This report is only one of many that will be of interest to the AAPOR audience. We want to point readers to a strategic initiative underway at the American Statistical Association to identify models for curriculum development, to engage in professional education, and to engage with external stakeholders1. Similarly, the United Nations2 Secretary-General has asked an Independent Expert Advisory Group to make concrete recommendations on bringing about a data revolution in sustainable development. The United Nations Statistics Economic Commission for Europe started a task team to work out key issues in using Big Data for Official Statistics, and the European Statistical System has put Big Data on their current roadmap for funding and development. Furthermore, we want to point readers to the three AAPOR task force reports on: the use of Social Media, mobile devices for survey research, and nonprobability sampling. All reports touch on related topics, but are distinct enough that reading all three might be necessary to get the full picture. These reports can be accessed on AAPOR’s website.
A dimension of Big Data not often mentioned in the practitioner literature, but important for survey researchers to consider is that Big Data are often secondary data, intended for another primary use. This means that Big Data are typically related to some non-research purpose and then re-used by researchers to make a social observation. This is related to Sean Taylor’s distinction between “found vs. made” data (Taylor 2013). He argues that a key difference between Big Data approaches, and other social science approaches, is that the data are not being initially “made” through the intervention of some researcher. When a survey researcher constructs an instrument there are levels of planning and control that are necessarily absent in the data used in Big Data approaches. Big Data sources might have only a few variables as compared with surveys that have a set of variables of interest to the researcher. In a 2011
Public Opinion Quarterly article and a blog post in his former role as director of the U.S. Census Bureau, Robert Groves described a similar difference between organic and designed data (Groves 2011a, Groves 2011b).
In the context of public opinion studies, a survey researcher could measure opinion by prompt responses about a topic that may never naturally appear in a Big Data source. On the other hand, the “found” data of social media are “nonreactive,” or “naturally occurring,” so that a data point, devoid of researcher-manipulation, may be a more accurate representation of a true opinion or behavior. ”Found” data may be a behavior, such as a log of steps drawn from networked pedometers or the previously mentioned recordings of travel patterns, which might be more accurate than what could be solicited in surveys given known problems with recall error (Tourangeau et al. 2000).
Before considering to the usability and use of Big Data it is worth exploring the paradigm shift happening in the presence of these new data sources. This change in paradigm stems from changes in many factors affecting the measurement of human behavior: the nature of the new types of data, their availability, and the way in which they are collected, mixed with other data sources, and disseminated. The consequences of these changes for public opinion research is fundamental in both the analysis that can be done and who the analysts might be. While the statistical community has moved beyond survey and even administrative data to begin understanding how to mine data from social media to capture national sentiment, from cellphone data to understand or even predict anti-government uprisings, and from financial data to examine swings in the economy, it is equally important to note that now some data are freely available and usable to anyone who wishes to mesh data points and series together and produce such analyses. With data readily accessible on the internet – this creates opportunities for amateur, rather than professional, data analysts.
The change in the nature of the new type of data is transformative. Its characteristics – its velocity, volume and variety – and the way in which it is collected, mean a new analytical paradigm is open to statisticians and social scientists (Hey et al. 2009). The classic statistical paradigm was one in which researchers formulated a hypothesis, identified a population frame, designed a survey and a sampling technique and then analyzed the results (Groves 2011a). The new paradigm means it is now possible to digitally capture, semantically reconcile, aggregate, and correlate data. These correlations might be effective (Halevy et al. 2009, Cukier and Mayer-Schoenberger 2013) or suspect (Couper 2013), but they enable completely new analyses to be undertaken – many of which would not be possible using survey data alone. For example, the new type of analysis might be one that captures rich environmental detail on individuals from sensors, Google Earth, videos, photos or financial transactions. Alternatively, the analysis might include rich and detailed information on unique and quite small subsets of the population (from microbiome data, or websearch logs), or the analysis could be on completely new units of analysis, like networks of individuals or businesses, whose connections can only be captured by new types of data (like tweets, cell phone conversations, and administrative records). As Kahneman (2011) points out, the new measurement can change the paradigm in its own right.
The change in data ownership has also transformed the way in which data are disseminated. The population of potential data analysts – trained and untrained – has dramatically expanded. This expansion can result in tremendous new insights, as the Sloan Digital Sky Survey and the Polymath project have shown (Nielsen 2012), and is reflected in Grey’s Fourth Paradigm (Figure 4) (Hey et al. 2009), but can also lead to the degradation of the quality of analysis that can be done and issues with the conclusions drawn and reported based on these data. AAPOR as an organization will need to find its place in giving guidance to the proper use of these data with respect to public opinion research.
Figure 5. A typical rectangular format for traditional data analysis
Row errors may be of three types:
- Omissions – some population elements are not among the rows
- Duplications – some population elements occupy more than one row
- Erroneous inclusions – some rows contain elements or entities that are not part of the population of interest
For survey sample data sets, omissions include nonsampled elements in the population as well as population members deliberately excluded from the sampling frame. For Big Data, selectivity is a common form of omissions. For example, a data set consisting of persons who conducted a Google search in the past week, necessarily excludes persons not satisfying that criterion. Unlike survey sampling, this is a form of nonrandom selectivity. For example, persons who do not have access to the internet are excluded from the file. This exclusion may be biasing in that persons with internet access may have very different demographic characteristics than from persons who do not have internet access. This problem is akin to non-coverage in sampling, depending on the population about which the researcher is attempting to estimate.
We can also expect that Big Data sets, such as a data set containing Google searches during the previous week, could have the same person represented many times. People who conducted many searches during that period would be disproportionately represented relative to those who conducted fewer searchers. Other erroneous inclusions can occur when the entity conducting a search is not a person but another computer; for instance, via a web scraping routine.
The most common type of column error is caused by inaccurate or erroneous labeling of the column data – an example of metadata error. For example, a business register may include a column labeled “number of employees” defined as the number of persons in the company that received a payroll check in the month preceding. Instead the column contains the number of persons on the payroll whether they received a check last month or not, including persons on leave without pay. Such errors would seem to be quite common in Big Data analysis given the multiple layers of processing required to produce a data set. For example, data generated from a source, such as an individual Tweet, may undergo a number of transformations before it lands in a rectangular file such as the one in Figure 5. This transformation process can be
quite complex; for example, it may involve parsing phrases, identifying words, and classifying them as to subject matter and then further as to positive or negative expressions about the economy. There is considerable risk that the resulting features are either inaccurately defined or misinterpreted by the data analyst.
Finally, cell errors can be of three types: content error, specification error, or missing data. A content error occurs when the value in a cell satisfies the column definition but is still erroneous. For example, value satisfies the definition of “number of employees” but the value does not agree with the true number of employees for the company. Content errors may be the result of a measurement error, a data processing error (e.g., keying, coding, editing, etc.), an imputation error or some other cause. A specification error is just as described for the column error but applied to a cell. For example, the column is correctly defined and labeled; however, a few companies provided values that, although otherwise highly accurate, were nevertheless inconsistent with the required definition. Missing data, as the name implies, is just an empty cell that should be filled. As described in Kreuter and Peng (2014), data sets derived from Big Data are notoriously affected by all three types of cell errors, particularly missing or incomplete data.
5.2 Extending the Framework for Big Data
The traditional TSE framework is quite general in that it can be applied to essentially any data set that conforms to the format in Figure 5. However, in most practical situations it is quite limited because it makes no attempt to describe the error in the processes that generated the data. In some cases, these processes constitute a “black box” and the best approach is to attempt to evaluate the quality of the end product. For survey data, the TSE framework provides a fairly complete description of the error generating processes for survey data and survey frames (see, for example, Biemer 2010). In addition, there has been some effort to describe these processes for population registers and administrative data (Wallgren and Wallgren 2007). But at this writing, very little effort has been devoted to enumerating the
error sources and the error generating processes for Big Data. One obstacle in this endeavor is that the processes involved in generating Big Data are as varied as Big Data are themselves. Nevertheless, some progress can be made by considering the generic steps involved. These steps include the following:
- Generate – data are generated from some source either incidentally or purposively.
- Extract/Transform/Load (ETL) –all data are brought together under a homogeneous computing environment in three stages:
- Extract Stage – data are harvested from their sources, parsed, validated, curated and stored.
- Transform Stage – data are translated, coded, recoded, aggregated/disaggregated, and/or edited.
- Load Stage – data are integrated and stored in the data warehouse.
- Analyze – Data are converted to information through a process involving:
- Filtering (Sampling)/Reduction – Unwanted features and content are deleted; features may be combined to produce new ones; data elements may be thinned or sampled to be more manageable for the next steps.
- Computation/Analysis/Visualization – data are analyzed and/or presented for interpretation and information extraction.
.jpg.aspx?width=500&height=284)
Figure 6. Big data process map
Figure 6 graphically depicts the flow of data along these steps. The severity of the errors that arise from these processes will depend on the specific data sources and analytic goals involved. Nevertheless, we can still consider how each stage might create errors in a more generic fashion.
