Essay

Revista Brasileira de Saúde Ocupacional

ISSN: 2317-6369 (online)

Rita de Cássia Pereira Fernandesa,b The critical use of statistical inference in occupational
https://orcid.org/0000-0002-3353-5365

Verônica Maria Cadena Limac

epidemiology: essay
https://orcid.org/0000-0003-2714-4525

Fernando Martins Carvalhoa O uso crítico da inferência estatística na epidemiologia

https://orcid.org/0000-0002-0969-0170
ocupacional: ensaio

Abstract
Introduction: Statistical inference (SI) methods are widely used in epidemiological
a
Universidade Federal da Bahia, Faculdade
de Medicina da Bahia, Programa de Pós-
data analyses to evaluate random variation as a probable source of error in the results.
graduação em Saúde, Ambiente e Trabalho. However, the assumptions necessary for its use are frequently unmet. Objective: To discuss
Salvador, BA, Brazil. assumptions for SI, implications of its use when individuals are non-randomly selected,
b
Universidade Federal da Bahia, Programa
de Pós-graduação em Saúde Coletiva.
different statistical approaches of epidemiological data and results extrapolation sceneries
Salvador, BA, Brazil. in occupational epidemiology. Methods: Essay based in the specialized literature and in the
c
Universidade Federal da Bahia, Instituto authors’ academic experience. Results: This essay presents situations of epidemiological
de Matemática e Estatística, Programa research in which the SI procedures are used without the support of the assumptions for its
de Pós-graduação em Saúde, Ambiente e
Trabalho. Salvador, BA, Brazil. appliance. Situations of generalization of epidemiological results are discussed, pinpointing
when the validity of using SI can be questioned. The considerations about themes
Contact: addressed in this essay aim to contribute to the critical management of data analysis and to
Rita de Cássia Pereira Fernandes the process of results extrapolation in occupational epidemiology research, delimiting the
E-mail: role of SI. Conclusion: The use of statistical methods and of the report of research results
[email protected]
should include essential provisos that allow the reader to evaluate its pertinence, possible
assumptions violation, and the validity of the reasoning and of the inference used.
Keywords: Epidemiology and Biostatistics; Sampling Studies; Confidence Intervals;
Statistical Analysis; Data Interpretation, Statistical; Occupational Health.

Resumo
Introdução: Métodos de inferência estatística (IE) são amplamente utilizados na análise
de dados epidemiológicos para avaliação da variação amostral como provável fonte de
erro dos resultados. Entretanto, pressupostos necessários para seu uso muitas vezes não
são verificados. Objetivo: Discutir pressupostos da IE, implicações do seu uso quando
os indivíduos são selecionados de forma não aleatória, diferentes abordagens estatísticas
de dados epidemiológicos e cenários de extrapolação de resultados em epidemiologia
ocupacional. Métodos: Ensaio com base em literatura especializada e na experiência
acadêmica dos autores. Resultados: São apresentadas situações da pesquisa epidemiológica
How to cite (Vancouver): nas quais os procedimentos de IE são utilizados sem apoio nos pressupostos para sua
Fernandes RCP, Lima VMC, aplicação. São discutidas as situações de generalização de resultados epidemiológicos,
Carvalho FM. The critical use of
statistical inference in occupational assinalando-se quando a validade do uso da inferência estatística pode ser questionada.
epidemiology: essay. Rev Bras Saude Ponderações sobre temas abordados neste artigo visam contribuir para a condução crítica da
Ocup. [Internet]. 2025;50:e1. Availabre análise de dados e no processo de extrapolação dos resultados da pesquisa em epidemiologia
from: https://doi.org/10.1590/2317-
6369/35622en2025v50e1
ocupacional, delimitando o papel da IE. Conclusão: O uso de métodos estatísticos e o relato
de resultados de pesquisa devem ser feitos com as ressalvas indispensáveis que possibilitem
ao leitor avaliar sua pertinência, possíveis violações de pressupostos e a validade do
raciocínio e da inferência aplicadas.
Palavras-chave: Epidemiologia e Bioestatística; Amostragem; Intervalos de Confiança;
Análise Estatística; Interpretação Estatística de Dados; Saúde do Trabalhador.

