Palabras clave: biostatistics, descriptive statistics, P values, confidence intervals, public health

Introduction

The development of research projects often requires the use of mathematical tools in order to express different magnitudes in numerical terms or graphics such as frequencies, parameters differences or associations, which have emerged in different fields of science as a resource to help visualize and answer research questions. These can be classified in descriptive or summary measures such as mean, standard deviation or frequencies, or in an inferential manner including the so called “null-hypothesis significance tests” and confidence intervals, to mention the basic and most widely used [1].

The application of descriptive measures has been widely proven, and they are of great use in helping to achieve many interesting and incisive results. However, the use of inferential resources such as null-hypothesis significance tests is more controversial.

Null-hypothesis significance tests were introduced in the 1920s as a resource that revolutionized inferential techniques. They were initially proposed by Ronald Fisher who introduced the concept of "null hypothesis", which generally states that there is no relationship between two measured phenomena, or no differences between groups. Fisher's contribution was centered on his famous concept of p values [2].

Jerzy Neyman and Egon Pearson published a paper in 1928 [3] suggesting a procedure intimately linked with Fisher´s proposal, although different both in an operational and epistemological manner. It was conceived in order to choose between two possibilities: the so-called null hypothesis and an alternative one. Since the nineteen fifties until today a combination of these two methods has been applied [1],[4]: ​​given some data obtained from an observation or experiment, a "p" value is calculated using resources associated with the corresponding statistical significance test. When the p value is below a certain pre-established threshold (usually equal to 0.05) the null hypothesis is implicitly rejected, it is stated that "significance has been found" and the actual p value obtained is presented.

Although its use has become popular in many disciplines, null-hypothesis significance tests have been extensively criticized by renowned specialists in the field of statistics, focusing on its logical weaknesses and limitations, as well as for epistemological reasons [5],[6],[7],[8],[9]. Silva et al [4] described some of them, which included:

• It is possible to obtain a p-value as low as desired by simply increasing sample size.  Given this, the conclusions depend more on the resources available, those that allow a large sample size, than on the reality being studied.
• In practice, null-hypothesis significance tests simply provide a binary response (acceptance or rejection; significance or non-significance), without requiring a reasonable interpretation of the results. This orthodox approach limited to a dichotomy, offers a simplistic view that ignores the assessment of effects as demanded in science.

Given that this refers to a mathematical concept, a statistically significant association may not be clinically relevant and may not be causal. Moreover, a non-statistically significant association is not necessarily irrelevant. Since it is well known that "significance hunting” seduces both authors and publishers, it is especially pertinent to note that it is common to find "statistically significant but conceptually spurious" associations, as recently emphasized in a highly recognized journal (Nature) [10]. Despite the above, the uncritical reproduction of the method is still part of a ritual that feeds many texts and statistics courses.

In light of these shortcomings, several alternatives to the use of null-hypothesis significance tests have been proposed in recent decades. Among them, the most prominent one is to present the magnitude of the effect accompanied by a measurement of the error associated with the estimate. Confidence intervals not only meet this requirement, but also provide more information than null-hypothesis significance tests, while operating under a logic that does not "force" us to dichotomize conclusions.

Although it is possible to cite dozens of articles that adhere to this suggestion, it is sufficient to note that, since 1988, the International Committee of Medical Journal Editors (ICMJE), also known as the "Vancouver group", incorporated in its recommendations for authors [11], the calculation of confidence intervals. Thus, implying that analysis which solely depend on null-hypothesis significance tests and p values ​​should be avoided.

Due to their potential in establishing trends in the use of statistics, and considering that they can serve as a reference for the authors to reconsider their inferential paradigms, the purpose of this study is to describe the recent use of simple statistical tools as well as the use of p values ​​and confidence intervals, for the presentation of research results in the articles published in a recognized Public Health Journal in Colombia.