For example, data generation error is somewhat analogous to errors arising in survey data collection. Like surveys, the generic data generating process for Big Data can create erroneous and incomplete data. In addition, the data generating sources may be selective in that the data collected may not represent a well-defined population or one that is representative of a target population of interest. Thus, data generation errors include low signal/noise ratio, lost signals, incomplete or missing values, non-random, selective source and meta-data that are lacking, absent, or erroneous.
ETL processes may be quite similar to various data processing stages for surveys. These may include creating or enhancing meta-data, record matching, variable coding, editing, data munging (or scrubbing), and data integration (i.e., linking and merging records and files across disparate systems). ETL errors include: specification error (including, errors in meta- data), matching error, coding error, editing error, data munging errors, and data integration errors.
As noted above, the analysis of Big Data introduces risks for noise accumulation, spurious correlations, and incidental endogeneity which may be compounded by sampling and nonsampling errors. Related to the former, data may be filtered, sampled or otherwise reduced to form more manageable or representative data sets. These processes may involve further transformations of the data. Errors include sampling errors, selectivity errors (or lack of representativeness), and modeling errors.
Other errors that may be introduced in the computation stage are similar to estimation and modeling error in surveys. These include modeling errors, inadequate or erroneous adjustments for representativeness, improper or erroneous weighting, computation and algorithmic errors.
In Section 4 we mentioned that all data collections suffer from error in the data generating process. AAPOR is promoting the transparency of these processes. A similar effort will be very valuable for Big Data driven research.
Summary
- Using Big Data in statistically valid ways is challenging. One misconception is the belief that the volume of the data can compensate for any other deficiency in the data (Big Data hubris).
- Size often dominates issues of selectivity in public perception.
- Many platforms that produce statistics with Big Data change their algorithms (algorithm dynamic). This can lead to ambiguous results for any kind of long term study.
- The massive size of Big Data can cause problems such as noise accumulation, spurious correlations, and incidental endogeneity.
- Each step in the Big Data process will generate errors that will affect the estimates and each Big Data source will have its own set of errors.
- We need to have a Total Error perspective when we consider using a Big Data source.
- It is the responsibility of Big Data analysts to be aware of the data’s many limitations and to take the necessary steps to limit the effects of Big Data error on their results.
- Many Big Data models currently in use are not measured for accuracy in a public way. So far models tend to fail privately.
- AAPOR's transparency initiative can be a model for data use beyond surveys.
6. What are the Policy, Technical and Technology Challenges, and How Can We Deal with Them?
Public opinion research is entering a new era, one in which traditional survey research may play a less dominant role. The proliferation of new technologies, such as mobile devices and social media platforms, are changing the societal landscape across which public opinion researchers operate. As these technologies expand, so does access to users’ thoughts, feelings and actions expressed instantaneously, organically, and often publicly across the platforms they use. The ways in which people both access and share information about opinions, attitudes, and behaviors have gone through perhaps a greater transformation in the last decade than in any previous point in history and this trend appears likely to continue. The ubiquity of social media and the opinions users express on social media provide researchers with new data collection tools and alternative sources of qualitative and quantitative information to augment or, in some cases, provide alternatives to more traditional data collection methods.
There is great potential for Big Data to generate innovation in public opinion research. While traditional survey research has a very important role, the addition of large-scale observations from numerous sources (e.g. social media, mobile computing devices) promises to bring new opportunities. To realize these potential advances we must address numerous challenges in a systematic way. This section examines several policy challenges for Big Data (ownership, stewardship, collection authority, privacy protection), technical challenges (multi-disciplinary skills required) as well as technology challenges (computing resources required).
6.1 Policy Challenge: Data Ownership
Many individuals now produce data that are potentially useful for research as part of their everyday participation in the digital world. There has always been a lack of clarity in legal guidance stemming from a lack of clarity as to who owns the data – whether it is the person who is the subject of the information, the person or organization who collects that data (the data custodian), the person who compiles, analyzes or otherwise adds value to the information, the person who purchases interest in the data, or society at large. The lack of clarity is exacerbated because some laws treat data as property and some treat it as information (Cecil and Eden 2003). The new types of data make the ownership rules even more unclear: data are no longer housed in statistical agencies, with well-defined rules of conduct, but are housed in businesses or administrative agencies. In addition, since digital data can be alive forever, ownership could be claimed by yet to be born relatives whose personal privacy could be threatened by release of information about blood relations. For the AAPOR community it will be important to stay informed about emerging rules and to be aware of differences in regulations across countries.
6.2 Policy Challenge: Data Stewardship
An eloquent description of statistical confidentiality is “the stewardship of data to be used for statistical purposes” (Duncan et al. 2011). Statistical agencies have been at the forefront of developing that stewardship community in a number of ways. First, on the job training is provided to statistical agency employees. Second, in the United States, academic programs such as the Joint Program on Survey Methodology, communities such as the Federal Committee on Statistical Methodology, and resources such as the Committee on National Statistics have been largely supported by the federal statistical community. In the past the focus was almost exclusively on developing methodologies to improve the analytical use of survey data, and to a lesser extent, administrative data. It is important to expand the training efforts to train scientists in developing an understanding of such issues such as identifying the relevant population and linkage methodologies. Around the US several programs are emerging. However it is important to integrate the training of these skills into the existing programs, in particular if the field is moving towards data integration from survey and non- survey data (see Section 7).
6.3 Policy Challenge: Data Collection Authority
When statistical agencies were the main collectors of data, they did so under very clear statutory authority with statutory protections. For example, Title 26 (Internal Revenue Service) and Title 13 (Census Bureau) of the US code provided penalties for breaches of confidentiality, and agencies developed researcher access modalities in accordance with their statutory authorization.
The statutory authorization for the new technology enabled collection of data is less clear. The Fourth Amendment to the Constitution, for example, constrains the government’s power to “search” the citizenry’s “persons, houses, papers, and effects.” State privacy torts create liability for “intrusion upon seclusion.” Yet the generation of Big Data often takes place in the open, or through commercial transactions with a business, and hence is not covered by either of these frameworks. There are major questions as to what is reasonably private, and what constitutes unwarranted intrusion (Strandburg 2014). Data generated by interacting with professionals, such as lawyers and doctors, or by online consumer transactions, are governed by laws requiring “informed consent” and draw on the Fair Information Practice Principles (FIPP). Despite the FIPP’s explicit application to “data,” they are typically confined to personal information, and do not address the large-scale data collection issues that arise through location tracking and smart grid data (Strandburg 2014)
6.4 Policy Challenge: Privacy and Reidentification
The risk of reidentifying individuals in a micro-dataset is intuitively obvious. Indeed, one way to formally measure the reidentification risk associated with a particular file is to measure the likelihood that a record can be matched to a master file (Winkler 2005). If the data include direct identifiers, like names, social security numbers, establishment id numbers, the risk is quite high. However, even access to close identifiers, such as physical addresses and IP addresses can be problematic. Indeed, the Health Insurance Portability and Accountability Act (HIPAA) regulations under The Privacy Rule of 2003 require the removal of 18 different types of identifiers including other less obvious identifiers such as birth date, vehicle serial numbers, URLs, and voice prints. However, even seemingly innocuous information makes it relatively straightforward to reidentify individuals, for example by finding a record with sufficient information such that there is only one person in the relevant population with that set of characteristics: the risk of re-identification has been increasing due to the growing public availability of identified data and rapid advances in the technology of linking files (Dwork 2011). With many variables, everyone is a population unique. Since Big Data have wide-ranging coverage, one cannot rely on protection from sampling (Karr and Reiter 2014). Indeed, as Ohm (2010) points out, a person with knowledge of an individual’s zip code, birthdate and sex can reidentify more than 80% of Netflix users, yet none of those are typically classified as Personally Identifiable Information (PII).
6.5 Policy Challenge: Meaning of “Reasonable Means” Not Sufficiently Defined
The statutory constraint on agencies such as the IRS and the U.S. Census Bureau makes it clear that the agencies, as data producers should take “reasonable means” to protect data, although these reasonable means are not defined. Trust clearly depends on people’s views on privacy, but these views are changing rapidly (Nissenbaum 2011). Nissenbaum (2011:34) also notes that it is increasingly difficult for many people to understand where the old norms end and new ones begin, as “Default constraints on streams of information from us and about us seem to respond not to social, ethical, and political logic but to the logic of technical possibility: that is, whatever the Net allows.” Yet there is some evidence that people do not require complete protection, and will gladly share even private information provided that certain social norms are met, similar to what Gerber reported in 2001. There are three factors that affect these norms: actors (the information senders and recipients or providers and users); attributes (especially types of information about the providers, including how these might be transformed or linked); and transmission principles (the constraints underlying the information flows).
.jpg.aspx?width=500&height=286)
Figure 7. Models for user-data interaction, from Kinney et al. (2009).
