cause or increase a person’s risk for a particular disease—Alzheimer’s or prostate cancer, for example. Ballman says researchers are able to compare patients’ genes with those of a control group, individuals believed to have a normal or healthy genome. “We can analyze one million SNPs in a single blood sample,” she notes.
The chief strength of such studies is also their chief problem: With a million comparisons per study, finding which aberrations are disease-related and which are benign is an immense undertaking. Besides determining which SNPs are actually disease-related, biostatisticians have to calculate the percentage of cases of a disease that could be avoided if the genetic vulnerability were removed.
Genome studies generate a volume of data that is four to five orders of magnitude greater than older studies, creating what Ballman calls “an avalanche of data” and a “quantum leap in information about the inherited component of certain diseases.”
Biostatisticians have to comb through it and make sense of it, which isn’t an easy task. Genome studies using SNPs are so statistically complex that the NIH often turns down proposals because they don’t show that enough senior statisticians will be part of the research team. “Not many people are graduating yet with the expertise needed for these studies,” Ballman says. Partly for that reason, she likes to hire “quick learners” who can work in statistical areas they didn’t train in. To remedy this knowledge gap, Connett says the Division of Biostatistics began offering a statistical genetics course at the master’s level five years ago. This year, they began offering one at the Ph.D. level.
Relatively conventional statistical techniques are adequate for analysis and interpretation of SNP data collected from blood samples, according to Ballman. But the next step, digging deeper using RNA and proteins from tissue samples to search for networks of interacting gene variants and how they, in turn, interact with lifestyle and environmental factors, will require more sophisticated methods of statistical analysis.
Clinical Trials Management
Genome studies are cutting-edge stuff most medical centers and biostatisticians don’t do. Instead, many biostatisticians help manage large clinical trials. Daniel Sargent, Ph.D., who has crunched numbers for clinical trials at Mayo Clinic for 11 years, emphasizes that biostatisticians don’t just run the numbers after the study is done. “I help the PI take a clinical idea, refine it, and formulate it into a mathematical question answered with data,” he says. Do they want to study how much longer a treatment lets people live? Or do they want to measure how well a treatment shrinks tumors or how much less toxic a treatment is? “Answering each of these questions calls for a different study design and a different statistical analysis,” Sargent says.
Because of such input, biostatistics has changed thinking about cancer. For years, physicians stuck to the empirical notion that tumor shrinkage equaled improvement. “When you look at large data sets, tumor shrinkage often doesn’t result in patients living longer,” Sargent says. “Statistics allow us to assess an empirical notion and determine that it is misleading. Instead of asking, Did the tumor shrink, we need to ask, Did the patient live longer?”
Sample size in any study is the key, Sargent says. “If a drug can reduce the risk of death after three years by 80 percent,” he explains, “that’s pretty obvious, and we wouldn’t have to treat many patients to prove that it works. But if it reduces the death risk by only 10 percent, that may still be significant; though it’s something most doctors wouldn’t notice, so we have to treat many patients and look at averages over those patients.”
In addition, Sargent designs data forms or case-report forms, which identify what type of information the PI needs to collect and how it should be collected in order to answer the research question. Toward the end of the study, he uses statistical analysis to tease out an answer to questions such as Did it work? How well? Did it work better in one gender or another? Did it work better in Caucasians or non-Caucasians?
Data from multi-center trials continually pour into Mayo from participating medical centers. Errors are common. Often it’s simple things such as mistyping a value. Sometimes it’s a more complex error such as mismultiplying bi-dimensional tumor measurements. Or, instead of submitting a percent change in tumor size, the treating site submits the absolute change in size. Biostatisticians and other trained data managers use computers to detect these errors.
Sargent runs monthly toxicity reports on the side effects of medications. Periodically during the course of a trial, he does what’s called an interim analysis. “If we find an experimental treatment to be statistically ineffective or harmful, we can stop the study and get patients on a traditional treatment,” he says. “If the numbers show the treatment works better than we’d hoped, we can end the study, cross over patients in the control group so they can get the new treatment, too, and get the drug to market quicker. We see it as an ethical obligation to review the data as it comes in. Statistics speed up the time it takes to research an experimental treatment.”
Herceptin became available for treatment of breast cancer earlier than expected after interim findings during a trial headquartered at Mayo Clinic showed it reduced breast cancer recurrence by about 55 percent—much better than anticipated—as compared with chemotherapy alone. Oxaliplatin for advanced colon cancer is another example. When combined with chemotherapy, it increased survival, delayed tumor progression, and increased tumor shrinkage much better than chemotherapy plus irinotecan, the conventional drug. “It’s empirically obvious in these cases that the new treatment is more effective,” Sargent says. “But the greater effectiveness must be objectified mathematically.”
Historically, the FDA required that survival rate be tested for five years for all colon cancer drugs designed to prevent recurrence of a tumor after surgical removal. That changed to three years after an international consortium of colon cancer experts led by Sargent analyzed data on 21,000 patients that showed three years of study was enough to accurately predict whether a patient would be alive in five years. So now colon cancer drugs get approved two years sooner. “This wouldn’t have been possible,” Sargent says, “without biostatistical involvement.”
Biostatisticians on Call
Being on call to answer questions is a big part of what most biostatisticians do, according to Connett. “I get a lot of spontaneous calls from physicians doing clinical trials who have a question about study design or the significance of some data. By far the most common question we get is, ‘What sample size do I need?’”
