Statistical literacy and knowledge is needed to read and understand the public health literature. The purpose of this study was to quantify basic and advanced statistical methods used in public health research. We randomly sampled 216 published articles from seven top tier general public health journals. Studies were reviewed by two readers and a standardized data collection form completed for each article. Data were analyzed with descriptive statistics and frequency distributions. Results were summarized for statistical methods used in the literature, including descriptive and inferential statistics, modeling, advanced statistical techniques, and statistical software used. Approximately 81.9% of articles reported an observational study design and 93.1% of articles were substantively focused. Descriptive statistics in table or graphical form were reported in more than 95% of the articles, and statistical inference reported in more than 76% of the studies reviewed. These results reveal the types of statistical methods currently used in the public health literature. Although this study did not obtain information on what should be taught, information on statistical methods being used is useful for curriculum development in graduate health sciences education, as well as making informed decisions about continuing education for public health professionals.
Citation: Hayat MJ, Powell A, Johnson T, Cadwell BL (2017) Statistical methods used in the public health literature and implications for training of public health professionals. PLoS ONE 12(6): e0179032. https://doi.org/10.1371/journal.pone.0179032
Editor: C. Mary Schooling, Hunter College, UNITED STATES
Received: November 25, 2016; Accepted: May 23, 2017; Published: June 7, 2017
This is an open access article, free of all copyright, and may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose. The work is made available under the Creative Commons CC0 public domain dedication.
Data Availability: All relevant data are within the paper and its Supporting Information files.
Funding: The authors received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
Public health practice relies on the peer reviewed public health literature for current research and findings that support an evidence basis for effective practice. Studies have shown that statistical literacy and knowledge are needed for understanding published research . The rapid growth and widespread availability in computing power and user-friendly statistical software packages in recent decades has led to the use of more advanced statistical methods and analyses being used and reported in the health literature . However, statistical training in public health may not have kept up with the modern data explosion and statistical complexities increasingly being applied in health studies and reported in scientific publications. A comprehensive understanding of statistical concepts and methods is essential for understanding current research and developing effective public health practice.
Biostatistics education is a core requirement in all graduate degree public health programs accredited by the Association of Schools and Programs of Public Health (ASPPH) in the United States . One of the core curriculum competencies in biostatistics education for the master of public health (MPH) degree is to develop skill and knowledge to critically evaluate the application, presentation, and interpretation of statistical analyses in public health studies [3,4,5]. Although this is a desired outcome of training, there are no known recent studies that quantify the types of statistical methods used in the public health literature. Information on methods used is needed to make informed decisions about curriculum development, continuing education, and training of public health professionals.
The purpose of this work is to quantify the use of basic and advanced statistical methods in the general public health literature. A critical question of interest is “What statistical concepts and methods do public health professionals need to know to read and understand the literature?” Our study provides the needed evidence basis for beginning to answer this question.
The data collection form used in this study was created by the study authors and designed to gather information on statistical methods described in each randomly selected article. The form was rigorously developed and tested prior to use on the study sample. Our data collection form consisted of a closed-coding system for quantifying statistical methods reported by the authors of each published paper. We developed the form through a process of four pilot studies in which three reviewers read and cataloged the statistical methods reported in random samples of articles from our selected journals for the 2014 publication year. The results presented in the following tables in this article are framed to correspond to the data collection form. In reviewing each article, the selection of any variable on the review form indicated that variable had been explicitly or implicitly reported within the text of the paper. The final form included domains with specific items in each for article type, study design, sampling technique, summary statistics, reporting of statistical inference, statistical tests, statistical models, reporting of missing data, causal inference, and statistical software.