What We Can Learn from Current Knowledge
Kinney et al. (2009) identify a variety of mechanisms for interaction between users and confidential data. As they note, in Figure 7 (above) “there are three major forms of interaction: direct access, dissemination-based access (public data releases), and query-based access. Direct access imposes the least interference between the users and the confidential data. Dissemination-based access refers to the practice of releasing masked data in public files. In the query-based interaction mode, users cannot directly access individual data records, but are able to submit queries, either electronically or manually.” (Kinney et al.2009:127). Thorough reviews of different approaches are provided in Duncan et al. (2011) and Prada et al. (2011).
The current statistical disclosure literature offers multiple ways of permitting access to microdata, but less relevant guidance about release.
6.6 Technical Challenge: Skills Required to Integrate Big Data into Opinion Research
Depending on the scale of the data being discussed, there can be significant challenges in terms of the skills and resources necessary to work with Big Data. In particular, most Big Data problems require a minimum of four roles:
- Domain Expert. A user, analyst, or leader with deep subject matter expertise related to the data, their appropriate use, and their limitations.
- Researcher. Team member with experience applying formal research methods, including survey methodology and statistics.
- Computer Scientist. Technically skilled team member with education in computer programming and data processing technologies.
- System Administrator. Team member responsible for defining and maintaining a computation infrastructure that enables large scale computation.
However, from our experience, many companies are trying to make do with only one person.
Domain expertise is particularly important with new types of data that have been collected without instrumentation, usually for purposes other than quantitative survey analysis. For example, looking at Big Data from social media sources requires an in depth understanding of the technical affordances and user behaviors of that social media source. Posting to Twitter, as an example, involves norms and practices that could affect the interpretation of data from that source. This could refer to the use of handles and hashtags, certain terminology and acronyms used, or practices such as retweeting, modifying tweets and favoring. Additionally it is important to understand to what degree different forms of new media may under-represent particular demographics (e.g. there may be a low number of citizens age 60 years and older using Twitter to express themselves).
Foundational research skills such as the application of classical survey methodology and the appropriate use of descriptive statistics remain critical for understanding Big Data. As the volume of digital data grows and the barrier to obtaining such data is continually lowered, there is an increasing risk of untrained engineers and computer programmers finding bogus associations in Big Data. To ensure Big Data are appropriately integrated into public opinion research, there remains an ongoing requirement for classically trained researchers to be involved throughout the entire process.
.jpg.aspx?width=500&height=329)
Figure 8. The different roles needed in a Big Data team
From the computer science skills standpoint, baseline competencies can include the ability to work in command line environments, some capability with programming languages, facility with databases and database languages, and experience with advanced analytical tools. The larger the dataset, the more important skills in databases and analytics become. Some researchers choose to partner with computer scientists, or skilled programmers, to cover these needed skills. While this has led to viable research partnerships, it creates a new need in terms of interdisciplinary collaboration. Major information technology components that are frequently used in the process of collecting, storing, and analyzing Big Data include:
- Apache Hadoop. A system for maintaining a distributed file system that supports the storage of large-scale (Terabytes or Petabytes of content), and the parallel processing of algorithms against large data collections. Requires a programming language such as Java or Python.
- Apache Spark. A fast and general purpose engine for large-scale data processing that works in support of Hadoop or in-memory databases. Requires a programming language such as Java or Python.
- Java programming language. A general purpose systems engineering language that supports the creation of efficient algorithms for data analysis.
- Python programming language. A general purpose systems engineering language that supports rapid prototyping and efficient algorithms for data analysis.
It is worth noticing that there are many different frameworks. Even though a framework such as e.g., Hadoop is commonly used today, given the fast development in this area this may very well change soon. It could therefore be helpful to think in more general terms of clusters and parallel processing of unstructured data.
System administrators play an important role in defining, creating, and maintaining computing environments for the storage and analysis of Big Data. Working with Big Data often requires additional computing resources. Depending on the size of the data being considered, resources can range from hardware and server stacks that are manageable by non- specialist IT staff, to very large scale computing environments that include high powered computing stacks of hardware and software that often require specialist IT training. As an example, many universities offer High Performance Computing Centers (HPC) that include networked servers, structuring software like Hadoop, as well as database and analysis packages. System administrators responsible for maintaining Big Data compute platforms often use one of three strategies:
- Internal compute cluster. For long-term storage of unique or sensitive data, it often makes sense to create and maintain an Apache Hadoop cluster using a series of networked servers within the internal network of an organization. Although expensive in the short term, this strategy is often the lowest cost in the long-term.
- External compute cluster. There is a trend across the IT industry to outsource elements of infrastructure to ‘utility computing’ service providers. Organizations such as the Amazon Web Services (AWS) division of Amazon.com make it simple for system administrators to rent pre-built Apache Hadoop clusters and data storage systems. This strategy is very simple to set up, but may be much more expensive than creating a long-standing cluster internally. Functional equivalents to Amazon Elastic Map Reduce Service are Microsoft HDInsight and Rackspace’s Cloud Big Data Platform. Other alternatives include Hadoop on Google’s Cloud Platform and Qubole.
- Hybrid compute cluster. A common hybrid option is to provision external compute cluster resources using services such as AWS for on-demand Big Data analysis tasks, and create a modest internal computer cluster for long-term data storage.
6.7 Technology Challenge: Computational Requirements
The formula “distance = rate x time” is well known by high-school math students. This formula may be applied to simplify the understanding of why large-scale parallel processing computer clusters are a requirement for Big Data analysis. In the analysis of a very large data set, the volume of data to be processed may be considered the distance (e.g. 10 Terabytes). Similarly the number of available central processing units and magnetic hard drives for storing the media may be considered directly related to the rate.
All other factors being held equal, a system with ten CPUs and ten hard drives (10 computation units) will process a batch of data 10 times faster than a system with one CPU and one hard drive (1 computation unit). If an imaginary data set consists of 50 million records, and systems with 1 computation unit can process 100 records per second, then it will take approximately 5.7 days (50,000,000 records /100 records per second) to finish the analysis of data - potentially an unacceptable amount of time to wait. A system with 10 computation units can compute the same result in just 13.9 hours, a significant time savings. Systems like Apache Hadoop drastically simplify the process of connecting multiple commodity computers into a cluster capable of supporting such parallel computations.
Although disk space may be relatively inexpensive, the cost of creating and maintaining systems for Big Data analysis can be quite expensive. In the past thirty years the cost of storing data on magnetic storage media such as hard drives has decreased dramatically. A hard drive with 3 Terabytes of storage capacity now costs less than $100 in the United States. However the total cost of ownership of a Big Data analysis system is the sum of several components including at a minimum:
- Disk based storage media
- Active computation components (computer central processing unit or CPU, Random Access Memory or RAM)
- Infrastructure elements such as server farm rackspace, electricity required, cooling costs, and network access and security fees.
When taken in aggregate these components may cost tens or hundreds of thousands of dollars. It may not be feasible to create a permanent Big Data computer cluster to support a single study. Within AAPOR there is the possibility to form public-private sector partnerships not only for sharing data but also for sharing analysis infrastructure.
.jpg.aspx?width=500&height=359)
Figure 9 - The Amazon Elastic MapReduce (EMR) Service remains one of the most popular utility compute cloud versions of Hadoop
Summary
- Data ownership in the 21st century is not well defined and there is no “one size fits all” policy. Researchers must carefully consider data ownership issues for any content they seek to analyze.
- We need to turn to additional sources of knowledge about how to collect and protect data when it comes to Big Data.
- There is no clear legal framework for the collection and subsequent use of Big Data. Most consumers of digital services (such as smart phone applications) have little or no idea that their behavior data may be re-used for other purposes.
- The removal of key variables as Personally Identifiable Information (PII) is no longer sufficient to protect data against reidentification. The combination of location and time metadata with other factors enables reidentification of “anonymized” records in many cases. New models of privacy protection are required.
- Current statistical disclosure literature offers multiple ways of permitting access to microdata, but less relevant guidance about release.
- Effective use of Big Data requires a multidisciplinary team consisting of e.g., a domain expert, a researcher, a computer scientist, and a system administrator. Many companies, however, are trying to make do with only one person.
- Organizations seeking to experiment with Big Data computer cluster technology can reduce their initial capital outlays by renting pre-built compute cluster resources (such as Apache Hadoop) from online providers like the Amazon Web Services organization.
- Systems such as Apache Hadoop drastically simplify the creation of computer clusters capable of supporting parallel processing of Big Data computations.
- Although the cost of magnetic storage media may be low, the cost of creating systems for the long-term storage and analysis of Big Data remains high. The use of external compute cluster resources is one short-term solution to this challenge.
