Leadership in Statistics: Increasing Our Value and Visibility

ABSTRACT

Scientists in every discipline are generating data more rapidly than ever before, resulting in an increasing need for statistical skills at a time when there is decreasing visibility for the field of statistics. Resolving this paradox requires stronger statistical leadership to guide multidisciplinary teams in the design and planning of scientific research and making decisions based on data. It requires more effective communication to nonstatisticians of the value of statistics in using data to answer questions, predict outcomes, and support decision-making in the face of uncertainty. It also requires a greater appreciation of the unique capabilities of alternative quantitative disciplines such as machine learning, data science, pharmacometrics, and bioinformatics which represent an opportunity for statisticians to achieve greater impact through collaborative partnership. Examples taken from pharmaceutical drug development are used to illustrate the concept of statistical leadership in a collaborative multidisciplinary team environment.

KEYWORDS:

• Big data
• Communication skills
• Drug development
• Influence
• Statistical leadership

1. Low Visibility Amid High Demand

Statistics is the science of learning from data and making decisions in the face of uncertainty (Rao 2001; Lindsay, Kettenring, and Siegmund 2004; Hand 2009; Davidian and Louis 2012). Virtually every research discipline generates data, whether it is planned data resulting from scientific experiments and carefully designed surveys, or massive observational data generated by online transactions, social media, or sensor technology. The need to design investigations, plan for data collection, and analyze and interpret the resulting data is why the field of statistics is valuable to every research discipline. Unfortunately, this high demand for statistical skills comes at a time when there is a problem of low visibility for our profession (Rodriguez 2015). While we see ourselves as “stewards of sound thinking for good decision-making,” the public too often sees us simply as number crunchers (Wasserstein 2015).

To resolve this paradox, it is useful to examine what Deming (1965) called the most important role of a professional statistician, namely that of a collaborator. The demand we face is for statisticians who can collaborate within multidisciplinary teams to design an investigation with the best chance to answer a research question, predict an outcome, or make a decision. Collaborative work within a multidisciplinary team should not be confused with short-term consulting, which is usually limited to brief transactions such as conducting a requested analysis or explaining minor technical details (Cox 1997). Strong technical capabilities alone are not sufficient to guide a research team toward the best design or decision; productive collaboration also requires statistical leadership.

Statistical leadership is the use of influence to guide a multidisciplinary research team to adopt the best design or decision based on the available data. The goal of this article is to characterize statistical leadership in a collaborative environment in terms of the essential competencies of listening, networking, and communication. The concepts of individual, enterprise, and policy leadership are introduced and illustrated using examples from drug development. The growth of big data is identified as an important opportunity for today's statisticians to demonstrate collaborative leadership working within an expanded multidisciplinary team that may also include experts from other quantitative disciplines. The article concludes with a discussion of the stakeholders that have an interest in building statistical leadership, namely, universities, employers, professional societies, and statisticians themselves.

While the illustrations are taken from my experience in drug development in an effort to make these leadership concepts more concrete, I believe they are applicable to statisticians working in any sector of academia, government, or industry. Statisticians in other sectors should be encouraged to share their own examples of leadership to contribute to the growth of our profession. Building statistical leadership is the key to successfully influencing multidisciplinary research across many scientific and quantitative disciplines, which will drive demand and increase visibility for the field of statistics.

2. Statistical Leadership in a Collaborative Environment

Statistical leadership in a collaborative environment is the use of influence without authority to guide the design, strategy, and decisions of a multidisciplinary team. Although statisticians in a research team may have no formal organizational or hierarchical authority over their colleagues, they can nevertheless have a significant impact on the research design and decisions based on their ability to influence their peers. Multidisciplinary research teams are a network of relationships that are not hierarchical in nature. Statisticians can employ statistical leadership in this setting regardless of their seniority or perceived place in the hierarchy. This requires a willingness to speak up and share ideas, together with the competencies necessary to translate those ideas into a clear and persuasive recommendation. Statisticians can also seek to lead at different levels beyond a multidisciplinary team as they develop further in their careers.

