Sometimes BMJ reviewers reject the articles because of poor statistical analysis. If the data of a manuscript submitted can be analyzed with advanced statistical techniques, then the editor will be saying that the manuscript should be shown to a statistician first. Is it necessary for the researcher to also know the basic concept of analysis as well as the statistician?
Statistics is the only tool we have to convert data into information, to quantify how much we can learn from data. Since learning is a major goal of all sciences and empirical sciences are dealing with data, an empirical scientist must understand the basic concepts of statistics. However, there is no need to be able to conduct an analysis, especially when it is complicated. A collaboration with a statistician is always a possible solution. The scientist must be able to communicate with the statistisian. Also for this the knowledge of the basic concepts is mandantory. Please note that statistics is involved already in planning a study! Starting to think about statistics after the data is obtained is often a kind of a post-mortem analysis.
To convert an individual data (raw data) to a meaningful data you need to do statistics. Once again science or research withou statistics is no science, As a researcher you have to have basic statistic information. And before starting a research work you have to know statistics in order to collect data or sample type and number. Overall, to read and understand other papers and to also to analyse your data, you should know statistics and sometimes you need to contact a profesional statisticians. Some methods of analysis are used wrongly and the outcome or interpretation is misleading.
Statistic is divided into descriptive and analysis. Knowing basic or elementary statistic some times is not eoungh to convert your data into meaningful information and indicative. Therefore, you need to share your discusion on analysis of data with statisticians in order to avoid misleading conclusion
One should have a sound knowledge of statistics as it will help you a justify your findings in a better way.
I will suggest you at least get some basic ideas of biostatistics since you will be the best candidate knowing what the manuscript is about. I had a few interesting experience discussing with statisticians before and realized that if I have no basic knowledge in statistics then the conversation will be extremely difficult. Not all high impact factor sicience journals are strict in statistics and I have seen numerous times that wrong statisctics were used in IF of 30+. Sometimes the content is more important than the numbers.
Especially in quantitative research, it is very essential to choose the statistical method properly and logically. The article rejection may be because of these reasons. A statistical test and result is actually a scientific support.
Thank you so much Sir's for your valuable answers. It is very helpful for other researchers who view the question and answers.
according to my point of view, knowledge of basic statistics is necessary, indeed crucial. His understanding is useful to prevent the scientific insight that must be the basis of the work, is subject to the alchemy of statistics itself.
The basic knowledge as well as interpretation skills are also essential along with the knowledge of Bio Statistics are essential for well documentation of the statistical significance.
Having some basic knowledge on these concepts will help you have a conversation with the statistician. Knowing and conveying the questions you want to examine will help the statistician run the right analyses for you. Many times, I find people have already collected the data and then post-hoc present the data without making sure that the data are cleaned or labelled to the statistician and ask them to do their magic. It usually doesn't work like that. Having a good collaborative relationship from start to finish will ensure that you have a good study and the findings are communicated correctedly if you knew what it is that you are asking and the statistician ran the correct analyses for the study. So in summary, it is very important for all of us researchers to have atleast basic knowledge in statistics, even if you are not running the analysis yourself.
Purnima, sadly I can give only a single upvote. So here are another 4 :)
One should use good stats from the point of conception of the study. It is risky to run an experiment without knowing first whether the methods you will use are sufficient to answer your question.
It depends on the research questions , that you are answering with your Manuscript.
If the Manuscript validates a fundamental question in biology, the importance focus on the mechanism.
If a manuscript validates a comparative study or analysis study ,here the statistics works. Even though the experiment functionally proved,it has to obey statistics, for example in genetics and evolutionary based manuscripts faces all these statistics oriented acceptance, evidently stats is a unit to show data to explanation or hypothesis.
It is necessary to learn statistics for doing research, publishing in higher impact journal is the least important of the problems...
My field is epidemiology and bio-statistics hence my research is involved in developing and appling new methods for statistical methods to problems in health related matters of research. I feel rewarded. Bio-statistics is an important tool that is important as an analytic tool so people can follow and understand implications.