Rev Bras Saude Ocup 2025;50:e1 https://doi.org/10.1590/2317-6369/35622en2025v50e1 1/10

 Introduction
In epidemiological studies, the use of convenience samples or censuses is common, whether they are cross-sectional
studies, cohort studies, or case-control studies. Although these are studies in which the selection of participants
was not based on a probabilistic strategy, there are plenty of published studies that present the results of statistical
tests and/or confidence intervals (CI) using a non-probabilistic sample. Considering the foundations of statistical
inference (SI) as the set of assumptions necessary for its application and purpose, generalizing conclusions based
on statistical tests and CIs implies compliance with these assumptions.

In this perspective, to cite just one of the renowned statisticians with an undeniable leading role in epidemiology,
Greenland, in a 1990 publication in the journal Epidemiology, entitled Randomization, Statistics and Causal
Inference1, states that

In most epidemiologic studies, randomization and random sampling play little or no role in the
assembly of study cohorts. I therefore conclude that probabilistic interpretations of conventional
statistics are rarely justified, and that such interpretations may encourage misinterpretation of
nonrandomized studies (p. 421).

Greenland and Poole2 emphasize that “distortions that arise solely from the play of chance are traditionally called
‘random errors.’ It is these distortions, and only these distortions, that P-values and confidence intervals address.”
(p. 115). These articles were preceded and followed by other contributions3-5, particularly the special edition of
The American Statistician in 2019, with 43 articles. In the editorial of this special issue, entitled Moving to a World
Beyond “p<0.05”6, the authors make it clear that CI results and p-values should be considered beyond binary
“statistical significance”, expressed as a significant result or not. The validity of conclusions in epidemiological
research depends on aspects other than SI methods. Valid scientific inferences involve considering different factors,
including study design, the process of recruiting and selecting participants, the quality of the measurements,
external evidence for the phenomenon under study, and the validity of the assumptions underpinning the analysis
of the data5,7.

The sources of uncertainty in epidemiological studies, systematic errors (biases), and random errors require
researchers to constantly seek to overcome them. The critical use of SI only concerns the procedures adopted
to deal with random errors. However, although the assumptions of SI are well-defined8, misinterpretations of
the results of hypothesis tests, using p-values and CIs, have long been mentioned9-11. Some authors attribute the
difficulty of using statistical methods to the complexity of the definitions of p-value, statistical significance, and
CI2. There seems to be a lack of knowledge, lack of understanding, or misunderstanding of the purpose and
assumptions of SI.

This essay aims to: critically discuss the use of SI, especially in occupational epidemiology studies; evaluate the
appropriateness of using statistics to approach epidemiological data, delimiting its different divisions - descriptive,
analytical, and inferential; and present and discuss situations of generalization of epidemiological results.

Stages of essay construction

A critical evaluation of the use of SI procedures is the axis that structures the text. A few authors from the fields
of statistics and epidemiology, especially in recent decades, have contributed their thoughts on the use of SI in
epidemiological studies. Initially, the essay presents a section on the principles and purposes of SI. It then critically
discusses situations commonly found in the analysis of epidemiological data, incorporated into the practice of
scientific dissemination, peer review, and journals in the field, which are rarely questioned.

Once this set of SI situations has been presented, the essay develops an argument about the statistical approach
to epidemiological data, with a view to delimiting its different divisions. It then characterizes the statistical
approaches to epidemiological data as descriptive, analytical, and inferential, with the aim of contributing
to the deconstruction of the conception that admits there are descriptive studies and analytical studies, but

Rev Bras Saude Ocup 2025;50:e1 https://doi.org/10.1590/2317-6369/35622en2025v50e1 2/10

 these would be analytical if, when presenting the measures of association resulting from the comparison of
groups (for example, exposed and unexposed), they necessarily made use of SI. According to this conception,
studies that do not use the generalization of data based on SI are not analytical and have compromised
external validity. Finally, the essay presents different situations of generalization of epidemiological data,
delimiting the role of SI.

SI principles and aims in Epidemiology

In epidemiology, studies of risk or exposure factors, or of the determinants of health conditions, or of the
effects of interventions, require the adoption of methods that make it possible to generalize, with a reasonable
margin of safety in the face of the limitations and uncertainties inherent in a population-based approach to
health problems.