2.1. Three Key Competencies

The keys to statistical leadership are competency in listening, networking, and communication. These competencies are often referred to as soft skills because they are personal attributes that enable effective interpersonal interactions. These skills are not easy to acquire and require a commitment to ongoing professional development. When coupled with strong technical capabilities, these attributes are essential for successful collaborative leadership from a statistician working as a part of a larger multidisciplinary research team. In the following I explore these three competencies further.

Understanding the rationale and objectives of the research team is the precursor to successful collaborative leadership, and requires both listening and asking questions. Active listening is a process that requires the statistician to explain their perception of the problem and ask various members of the research team to identify any misconceptions or additional relevant information to ensure an accurate understanding of the problem from multiple viewpoints. This should be underpinned by developing additional subject matter knowledge of the discipline at the center of the research. An understanding of the discipline and the research objectives is essential for designing the investigation, collecting the right data, and determining the best analysis of the data. This process may also offer opportunities for the statistician to help the team define or further refine the research question. As George Box (1976) noted, successful collaboration requires “the wit to comprehend complicated scientific problems, the patience to listen, the penetration to ask the right questions, and the wisdom to see what is, and what is not, important.”

The multidisciplinary team is a network of relationships which need to be proactively developed to facilitate successful statistical leadership. Influencing a team is different from influencing a single individual, and requires an understanding of how the team members fit together in a research environment. Respect for the members of the team and their expertise is the foundation of any good collaborative relationship. Statisticians should meet with their colleagues regularly to understand their needs and objectives, and also to learn about their expertise and how they communicate. Nowadays, too many statisticians wait to be approached by members of the team, which limits their interaction to a brief consultation on an investigation that is already underway. The most creative aspects of collaboration often occur at the early stages of the project, which makes proactive networking essential to establish relationships with the team.

Communicating with impact to a multidisciplinary audience represents perhaps the greatest opportunity for statisticians. Whether in verbal or written communications, statisticians too often get lost in providing details that are not pertinent to their intended audience and misjudge the ideal length. It is better to strive for simplicity using “as much, but only as much, statistical complication as the circumstances warrant” (Marquardt 1979). Preparation is essential for statisticians to adequately organize their thoughts and ideas before meeting with the research team. It is important to convey respect when communicating with the team and to adapt to the audience, such as condensing key messages into their essentials when speaking to senior leaders within a company or organization. Of course, the implicit element of communication is a willingness to speak up and share ideas. Statisticians often fail to realize that members of the research team really want to hear their input—in a manner they can understand. As former American Statistical Association (ASA) President Nancy Geller (2011) stated, “in statistics, the three most important things are communication, communication, communication!”

2.2. Influence Requires a Clear Recommendation

Statisticians can leverage these competencies to provide recommendations that influence the direction of the research team. Changing the design of an investigation, sampling from a different population, or making a different decision based on the data are all examples of statistical recommendations. Each requires the statistician to guide the team to an understanding of the relative advantages of alternative strategies or decisions, and to answer questions along the way. Patience is also needed to adequately develop a relationship with the team, so that the intended level of influence does not get ahead of the relationship. Influence is a process and not an event, and the network of relationships are important in persuading the team to adopt changes in the research plan. Statisticians should try to avoid equivocation when offering a recommendation to the team—a common frustration from multidisciplinary teams arises when their desire for a clear recommendation is never realized. If two or more statisticians are on the team together, or handing off from one to another, they should take care to align their recommendations in advance. The research team will not agree with every recommendation, and judgment is important in determining when to press further. In some cases, a clear recommendation is simply the courage to state clearly that there is not enough data to make the desired determination.

Statisticians should take care not to be too quick to recommend against proposals from within the team. We are too often perceived as “the naysayer who just throws cold water” (Geller 2011). This pessimism is often regarded as a “can't do” attitude and undermines successful collaboration. Statistical leadership should not drift into policing the work of others, which erodes the respect within the team necessary for constructive collaboration. A research team with a good understanding of the risks and limitations of different approaches will usually choose the best design or decision within the constraints of feasibility. However, when statisticians are focused too intently on the risks associated with a design, they may react too sharply without taking the time to discuss alternative designs and their relative advantages. Guiding teams to adopt the best strategy based on data by listening and sharing ideas and recommendations will lead to better research, and conveys a more constructive “can do” attitude.