Dear Magnus, knowing statistics is not to have a detailed knowledge of the differences between Type III and Type II errors in ANOVA or to go in depth into the philosophical subteleties of Bayesian vs. Frequentist approaches, to know statistics is to be able to understand that the statements appearing in many clinical journals like ' The statistical methods approved for the comparison og life expectency curves are Cox regression model, Weibull distribution..' are pure rubbish and that the essential is to have a good measure of BOTH entity and statistical significance of an outcome and that for a survival comparison between two groups a Mann-Whitney U's between the two groups having as measured variables the survival time for each patient is more than enough. Like Robert Laughlin, the 1998 Nobel prize for physics was used to state 'If a property is emergent you can see it even with a wrong mathematical method'. Anyting can be statistically significant with a sufficient number of subjects , this comes from the fact that at the denominator of the Standard Error you find the square root of the number of cases, this I think answers the second question about 'unrealistic statistically valid results'...
As you probably have understtod, my point is that you need to know statistics in order not to idolatrate statistics (like the editorial boards of some clinical journals that treat statistics as a magic recipet that cannot be varied in any tiny particular) , to know statistics corresponds to exercitate the common sense along a quantitative path, for a scientist is like to have a precise kick for a soccer player, it has nothing to do with paradigms conservation or any other thing, is the essence of the game.
Yes it is necessary because any research without hypothetical base has no reason to be researched, because it is assumed that certain confounding factors may affect the truth of the findings. so statistical significance is highly regarded to ensure high impact and check for variability in each of the factor studied in the research.
I quoted late (Florida State University) Dr. Paul Dirac in a statistical paper I wrote two years ago. he added:
" According to Paul Dirac “Some time before the discovery of quantum mechanics people realized that the connection between light waves and photons must be of a statistical character” (Dirac, 1958)."
it makes sense to make sense statistics as a tool can be very helpful as a result in many fields.
Thank you
"Error is all around us and creeps in at the least opportunity. Every method is imperfect". Charles Nicolle
Magnus, inferential statistics histroically came from this "Schrödeinger effect"-side. Probability was a measure for believes. RA Fisher still took it that way. Popper introduced some confusion by stating philosophically than positive proves are impossible and that only negative proves (disproves) can hold theoretically. This initiated the development of the habit to "disprove" null-hypotheses. Unfortunately, also a disprove requires factual knowledge, it won't work with *estimates* about facts that itself remain unknown can only be judged by probabilities. A work around was not to consider a particulat "prove" at all but focus only on a long-run properties of decisions (->Newman-Person). Given some assumptions about the data a true (a-priori and unproven!), then the error rates (types I & II) of yes/no decisions based on such data can be effectively controlled. This enforced many researchers to make such yes/no decisions, and additionally the black-and-white picture of the world seems so much simpler and clearer than the fuzzy gray shades of our beliefes.
And there was 2nd strand contributing to this development: beliefes were often thought to be unscientific and thus science should not be based on something that itself is based on beliefes. "Probability" should be defined in a different way. After many nor very successful attempts, von Mises founded a (pseudo-)definition of probability as being the limiting relative frequency. Seen this way, it was possible to estimate probability values from observed relative frequencies. Probabilities became a part of the objective nature around. (Fisher was not directly following this, he was well aware of the problem that probabilities are also assigned to unique events that in principle won't be repeatable ans thus can not have a relative frequency, but pressed by the common desire to forbit subjectivity he claimed the existense of what he called a "propensity" as an objective property that would translate/express into a relative frequence if the experiment was repeatable) This point of view lead to focus on repeatability of experiments to estimate objectively true values. There is a large subset of experiments where this is reasonable, but for many it is not. This was largely ignored.
Another force moving our understanding of probability and statistics into the same direction was established by the fact that probability theory was developed in relation to gambling, where you have a well-defined, invariant population of possible events from which one takes samples. This was transferred to problems where no such population is defined (sometimes virtual populations in thinkable parallel universes were postulated!) and a limited amount of data was interpreted as being a sample from a population, and such samples served to estimate population parameters. If there is a well-defined population, then there is a "true" (well-defined) value of any population parameter, so an estimate (or better an intervall estimate like the confidence interval) could really and objectively be correct or wrong.
Yes, Jan, you are right with Popper. His work was never about statistical inference. But as far as I know (I miss a porper source, though! *) his work influenced the statisticians. Popper wrote about the "scientific method", and his falsianism was actually about theories, not about statistical hypotheses.