The set of methods available to support the argument in favor of a study hypothesis includes SI, i.e. inferential
statistical procedures constitute an auxiliary axis to support the study hypothesis. The statistical argument, put
in this way, would not be the most important or central in epidemiological reasoning and in the discussion of
the results of epidemiological studies. Especially important for epidemiological studies is the magnitude of the
numerical difference in the frequency measures of the comparison groups, or the magnitude of the ratio of the
frequency measures, the directionality of the differences, or the ratio in the direction of the study hypothesis.
However, once SI is applied, its assumptions must always be considered.

In any epidemiological study, it is essential for researchers to delimit the population of interest (target population),
describe the actual population studied and how the individuals were selected to take part in the study. One of the
reasons for this is to provide the scientific community with enough information to judge the merits of the study
and the applicability of the results. Regarding the selection of individuals, the epidemiologist often does not have
the necessary resources to study all the individuals in the population to which they wish to generalize their results.
In these situations, they are led to study a part (sample) of this population.

In this essay, it is assumed that the population is a set of individuals who have at least one characteristic in common,
and from whom we are interested in analyzing the behavior of certain other characteristics that vary between the
subjects that make up the population. The target population is the population of interest, which may coincide with
the population under investigation if all its components are studied. If this is not possible, the target population
will still be the population of interest, but the population being investigated will consist of a sample drawn from
it. A sample can be either probabilistic or non-probabilistic depending on the method for selecting individuals
from the population.

Using probability theory and SI methods, the researcher will draw conclusions about the parameters of interest
in the target population, based on the results obtained from the sample. To do this, they will adopt the most
appropriate procedures so that all the characteristics of interest of the individuals who make up the target
population are represented in the sample. Thus, they will have to apply probability sampling procedures, by which
they will try to ensure that all the individuals in the population have an a priori known probability of being
included in the sample.

This process implies that if numerous probabilistic samples were taken from the same target population, there
would be a variation in the sample results, as many of these samples would present different results, since the
individuals making up the target population are different, and each sample taken at random could contain a
different set of individuals. As the epidemiologist will only be selecting one sample in their study, which could be
any one of a number of samples that could be selected, the fact that the results obtained from different possible
samples will vary between them will force them, faced with the only sample selected, to ask themself: what
information will I be able to obtain about the parameters of the target population through the results obtained
from this single sample studied? The process of SI - with the calculation of p-values or CIs, in the case of classic
SI - has the sole purpose of helping the epidemiologist answer this question.

Rev Bras Saude Ocup 2025;50:e1 https://doi.org/10.1590/2317-6369/35622en2025v50e1 3/10

 With hypothesis testing, in which the p-value is calculated, the epidemiologist will test whether or not a given
statement about a population parameter is refuted by the results of the sample. For example, in a cross-sectional
study, one could test the null hypothesis that the prevalence ratio is equal to 1, i.e. the hypothesis that there is no
association between the independent variable and the outcome.

With the help of the CI, the epidemiologist will assess the uncertainty surrounding the point estimate obtained from
the sample. The smaller the amplitude of the CI, the greater the precision with which the population parameter
of interest is known. A 95% confidence interval means that in 100 replications, 95% of them would contain the
population parameter. The binary interpretation of the CI should be avoided, dichotomized into significant or
non-significant, a practice that is unfortunately quite common, based solely on checking whether the CI includes
the null value6,10,12,13. We will return to this criticism later.

Common situations of questionable use of SI

Consider that without SI it is not possible to assess differences between the

subgroups studied
Some epidemiologists argue that they need to use SI because without it they wouldn’t be able to assess whether
or not there is a statistical difference in some outcome of interest between the subgroups studied (for example,
the prevalence of hypertension among older and younger workers in the study population), even if it is a census
survey or with a non-probabilistic sample. But why does the epidemiologist, in this case, think they should assess
whether there is a statistical difference between the elderly and young subgroups?

Statistically evaluating a difference between subgroups involves evaluating the sampling distribution for the
estimator, in the example in question, the difference in the prevalence of hypertension between older and
younger workers, or evaluating the test statistic, in the case of the hypothesis test for comparing prevalence
between subgroups. If the assumptions about the probabilistic model and probabilistic sampling are not met,
what guarantee can there be for the results of the CI or hypothesis test? Clearly, there are limitations to the use
of SI methods in these situations which need to be considered and discussed critically. In census studies, as the
entire target population has been investigated and there has been no sampling, the numerical difference (not the
statistical difference) observed between the prevalence of hypertension between the subgroups is the difference
that exists in the target population, provided that the study does not contain significant biases. The relevance
of this numerical difference obtained can be assessed in terms of its magnitude and epidemiological, clinical,
and/or pathophysiological significance.