2.3. Leading at Different Levels

A statistician can function as an individual leader, an enterprise leader, or a policy leader, depending on the level of influence necessary to achieve the desired outcome. An individual leader seeks to influence the design, strategy, and decisions of a multidisciplinary research team. Statisticians at every level should strive to employ individual statistical leadership when working within a collaborative research team. Statisticians do not begin their careers as highly skilled individual leaders. Instead, they become leaders through experience and professional development, but only when they are willing to actively focus the appropriate effort on this dimension.

In contrast, enterprise leaders influence an entire organization and are often senior statisticians in a position of scientific or organizational leadership. Enterprise leaders work to change or improve the way a business, function, or division learns from data and makes decisions in the face of uncertainty. An enterprise leader drives influence and alignment beyond their immediate organization to help a business use statistical reasoning to create value based on data. Implementing statistical machine learning to optimize operations across business divisions is an example of enterprise leadership. Providing a quantitative framework for ranking programs across a research portfolio to facilitate better investment decisions is another example of enterprise leadership from statisticians.

Policy leadership is demonstrated by statisticians working across academia, industry, and government to advocate for the acceptance of new and innovative statistical methodology in the design and development of public and scientific policy. Long-standing methodological approaches used in policy domains such as environmental and climate science, criminal justice, and health care are often resistant to change. Janet Norwood (1990) advised that no statistical method in public policy is absolutely perfect, and thus “we should always be searching for improvements, and we should not be afraid to adopt them.” The unmet need for statistical thinking to play a greater role in shaping public and scientific policy is well-recognized (Bailar 1988; Norwood 1990; Keller-McNulty 2007; Morton 2010). The recent White House Precision Medicine and Cancer Moon Shot initiatives are examples of opportunities for statisticians to demonstrate policy leadership.

3. Examples of Statistical Leadership from Drug Development

The concept of statistical leadership at different levels is illustrated using examples taken from pharmaceutical drug development. A brief overview of drug development is provided to give the necessary context for the subsequent examples. For each of these examples, the background of the problem is briefly given, as well as the methodological or strategic elements involved in the solution, and a description of how statistical leadership was demonstrated. The examples are intended to make leadership concepts more tangible to statisticians seeking to understand and develop their own statistical leadership ability. While the examples predominantly reflect my own work experience, they illustrate key aspects that apply beyond the specifics of drug development. As such, I believe that the following examples are useful for statisticians working in any sector of academia, government, or industry.

The drug development process spans discovery in the laboratory, preclinical safety testing, first in human trials of safety and tolerability (phase 1), dose ranging trials to identify the best dose or doses (phase 2), confirmatory testing of those doses to demonstrate safety and efficacy for regulatory approval (phase 3), and post-approval trials of other uses (phase 4). The early phases of drug development are learning phases that tend to be more exploratory in nature as compared to the confirmatory phase which focuses on large trials designed to demonstrate evidence to support regulatory approval and market access. Large pharmaceutical companies often manage a portfolio of several hundred drug development projects for different experimental medicines, resulting in a large number of clinical trials conducted simultaneously across all phases of development.

Drug development projects are typically run by cross-functional teams with experts representing a variety of disciplines and functions such as medical, epidemiology, pharmacometrics, chemistry, quality assurance, commercial, health economics, legal, data management, and project management. Statisticians are part of the core team responsible for the design and planning of clinical trials necessary for a successful drug development program. The opportunities for statistical leadership across the many phases of drug development are considerable as statisticians collaborate with other members of the project team and provide input or guidance on design, data collection, analysis, interpretation, and decisions related to clinical trials. The following examples cover experiences in emerging safety signals, a quantitative framework for portfolio decisions, and model-based methods for dose selection.