*Hilborn & Mangel: "Popper supplied the philosophy and Fisher, Person and colleagues supplied the statistics" (The ecological detective: confronting models with data, 1997)
The link between Popper's logic and hypothesis testing is under debate, e.g. in Mayo (Error and the growth of experimental knowledge, 1996) and Oakes (Statistical Inference: a commentary for the social and behavioural sciences, 1986).
This also is a nice paper: http://www.laeuferpaar.de/Papers/HypothesisTesting_v5.pdf
In my opinion the 'big misunderstanding' is in the blurring of two very different concepts that instead must be kept well seprated: 1) relevance of an effect and 2) probability that the observed effect can be ascribed to chance and not to the 'order parameter' (drug, time, mutation..) I'm testing.If I had a drug that provokes a change of 1% in the hypertension of ALL THE TREATED PATIENTS it is sufficient a not so large sample of paired samples to observe the treatment is highly statistically significant, nevertheless the result is totally not-relevant because the 'owner' of the pressure (the patient) has no detectbale change in its health status from such a small effect. On the same coin if I will divide a set of 1000000 measurements of the length of my table into two A and B sets 500000 elements each, the difference in length between A and B groups will be surely statistically significant for the pure effect of number. This is not silly, it is perfectly rational given we keep separate the two issues, as a matter of fact the two A and B groups of the last example are not statistically significant by chance but by a very important effect called 'truncation error', nevertheless the REAL TABLE has obvioulsy the same length, Now we live in somewhat silly times in which none (especially in biomedical science) dares to accept the risk to be considered as 'subjective' (very strange fear to be hones, anyone can only be subjective given he is a subject, that means someone that does something) and simply saying 'well, I know the meaning of the measure I'm taking, and I can proudly affirm that a variation of less than 30% of this measure is totally not relevant' .. oh no, too antiscientific !!!! Let's go to the statistics and if I get a p < 0.05 than what I observed is TRUE !!! None wants to hear that if I use 200 subjects anyhing becomes significant (on the contrary the vulgata is 'my experimentation is funded on a very large data base, thus it is very reliable'), statistics is a very usefull tool for thinking (probably the most useful in theoretically poor sciences and especially inserting in statistics the development of a good measure, of a good sampling strategy, of the correct question and not simply the relatively trivial last step of choosing the inferential test), but while Universities are plenty of professors teaching difficult subjects they are most totally depleted of professors teching simple things...that in general are the most useful ones...
Well said, Allessandro: "but while Universities are plenty of professors teaching difficult subjects they are most totally depleted of professors teching simple things...that in general are the most useful ones..."
May I cite you? :)
Additionally, many things are presented in a complicated way just because
a) the teacher wants to be adored for understanding such complicated stuff and
b) the teacher did not understand the topic deeply enough to be able to present it in a simple way
Taken as a whole I would say: it is much, much more important to have a good name (at least of the senior aoutor) and a "sexy" topic. There is no other way for me to explain why so many articles published in high-impact journals are so bad (in many ways).
Thank you dear Jochen, yes you can quote me ;), and thank you dear Jan for reminding us the 'second part' of the question relating to 'high impact journal'. In a certain sense the fact the discussion is shifted to the first side of the question 'the statistics side' is a good sign given makes us to hope we are more interested to the core of our work (knowledge, contemplation of Nature, and satisfaction to share our views with the friends) than to the need of publishing on high impact factor journals, but we must face this point because, you are totally right, it is a crucial point.