Consider that without the use of SI the article will not have scientific validity or
will not be publishable
Take the example of a study with a probabilistic sample that concluded, based on SI, that the difference found
between the incidences of lung cancer among smokers and non-smokers was not due to simple sampling variation
of the results, so there was a statistically significant result. In this example, the difference in incidence exists, it is
not zero, in the target population from which the sample was taken. At the end of the study, the researcher will be
able to conclude that a statistical association has been observed between smoking and lung cancer. But concluding
that smoking causes lung cancer would not be the result of a single study.

The evidence obtained from a specific study can only add to the body of evidence already available. And the
relationship between smoking and lung cancer is a classic example of the long process of attributing causality,
given the accumulated body of evidence thanks to the various epidemiological investigations that have allowed
this causal relationship to be consolidated. It is therefore important to consider that the relevance of the
evidence from a single study will depend on the appropriate study design; the biological, pathophysiological,
and epidemiological plausibility of the hypotheses formulated and the results observed; the strength of the

Rev Bras Saude Ocup 2025;50:e1 https://doi.org/10.1590/2317-6369/35622en2025v50e1 4/10

 associations found; and the consistency of the results when comparing them with those found in other studies.
In the long term, this process of assessing causality will continue, through systematic reviews of the literature
and meta-analyses, considering the results of various studies of good scientific quality and assessing whether or
not they meet the criteria for attributing causality14,15. In view of the above, SI procedures, when applicable, will
be important in each specific study to address the influence of sampling variation on its results. It is therefore
necessary to distinguish the process of SI from the process of causal inference in epidemiology, which involves many
other assessments.

Another point to consider, which can contribute to authors’ misconceptions, is that many editors of scientific
journals also mistakenly consider the use of SI procedures in articles to be essential for publication. According to
Rothman, Greenland, and Lash13:

Many researchers still believe that it would be fruitless to submit for publication any paper that lacks
statistical tests of significance. Their belief is not entirely ill-founded, because many journal editors
and referees still rely on tests of significance as indicators of sophisticated and meaningful statistical
analysis as well as the primary means of assessing sampling variability in a study (p. 151).

Again, SI procedures were developed to help the researcher carefully extrapolate the results obtained from the
sample to a target population.

Not considering the specifics of using SI in a randomized clinical trial

In a randomized clinical trial (RCT), when the study is well planned and conducted, two draws are made: one to
select possible study participants and another (called randomization) to separate the groups to be compared. If the
first draw is not carried out, but only the second, the epidemiologist will still be faced with the following question:
was the difference observed in the occurrence of the outcome between the compared groups due to the different
interventions to which these groups were subjected? Note that if another randomization were carried out, the
groups selected might not be exactly those obtained in the only randomization carried out and, consequently,
the difference found in the occurrence of the outcome between the groups might not be exactly the same, i.e. if
numerous randomizations were carried out, there would be a variation in the results observed. The epidemiologist
should then ask themself: based on the difference obtained in the compared groups, which difference is more
likely to be the real difference between the groups? Thus, SI can help them answer this question, concluding with
the use of SI - CI or p-values - on the difference most likely to be the true difference between the groups. However,
as the method of selecting possible participants in the target population was not probabilistic, i.e. there was not
the first draw mentioned above, the epidemiologist will have to judge how appropriate it is to generalize the results
of the study to the target population and to populations with comparable clinical conditions, noting possible
limitations in doing so16,17.

Statistical approaches to epidemiological data: the necessary delimitation of

each stage
Epidemiology is classically assumed to be based on three pillars: clinical, social sciences, and statistics18. And
the role of statistics is well recognized, identifying three different statistical approaches to epidemiological data,
according to the purpose of the study: descriptive, analytical, and inferential17.

Through the descriptive approach, it is possible to quantitatively characterize a reality or the individuals studied,
using simple, relative, and accumulated frequencies; calculating measures of central tendency, dispersion, and
position; graphs; and evaluating the shape of distribution curves. There are various types of statistical tools used
to characterize the population studied17.