3.1. Interpreting an Unexpected Safety Signal

Individual leadership from statisticians is critical in evaluating emerging safety issues, which require a strategy for the ideal data and methodology to characterize unexpected adverse events in the right context. An example comes from a single case of male breast cancer reported in a large cardiovascular outcomes trial of a drug for which cancer had never been a prior concern. Medical colleagues wanted to know whether “this case was statistically significant.” The statistician consulted with an epidemiologist to demonstrate that a single observed case was consistent with the number of cases expected in an age-matched group followed for the same period of time. Regulatory authorities immediately requested a cumulative meta-analysis of cancer incidence across all completed clinical trials according to an analysis plan specified in advance. One of the key strategic issues was how to select the trials for inclusion in the analysis from the cumulative database of approximately 100 clinical trials conducted in many different populations and under varying conditions. The multidisciplinary team held a very strong opinion that literally every trial ever conducted should be included in the cumulative meta-analysis.

Instead of simply delivering the requested analysis, the statistician demonstrated individual leadership by guiding the team to understand that the scope of the trials under consideration were very different, such as food effect trials, pharmacokinetic trials, drug–drug interaction trials, open-label trials, comparative trials, and large outcome trials. A meta-analysis conducted across such heterogeneous trials and populations would have been very difficult to interpret. The statistician instead led the team to develop and agree upon inclusion criteria that consisted of long-term comparative trials that followed participants for a minimum period of time, which resulted in a more homogeneous set of trials better suited to characterize cancer incidence by body system. This reduced list of trials captured the majority of person-years of exposure in the database of trials, that is, the cumulative time at risk for adverse events across studies for participants that received the drug. The short-term trials which were excluded contributed relatively little additional person-years of exposure and were the most heterogeneous in terms of design and population. This analysis was accepted by the regulatory authority, confirming that there was no evidence of cancer being a safety issue. In this case, statistical leadership was critical in helping the team understand best practices in meta-analysis and identify the right data necessary to address an important question (Sutton and Higgins 2008).

3.2. Data-Driven Portfolio Management

Statisticians can demonstrate enterprise leadership by influencing the strategy or decisions beyond their immediate organization. In this example, statisticians seized an opportunity to develop a framework for the evidence-based management of an entire portfolio of drug development projects. Managing a portfolio of drug development projects involves a series of decisions that lead to investments associated with substantial financial risks. Most large pharmaceutical companies review accumulating data at key development milestones to maintain an ongoing prioritization of projects across the portfolio in terms of strength of evidence, risk, and return on investment. This portfolio analysis is used to determine which projects will receive additional funding for continued development. The senior leader responsible for the portfolio review asked the statistics team to develop a simple quantitative framework to facilitate the objective evaluation of complex clinical evidence in the decision making process. Developing the framework spanned six months, during which several different approaches were presented and modified based on input from the network of senior leaders involved in the portfolio review.

Multi-criteria decision analysis was used to evaluate accumulating project data relative to a range of outcomes from a base case to the ideal result. This begins with a quantitative target product profile (TPP) that specifies the key efficacy and safety endpoints necessary to characterize the viability of the drug under development. The TPP provides a range of outcomes against which the cumulative clinical efficacy and safety data are compared at each milestone between development phases, such as the transition from phase 2 to phase 3. The posterior distribution of the treatment effect for each endpoint based on the accumulated data is used to calculate the likelihood of exceeding either of the targets specified in the TPP (Nixon et al. 2016). This quantifies the cumulative evidence relative to the TPP objectives and accounts for the uncertainty depending on the amount of available data. Tailored graphical displays, such as Nightingale plots, can be used to capture the multivariate evidence for the project in a format that is more accessible to the stakeholders and decision makers.

These probabilities are used to make a qualitative evaluation of the data based on multidisciplinary input that translates evidence from multiple factors to a common qualitative scale. An overall recommendation is then developed for each project in light of the strength of evidence of the data. A key success factor in this framework is individual leadership from each project statistician to drive discussions on defining the quantitative limits in the TPP, synthesizing the available evidence, guiding a robust multidisciplinary discussion of the results, and contributing to the decision options. When this framework is applied consistently to all of the projects in the portfolio, the result is a more disciplined approach to portfolio management with evidence-based decisions on each project, which hopefully leads to smarter drug development. In this case, enterprise leadership was demonstrated by influencing stakeholders across the business to see the compelling value in adopting a new framework for better portfolio decisions based on data.