I do not know if there exist low impact journals in which you can publish without knowing statistics (probably yes, but I never met a journal that at least not ask you some statistics, yes they could ask you the wrong questions but I never saw a so low impact journal that never asks statistics, and I have a large experience (I'm 54 and published in a lot of low impact journals..) under the same heading I am not so sure high impact journal are so 'statistically sophisticated' to ask for stellar statistics (I published in some high impact journal too and did not note some basic difference in this aspect, given my work is statistics this is the first thing I note and sincerely I do not rememeber such a particular interest). What I noted are two interesting tribal uses that differentiate medical and nowadays 'computational biology' journal: medical journals idolatrate statistics because in the depth of their soul they hate it , thus in order to have the deity under control feed her with human sacrifices like inferential tests with very strange names, particularly complicated strategies, and reiteration of myhologies like 'You did not check for normality, you cannot use a parametric test' (when at the first year of statistics you learn that inferential tests deal with sample (and not population) distribution that gets normal with samples around 30 units independently of the original population distribution form..), like in any ritual no deviations from the orthodoxy is accepted (only Cox analysis of variance, no principal components..). On the othe hand computational biology journals have the myth of 'the most powerful and innovative technique' they say they are interested to apply math and in geenral quantitative approach to biology, if this should be true any interesting (from the biological point of view) feature discovered by mathematical methods should be welcomed even if you use a mathematical method 100 years old, not at all ! If your method is not brand-new your paper must be rejected, you could say 'come on, this a cluster technique like any other' but you cry in a desert, you must apply some brand new clustering technique , it is sufficient it gives a clusterization R-square = 0.91 while the old one is 0.90 and you have a computational valid method.....I have still not decide which of the two tribes is the worst, but the point, so to answer to the original question is in any case the same 'Learn statistics, so you can both understand what you are doing and publish on high impct journal both becasue you have good results and because you can manage sufficiently good statistical rethorics to convince the tribal persons'.
..for sure not statisticians...at least in Italy they are much more interested in much more lucrative fields....
I certainly agree in this point, however if you do not love statistical analysis much. I think the way out is inviting your statistician friend to be one of your co-authors
If your work presents the necessity to develop stats hence you will need it. Otherwise with elementary level of them is enough. Finally if the bunch of datasets is huge you can obtain collaboration with a more experienced colleague in this line to avoid diplomats and so on. Hope this helps. Best regards.
My views about this whole issue are that there should be several expert perspectives to produce high quality research reports. I have a good understanding of what the number mean, how they are applied, and when they are/are not relevant to either the question at hand or the inferences drawn from them. I find that higher education places too much emphasis on all students being able to crunch the numbers and not enough on how the numbers should/should not be applied. The truth in academia is that it is slow to change, which is a double edged sword. Ideally a research team should have members representing a range of perspectives- clinical, epidemiological, pharmacology, sociology, economic, public policy, to name a few. I think if you have a multidisciplinary team that includes a good statistician you should be OK. I am a reviewer for an academic journal, and when I am asked to comment on the statistical analysis (the actual crunching of the numbers) I defer to the statistics experts and only comment on that aspect if I see something that either doesn't make sense or I believe the inferences being drawn are not supported by the data presented. If anything, I see plenty of high quality statistical analyses but poor application or interpretation of what they represent. If you do not have a statistician on your team, there are agencies you can hire to assist in this.
No, but you need someone on the article that can... no doubt about that
As a statistician, i would say that you don't need to learn statistics, instead you should try to find a strong statistician to collaborate with you. Once you find such a person, try to create a good environment for a him/her, be colleagues and i am sure you won't have set backs related to revisions of stats once you submit your research to strong journals.
In fact, no harm of adding statistician, but knowing the basics of statistics will help more in understanding andwritting your research in far better way. It is not a kind of buzzle as long as you understand the concepts of statistics and understand the way when you translate the findings from numbers to words.
I think it's crucial to be able to understand and stand over all components of your papers - that's what you're signing off on when you agree to publication - while you may not need to sit at the computer actually doing the analysis (though sometimes this is the best bit of doing research - so it is totally recommended) you do need to fully understand what was done and be able to ensure that it was conducted correctly. Ethically this is essential to maintain the integrity of your analyses. Sure get specialist input, but spend the time required to learn about the analyses conducted - it will certainly improve the rest of the paper
I think you should at least understand the basics of statistics and able to understand all term used in your work and able to interpret it wisely
The key is to understand what your statistical method answers on. Too often conclusions are drawn based on what you want to test and not what you tested for. The basics is not to understand exactly how theoretically a test works but to know enough about what kind of answers that the test is solving. If you fairly well can describe the limitations of your test then you must not necessarily use a test that is perfectly suited to your data.