The analytical approach investigates the existence of relationships between the characteristics or attributes of
individuals. Thus, the analytical approach makes it possible to obtain epidemiological measures of association

Rev Bras Saude Ocup 2025;50:e1 https://doi.org/10.1590/2317-6369/35622en2025v50e1 5/10

 between two variables, expressed as mean ratios, prevalence ratios, incidence ratios (relative risk), and odds ratios.
Other common techniques of the analytical approach provide correlation and linear regression coefficients, to
name a few of the most common13. An illustration of the appropriate use of the analytical approach can be seen
in the following excerpt from a study: “The sampling plan was non-probabilistic, so inferential statistics were not
calculated due to the inadequate estimation of the standard error in such a circumstance [...]. To assess correlation,
[...] Spearman’s correlation coefficient was used”19, without using the p-value or CI.

The inferential approach will take place when, based on data obtained from a probabilistic sample, it is intended
to statistically infer results for the target population, as already presented in the section about SI principles and
aims in Epidemiology. In this case, many statistical tests are available and will be selected based on the specific
hypothesis to be tested, both in terms of the type and characteristics of the data collected17,20. In addition, the
estimation theory will allow the CIs to be obtained.

Note that, in adopting these three different approaches, we are starting from the contribution of Silvany-Neto17,
who maintains the delimitation of Descriptive Statistics, but distinguishes Analytical Statistics from Inferential
Statistics. The most common delimitations in classic literature are Descriptive Statistics and Inferential Statistics.
In this sense, the proposed distinction seems crucial to deconstruct misconceptions regarding the use of SI based
on the understanding that a study is analytical only if it makes use of Inferential Statistics.

These approaches to epidemiological data are used in isolation or in combination, according to the objective and
nature of each study. For example, the sick leave rate was 19.9 per 100 people-year in a cohort study of oil industry
workers, obtained through a census21. Or the prevalence of low back pain in plastic industry workers, a study with
a probabilistic sample, was described as 28.9%, followed by a 95% CI (CI 22.0-35.8)22. In the first case, a descriptive
approach and in the second, descriptive and inferential approaches.

Furthermore, when the aim is to assess the existence, direction, and magnitude of an association between
variables, i.e. in an analytical study, if the sample is non-probabilistic - a census or convenience sample - a
measure of association will be obtained. In the case of the cohort described, women had a 79% higher absence
rate than men (HRaj: 1.79)21. And in a cross-sectional study of prison guards, female workers had 2.77 times
the prevalence of major depressive disorder of their male colleagues23. In the case of an analytical study with a
probabilistic sample and which also aims to investigate the association between variables, the CI will be added
to the measure of association: for example, the prevalence of pain in the proximal upper extremities among
those exposed to high psychological demand in the footwear industry is 63% higher when compared to workers
under low demand (RPaj: 1.63, 95%CI: 1.29-2.06)24. On the other hand, in the cohort with oil industry workers21
and in the survey among prison guards23, an analytical approach more suited to census studies was used. The
study in the footwear industry used both an analytical and inferential approach, investigating musculoskeletal
disorders in a random sample24.

Generalization issues in epidemiological studies

The sources of uncertainty in the scientific research process have been recognized as those related to sampling
variation, which is the subject of this essay, and those resulting from systematic errors, as suggested by Altman
and Bland25. The first concerns the accuracy or reliability of the study, the role of chance, and the second source
refers to biases related to the validity of the study or accuracy20. Although this text discusses issues relating to the
precision of the study and the various ways of approaching it, it is important to mention the validity of the study
and its link to the sample selection process.

The validity of a study is presented, in theory, as the absence of systematic errors or biases, whether of information
or selection, and confounding. It is preferable to assume that the role of the epidemiologist is not to eliminate any
and all biases; however, it is their role to try to do so and, more than that, to assess, explain, and document the
impact of imperfections that have not been overcome.

Once the validity of the study has been ensured, the internal validity of the sample or study population, the aim
is to extrapolate the results. The validity of the study is therefore the primary objective of an investigation: “The

Rev Bras Saude Ocup 2025;50:e1 https://doi.org/10.1590/2317-6369/35622en2025v50e1 6/10

 usefulness of research lies primarily in the generalisation of the finding rather than in the information gained about
those particular individuals “26 (p. 409).