7. How can Big Data be Used to Gain Insights?
The recent literature on developments in Big Data can give the reader the impression that there is an ongoing, head-to-head competition between traditional research based on data specifically designed to support research and new research methods based on more organic data or found data. Researchers who have created a career around analysis of survey data are particularly anxious about the rise of Big Data, fearful that the skills they have developed throughout their career may become obsolete as Big Data begins to crowd out survey data in supporting future research.
We have seen similar debates on statistical methods. The predominant theory used in surveys emanates from the Neyman-Pearson framework. This theory states that survey samples are generated from a repeatable random process and governed by underlying parameters that are fixed under this repeatable process. This view is called the frequentist view and is what most survey researchers are most familiar with. An alternative theory is the Bayesian view that emanates from Bayes, Savage, deFinetti and others. In this theory, data from a realized sample are considered fixed while the parameters are unknown and described probabilistically. Typically a prior distribution of the parameter is combined with the observed data resulting in a posterior distribution. The discussions of these views have successively moved from controversy to more pragmatic standpoints. A survey statistician's job is to make the most valid inferences about the finite population and therefore there is room for both views. Both frequentist and Bayesian statistics play key roles in Big Data analysis. For example, when data sets are so large that the analysis must be distributed across multiple machines, Bayesian statistics provides efficient algorithms for combining the results of these analyses (see, for example. Ibrahim and Chen 2000, Scott et al. 2013). Sampling techniques are key in gathering Big Data and for analyzing Big Data in a small computing environment (Leek 2014a, Leek 2014b).
In general, framing the rise of Big Data as a competition with survey data or traditional research is counterproductive, and a preferred route is to recognize how research is enhanced by utilizing all forms of data, including Big Data as well as data that are designed with research in mind. Inevitably, the increased availability of the various forms of Big Data discussed in Section 3 will supplant survey data in some settings. However, both Big Data and survey data have advantages and disadvantages, which we describe in more detail below. An effective and efficient research strategy will be responsive to how these advantages and disadvantages play out in different settings, and deploying blended research methods that maximize the ability to develop rigorous evidence for the questions of interest for an appropriate investment of resources.
Research is about answering questions, and the best way to answer questions is to start by utilizing all of the information that is available. The availability of Big Data to support research provides a new way to approach old questions as well as an ability to address some new questions that in the past were out of reach. However, the findings that are generated based on Big Data inevitably generate more questions, and some of those questions tend to be best addressed by traditional survey research. As the availability and use of Big Data increases, there is likely to be a parallel growth in the demand for survey research to address questions raised by findings from Big Data. The availability of Big Data liberates survey research, in the sense that researchers no longer need to generate a new survey to support each new research endeavor. Big Data can be used to generate a steady flow of information about what is happening - for example, how customers behave - while traditional research can focus instead on deeper questions about why we are observing certain trends or deviations from trends - for example, why customers behave as they do and what can be done to change their behavior.
In thinking about how to blend Big Data with traditional research methods, it is important to be clear about the relevant questions to be addressed. Big Data can be especially useful for detecting patterns in data or for establishing correlation between factors. In contrast, establishing causality between variables requires that data be collected according to a specific design in order to support models or research designs intended to isolate causality. Marketing researchers use Big Data for so called A/B testing to establish causality, though even this can be problematic, for example since it relies on cookies. In the public sector, traditional research based on designed data is likely to continue to play a primary role in supporting policy development, particularly when customized data and research designs are necessary to ensure that we can identify causality between variations in public interventions and the outcomes that they affect. At the same time, research based on Big Data can be best utilized to meet the needs of program administrators, who are focused on monitoring, maintaining, and improving program operations within an ongoing policy regime. In this setting, measuring trends and correlations and making predictions may be sufficient in many cases - isolating causality is not essential - and the administrative data and related Big Data sources can best meet these needs. However, when causation is ignored and the focus is on predictions using models that are based on historical training data, there is a risk to perpetuate what happened in the past, for example embedding racism, sexism, or other problematic patterns in the models.
7.1 Relative Advantages of Survey Data and Big Data to Support Research
For many years research has depended on data collected through surveys because there have been few alternatives. Even as alternative sources of data begin to proliferate, survey data retain some critical advantages in facilitating social science research. The primary advantage of basing research on survey data is the control it provides for researchers - the survey can be designed specifically to support the needs of the research. Use of a survey allows for customizing outcome measures to closely match the primary questions to be addressed by the research. For example, if a research project is designed to address hourly wage compensation as a key outcome of interest, the supporting survey can be designed to measure hourly compensation rather than use a proxy or impute hourly compensation from some pre-existing data source.
The control afforded by using a survey to support research also allows for generating estimates for samples that are representative of a specific population of interest. By using a specific population to create a probabilistic sample frame for a survey, researchers can use data from the survey sample to generate estimates that apply to the population with a known degree of precision. Researchers have fully developed the theory and practice of probability sampling and statistical inference to handle just this type of data collection and use these data effectively in addressing questions of interest.
In contrast, since most Big Data sources are organic and beyond the control of researchers, researchers using Big Data sources take what they get in terms of the population that is represented by the data. In many cases, the population represented by a Big Data source does not exactly match the population of interest. For example, databases based on Google searches are constrained to represent the searches conducted by Google users rather than the general population or some other population of interest. It’s difficult to assess the degree to which this may bias estimates relative to a given research question. Research on television audience measurement and viewing habits in the UK offers a choice between research based on a 5,100-household sample that is representative of the UK population, compiled by the Broadcasters’ Audience Research Board (BARB), and research based on the SkyView 33,000-household panel, developed by Sky Media based on Sky Digital Homes (homes that subscribe to this particular service). While the SkyView panel is considerably larger than the BARB panel, the BARB panel can be used to generate estimates that are directly representative of the UK population. Another example of this is that sometimes people want to estimate TV viewership by twitter feeds. The problem is that people never tweet which news channel told them some piece of news, just the news itself, whereas people tweet which show they're watching if it's a drama like House of Cards. In this case TV news will be underreported by twitter analysis.
Regardless, Big Data have a number of advantages when compared with survey data. The clearest advantage of Big Data is that these data already exist in some form, and therefore research based on Big Data does not require a new primary data collection effort. Primary data collection can be expensive and slow, which can either greatly delay the generation of new research findings or make new research prohibitively expensive. Problems may also arise with survey data collection as response rates trend down, particularly in research settings that would require lengthy surveys.
Compared with survey data, Big Data usually require less effort and time to collect and prepare for analysis. However, the effort associated with the creation and preparation of a Big Data set for analysis is not trivial. Even though Big Data already exist, it may still require substantial effort to collect the data and link digital data from various sources. According to expert estimates, data scientists spend 50 to 80 percent of their time collecting and preparing data to make it ready for investigation (Lohr 2014). The task of analyzing Big Data often involves gathering data from various sources, and these data - including data from sensors, documents, the web, and more conventional data sets - come in different formats. Consequently, start-ups are developing software to automate the gathering, cleaning, and organizing of data from different sources, so as to liberate data scientists from what tend to be the more mundane tasks associated with data preparation. There will however always be this type of routine work because you need to massage data one way for one study and another way for the next.
Big Data also are often available in high volumes, and with current technology, these high volumes of data are more easily processed, stored, and examined than in the past. For years, researchers have worked with data sets of hundreds or thousands of observations, which are organized in a relatively straightforward rectangular structure, with n observations and k variables. While these datasets are straightforward to deal with, the constrained volume of the data created limitations with respect to statistical power. In contrast, Big Data come in many different forms and structures, and the potential for huge volumes of observations implies that statistical power is less of a concern than in the old days. As mentioned in Section 5, however, huge volumes cause their own sets of problems. The varied structure (or lack of structure) and large volumes of observations in Big Data can be a challenge for processing and organizing the data, but the volume of observations in Big Data also translates into a more comprehensive and granular picture of the processes that are represented by the data. More granular and comprehensive data can help to pose new sorts of questions and enable novel research designs that can inform us about the consequences of different economics policies and events. Finally, enhanced granularity allows researchers to examine behavior in greater detail, and also to examine much more detailed subgroups of the population with adequate statistical power. For example, traditional research may identify the impact of class size on student performance, but Big Data could allow us to investigate how it varies by grade, school, teacher, or student mix; assuming all other confounders can be removed. Similarly, section 4 of this report also talks about using Big Data to study the tails of a distribution, which is not possible with a small data set.
Big Data also are often available in real time, as it is created organically as individual behavior (for example, phone calls, internet browsing, internet shopping, etc.) is occurring. This characteristic of Big Data has made it particularly appealing in the private sector, where businesses can use data to support management decision making in a timely manner. Traditional research, which relies on primary data collection, is slow, and so it generally cannot support making decisions quickly. One analyst characterizes traditional research as being built for “comfort not speed” - it generates sound findings that can instill confidence in the resulting decisions, but it generates them slowly and therefore cannot support quick decision making. In contrast, the timing of Big Data is more aligned with the cadence of decision making in a private or public sector management setting, where there is a premium on responding quickly to rapid changes in consumer demand or client need.