3.3. Model-Based Dose Selection

Driving the increasing use and acceptance of model-based methods for dose selection within the scientific and regulatory community is an example of policy leadership from statisticians and pharmacometricians. Selecting the optimal dose for confirmatory testing in large phase 3 trials is one of the most important decisions in the drug development process. A common method of dose selection is the pairwise comparison of two or three fixed dose levels with a common control (e.g., placebo). Pairwise comparison designs are based on a traditional mindset of hypothesis testing with large samples, which are better suited for confirmatory goals and may not match the scientific objective of characterizing the dose-response relationship. Using only a limited range of fixed doses often fails to capture the steep part of the dose–response curve crucial for determining the best dose, and often results in the need for additional trials to explore lower or higher dose levels.

Model-based designs and analyses for dose selection are more efficient than traditional pairwise comparisons. Such approaches often recommend the use of a 10-fold (or larger) dose range of 4–7 doses with the high dose close to the maximum tolerated dose and more doses at lower exposures, within practical constraints. Integrated statistical and pharmacometric methods have been extensively developed and are available to characterize the dose–exposure–response (D-E-R) relationship. MCP-Mod is an example of an efficient statistical methodology for model-based design and analysis of dose finding trials under model uncertainty (Bretz, Pinheiro, and Branson 2005). The method consists of multiple comparisons (MCP) of a set of contrasts to determine a significant dose–response signal based on a set of candidate dose–response models, and then uses either model selection or model averaging (Mod) for target dose estimation to inform dose selection for the subsequent confirmatory development phase.

The MCP-Mod method received a positive qualification opinion from the Committee for Medicinal Products for Human Use (EMA 2014) and a fit for purpose determination from the U.S. Food and Drug Administration (FDA 2016) as a result of statistical leadership and collaboration across industry, academia, and regulatory bodies spanning a decade. This collaboration also included the 2011 EMA workshop on “Modeling and Simulation” and the 2014 EMA workshop on “Dose-Finding.” In particular, the qualification opinion noted that MCP-Mod promotes better trial designs with a wider dose range and more dose levels and enhances multi-disciplinary discussions of trial designs and risks. Both regulatory determinations are examples of changes in scientific and regulatory policy that advocate for the greater acceptance and use of a model-based design and analysis for dose selection in drug development.

4. Opportunities in Big Data

The growth of big data has been identified as the greatest opportunity for today's statisticians, driving an increasing demand for statistical skills (Madigan and Wasserstein 2014) that exceeds the available supply of statisticians. This shortage is amplified by the decreasing visibility for the field of statistics, as noted by Speed (2014), Rodriguez (2015), and Morganstein (2015) among others. To seize the opportunity of big data, statisticians will need to invest in a better basic understanding of this new area as part of their own training, while embracing the unique expertise of other quantitative disciplines to successfully apply collaborative leadership within an expanded research team.

Research with big data looks for patterns in datasets that do not originate from a planned sample or designed experiment. In addition, such data are often too large for processing and analysis with traditional software tools. In health care, for example, big data may consist of electronic health records from a network of hospitals, health claims data representing health care costs from different providers, or genomic sequencing consisting of millions of variables. The increasing volume of health care data may have the potential to help predict important health outcomes or develop precision medicines. The rapid growth of big data creates new opportunities for statisticians to collaborate on issues relating to minimizing bias, false discovery, and generalizability of results from data that is not sampled and may represent almost the entire population.

There are not enough available employees with the statistical skills needed to capitalize on the opportunity of big data, and the shortage is projected to grow to between 140,000 and 190,000 employees by 2018 (Manyika et al. 2011). This anticipated shortfall was confirmed in the recent national survey on the supply and demand of data analysis skills for jobs of the future, sponsored by the ASA and conducted by the Society for Human Resource Management (Mulvey, Esen, and Schramm 2016), with 78% of employers who filled a position requiring data analysis skills within the last year reporting difficulty in recruiting qualified candidates. Employers are therefore increasingly looking to graduates from alternative quantitative disciplines to fill these positions. Data scientists, for example, have diverse training that includes database management, statistical modeling and machine learning, and distributed and parallel computing (Cleveland 2001; Rodriguez 2013). Although the number of new graduate degrees in statistics has grown to 3500 annually in the U.S. (Rodriguez 2012b; Pierson 2015), it will not be adequate to keep pace with the increasing demand for statistical skills, and traditional statistical training is not sufficient to meet the needs of research in big data.