Methods

All original articles published in the National Journal of Public Health (Revista Nacional de Salud Pública) of the National University of Colombia during 2013 and 2014 were included. This is a unique set of items, limited to a given biennium and to a specific journal. However,  it can be understood as a sample of a superuniverse as defined by Hagood [12]: the work currently being carried out in this field and published after rigorous peer review in the Latin American context.

All original papers which presented results of experimentation or observation were included in the study. Review articles and theoretical or methodological articles were excluded.

The presence or absence of 12 statistical resources was identified for each article: rates, frequencies (absolute or relative ones), means, medians, standard deviations, charts (bar, pie, line, trend or other), p values ​​and confidence intervals.

Conditions under which each article was assessed by researchers were established, and subsequently, a pilot study with 20 papers was carried out. Every article was assessed by the first four authors. A Kappa coefficient in its variant for multiple classifiers was calculated using EPIDAT (version 4.1) [13] for each of the 12 variables. Results are presented in Table 1.

Table 1. Kappa values corresponding to several reviewers of binary variables for each statistical resource considered in the study (pilot study: n=20 articles).

This was certainly a poor result which reveals the difficulty in achieving consistency between observers even in apparently simple cases. A careful examination of the discrepancies led to additional adjustments. Since discrepancies were almost entirely due to the fact that the operational rules of the concepts involved were interpreted differently by observers, the adjustments consisted basically in refining or clarifying these operational definitions in order to ensure adequate consistency of the information that was going to be collected.

In the final assessment, 84 articles that met the conditions for inclusion in the biennium were considered. Each investigator performed a complete reading of all articles in the sample.

In all cases, a positive result was considered in cases in which the resource in question appeared at least once in the results, and negative when it does not appear or when it was only mentioned in the introduction or in an another part of ​​the article referring to previous work.

Results

Upon review of the journal website (http://www.revistas.unal.edu.co/index.php/revsaludpublica), all issues of the 2013 volume, and three of the six issues of the 2014 volume were available (Table 2).

Table 2. Original papers by year of publication in the selected journal.

The most frequently statistical resource used in the presentation of research results were absolute or relative frequencies (89.3 %) usually expressed in percentages, followed by p values ​​(65. 4%) and confidence intervals (53.6 %). Moreover, rate was the resource less frequently used (14.3 %).

Regarding graphics, bar and line charts or trends were the most commonly used (15.4 %), although generally the use of graphics was scarcely noticeable when presenting results of studies (Table 3).

Table 3. Frequency of the employment of descriptive resources. National Journal of Public Health (RNSP) of the National University of Colombia, 2013-2014 (n = 84).

In almost half of the articles (48.9 %; 95% CI = 37.7 to 59.9) both confidence intervals and p values were present. The isolated use of p value was present in one of six papers (16.7 %; 95% CI = 9.9 to 26.4) and the use of confidence intervals without p values ​​was extremely small (4.8 %; 95% CI = 1.3 to 11.7) (Table 4).

Table 4. Frequency of use of p values and confidence intervals. National Journal of Public Health (RNSP) of the National University of Colombia, 2013-2014 (n = 84).

Discussion

One of the interesting findings provided by this study concerns the use of descriptive statistics as a sole tool to present the results of research. It is assumed that the journal publishing team is zealous in corroborating that the questions raised by the authors have been effectively answered. The fact that 29.8 % of the articles used neither null-hypothesis significance tests nor confidence intervals reveals the expressive power of descriptive statistics. The use of purely descriptive measures allows many interesting and insightful results; somehow our results validate this reality.

One role of statistics is to provide objective quantitative alternatives in order to avoid, as far as possible, subjectivity and bias in the process of obtaining knowledge. Significance tests are often seen as less than ideal when seeking objectivity, since they, as it is believed, can generate conclusions independently of the people who use them [14].

Although objectivity is a natural and legitimate desire, in the strict sense, it is an unattainable goal. Statistics can not fully solve this conflict given that all inferential processes will always have a subjective component. While statistical techniques can be very useful, they tend to promote a belief, which many times disregard the need to examine reality through comprehensive thinking [15].