At its Twin Cities campus, the University of Minnesota employs more than 150 biostatisticians who do a variety of tasks. Connett crunches numbers for NIH-funded multi-center COPD studies searching for ways to prevent progression to emphysema. He designs uniform protocols so every center submitting patient data submits the same information in the same way. Then he and his colleagues analyze the data using statistical packages—a bundle of mathematical methods (algorithms) for analyzing data in a limited number of steps.
Biostatisticians at the university’s Biostatistical Analysis and Design Center help physician researchers with data analysis and study design. At the design stage, they help investigators clarify their study goals. That means asking questions: What kind of patient improvement do you expect your experimental technique to have? If, for example, a disease causes a 50 percent mortality rate, do you hope to decrease that by 10 percent? What kind of statistical power will your study have? In other words, what is the probability that the study can detect a meaningful improvement in patient outcomes? “Ideally,” Connett says, “studies should have power levels of 80 or higher—an 80 percent chance or better of finding improvement if the new treatment really is better.”
Another consideration is setting a target significance level. “It’s important to establish this P value—or significance level—goal in advance and use it as a basis for designing your study,” Connett says. P value is a measure of how well you’ve proven what you set out to prove. The lower the P value, the more significant the findings. Results with a P value of less than .05 are often considered statistically significant.
Many Hats
Biostatisticians who don’t work at big research centers such as Mayo Clinic or the University of Minnesota tend to wear many hats—helping physicians in all specialties on an array of projects. At Regions, Flottemesch uses differential equations and queuing theory to study patient wait times in the emergency department in order to make staffing more efficient and avoid crowding. He calculates how many physicians the emergency department needs during a particular six-hour shift, assuming 10 new patients per hour. He does similar patient-flow analysis for the post-op care unit.
Flottemesch also consults with physicians who have an idea for a research project and want to know how to design the study and reviews clinical trials before HealthPartners agrees to enroll patients. At journal clubs or one on one, he teaches HealthPartners doctors how to read papers from a design point of view. Is it a solid design? Did the researchers’ analysis support their conclusions, or are they going beyond their data? “Typically, physicians don’t think about these things,” he says.
Melissa Skeans, M.S., works on population studies for the Minneapolis Medical Research Foundation at Hennepin County Medical Center. As part of the foundation’s Chronic Diseases Research Group, which contracts with the U.S. Renal Data System coordinating center, she specifically crunches kidney transplant data. Using Medicare records, she analyzes what types of patients need transplants, how long they wait, their post-transplant outcomes, and factors affecting those outcomes. “In a nutshell,” she says, “I take huge piles of data on transplant recipients and produce tables and graphs that are easy to digest.”
Like Flottemesch, she also serves as a resource for physicians doing research. Skeans helped a cardiologist who wanted to know the incidence of bacterial endocarditis in people waiting for a kidney transplant versus those who’d already had a transplant. “I used Medicare data and standard statistical techniques to answer his question,” Skeans says.
Colleen Renier, B.S., serves as the lone biostatistician for the entire SMDC Health System in Duluth. A biostatistician for 27 years, she earned her degree from the University of Minnesota before master’s programs were as common as they are now. She moved from the university’s Duluth campus to SMDC nine years ago.
Renier helps physicians prepare grants for submission to the NIH or medical foundations. She’s run numbers for studies on topics ranging from preventing cardiovascular disease in women to communitywide screening for blood-lead levels. She helped design a new research model to study internal defibrillators. And she’s collaborating with other colleagues on a five-state CDC-funded study of agricultural injuries.
Renier keeps researchers from straying off course when drawing conclusions. “I help them write their paper,” she says, “so they don’t make inaccurate statements about what the data say. Sometimes principal investigators see things in the results that aren’t there because they saw it anecdotally in their patients.”
Renier says most physicians don’t know what biostatisticians do. “They think we’re geeks who wear pocket protectors,” she says. (She’s not, and she doesn’t.)
Team Players
Gaining the esteem of physicians can be a challenge for the people who are sometimes bearers of bad news—interpreting statistics that disprove a hoped-for finding. That can create friction. “The key to biostatisticians earning respect and trust,” Connett says, “is clearly explaining the numbers or why certain design features in a study are necessary.” Meanwhile, physicians need to understand and accept that there are best practices for study design. “There needs to be true collaboration between all involved in a project,” he says.
Sargent says close collaboration with physicians helps biostatisticians understand the important issues in a particular diagnosis so they can better design a study. “The relationship is not confrontational or argumentative,” he says. “If a PI questions the numbers, we show them the raw data. We’re transparent.”
Biostatisticians working on clinical trials never work alone but rather in groups that continually cross-check each other’s work. “We hold ourselves to a very high level of rigor,” Sargent says. “Our reputations are on the line. So are our patients.” Although ultimate responsibility for the quality of a study rests with the PI, biostatisticians do have an ethical responsibility to inform oversight committees or research sponsors if there is anything suspicious about data integrity or conclusions drawn from data.
And that means number crunchers need good people skills to do their work. When Ballman is hiring new biostatisticians, she looks for people who are plainspoken and good at explaining complicated concepts in ways nonbiostatisticians understand.
Having a sense of humor helps, too, which brings us to the biostatistician whose wife had twins. He called their minister to share the good news. “Bring them to church on Sunday, and we’ll baptize them,” said the minister. “No,” replied the statistician. “Baptize one. We’ll keep the other as a control.” MM
Howard Bell is a medical writer in Onalaska, Wisconsin.