The shortfall is magnified by poor visibility for the field of statistics, a long-standing issue resulting from the failure of statisticians to communicate clearly how they add value, as highlighted by former ASA Presidents such as Marquardt (1987). Many years later, ASA 2012 President Bob Rodriguez warned “not only are we invisible, we don't even know what visibility would look like” (Wasserstein 2015). As a result, researchers often develop misconceptions that statisticians are compilers of data and number crunchers who deliver analysis requests for defined problems. Their experience suggests that statisticians tend to be very conservative, whereas computer scientists and data scientists are seen as more flexible and adventurous when working on problems, especially with big data (Brown and Kass 2009). Leo Breiman (2001) referred to these different attitudes as “the two cultures.” This may explain the absence of any statisticians among the nineteen participants at the White House Big Data Partners Workshop as noted by Morganstein (2015).

This is compounded by statisticians who become apprehensive and competitive when working alongside other quantitative scientists. Many statisticians feel these alternative quantitative disciplines should belong to statistics because they also apply statistical methods to data. This anxiety first emerged with the popularity of data mining (Kettenring 1997) and remains a concern today as statisticians feel they are losing ground to computer scientists (Matloff 2014). Kowalski (2015) described a similar long-standing “tension and skepticism between biostatisticians and pharmacometricians.” While these other quantitative disciplines share some common elements with statistics, they are much more specialized with many unique capabilities. Data scientists are trained in database management, distributed and parallel computing, and machine learning with greater focus on working with real data. Pharmacometricians routinely fit nonlinear mixed effects models to longitudinal pharmacokinetic data and refine their skills to an art. Computational biologists combine molecular biology, genetics, computer science, statistics, and machine learning techniques (Zelen 2003). These are all neighboring disciplines with whom statisticians share some common language and customs, but it is important to appreciate the elements that make them unique.

Statistical leadership is essential for statisticians to realize the opportunity of big data and achieve greater impact through collaboration in a team environment that has expanded to include other quantitative disciplines. Influence within this type of multidisciplinary team requires a greater appreciation of the unique capabilities of alternative quantitative disciplines such as machine learning, data science, and bioinformatics. It also requires adopting a more flexible exploratory attitude as opposed to a purely confirmatory mindset (Tukey 1980).

5. Building Statistical Leadership

The stakeholders that should have the greatest interest in building statistical leadership are universities, employers, professional societies, and statisticians themselves. Universities are increasingly integrating leadership training into their programs, as well as offering their students an introduction to collaborative leadership through their consulting centers. Internships in the private sector and government represent opportunities for graduate students to gain exposure to real-world multidisciplinary research. Professional societies are also beginning to offer programs for professional development that include leadership training. Finally, statisticians themselves must invest more in actively seeking out mentoring and coaching from others to further develop their own individual leadership capability.

DeMets et al. (2006) recommended that training in leadership and communication should be integrated into graduate biostatistics programs, and graduate schools are beginning to offer these courses. The University of North Carolina developed a course entitled “Leadership in Biostatistics” to help prepare biostatistics students for leadership roles in academic and non-academic public health settings. The course consists of modules including leadership concepts, management skills, leadership styles, and guest leader presentations (LaVange et al. 2012; Rodriguez 2012a). Students are presented with real examples of statistical leadership as described by leaders working in the field. The University of Pittsburgh offers a course entitled “Scientific Communication Skills” to strengthen written and oral communication abilities to convey statistical concepts to a nonstatistical audience (Buchanich 2012). Presenting in a contributed session at a professional meeting is another opportunity for students in statistics to develop their communication skills.

Most graduate programs in statistics have a consulting center that provides statistical consulting services to the university research community while also training graduate students. These are often short-term consultations to offer guidance on design or sampling plans, advice on statistical methods, help with an analysis in SAS or R, or interpreting the results. In some cases, the consultations result in long-term collaborations with researchers from other disciplines. Undergraduate programs such as the University of Georgia offer a “capstone” course that provides students with hands-on data analysis experience through exposure to real interdisciplinary work with researchers from other fields (Lazar, Reeves, and Franklin 2011). The course concludes with a poster session, in-class presentations, and a written report to provide students the opportunity to develop their written and oral communication skills. Graduate programs are also beginning to include interdisciplinary training as part of their curriculum. Begg and Vaughan (2011) recommended that graduate students undergo “a longer-term immersive experience with extra-disciplinary colleagues at various levels of seniority.” Teamwork, interdisciplinary communication, integration skills, and developing subject matter expertise for a research project are key learning objectives.