We aimed to obtain a list of influential general public health journals from which to sample articles. We sampled articles from seven top tier public health journals using the following method. Journals were selected based on a multi-faceted process. First, to gauge a general familiarity with general public health journals, we conducted an online internet search using the term “most influential public health journals.” From this, we compiled a master list of fourteen journals appearing on three or more lists identified from our online search. Next, one of the authors informally surveyed three experienced public health faculty members for suggestions of reputable public health journals. None were added, as all journals suggested were on our list. We next checked that all journals were recognized and included in PubMed. “PubMed comprises more than 26 million citations for biomedical literature from MEDLINE, life science journals, and online books.” (pubmed.com)” We next examined impact factors, deciding in advance to only include journals with exceptional impact factors. The cutoff was set at 3.0. After eliminating journals that were deemed more medically-focused and those that were specific to public health topics (e.g., policy, environmental health), we had seven remaining journals. Table 1 displays the 5-year impact factors for the 7 selected journals. The lowest impact factor was 4.245 for the European Journal of Epidemiology which we agreed was acceptable. We considered this set of 7 journals to comprise a representative sample of the top-tier general public health literature.
Sample size determination
The goal of this study was to quantify the types and frequencies of use of statistical methods in the public health literature. For purposes of determining the number of articles to be sampled to adequately accomplish this, we considered statistical methods that were rarely used. Thus, we concluded that if we calculated the sample size needed to detect rarely used methods, we would have a sufficient sample size to also cover the other more frequently occurring methods. Some advanced statistical techniques (such as nonlinear regression) were reported in only 1 of 42 articles in our pilot work. We therefore used ≈2.4% (= 1/42) as our estimate for a rarely used method to determine sample size with a precision estimation approach. Detection of a proportion of occurrence of 0.024 for an infrequently occurring statistical method and a given precision (interval width) of 0.05 resulted in a needed sample size of 188 articles. This is a reasonable precision within which we can be confident in our detection of rarely used statistical methods in the public health literature. We equated the notion of ‘attrition’ in our study to inappropriate articles that we agreed should be excluded from review (e.g., qualitative studies, editorials, etc.). In our pilot work, we had 5 articles appearing in the Research section of the journal that we deemed unsuitable for review. This equated to an ‘attrition’ rate of ≈12% (5/42). Assuming an attrition rate of 12%, we estimated a target sample size of 211 articles. Since we had 4 reviewers, for purposes of rounding, we decided to sample a total of 216 articles
The journals selected were American Journal of Public Health, American Journal of Preventive Medicine, International Journal of Epidemiology, European Journal of Epidemiology, Epidemiology, American Journal of Epidemiology, and Bulletin of the World Health Organization. All research based articles published in 2013 in these journals were eligible for review. Table 1 displays the names of our selected study journals, descriptions of the sections from which articles were sampled, and the number of eligible and sampled articles. There were a total of 1,023 research articles published in total across all seven study journals.
Data collection and analysis
We randomly sampled with probability proportional to the number of articles contributed by each journal [S1 File]. The 216 articles that comprised the study sample were then randomly allocated to four article groupings of 54 articles each. Each of the four reviewers was randomly assigned to review two of these 54-article groups (for a total of 108 articles per reviewer) and paired with one other reviewer for each article group. Review pairs consisted of one senior author (BC, MH) and one junior author (AP, TJ), such that both senior authors worked with both junior authors but not with one another and vice versa. Reviewers read and cataloged each article individually and a final consensus was reached in review pairs. Each pair met to review their assigned articles in three waves (wave 1 = 15 articles, wave 2 = 19 articles, and wave 3 = 20 articles), and the ordering of review and discussion between pairs was randomly ordered to mitigate learning and other group interaction effects on data collection. All four reviewers met as a large group periodically throughout the review process to discuss flagged articles and to ensure procedural consistency. Criteria for flagging articles included articles questionable for inclusion in our study (e.g., qualitative studies, program evaluation, study design overview reports).
Data entry was conducted using EpiInfo7  and data analysis performed with the SAS Software System (SAS Institute, Cary NC). The online database created in EpiInfo7 was designed to match the paper form used throughout the review process for ease of data entry and efficiency. Master copies of the paper forms drafted during each pair reviewer meeting were collected and hand-entered by one of the reviewers. Upon completion of data entry, 10% of the records were randomly sampled, and a second reviewer cross-checked the entered records with the master copies. Percent agreement was near 100%, indicating a high confidence with the accuracy of the data entry process.