7.2 Research Methods that Exploit Availability of Big Data
As discussed above, Big Data are particularly advantageous in situations where decision makers want to use evidence to drive critical decisions. For a given organization interested in utilizing Big Data analysis to support effective operation of a program or set of programs, one can imagine at least three ways in which this would happen. First, Big Data can be used to match the right people to the right programs. For example, an employer engaged in a health management program to promote better employee health would want to be able to direct employees to the appropriate services given their needs, which would require collecting, processing and analyzing data on individual health and behaviors. Second, Big Data can be used to facilitate better operations. In the case of an employee health management program, this might amount to using Big Data to support building and facilitating healthful interactions between employees, their interpersonal networks, care providers, and insurers. Third, Big Data can be used to measure the outcomes among participants, the impacts of the program on those outcomes, and the net value of the program. In the case of an employee health management program, this might entail measuring key health and work outcomes, extrapolating to future outcomes, estimating the impact of the program on these outcomes, and monetizing the impact estimates so as to estimate the net value of the program investment. Based on these estimates, managers could make informed decisions on how the program would evolve over time in order to best meet the needs of employees and the employer. Of course any of these examples carry the risk that the information is used not in the employees’ best interest, which gets back to the ethical challenges discussed before.
Given the potential benefit of Big Data in driving evidence-based decisions, private sector organizations have quickly gravitated towards greater reliance on Big Data and have adopted research methods that exploit the advantages of these data. Predictive analytics and rapid- cycle evaluation are two of the big-data supported research methods that have become much more popular in the private sector in recent years. These methods allow managers to not only track ongoing activity, but also support decision making regarding how to respond tactically to a changing environment and customer base.
Predictive analytics refers to a broad range of methods used to predict an outcome. For example, in the private sector, predictive analytics can be used to anticipate how customers and potential customers will respond to a given change, such as a product or service change, a new marketing effort, establishment of a new outlet, or the introduction of a new product or service. Businesses can use predictive analytics to estimate the likely effect of a given change on productivity, customer satisfaction, and profitability, and thereby avoid costly mistakes. Predictive analytics can be conducted based on data that are collected as part of routine business operations and stored so as to support ongoing analytics, and these data can also be combined with other Big Data sources or survey data drawn from outside the organization.
Predictive analytics modeling also has been used to support new information products and services in recent years. For example, Amazon and Netflix recommendations rely on predictive models of what book or movie an individual might want to purchase. Google’s search results and news feed rely on algorithms that predict the relevance of particular web pages or articles. Predictive analytics are also used by companies to profile customers and adjust services accordingly. For example, health insurers use predictive models to generate “risk scores” for individuals based on their characteristics and health history, which are used as the basis for adjusting payments. Similarly, credit card companies use predictive models of default and repayment to guide their underwriting, pricing, and marketing activities.
Rapid-cycle evaluation is the retrospective counterpart to predictive analytics - it is used to quickly assess the effect of a given change on the outcomes of interest, including productivity, customer satisfaction, and profitability. As with predictive analytics, rapid-cycle evaluation leverages the available operations data as well as other Big Data sources. The exact statistical methods used in rapid-cycle evaluation can vary according to the preferences and resources of the user. For example, rapid cycle evaluation can be based on experimental methods in which a change is implemented in randomly chosen segments of the business or customers are randomly selected to be exposed to the change. In this way, the evaluation of a given change can be conducted by comparing outcomes among a “treatment group,” which is exposed to the change, and a “control group,” which is not exposed to the change.
Private businesses have begun to invest heavily in these capabilities. For example, Capital One has been a pioneer in rapid-cycle evaluation based on their transactions data to support business decisions, running more than 60,000 experiments and related analytics addressing a range of questions related to their operations or product offerings. Many other companies are moving in this direction as well (Manzi 2012).
While the public sector is not moving as fast as the private sector in adopting Big Data and data analytics techniques, public administrators are beginning to appreciate the value of these techniques and experiment with their use in supporting administrative decisions and improving public programs (Cody and Asher 2014). At the broadest level, some government agencies at all levels are collecting available data and examining data patterns related to their operations, in the hope of generating insights. For example, a recent New York Times editorial (from Aug 19, 2014) highlights this trend in New York City by focusing on the ClaimStat initiative, which was begun recently by NYC comptroller Scott Stringer. ClaimStat collects and analyzes data on lawsuits and claims filed each year against the city. By identifying patterns in payouts and trouble-prone agencies and neighborhoods, city managers hope to learn from these patterns and modify operations so as to reduce the frequency and costs of future claims (New York Times Editorial Board 2014).
Predictive analytics can be used in the government sector to target services to individuals in need or to anticipate how individuals or a subset of individuals will respond to a given intervention, such as the establishment of a new program or a change in an existing program (Cody and Asher 2014). For example, program administrators can use administrative data and predictive analytics to identify clients who are at risk of an adverse outcome, such as unemployment, fraud, unnecessary hospitalization, mortality, or recidivism. By knowing which participants are most likely to experience an adverse outcome, program staff can provide targeted interventions to reduce the likelihood that such outcomes will occur or reduce the negative effect of such an outcome.
With information from predictive analytics, administrators may also be able to identify who is likely to benefit from an intervention and identify ways to formulate better interventions. As in the private sector, predictive analytics can exploit the operational data used to support the day-to-day administration of a program, and the analytics may even be embedded directly in the operational data systems to guide real-time decision making. For instance, predictive analytics could be embedded in the intake and eligibility determination systems associated with a given program so as to help frontline caseworkers identify cases that may be have eligibility issues or to help customize the service response to meet the specific needs of individuals. In some state Unemployment Insurance systems, for example, program administrators use statistical models to identify new applicants who are likely to have long unemployment spells and refer the applicants to reemployment services. With any of the predictive models it is important that ethical and legal requirements are still met, which unfortunately is not always the case (for a discussion of unconstitutional sentencing see
http://bit.ly/1EpKt2j ).
7.3 Combining Big Data and Survey Data
Despite the theoretical and practical advantages of Big Data analysis described above, a preferred strategy is to use a combination of new and traditional data sources to support research, analytics, and decision making, with the precise combination depending on the demands of a given situation. As described in the introduction, traditional research that relies on primary data can be deployed to address the questions that are not adequately or easily addressed using Big Data sources. In many cases, this will entail going beyond the observed trends or behaviors that are easily captured using Big Data to more systematically address questions regarding why those trends or behaviors are occurring. For example, imagine a large advertiser has constant, real time-monitoring of store traffic and sales volume. Traditional research designs, which probe survey panelists on their purchase motivation and point of sales behavior, can help a retailer better target certain shoppers. Alternatively, the analytic design can be expanded to bring in the data on store traffic and sales volume so that these data become the primary monitoring tool, and surveys are utilized to conduct deeper probing based on trends, changes in trends or anomalies that are detected in the primary monitoring data.
Researchers recently have formulated ideas for blending Big Data with traditional research in the area of market research, which has traditionally been heavily reliant on data collected through surveys. For example, Duong and Millman (2014) highlight an experiment based on the premise that the use of behavioral data collected online can be used in combination with survey data on brand recognition to enhance learning regarding advertising effectiveness. In their experiment, data collected on users’ interactions with a website combined with data from a traditional online survey provided a clearer picture regarding the effect of different types of advertising than relying on the survey alone. Similarly, Porter and Lazaro (2014) describe a series of business case studies to illustrate how survey data can be blended with data from other sources to enhance the overall analysis. In one case study, the authors highlight the use of a blended data strategy to make comparisons by respondent. In the case study, consumer behavior data from website activity and transactions is combined with survey data capturing perceptions, attitudes, life events, and offsite behavior. By using respondent-level models to relate customer perceptions (from survey data) to behaviors for the same customers (from data on website activity), they were better able to understand the whys behind online behavior, and prioritize areas for improvement based on an understanding the needs of different individuals.
Blending strategies are also being pursued by government agencies. For example, the
National Center of Health Statistics (NCHS) is developing a record linkage program designed to maximize the scientific value of the Center’s population based surveys
6. The program has linked various NCHS surveys to administrative records from CMS and the Social Security Administration (SSA) under an interagency agreement among NCHS, CMS, SSA, and the Office of the Assistant Secretary for Planning and Evaluation, so these linked data can be used to support analysis of the blended data. For example, Day and Parker (2013) uses data developed under the record linkage program to compare self-reported diabetes in the National
Health Interview Survey (NHIS) with diabetes identified using the Medicare Chronic Condition Summary file, derived from Medicare claims data. Ultimately, linked data files should enable researchers to examine in greater detail the factors that influence disability, chronic disease, health care utilization, morbidity and mortality.