Internships offer an important opportunity for graduate students in statistics to gain hands-on experience collaborating on real problems under the direction of an experienced statistician prior to starting their career. Over 30 organizations advertised 2016 summer internship opportunities in the November 2015 AmStat News including AbbVie, Chevron, Eli Lilly, FDA, Federal Reserve, Genentech, Google, Hartford Financial Services, Institute for Defense Analyses, Liberty Mutual, Lubrizol, Mayo Clinic, Merck, National Cancer Institute, Novartis, Pew Research Center, Pfizer, RAND Corporation, SAS Institute, Travelers, and the US Census Bureau. The key to successful internships is matching interns to mentors who can guide them as they work on a well-defined research project. Internships provide experience in a real-world setting and facilitate a more seamless transition from graduate school to professional work.

There are also resources for continued professional development after graduation, such as the career success factors initiative introduced by 2012 ASA President Robert Rodriguez. This initiative began with a survey of working statisticians and their employers, which identified communication and teamwork as the most important non-technical skills for success in their careers (Rodriguez 2012c, 2013). The ASA designed a professional course to help address these needs entitled “Effective Presentations for Statisticians” as well as a webcast resource on “Leadership Development for Statisticians.” Janet Buckingham, chair of the ASA working group on “Developing Training in Statistical Leadership,” implemented a course first offered at the 2014 JSM entitled “Preparing Statisticians for Statistical Leadership: How to See the Big Picture and Have More Influence” (Utts 2015). Expanding the resources for professional development continues to be an area of focus for the ASA, and employers should also invest in resources tailored to the needs of their organizations including materials to support self-development, internal training courses, and presentations to statisticians from leaders in other functional and technical areas.

Finally, statisticians should actively seek mentorship opportunities and cultivate mentor–mentee relationships with individual and organizational leaders to develop their own individual leadership. For example, a talented statistician initiated a career development discussion and described how she was already functioning as a statistical leader capable of challenging and guiding very senior medical colleagues in the design of clinical trials and development programs. Her mentor helped her to develop a self-assessment of her strengths relative to her peers in terms of technical ability, communication, influence, and visibility. The statistician came to realize that she was comparable to her peers and very strong in the first two competencies. However, her influencing skills could have been developed further, and opportunities for more visibility were needed. Before concluding the discussion, several concrete steps were identified to increase her exposure to senior decision makers in the company.

6. Conclusion

Rodriguez (2015) asked “how do we resolve the paradox of a profession that is both invaluable and invisible?” It is encouraging that he and other leaders in the field have already begun to offer meaningful solutions (Wasserstein 2015). Resolving this paradox requires statistical leadership, effective communication of the unique capabilities of statisticians, and increased partnership with other quantitative disciplines.

A greater focus on statistical leadership and the essential competencies of listening, networking, and communication as part of the training and development for statisticians will cultivate successful multidisciplinary research and improve our visibility. Building statistical leadership will also help to “ensure that statisticians are at the decision-making tables of the future” (LaVange 2014). The courses in “Leadership in Biostatistics” and “Scientific Communication Skills” developed for graduate students in biostatistics at the University of North Carolina and the University of Pittsburgh, respectively, are prime examples that could be followed by other universities. Universities should also prepare graduate students for leadership roles by inviting outstanding statistical leaders (e.g., alumni) to give seminars in which they share their own stories of how they developed into leaders during their careers. The ASA has invested in several career development resources for statisticians, and should expand its training in statistical leadership by encouraging chapters and university departments to sponsor workshops led by ASA volunteers. Employers can also develop additional resources including self-paced learning, classroom training, and executive presentations from functional and technical leaders in other areas within the organization.