As been mentioned previously in the thread it is also very important to think of what your results really prove. If you can prove a difference in weight between smokers and non-smokers is this really interesting? I think that too many are just happy with a significant p-value and does not think of if there is a clinical relevance in the significance. I think that more of the research needs to think in terms of clinical relevance and not focus as much on the p-value as currently seems to be the case. In this thinking I also include a higher valuation of descriptive statistics, i.e. that you do not need to only focus on what can be proven but also provide valuable information of what else you can see.
Agree - clinical significance (relevance) is as importance as statistical significance!
Clinical significance is more important than statistical significance. but actually you need both. You will never be certain based on statistical significance alone.
Definitely yes.
Although Statistics has many problems, sometimes it is the only available scientific or near to scientific approach.
Of course it is not necessary to be an expert in order to write a research article, but you have to study the basic theory before you try to apply it.
Otherwise, you will probably write something that seems not to 'came out' from you.
The question has come from a person in College of Medical Science. Thus, if the person is writing a research paper in a medical subject that is only explaining the case, there may not be any need for statistics whatsoever. Even, in case that there is a case study that needs to be supported by statistical analysis of some data, a statistician as a consultant may be hired to help. In my opinion, weak knowledge of statistics and applying it to a statistical analysis and inference is very dangerous. A person needing a research based on statistical analysis should know statistics very well so that he/she can interpret the result rather than reporting the results. This is where the line must be drawn.
I agree with Haghighi, Yes statistician can be hired to help you in doing and writting the statistical part in your research paper. BUT you will never feel what you write, sometimes...you might not understand it, if- and only if- you have a prior knowledge of statistics
True Ahmed, but the reality is that it is not practically possible that we know everything we need. That is why we collaborate in conducting research and writing books. However, a good statistician should be able to write his/her statistical analysis results so easy and without much tecnical terms that any reader including the author of the paper can easily understand and be able to relate it to the topic of the paper.
Slightly off-topic: (Alikabar) "However, a good statistician should be able to write his/her statistical analysis results so easy and without much tecnical terms that any reader including the author of the paper can easily understand"
This is a very good point, but it is not only true for statisticians. EVERY researcher should be able to express his ideas, results and interpretations in an understandable way, not only for those people being deep in the topic.
Very true Jochen. This is infact usually recorded in the Absract and Conclusion of a paper. Indeed, I do require authors do so in my Journal AAM.
I think so, if you do not know statistics at all, the plan of the project and methods of collection of samples as samples number and type will certainly delay and affect the project results as well as time.
No. In fact, knowing what the hell you're doing may be a disadvantage because it takes longer to analyze and interpret data correctly than to throw together something that sounds good and confirms your and the reader's biases. This blissfully unaware confidence can even rub off on a statistician if you do decide to talk to one. Try it sometime: present your data as a cut-and-dry case and don't voice any concerns about confounding variables, model selection, power, and alternative interpretations of your results. The less you know about these things the more convincing you'll sound. Unless the statistician has experience in your field, they will often take what you say at face value and breathe a sigh of relief because nobody likes complications or (god forbid) negative results.
The Dunning-Kruger effect is alive and well in research science.
Recientemente la IARC anuncio que el Diesel era una sustancia carcinogénica y a reglón seguido refiere que el DIESEL contiene 45 químicos con ese efecto. No me imagino como se pudiera llega a esa conclusión si un excelente asesor estadístico.
Statistics are an essential tool in any research area. Thus, If you want to have a strong and excellent article, you have to get a good statistical knowledge.
Statistics is not only used to analyse and present the results, it is necessary in the planing and design of the methodology. So, it is mandatory for a good researcher to have at less a basic knowledge of statistics. Also, I think that it will be better to have a statistician to advise your research, even if you have a good knowledge in statistics.
As a novice researcher, about to embark on a PhD, i would agree with most that statistical know-how is very important.
Firstly, one needs to understand how to read a paper and without some knowledge of stats, it makes it very difficult at times, as i have found on my own personal journey. Acceptance of the interpretations of authors is a very dangeorus game to play, when you are then using to endorse your own work.
Secondly, without stats knowledge, it is very difficult to put together one's own research in place for obvious reasons.
As a clinician too, learening how to read papers so that you can guide practice using available evidence is also important.