Data analysis consisted of frequency distributions for all study variables.
A total of 216 articles were reviewed. Table 2 displays the frequency of reported study types, as well as occurrence of descriptive and inferential statistics. The majority of articles were substantively focused (93.1%, n = 201) and reported an observational study design (81.9%, n = 177). Descriptive statistics (91.7%, n = 198) and tables (95.4%, n = 206) were reported in the vast majority of articles. Visual displays of data in the form of charts, figures, or graphs, were reported in 61.6% (n = 133) of the articles. The odds ratio was the most commonly reported epidemiological statistic (40.7%, n = 88). P-values and confidence intervals were the most commonly reported results from the use of inferential statistics, appearing in 72.2% (n = 156) and 76.4% (n = 165) articles, respectively. The reporting of more than one level of significance, indicated by a hierarchy of ‘*’ symbols (e.g., p<0.10*, p<0.05**, p<0.01***), was used in 18.1% (n = 39) of the studies. Adjustments for multiple testing were only reported in 5.1% (n = 11) of the studies reviewed. The Pearson’s Chi-Square or Fisher’s Exact statistical test were used in 25.9% (n = 56) of the reviewed articles.
Frequency of reported use of statistical models in the public health literature are reported in Table 3. We classified all types of logistic regression analyses (including binomial, ordinal, and multinomial) that assumed independent observations into a single category labeled simply as “Logistic Regression.” This was the most commonly reported statistical modeling technique used in the articles reviewed (38.4%, n = 83). Linear regression and Cox Proportional Hazards Regression were reported in 19.4% (n = 42) and 15.3% (n = 33) articles, respectively.
Advanced statistical models that accommodate an independence assumption violation required of classical statistical methods are displayed in Table 3 as dependent statistical models. The general linear mixed model, which assumes a normal distribution, was reported in 6.9% (n = 15) articles, and the generalized linear mixed model, which includes an extension of logistic and Poisson regression models to allow for dependent data, were reported 10.2% (n = 22) of the time. Complex statistical modeling techniques, including structural equation modeling and latent variable models, were reported in less than 5% of the study sample. Missing data was handled most often with casewise deletion (30.6%, n = 66). Multiple imputation was only used in 5.6% (n = 12) of the studies reviewed.
Statistical software packages cited in the reviewed articles is described in Table 4. The most common statistical software package cited as used by study authors was the SAS Software System. STATA was the second most commonly used software package (25.5%, n = 54). R was used in (8.3% n = 18) of the studies.
In order to properly and adequately train public health professionals to access scientific publications, it is essential to, at a minimum, be teaching statistical methods actually used and reported in top tier public health journals. Classical statistical frameworks, including hypothesis testing, confidence intervals, and statistical models, are essential and need to be taught in order for a student to read and comprehend what is being published. Our study results show that descriptive statistics were reported in a tabular or graphical format in more than 95% of the articles reviewed. Somewhat surprisingly, when statistical techniques were used, classical statistical modeling techniques were infrequently used, with logistic regression as the most commonly reported type of model applied in the articles reviewed. While these study data only quantify the methods used in the literature, based on its frequent use we advocate for logistic regression to be included in biostatistics education for graduate public health students. It is not specifically mentioned in the current ASPPH competency guidelines for MPH students .
Less than half of the studies reviewed mentioned anything about missing data. It is extremely unlikely that missing data is not encountered in the majority of public health research. This lack of reporting about missing data, including attrition, non-response, and dropouts, may reflect a need for journal submission guidelines to require mention of missing data, including its frequency, and how it was addressed in the statistical analysis. About a third of the studies reported using casewise deletion, a relatively outdated and biased approach for analyzing missing data. Missing data is a well-recognized challenge with human subject research. Modern methods for handling missing data (e.g., multiple imputation) were rarely used. This indicates several possible needs. On one hand, in order for newly developing public health professionals to read and understand the limitations of inadequately handling missing data in a statistical analysis, biostatistics education needs to include training on this topic. And on the other, public health professionals may benefit from an introduction to modern methods for handling missing data in a short course or continuing education workshop.