I have shared a few examples from drug development, including enterprise leadership in influencing portfolio decisions and policy leadership in driving the acceptance of model-based methods for dose selection within the scientific and regulatory communities. Examples of statistical leadership are also demonstrated by professional statistical organizations and from scientific journals, including the ASA and ISI response (Pantula, Teugels, and Stefanski 2010) to the highly publicized Science News article (Siegfried 2010) in which the author claimed that science had been “seduced” by a “mutant form of math” known as statistics. The journal Science recently established the Statistical Board of Reviewing Editors (SBoRE) to review data analyses and methods to increase the confidence in the conclusions published in the journal (McNutt 2014). More recently, the ASA issued a position statement for the scientific community on the “proper use and interpretation of the p-value” (Wasserstein and Lazar 2016). We need to seize more opportunities to demonstrate statistical leadership, especially in scientific and public policy.

We also need more effective communication of the unique capabilities of statistics to researchers in other disciplines. We miss the first opportunity to convey the value of statistics to an increasingly large number of nonstatisticians enrolled in our service courses. The last Conference Board of the Mathematical Sciences (CBMS) survey found that 192,000 students took an elementary statistics course during the fall 2010 semester at 4-year college mathematics and statistics departments, an increase of 37% from 2005 to 2010 (Blair, Kirkman, and Maxwell 2013). These courses need to be redesigned to feature more interesting aspects of data analysis and real-world problems that convey the value of the statistician in scientific research (Iman 1995; Meng 2009). Without any change, these courses may continue to be viewed by many students as their worst course in college and result in 384,000 missed opportunities annually to convey the value of statistics in the modern world, a number likely to exceed 500,000 when the 2015 CBMS survey results are available. Moore (1998) described this as “one of our most serious public responsibilities,” and suggested we consider statistics as one of the “liberal arts,” since statistical reasoning is a mode of reasoning about data, variation, and chance which are so prevalent in modern society.

There have been several recent initiatives to promote the contributions of statistics, such as the International Year of Statistics in 2013 and the founding of the World of Statistics organization. The ASA launched the national “This is Statistics” campaign in 2014 to promote careers in statistics (Myers 2015). The STATS.org initiative was launched in 2015 to provide information and resources to help journalists and others understand data and statistical issues relevant to their articles or reports. Similar resources for journalists are offered by the Royal Statistical Society at www.StatsLife.org. The FiveThirtyEight website developed by Nate Silver became popular during the 2008 elections, and expanded to present fascinating but clear statistical analysis of topics from sports, science, economics, and other popular subjects. There is also Hans Rosling, who became famous for bringing data to life and promoting statistical reasoning, and whose work lives on today through the Gapminder Foundation (Rosling and Zhang 2011) which he helped establish. Similar efforts are needed to communicate a more complete image of statistics within the many different organizations that employ statisticians.

Increased partnership between statisticians and colleagues in other quantitative disciplines is a significant opportunity to achieve greater impact in scientific research. Instead of seeking to pull data scientists under the discipline of statistics, we should invest more in understanding the elements that make data scientists unique, such as their capabilities in database management and distributed or parallel computing. The prevalence of big data will continue to drive demand for data scientists, but it will also create new opportunities for statisticians to collaborate on important issues such as bias inherent in data that may not be sampled, multiplicity, overfitting, and linking the plans for data integration to the research questions. There are also opportunities to adapt the traditional training of statisticians to better fit today's world, and to consider whether “statistical scientist” may be a better description of our capabilities.

Building leadership in statistics will drive increased value and visibility for the discipline of statistics. It begins with speaking up and clearly communicating our ideas on the best way to design investigations, plan for data collection, and analyze and interpret the data. I believe the prediction of Google's Chief Economist (Lohr 2009) for statistics to become the “sexy job” of the future can be realized through statistical leadership.

Acknowledgments

The author thanks the Editor, Associate Editor, and two Reviewers for their careful review and helpful suggestions which significantly improved this article. The author is also grateful to his colleagues Steffen Ballerstedt, Frank Bretz, Paul Gallo, Byron Jones, and Michael Looby for their helpful suggestions. The author would also like to thank Gary Koch, Lisa LaVange and Bob Rodríguez for their comments and suggestions. This article was inspired by the life and work of my colleague Richard Nixon.