In ​​Public Health, non-systematic reflections have been made about the use of statistical significance tests and their ability to measure, quantify or assess differences between objects under investigation. However, the value of p itself does not provide information regarding the global importance or qualitative meaning when results are applied, and provides no information about possible future outcomes in the general population [16].

In this regard, it could be a fallacy to consider in a complex world, that the calculation of p values ​​can provide enough information to answer research questions, given that it ​​only refers to dichotomous conclusions. Consequently, its use may threaten the construction of useful knowledge for the decision making process in the real world [17].

In the first place, research questions that are addressed based on the logic of p values ​​are only those that inquire about differences between groups being compared. On the other hand, p appeal as a sole tool to present results and values ​​is insufficient, given that it is the only information that could be useful in constructing a biased, fragmented and often unconnected with life discourse.

In this regard, Manterola and Pineda´s concept [17] is worth noting. They claim that statistical significance is nothing more than that, "statistical significance", meaning that it can sometimes be positive and clinically irrelevant, or negative without necessarily signifying that there are no real differences between the variables being studied.

Despite the recommendation of the International Committee of Medical Journal Editors (ICMJE) back in 1993, on avoiding sole reliance on hypothesis testing for the presentation of research results, given that it fails to transmit information on the effect size [11], and after decades of objections by various authors [5],[8],[17],[18],[19],[20],[21]; null-hypothesis significance tests not even accompanied by confidence intervals, are still regularly used.  Cohen [22], in a paper intensely cited during the last two decades, considers that the use of the p-value approach to hypothesis testing is nothing more than a current ritualization that has been occurring over the last 40 years; the present review extends this reality over another 20 years. Our results reveal that such warnings are not fully taken into account in Public Health research.

A previous study, similar to this investigation [4] found that 21 % (95% CI = 16-26) (n = 306) of items reviewed during the 2005-2006 period, only use p values ​​for the presentation of research results. This result is similar to ours (16.7 %, 95% CI: 9.9 to 26.4) in which the lower end of the confidence interval confirms that at least 10 % of the studies seem to incur in this practice that is explicitly condemned by the most recognized methodological authorities.

In the same study [4], they also reported that only 14 % (95% CI = 9-17) of the articles reviewed in the 2005-2007 period did not use  confidence intervals without appealing to p values, while in this study the results were even lower (4.8 %; 95% CI = 1.3 to 11.7).

A limitation of the present study is that data were collected only for a short period and a single journal was evaluated. However, the idea of ​​offering a portrait of current practices is essentially valid despite this limitation.

Conclusions

Descriptive statistics is a commonly used tool. In the inferential realm however, criticisms and warnings by the International Committee of Medical Journal Editors, which condemn the use of null-hypothesis significance tests for the analysis and presentation of research results, are still not taken into account by a significant number of authors. Moreover, it also reveals that journal editors do not always demand compliance with this recommendation.

Notes

From the editor
This article was originally submitted in Spanish and was translated into English by the authors. The Journal has not copyedited this version.

Conflicts of interests
The authors have completed the conflict of interests declaration form from the ICMJE, and declare not having any conflict of interests with the matter dealt herein. Forms can be requested from the responsible author or the editors.

Table 1. Kappa values corresponding to several reviewers of binary variables for each statistical resource considered in the study (pilot study: n=20 articles).

Table 2. Original papers by year of publication in the selected journal.

Table 3. Frequency of the employment of descriptive resources. National Journal of Public Health (RNSP) of the National University of Colombia, 2013-2014 (n = 84).

Table 4. Frequency of use of p values and confidence intervals. National Journal of Public Health (RNSP) of the National University of Colombia, 2013-2014 (n = 84).

Esta obra de Medwave está bajo una licencia Creative Commons Atribución-NoComercial 3.0 Unported. Esta licencia permite el uso, distribución y reproducción del artículo en cualquier medio, siempre y cuando se otorgue el crédito correspondiente al autor del artículo y al medio en que se publica, en este caso, Medwave.