Similarly, the U.S. Census Bureau is identifying ways in which Big Data can be used to improve surveys and census operations to increase the timeliness of data, increase the explanatory power of Census Bureau data, and reduce operational costs of data collection (Bostic 2013). For example, the Bureau is planning to use data on electronic transactions and administrative data to supplement or improve construction and retail and service statistics that the Bureau maintains. In construction, the agency is examining the value of using vendor data on new residential properties in foreclosure to aid analysis of data on new construction and sales. The agency is also looking at ways to incorporate the use of online public records that are maintained by local jurisdictions and state agencies. In retail, the agency is evaluating electronic payment processing to fill data gaps such as geographical detail and revenue measures by firm size. All the Nordic countries have a system of statistical registers that are used on a regularly basis to produce statistics. The system is shown in Figure 10 and has four cornerstones: population, activity, real estate and business registers (Wallgren and Wallgren 2014).
Figure 10. A system of statistical registers - by object type and subject field, from Wallgren and Wallgren (2014).
6 http://1.usa.gov/1IlwiLW
Summary
- Surveys and BD are complementary methods, not competing methods. There are differences between the approaches, but this should be seen as an advantage rather than a disadvantage.
- Research is about answering questions and the best way to start is to study all available information. Big Data is one such source that provides new ways of approaching old questions and to address new questions.
- In the private sector Big Data is used to manage work and to make decisions. Examples of research techniques used are predictive analytics and rapid cycle evaluation.
- The use of data analytics to improve operations and public management decisions has been much less prevalent.
- Big Data can be used to detect patterns and establishing correlation between factors.
8. Conclusions and Research Needs
In this section we revisit the questions in the task force mission.
A. Can/Should Big Data be used to generate population statistics related to knowledge, opinion and behavior?
There are many different types of Big Data (Section 3). In this report we include
administrative data as one of them. The different types of Big Data differ due to the amount of researcher control and the degree of potential inferential power associated with each type (Kreuter and Peng 2014). On one side of the spectrum we have administrative data that has been used in some countries for many years to derive population estimates e.g., in the Nordic countries their population censuses are based on administrative data. Statistical agencies form partnership with owners of administrative data and can influence the design of the data. On the other side of the spectrum we have Big Data from social media platforms, where the researcher has no control of or influence on the data. During the last few years we have seen examples of statistics based on social media data. We have also seen studies that compare estimates from Big Data sources to estimates from traditional surveys. At the moment, however, there is not enough research to allow best practices to be developed for deriving population estimates from these social media types of Big Data. In between we have examples of Big Data sources where researchers actually can exert some control, for instance by positioning sensors in preassigned places to measure traffic flows and thereby also to some extent measure travel behavior.
One of the main criticisms regarding the use of Big Data is that there is no theory for making inference from it. The fact that Big Data is big is not enough, albeit some argue just that. The sampling theory that many statistical agencies rely on today was developed at a time when the only way to get data was to collect information from the total population. This was a very expensive endeavor and the sampling theory came as the rescue. Today we have a situation where a lot of Big Data is generated as byproducts from various processes or even the product from these processes. At the same time it is difficult to get participation in surveys, costs for surveys are rising and many of the assumptions from the sampling theory are violated due to nonresponse and other nonsampling errors. We are moving from the traditional survey paradigm to a new one where multiple data sources might be used to gain insights. Big Data is one of these data sources that can contribute valuable information. It is essential that theory and methods be developed so that the full potential of Big Data can be realized, in particular for “found” data that lack purposeful design. We are not there yet.
The gathering or collection of Big Data contains errors that will affect any estimates made and each Big Data source will have its own set of errors. The potential impact of each error source will vary between different Big Data sources. Just like we do for “small data”, we need to have a total error perspective when we consider using a Big Data source (Section 5) and a Big Data Total Error framework would help guide research efforts (Biemer 2014).
B. How can Big Data improve and/or complement existing ‘classical’ research methods such as surveys and/or censuses?
The availability of Big Data to support research provides a new way to approach old questions as well as an ability to address some new questions that in the past were out of reach (Section 7). Big Data can be used to generate a steady flow of information about what is happening - for example, how customers behave - while traditional research can focus instead on deeper questions about why we are observing certain trends or deviations from them - for example, why customers behave the way they do and what could be done to change their behavior.
Administrative data are used in several countries as sampling frames, in the estimation process in order to improve precision and in combination with surveys in order to minimize respondent burden. Other types of Big Data can be used in similar ways. Social media platforms can be used to get quick information about how people think about different concepts and to test questions.
Administrative data is also used as the gold standard in some methodological studies. For example, Day and Parker (2013) use data developed under the record linkage program to compare self-reported diabetes in the National Health Interview Survey (NHIS) with diabetes identified using the Medicare Chronic Condition Summary file, derived from Medicare claims data.
If we go beyond administrative data and look at other types of Big Data we see the opposite. Now survey data is used as a benchmark. There are a number of studies that look at estimates from a Big Data source and compare those results with estimates from a traditional survey (Section 3). The correlation between the two sets of estimates is of interest in these types of studies. If the correlation is high (and does not suffer from unknown algorithmic changes) the Big Data statistics can be used as an early warning system (e.g., Google Flu) since they are cheap and fast. For this to work transparency of algorithms is key, and agreements need to be found with the private sector to ensure they are stable and known
In the private sector Big Data is used to manage work and to make decisions. Examples of research techniques used are predictive analytics and rapid cycle evaluation.
C. Can Big Data outperform surveys? What if any current uses of Big Data (to learn about public knowledge, opinion and behaviors) appear promising? Which types of applications seem inappropriate?
Big Data has a number of advantages when compared to survey data. An obvious advantage is that these data already exist in some form, and therefore research based on Big Data does not require a new primary data collection effort. Primary data collection is usually expensive and slow, which can either greatly delay the generation of new research findings or even make new research prohibitively expensive.
As mentioned earlier administrative data is being used in many countries. The Nordic countries have a system of statistical registers that are used on a regular basis to produce statistics about the population, businesses, as well as economic and real estate activities.
A useful strategy is to combine new and traditional data sources to support research, analytics, and decision making, with the precise combination depending on the demands of a given situation. Scanner data from retailers is one example of a type of Big Data source that combined with traditional survey methods can both increase data quality and decrease costs. Scanner data are, for example, used in the production of the Consumer Price Index (CPI) in several countries. Another example is Big Data obtained from tracking devices such as a log of steps drawn from networked pedometers which might be more accurate than what could be solicited in surveys given known problems with recall error. Other Big Data sources with a similar potential include sensor data and transactional data. As these examples show, so far
the integration of the data sources is more straightforward if both, small and Big Data, are designed data. However, we are hopeful that the work of AAPOR and others in this area will expand the integration to found data as well.
D. What are the operational and statistical challenges associated with the use of Big Data?
The current pace of the Big Data development is in itself a challenge. It is very difficult to keep up with the development and research on new technology tends to become outdated very fast. Therefore a good strategy for an organization is to form partnerships with others so that multidisciplinary teams can be set up in order to make full use of the Big Data potential (Section 6).
Data ownership is not well defined and there is no clear legal framework yet for the collection and subsequent use of Big Data. Most users of digital services have no idea that their behavior data may be re-used for other purposes. Researchers must carefully consider data ownership issues for any content they seek to analyze. The removal of key variables as Personally Identifiable Information (PII) is no longer sufficient to protect data against reidentification. The combination of location and time metadata with other factors enables reidentification of “anonymized” records in many cases. New models of privacy protection are required.
Organizations seeking to experiment with Big Data computer cluster technology can reduce their initial capital outlays by renting pre-built compute cluster resources (such as Apache Hadoop) from online providers. Systems such as Apache Hadoop drastically simplify the creation of computer clusters capable of supporting parallel processing of Big Data computations.
Although the cost of magnetic storage media may be low, the cost of creating systems for the long-term storage and analysis of Big Data remains high. The use of external computer cluster resources is one short-term solution to this challenge.
9. References
Acquisti, Alessandro. 2014. “The Economics and Behavioral Economics of Privacy.” Pp. 98- 112 in Privacy, Big Data, and the Public Good: Frameworks for Engagement, edited by J. Lane, V. Stodden, S. Bender, and H. Nissenbaum. Cambridge: Cambridge University Press.
Antenucci, Dolan, Michael Cafarella, Margaret Levenstein, Christopher Ré, and Matthew D. Shapiro. 2014. “Using Social Media to Measure Labor Market Flows.” NBER working paper series. doi:10.3386/w20010.
Barocas, Solon and Helen Nissenbaum. 2014. “Big Data’s End Run around Anonymity and Consent.” Pp. 44-75 in Privacy, Big Data, and the Public Good: Frameworks for Engagement, edited by J. Lane, V. Stodden, S. Bender, and H. Nissenbaum. Cambridge: Cambridge University Press.
Biemer, Paul P. 2010. “Total survey error: Design, implementation, and evaluation.” Public Opinion Quarterly, 74(5): 817-848.