About 18% of studies reported significance testing results with a notation of some variation of the following format: *p < .05, **p < .01, ***p < .001. A result with two asterisks is mistakenly interpreted as more significant than a result with one asterisk . The level of significance in a scientific investigation, also known as alpha (α), is a fixed quantity determined before observing the data. In fact, all that is meaningful is whether or not the p-value is less than alpha. Use of the asterisks notation indicates a possible misunderstanding of p-values and the classical null hypothesis significance testing process used in determining statistical significance . The relatively high frequency of this problematic reporting could be avoided with education and training on appropriate statistical reporting of inferential statistics.
Statistical software is needed to analyze data. SAS and STATA were the two most commonly used packages reported. Exposure to one or both of these packages may be beneficial. Online training courses in statistical methods and statistical software have grown in popularity and may be an option for many working professionals seeking additional training in a format that is manageable with a full time position.
About 82% of studies were observational and less than 6% experimental. As the modern data age continues to evolve, with the increasing use of administrative and other large data sources, it seems plausible to expect more observational data not originally intended for research to become available and used in public health research. Avoiding misuse and ensuring scientific validity of health-related findings from such sources depends on well-educated and trained public health professionals. Although experimental studies remain as the gold standard for enabling causal inference, only a handful were reported. And while there are statistical methods that make causal inference with observational data possible, these approaches were scarcely used in our study sample.
When statistical techniques were used, the vast majority of statistical methods seen in our sample were classical statistical techniques commonly taught in a first or second course in introductory and intermediate statistics. Classical statistics is based on normal theory and rooted in the general linear model (GLM), a framework that includes the three t-tests, linear regression, and ANOVA. The GLM paradigm assumes independence between observations. When this assumption is violated, as is the case with repeated measures data, more advanced statistical techniques are needed to account for the data dependencies that arise. Advanced statistical modeling techniques, including mixed and marginal models, are such methods. However, these techniques, as well as complex statistical modeling techniques such as structural equation modeling and factor analysis, were rarely applied and reported.
The scarce reporting of advanced methods could be an indication that these methods are not of importance or relevance in public health studies. However, since training in these methods has only become available in more recent years, we postulate this may be due to the historic lack of education and training availability on these topics. Many of the advanced statistical techniques rarely observed in our study are methods that were not available in mainstream statistical software ten to twenty years ago. For example, seasoned researchers may not have been exposed to modernized statistical modeling techniques which now available and appropriate for analyzing dependent or multilevel data .
Education in modernized statistical methods, including advanced modeling and computationally intensive statistical techniques, is necessary for staying current and implementing new advanced and methods. In addition to solid training in classical statistics, we suggest that graduate public health programs may also benefit from providing advanced biostatistics education and training opportunities to their students. Statistical software and computing power now enables researchers to readily access and make use of advanced statistical methods. Public health professionals may benefit greatly from continuing education training opportunities that provide a structured foray into such methods, coupled with statistical software training to show how to apply these methods to real world data.
Reporting of a statistical method does not necessarily mean its use was appropriate or correct. We did not evaluate the appropriateness or correctness of application. The work presented here is limited to an assessment of statistical methods currently used in the general public health literature. Methods applied in research studies may not be adequate, correct, or appropriate. Previous work estimates that up to 80% of published research is wasted due to poor methods . Our work did not assess these aspects, and instead focused on quantifying which methods were used.
It is also important to note that the language used by authors to describe some statistical methods varied. For example, classical linear regression was referred to in many ways, including fixed-effects regression, linear regression, least-squares regression, and general linear model. In a few cases, the description of statistical methods used was unclear and necessitated group discussion to come to a consensus. Finally, our study is limited to studies accepted for publication. It would be interesting to assess any possible publication bias resulting from statistical methods used in accepted as compared to rejected manuscripts. Since articles were selected only from 2013, the cross-sectional nature of this study limits an ability to consider how the use of statistical methods has changed over time.