In epidemiology, a distinction is made between internal validity and external validity. The former refers to the
degree of reliability of the conclusions reached for the sample or study population itself. As for external validity,
which is the extrapolation of the results of the sample beyond it, Pereira27 distinguishes as the first extrapolation
that of the sample studied for the target population and, as the second extrapolation, of the population studied
(sample or census) for other populations. He also points out that the second extrapolation is “based on subjective
judgments”. He states that: “A margin of uncertainty always surrounds any extrapolation, the interpretation of which
is a mixture of science and art”27. Similarly, Rothman et al.28 state about extrapolating results to other populations:
“That process is uncertain, along with everything else in empirical Science” (p. 1012).

The process of extrapolating the results obtained in a study will take place in different research situations
which will imply different types of extrapolations, which relate to the external validity of the study, as will be
detailed below.

Referring to the different research situations and the basis for extrapolating their results makes it clear that the use
of SI - which plays a leading role in health research - if its assumptions are taken into account, should be restricted
to a single extrapolation situation, the extrapolation of the results of a probabilistic sample to the target population
from which the sample was taken. In this situation, and assuming that the sample selection does not involve
significant systematic error, it is possible to use probability theory and SI methods to extrapolate the results of the
sample to the population, as presented in the section about SI principles and aims in Epidemiology. In other cases,
the use of SI can be questioned.

In a study with a probabilistic sample, a statistical test will evaluate the p-value of the test, i.e. the probability
of more extreme values occurring for the test statistic than those observed with the sample, assuming that the
null hypothesis is true. If the p-value is small enough, it indicates, for example, that the difference in means is
most likely independent of chance, given a previously set significance level (the so-called “alpha” value), or if this
probability is not small enough to allow you to rule out chance as an explanation for the difference obtained. Just
to illustrate, we can note the frequent use of an “alpha” of 5%, but there are more demanding situations in the
control of this type I error (rejecting a true null hypothesis), which use an “alpha”, for example, of 1%. In addition
to statistical tests, SI can be based on CIs, as already mentioned.

A relevant aspect concerns the interpretation of CIs25. Many authors have now reinforced the need to
overcome the binary interpretation of the CI, which we mentioned earlier, i.e. to conclude that the result is not
statistically significant (the one that includes the “null value”, 1, for measures of association - risk ratios or odds
ratios - or 0 for difference measures), or statistically significant (the one that does not include it). For Schmidt
and Rothman10:

Using CIs to judge whether the null value is within or not within the interval is equivalent to using
the CI as a significance test, rather than as a quantitative tool for determining the size of an effect and
the precision with which it is measured (p. 1090).

Therefore, in this case, refusing to interpret the CI as a dichotomous indicator of statistical significance, the
researcher will describe the practical implications of the values within the interval. This will make it clear that all
values between the limits of the interval are compatible with the data, given the statistical assumptions used in
its calculation. For example, a 95% CI between 0.99 and 4.20, with a relative risk (RR) of 1.8, between a certain
outcome and exposure, will allow the researcher to conclude that it is much more likely that there is this association,
than obtaining a 95% CI between 0.76 and 3.50, considering the same RR11,12.

Consider the assumptions of SI is the procedure envisaged by researchers who are critical of its indiscriminate use.
According to Daniel and Cross29, “in many instances in real-world applications it is impossible or impractical to use
truly Random Samples” (p. 164) and they postulate, regarding the extrapolation of the results thus obtained that:
“nonstatistical considerations must play a part in the generalization process “ (p. 165).

Rev Bras Saude Ocup 2025;50:e1 https://doi.org/10.1590/2317-6369/35622en2025v50e1 7/10

 The success of this extrapolation process will depend on the degree of similarity between the populations being
compared, when “there is no reason to believe that the material actually used is not representative of the population
about which inferences are desired” (p. 165). And they reaffirm: “Again, generalizations must be made on the basis
of nonstatistical considerations”29 (p. 165).

It is important to emphasize once again that, even if a study complies with the assumptions of SI, it can incur
systematic errors that can compromise its validity, interfering with the representativeness of the group initially
defined as the target population: the researcher must evaluate these limitations. But the validity of a study - the
chapter on biases, which should be reasonably well understood by epidemiologists - implies another subject for
discussion, which is not the aim of this essay.