Biemer, Paul P. 2014. “Toward a Total Error Framework for Big Data.” Presentation at the American Association for Public Opinion Research (AAPOR) 69th Annual Conference, May 17. Anaheim, CA.
Biemer, Paul P. and Dennis Trewin. 1997. “A review of measurement error effects on the analysis of survey data.” Pp. 603-632 in Survey measurement and process quality, edited by L. Lyberg, P. P. Biemer, M. Collins, E. de Leeuw, C. Dippo, N. Schwarz and D. Trewin. New York: Wiley & Sons.
ten Bosch, Olav and Dick Windmeijer. 2014. “On the Use of Internet Robots for Official Statistics.”
Presentation at the Meeting on the Management of Statistical Information Systems (MSIS), April 16. Dublin, Ireland.
Bostic, William G. 2013. “Big Data Projects at the Census Bureau.”
Presentation to the Council of Professional Associations on Federal Statistics (COPAFS), March 1. Washington, DC.
Brynjolfsson, Erik, Lorin M. Hitt, and Heekyung Hellen Kim. 2011. “Strength in Numbers: How Does Data-Driven Decision-Making Affect Firm Performance?”
ICIS 2011 Proceedings, paper 13.
Butler, Declan. 2013. “When Google Got Flu Wrong.”
Nature, 494(7436): 155-156.
Cecil, Joe and Donna Eden. 2003. “The Legal Foundations of Confidentiality.” cited by Julia Lane.
Key Issues in Confidentiality Research: Results of an NSF Workshop. National Science Foundation. Retrieved January 28, 2015 (
http://1.usa.gov/1Eq58Df).
Cody, Scott and Andrew Asher. 2014. “Smarter, Better, Faster: The Potential for Predictive Analytics and Rapid-Cycle Evaluation to Improve Program Development and Outcomes.”
Mathematica Policy Research. Retrieved January 28, 2015 (
http://brook.gs/1AMaW5t).
Couper, Mick P. 2013. “Is the Sky Falling? New Technology, Changing Media, and the
Future of Surveys.”
Survey Research Methods, 7(3): 145-156.
Cukier, Kenneth and Viktor Mayer-Schoenberger. 2013. “Rise of Big Data: How it's
Changing the Way We Think about the World.”
Foreign Affairs 92(3). Retrieved January 28, 2015 (
http://fam.ag/1bKTv6t).
Daas, Piet J.H. and Marco J.H. Puts. (2014) “Social Media Sentiment and Consumer
Confidence.”
European Central Bank Statistics Paper Series No. 5.Frankfurt, Germany.
Daas, Piet J.H., Marco J.H. Puts, Bart Buelens, and Paul A.M. van den Hurk. 2013 “Big Data and Official Statistics.”
Presented at the 2013 New Techniques and Technologies for Statistics conference (NTTS), December 21. Brussels, Belgium.
Day, Hannah R. and Jennifer D. Parker. 2013 “Self-report of Diabetes and Claims-based Identification of Diabetes Among Medicare Beneficiaries.”
National Health Statistics Reports, number 69, November 1, 2013. Centers for Disease Control and Prevention. Washington, DC.
Dinan, Kinsey. 2013 “Local Agency Lessons on Implementing Random Assignment: An Example from NYC’s Child Support Program.”
Presented at the Association for Public Policy Analysis and Management (APPAM) Annual Fall Research Conference, November 8. Washington, DC.
Donohue, John J. and Justin Wolfers. 2006. “Uses and Abuses of Empirical Evidence in the Death Penalty Debate.”
Stanford Law Review, 58(3): 791-846.
Duncan, George T., Mark Elliot and Juan Jose Salazar-Gonzalez. 2011.
Statistical
Confidentiality, Principles and Practice. New York: Springer.
Duong, Thao and Steven Millman. 2014. “Behavioral Data as a Complement to Mobile Survey Data in Measuring Effectiveness of Mobile Ad Campaign.”
Presented at the CASRO Digital Research Conference 2014. Retrieved January 28, 2015 (
http://bit.ly/1v4fFBc).
Dwork, Cynthia. 2011. “A Firm Foundation for Private Data Analysis”
Communications of the ACM, 54(1): 86-95.
Evans, David S. 1987. “Tests of Alternative Theories of Firm Growth”
The Journal of
Political Economy, 95(4): 657-674.
Executive Office of the President. 2014. “Big Data: Seizing Opportunities, Preserving
Values.” Washington DC. Retrieved January 28, 2015 (
http://1.usa.gov/1hqgibM).
Fan, Jianqing, Fang Han, and Han Liu. 2014. “Challenges of Big Data analysis.”
National Science Review, 1: 293-314.
Fan, Jianqing and Yuan Liao. 2012. “Endogeneity in ultrahigh dimension.”
Annals of
Statistics 2014, 42(3): 872-917.
Fan, Jianqing, Richard Samworth, and Yichao Wu. 2009. “Ultrahigh dimensional feature selection: beyond the linear model.”
The Journal of Machine Learning Research, 10: 2013-2038.
Gelman, Andrew, Jeffrey Fagan, and Alex Kiss. 2007. “An Analysis of the New York City Police Department’s “Stop-and-Frisk” Policy in the Context of Claims of Racial Bias.
” Journal of the American Statistical Association, 102(479): 813-823.
Gerber, Eleanor R. 2001. ”The Privacy Context of Survey Response: An Ethnographic Account.” Pp. 371-395 in
Confidentiality, Disclosure and Data Access: Theory and Practical Applications for Statistical Agencies, edited by P. Doyle, J. Lane, J. Theeuwes and L. Zayatz. Amsterdam: Elsevier.
Greenwood, Daniel, Arkadiusz Stopczynski, Brian Sweatt, Thomas Hardjono, and Alex Pentland. “The New Deal on Data: A Framework for Institutional Controls.” Pp. 192-200 in
Privacy, Big Data, and the Public Good: Frameworks for Engagement, edited by J. Lane, V. Stodden, S. Bender, and H. Nissenbaum. Cambridge: Cambridge University Press.
Griffin, Jane. 2008. “The Role of the Chief Data Officer.”
DM REVIEW, 18(2): 28.
Groves, Robert M. 2011a. “Three Eras of Survey Research.”
Public Opinion Quarterly, 75(5): 861-871.
Groves, Robert M. 2011b. ““Designed Data” and “Organic Data.”
Directors Blog, May 31. U.S. Census Bureau. Retrieved January 28, 2015 (
http://1.usa.gov/15NDn8w ).
Halevy, Alon, Peter Norvig, and Fernando Pereira. 2009. “The Unreasonable Effectiveness of Data.”
Intelligent Systems, IEEE, 24(2): 8-12.
Hall, Peter and Hugh Miller. 2009. “Using generalized correlation to effect variable selection in very high dimensional problems.”
Journal of Computational Graphical Statistics, 18(3): 533-550.
Hey, Tony, Stewart Tansley, and Kristin Tolle. 2009.
The Fourth Paradigm: Data Intensive Scientific Discovery. Microsoft Research.
Ibrahim, Joseph and Ming-Hui Chen. 2000. “Power Prior Distributions for Regression
Model.”
Statistical Science, 15(1): 46-60.
Jovanovic, Boyan. 1982. “Selection and the Evolution of Industry.”
Econometrica: Journal of the Econometric Society, 50(3): 649-670.
Kahneman, Daniel. 2011.
Thinking Fast and Slow. New York: Farrar, Straus and Giroux.
Karr, Alan F. And Jerome P. Reiter. 2014. “Using Statistics to Protect Privacy." Pp. 276-295 in
Privacy, Big Data, and the Public Good: Frameworks for Engagement, edited by J. Lane, V. Stodden, S. Bender, and H. Nissenbaum. Cambridge: Cambridge University Press.
Keller, Sallie Ann, Steven E. Koonin, and Stephanie Shipp. 2012. “Big Data and City Living – What Can It Do For Us?”
Significance, 9(4): 4-7.
Kinney, Satkartar K., Alan F. Karr, and Joe Fred Gonzalez Jr. 2009. “Data Confidentiality: The Next Five Years Summary and Guide to Papers.”
Journal of Privacy and Confidentiality, 1(2): 125-134.
Koonin, Steven E., and Michael J. Holland. 2014. “The Value of Big Data for Urban Science.” Pp. 137-152 in
Privacy, Big Data, and the Public Good: Frameworks for Engagement, edited by J. Lane, V. Stodden, S. Bender, and H. Nissenbaum. Cambridge: Cambridge University Press.
Kreuter, Frauke, and Roger D. Peng. 2014. “Extracting Information from Big Data: Issues of Measurement, Inference and Linkage.” Pp. 257-275 in
Privacy, Big Data, and the Public Good: Frameworks for Engagement, edited by J. Lane, V. Stodden, S. Bender, and H. Nissenbaum. Cambridge: Cambridge University Press.
Lane, Julia and Victoria Stodden. 2013. “What? Me Worry? What to Do About Privacy, Big Data, and Statistical Research.”