Statistics knowledge is essential for reading and understanding public health research. Review of a random sample of publications from top tier general public health journals showed descriptive statistics and tabular results were reported in more than 95% of the articles. About three quarters of the articles reviewed reported inferential statistics (e.g., p-value, confidence interval). In addition, classic and advanced statistical models were reported in more than a third of the publications. A working knowledge of descriptive and inferential statistics is essential to comprehend, evaluate, and interpret the results for most research studies. Graduate training for public health students and continuing education in biostatistics education for public health professionals are essential for acquiring and maintaining statistics knowledge, as well as continuing to develop new skills as more complex methods are increasingly used and reported.
There is a noticeable lack of an evidence basis to make curricula decisions about biostatistics education. Biostatistics competencies in graduate public health education include developing and cultivating a student’s ability to read and understand the public health scientific literature. However, little is known about the methods used in the literature. The work presented here may be useful to curriculum committees deciding on course and content offerings.
- Conceptualization: MJH AP TJ BC.
- Data curation: MJH AP TJ BC.
- Formal analysis: MJH.
- Investigation: MJH AP TJ BC.
- Methodology: MJH AP TJ BC.
- Project administration: MJH AP TJ BC.
- Resources: MJH AP TJ BC.
- Software: MJH AP TJ BC.
- Supervision: MJH AP TJ BC.
- Validation: MJH AP TJ BC.
- Visualization: MJH.
- Writing – original draft: MJH AP TJ BC.
- Writing – review & editing: MJH AP TJ BC.
- 1. delMas R., Garfi eld J., Ooms A., & Chance B. 2007. Assessing students’conceptual understanding after a first course in statistics. Statistics Education Research Journal, 6(2), 28–58.
- 2. Arnold LD, Braganza M, Salih R, Colditz GA. 2013. Statistical trends in the Journal of the American Medical Association and implications for training across the continuum of medical education. PLoS One, 8(10):e77301. pmid:24204794
- 3. ASPPH Education Committee. 2006. Master’s Degree in Public Health Core Competency Development Project, Version 2.31, downloaded 8/25/2016 at http://www.aspph.org/educate/models/mph-competency-model.
- 4. Council on Education for Public Health (2016). Accreditation Criteria: Schools of Public Health and Public Health Programs. Silver Spring, MD: Author. https://ceph.org/assets/2016.Criteria.pdf
- 5. Sullivan LM, Hooper L, Begg MD. 2014. Effective practices for teaching the biostatistics core course for the MPH using a competency-based approach. Public Health Rep 129:381–92 pmid:24982544
- 6. Dean AG, Arner TG, Sunki GG, Friedman R, Lantinga M, Sangam S, Zubieta JC, Sullivan KM, Brendel KA, Gao Z, Fontaine N, Shu M, Fuller G, Smith DC, Nitschke DA, and Fagan RF. Epi Info™, a database and statistics program for public health professionals. CDC, Atlanta, GA, USA, 2011.
- 7. Slakter MJ, Wu YW, Suzuki-Slakter NS. 1991. “*, **, and ***; statistical nonsense at the .00000 level.” Nursing Research, 40(4): 248–9. pmid:1857654
- 8. Hayat M.J. 2010. "Understanding Statistical Significance," Nursing Research, May/June: 59(3), 219–223. pmid:20445438
- 9. Hayat M.J., Hedlin H. 2012. “Modern Statistical Modeling Approaches for Analyzing Repeated-Measures Data.” Nursing Research, 61(3), 188–194. pmid:22551993
- 10. Chalmers I, Bracken MB, Djulbegovic B, Garattini S, Grant J, Gulmezoglu AM, et al. How to increase value and reduce waste when research priorities are set. Lancet. 2014;383(9912):156–65. pmid:24411644