In addition, it is important to check the state of the art on the subject under study. Ultimately, in any of these
situations, whether using a probabilistic sample, a convenience sample, or a census study, epidemiologists
should seek support for their conclusions in knowledge produced in the medium and/or long term, through
repeated studies in different locations and populations, using different study designs, and which provide
consistent results5.

Conclusion
In this essay, the foundations for the application of SI were presented, which allow statistical approaches to
epidemiological data to be organized as descriptive, analytical, and inferential. It is important to note that a study
is analytical when it investigates associations between variables and not because it uses SI. This distinction is
essential and timely for the necessary criticism of the use of SI.

When the sample is probabilistic, the results of the sample are extrapolated to the target population through
the SI process. In the case of non-probabilistic samples - judicious or intentional - or even in censuses, the use
of CIs or hypothesis tests can be questioned as to the possibility of inadequate results in such circumstances.
Considerations about the limitations of these methods for making generalizations in these situations should guide
the interpretation of the results.

It is hoped that this essay will help to stimulate the critical and parsimonious use of SI in epidemiological studies,
recognizing its contribution to the causal inference process, but distinguishing SI from it. Furthermore, given
the usual use of studies with non-probabilistic samples or censuses, as presented in this essay, it is necessary for
epidemiologists to note possible limitations, adopting the necessary parsimony in interpreting their findings and
the impact of using SI in these situations. This attitude should encourage debate about the scientific validity of
epidemiological studies.

References
1. Greenland S. Randomization, statistics, and causal inference. Epidemiol.1990 Nov;6(1):421-9.
https://doi.org/10.1097/00001648-199011000-00003
2. Greenland S, Poole C. Problems in common interpretations of statistics in scientific articles, expert reports, and testimony.
Jurimetrics. 2011;51(2):113-29.
3. Hahn GJ, Meeker WQ. Assumptions for statistical inference. Amer Stat. 1993;47(1):1-11.
https://doi.org/10.1080/00031305.1993.10475924
4. Strasak AM, Zaman Q, Pfeiffer KP, Göbel G, Ulmer H. Statistical errors in medical research: a review of common pitfalls.
Swiss Med Wkly. 2007 Jan;137(3-4):44-9. https://doi.org/10.4414/smw.2007.11587
5. Hubbard R, Haig BD, Parsa, RA. The limited role of formal statistical inference in scientific inference. Amer Stat.
2019;73(sup1):91-8. https://doi.org/10.1080/00031305.2018.1464947
6. Wasserstein RL, Schirm AL, Lazar NA. Moving to a World Beyond “p < 0.05”. Amer Stat. 2019;73(sup1):1-19.
https://doi.org/10.1080/00031305.2019.1583913

Rev Bras Saude Ocup 2025;50:e1 https://doi.org/10.1590/2317-6369/35622en2025v50e1 8/10