AMStat News, December 1. Retrieved January 28, 2015
(http://bit.ly/15UL9OW).
Lane, Julia, Victoria Stodden, Stefan Bender, and Helen Nissenbaum. 2014. “Editors‘
introduction.“ Pp. xi-xix in
Privacy, Big Data, and the Public Good: Frameworks for Engagement, edited by J. Lane, V. Stodden, S. Bender, and H. Nissenbaum. Cambridge: Cambridge University Press.
Laney, Douglas. 2001. “3-D Data Management: Controlling Data Volume, Velocity and Variety.”
META Group Research Note, February, 6. Retrieved January 28, 2015
(
http://gtnr.it/1bKflKH).
Laney, Douglas. 2012. “The Importance of 'Big Data': A Definition.”
Gartner Inc.
Lazer, David M., Ryan Kennedy, Gary King, and Alessandro Vespignani. 2014. “The parable of Google Flu: Traps in big data analysis.”
Science, 343(6176): 1203-1205.
Leek, Jeff. 2014a. “Why big data is in trouble: they forgot about applied statistics.”
Simplystats blog, May 7. Retrieved January 28, 2015 (
http://bit.ly/1fUzZO1).
Leek, Jeff. 2014b. “10 things statistics taught us about big data analysis.”
Simplystats blog, May 22. Retrieved January 28, 2015 (
http://bit.ly/S1ma4Z).
Levitt, Steven D. and Thomas J. Miles. 2006. “Economic Contributions to the Understanding of Crime.”
Annual Review of Law and Social Science, 2: 147-164.
Lohr, Steve. 2012. “The Age of Big Data.”
New York Times, February 11. Retrieved
January
28, 2015 (http://nyti.ms/1f7WKqh).
Lohr, Steve. 2014. “For Big-Data Scientists, 'Janitor Work' Is Key Hurdle to Insights.”.
New York Times, August 17
. Retrieved January 28, 20
15 (http://nyti.ms/1Aqif2X).
Manzi, Jim. 2012.
Uncontrolled: The surprising payoff of trial-and-error for business, politics, and society. Basic Books.
McAfee, Andrew and Erik Brynjolfsson. 2012. “Big Data: The Management Revolution.”
Harvard Business Review, 90(10): 61-67.
Murphy, Joe, Michael W. Link, Jennifer Hunter Childs, Casey Langer Tesfaye, Elizabeth Dean, Michael Stern, Josh Pasek, Jon Cohen, Mario Callegaro, Paul Harwood. 2014. “Social Media in Public Opinion Research: Report of the AAPOR Task Force on Emerging Technologies in Public Opinion Research.”
AAPOR Task Force Report. Retrieved January 28, 2015 (
http://bit.ly/15V7coJ).
New York Times Editorial Board. 2014. “Better Governing Through Data.”
New York Times, August 19
. Retrieved January
28, 2015 (http://nyti.ms/1qehhWr).
Nielsen, Michael. 2012.
Reinventing Discovery: The New Era of Networked Science. Princeton, NJ: Princeton University Press.
Nissenbaum, Helen. 2011. “A Contextual Approach to Privacy Online.”
Daedalus, 140(4): 32-48.
Norberg, Anders, Muhanad Sammar, and Can Tongur. 2011. “A Study on Scanner Data in the Swedish Consumer Price Index.”
Presentation to the Statistics Sweden Consumer Price Index Board, May 10-12, 2011. Stockholm, Sweden.
Ohm, Paul. 2010. “Broken Promises of Privacy: Responding to the Surprising Failure of Anonymization.”
UCLA Law Review, 57(6): 1701-1818.
Pardo, Theresa A. 2014. “Making Data More Available and Usable: A Getting Started Guide for Public Officials.”
Presentation at the Privacy, Big Data and the Public Good Book Launch, June 16. Retrieved January 28, 2015
(http://bit.ly/1Czw7u4).
Porter, Scott and Carlos G. Lazaro. 2014. “Adding Big Data Booster Packs to Survey Data.”
Presented at the CASRO Digital Research Conference 2014, March 12. San Antonio, TX.
Prada, Sergio I., Claudia González-Martínez, Joshua Borton, Johannes Fernandes-Huessy, Craig Holden, Elizabeth Hair, and Tim Mulcahy. 2011. “Avoiding Disclosure of Individually Identifiable Health Information. A Literature Review.”
SAGE Open. doi:10.1177/2158244011431279.
Schenker, Nathaniel, Marie Davidian, and Robert Rodriguez. 2013. “The ASA and Big Data.”
AMStat News, June 1. Retrieved January 28, 2015
(http://bit.ly/15XAzX8).
Scott, Steven L., Alexander W. Blocker, Fernando V. Bonassi, Hugh A. Chipman, Edward I. George, and Robert E. McCulloch. 2013. “Bayes and Big Data: The Consensus Monte Carlo Algorithm.” Retrieved January
28, 2015 (http://bit.ly/1wBqh4w).
Shelton, Taylor, Ate Poorthuis, Mark Graham, and Matthew Zook. 2014. “Mapping the Data Shadows of Hurricane Sandy: Uncovering the Sociospatial Dimensions of ‘Big Data’.”
Geoforum, 52: 167-179.
Squire, Peverill. 1988. “Why the 1936 Literary Digest Poll Failed.”
Public Opinion Quarterly, 52(1): 125-133.
Stanton, Mark W. 2006. “The High Concentration of U.S. Health Care Expenditures.”
Research in Action, 19: 1-10.
Stock, James H. and Mark W. Watson. 2002. “Forecasting using principal components from a large number of predictors.”
Journal of the American Statistical Association, 97(460): 1167-1179.
Strandburg, Katherine. J. 2014. “Monitoring, Datafication, and Consent: Legal Approaches to Privacy in the Big Data Context.” Pp. 5-43 in
Privacy, Big Data, and the Public Good: Frameworks for Engagement, edited by J. Lane, V. Stodden, S. Bender, and H. Nissenbaum. Cambridge: Cambridge University Press.
Tambe, Prasanna and Lorin M. Hitt. 2012. “The Productivity of Information Technology
Investments: New Evidence from IT Labor Data.”
Information Systems Research,
23(3-part-1): 599-617.
Tapia, Andrea H., Nicolas LaLone, and Hyun-Woo Kim. 2014. “Run Amok: Group Crowd Participation in Identifying the Bomb and Bomber from the Boston Marathon Bombing.”
Proceedings of the 11th Int. ISCRAM Conference. Pennsylvania, PA.
Taylor, Sean J. 2013. “Real Scientists Make Their Own Data.”
sean j taylor blog, January 25. Retrieved January 28, 2015
(http://bit.ly/15XAq5X).
Thompson, William W., Lorraine Comanor, and David K. Shay. 2006. “Epidemiology of seasonal influenza: use of surveillance data and statistical models to estimate the burden of disease.”
Journal of Infectious Diseases, 194(Supplement 2): S82-S91.
Tourangeau, Roger. Lance J. Rips, and Kenneth Rasinski. 2000.
The psychology of survey response. Cambridge University Press.
Varian, Hal R. 2014. “Big Data: New Tricks for Econometrics.”
The Journal of Economic Perspectives, 28(2): 3-27.
Wallgren, Anders, and Britt Wallgren. 2007.
Register-based statistics: Administrative Data for Statistical Purposes. New York: Wiley & Sons.
Wallgren, Anders, and Britt Wallgren. 2014.
Register-based statistics: Statistical Methods for Administrative Data. New York: Wiley & Sons.
Winkler, William E. 2005. “Re-Identification Methods for Evaluating the Confidentiality of Analytically Valid Microdata.”
Research Report Series, Statistics #2005-09. U.S. Census Bureau.
Big data: Data that is so large in context that handling the data becomes a problem in and of itself. Data can be hard to handle due to its size (volume) and/or the speed of which it’s generated (velocity) and/or the format in which it is generated, like documents of text or pictures (variety)
Data-generating process: Also known as the likelihood function, the process from which the data is generated (i.e. where did the data come from)
Found data: Also known as organic data, data that is created as a by-product of another process or activity (for example sensor data from a production line or timestamps and geo- data created from a tweet)
Hadoop: An open source distributed file system that can store both structured and unstructured data. Further, all data is duplicated so that no data is lost even if some hardware would break
Made data: Also known as designed data, data that is created with an explicit purpose (for example survey data or data from an experiment)
Map-Reduce: A divide-and-conquer data processing paradigm which distributes a heavy computation between several computers, speeding up the total time of the computation (for example, having ten computes searching 1 billion records each takes less time than having one computer searching 10 billion records by itself)
Structured data: Numerical and categorical data that fits into traditional relational databases. Most data that ‘feels natural’ to work with can be considered structured data
Unstructured data: Data that does not follow a clear structure (for example text in PDF files, sequences of video from security cameras, etc.) and that would need to be processed and organized in order to be worked with