 7. Wasserstein RL, Lazar NA. The ASA statement on p-values: context, process, and purpose. Amer Stat. 2016; 70(2):129-133.
https://doi.org/10.1080/00031305.2016.1154108
8. DeGroot MH, Schervish, MJ. Probability and statistics. 4th ed. Boston: Addison-Wesley; 2012.
9. Greenland S, Senn SJ, Rothman KJ, Carlin JB, Poole C, Goodman SN, et al. Statistical tests, P values, confidence intervals, and
power: a guide to misinterpretations. Eur J Epidemiol. 2016;31(4):337-50. https://doi.org/10.1007/s10654-016-0149-3
10. Schmidt M, Rothman KJ. Mistaken inference caused by reliance on and misinterpretation of a significance test. Int J Cardiol
2014 Dec;177(3):1089-90. https://doi.org/10.1016/j.ijcard.2014.09.205
11. Szklo M. Quality of scientific articles. Rev Saude Publica. 2006;40(spec):30-5.
https://doi.org/10.1590/s0034-89102006000400005
12. Amrhein V, Greenland S, McShane B. Scientists rise up against statistical significance. Nature. 2019;567(7748):305-7.
https://doi.org/10.1038/d41586-019-00857-9.
13. Rothman KJ, Greenland S, Lash TL. Precision and statistics in epidemiologic studies: modern epidemiology. 3rd ed.
Philadelphia: Lippincott Williams & Wilkins, 2008.
14. Hill AB. The environment and disease: association or causation? Proc R Soc Med 1965 May;58(5):295-300.
15. Rothman KJ, Greenland S. Causation and causal inference in epidemiology. Am J Public Health. 2005;95 (Suppl 1): S144-50.
https://doi.org/10.2105/AJPH.2004.059204
16. Friedman LM, Furberg CD, DeMets D, Reboussin DM, Granger CB. Fundamentals of clinical trials. 5th ed. Switzerland:
Springer Cham; 2015.
17. Silvany Neto, AM. Bioestatística sem segredos. Salvador: Annibal Silvany Neto; 2008.
18. Susser M. Epidemiology, health & society: selected papers. New York: Oxford University Press; 1987.
19. Oliveira JL, Camelier FWR, Jesus FR, Barreto RMA, Ferreira PP, Terse-Ramos R et al. Perfil ventilatório e capacidade
funcional de pacientes com mucopolissacaridoses. Fisioter Bras. 2017;18(1):69-79. https://doi.org/10.33233/fb.v18i1.757
20. Henneckens CH, Buring JE, Epidemiology in medicine. Boston: Little, Brown and Company; 1987.
21. Almeida CGDSTG, Silvany Neto AM, Fernandes RCP. Risk factors for sick leave in Brazilian oil industry workers. Int Arch
Occup Environ Health. 2021 Aug Aug;94(6):1223-31. https://doi.org/10.1007/s00420-021-01650-w.
22. Fernandes RCP, Assunção AA, Carvalho FM. Prevalence of musculoskeletal disorders among plastics industry workers. Cad
Saude Publica. 2011 jan;27(1):78-86. https://doi.org/10.1590/S0102-311X2011000100008
23. Santos SN, Santos KOB, Carvalho FM, Fernandes RCP. Major depressive disorder in detention officers. Rev Saude Publica.
2021 Apr 14;55:11. https://doi.org/10.11606/s1518-8787.2021055002507.
24. Carvalho RLRB, Fernandes RCP, Lima VMC. Demandas psicológicas, baixo apoio social e repetitividade: fatores
ocupacionais associados à dor musculoesquelética de trabalhadores da indústria de calçados. Rev Bras Saude Ocup.
2019;44:e6. https://doi.org/10.1590/2317-6369000001517
25. Altman DG, Bland JM. Uncertainty beyond sampling error. BMJ 2014;349: g7065. https://doi.org/10.1136/bmj.g7065
26. Altman DG, Bland JM. Generalisation and extrapolation. BMJ.1998;317(7155):409-10.
https://doi.org/10.1136/bmj.317.7155.409
27. Pereira MG. Epidemiologia: teoria e prática. Rio de Janeiro: Guanabara Koogan; 1995.
28. Rothman KJ, Gallacher JEJ, Hatch EE. Why representativeness should be avoided. Int J Epidemiol. 2013;42(4):1012-4.
https://doi.org/10.1093/ije/dys223
29. Daniel WW, Cross CL. Biostatistics: a foundation for analysis in the health sciences. 10th ed. Hoboken: John Wiley; 2013.

Acknowledgments: The authors would like to thank Professor Annibal Muniz Silvany Neto, an epidemiologist with an outstanding
dedication to statistics, for his important contributions.

Authors’ Contributions: Fernandes RCP, Lima VMC contributed to the conception and design of the study; to the survey, analysis,
and interpretation of the data and to the drafting and critical revision of the article. Carvalho FM contributed to the survey, analysis,
and interpretation of the data and to the drafting and critical revision of the article. All the authors contributed to the approval of
the final version and take full responsibility for the work carried out and the content published.

Data availability: The entire data set supporting the results of this study has been published in the article itself.

Rev Bras Saude Ocup 2025;50:e1 https://doi.org/10.1590/2317-6369/35622en2025v50e1 9/10

 Funding: The authors declare that the study was not funded.

Competing interests: The authors declare that there are no conflicts of interest.

Presentation of the study at a scientific event: The authors inform us that the study has not been presented at a scientific event.

Received: September 28, 2022

Revised: February 09, 2023
Approved: February 14, 2023

Editor-in-Chief:
Eduardo Algranti

Rev Bras Saude Ocup 2025;50:e1 https://doi.org/10.1590/2317-6369/35622en2025v50e1 10/10