So in summary, knowledge of stats is a great tool to have wheteher involved in research or using it to guide clinical practice.
No, therefore we hace statistician. It make also no sense not do go to a dentist and just make it by your own.
Yes, a researcher can always keep a statistician as part of the team, but the research concept, design, implementation, analyses, and conclusions will be improved if the original researcher (whatever the scientific field of study) has a knowledge of basic biostatistics and can then be more informed to communicate and understand what the statistician is trying to contribute to the research study. There will be less probability for misunderstanding and inappropriate analyses and conclusions.
I think a researcher should know some degree of statistics in order to position its paper and understand what the meaning of all statistic results!
I agree the fore answers and i want to tell that the researchers are well aware of statistics, it will help for understanding and answering the quries raised by reviewers at the time of presentation.The statistician will not help the researcher at this context. Hence my humble appeal is all researchers should at least learn basic statistical concepts and knowledge..
Stand alone researches are to be discouraged. Collaborative studies are deemed more useful and honest scientifically. What I advocate is forming a group as scientists/researchers and having all sorts of professionals to consult from the planning stage to completion and publishing or presentation at conferences. No one can know ALL aspects of research at expart level, thus do your best and live the rest to others in the group but try and cross-check often for the integrity of the work. You need not be an expert in statistics to publish good papers. Some do the job for a fee and I tell you that it worths paying for the services while you do other meaningful things.
I believe that a deep knowledge of statistics is not necessary provided you know well the basic. If the research involves major statistical studies and interpretations is better you invite a colleague in such area to colaborate with you.
Dear Kayode,
the by far major part of the very relevant advancements in science come from stand alone researches (and stand alone scientists). This does not mean we do not need to collaborate with other people but simply that a scientific theme can (and MUST) be comprehensible by a single person that has the idea, follows his path of experimental evidence and evaluates the evidence by using statistics that is nothing more nothing less than 'quantitative common sense', a scientist with no quantitative common sense is like a soccer player with no legs. It is important we stuggle gainst the tendency to the ever increasing fragmentation of scientific work in which a single scientist only understands a tiny piece of an opera that trascends him and that, at the end of the day, like Babel tower goes down for the simple reason none reminds waht was its aim and geenral project.
Of course we need to have basic ideas of statistics, such as mean, standard deviation, median, IQR, odds ratio, correlation, regression, 95%CI for writing a research article, for any journal and not only an high impact factor reviews, even if you are not running the analysis yourself.
Of course it will be better to have a statistician to advice your study, but his collaboration is not always evident. In practice, at least in my university and institution, it is very easy to have a statistician in your team for a prospective multicenter randomized trial with a national funding, but, for your “basic clinical research” statisticians are not “helpful” since their department has not previously validated your methodology and your data. Furthermore, any of them can be a co-author, even if he or she had validated your “post-mortem” analysis. Thus, for the clinical research of my university medical intensive care team, I disagree with Jochen Wilhelm, collaboration with a statistician is Not always a possible solution. So, it is very important for all the clinicians and the researchers to have basic knowledge in statistics, since, as we done, they will run their analyses for more than 8O% of their publications.
Furthermore, basic concepts in statistics are mandatory to perform a good analysis of the literature. Moreover, Evidence Based Medicine principles (based on statistics) should be used for the examination of the study methods. For instance, in the 90’s, I remember the proposition of a firm to use a endotracheal tube with “continuous aspiration of subglottic secretions in preventing ventilator associated pneumonia”, a new device which may have some benefit according to the Valles et al publication (Ann Intern Med 1995, 122(3) 179-186). Unfortunately for the firm, this new tube was more expansive and furthermore this paper was used for the “how to use an article on therapy or prevention” section of Evidence Based Medicine group of Deborah Cook ( CCM 1997,25: 1502-1513) which demonstrated that benefits may be discussed and harm not eliminated according to biases and 95%CI. Thus, as others, we never used subglottic continuous aspiration endotracheal tubes. So, statistics are required for clinicians and researchers, but they do not need to be an expert of propensity scores, for instance.
Statistical methods when appropriately used will bring in a lot of life to the research that you have performed
As I phrased in my previous response in the thread is the key to understand what your statistical method actually answers to. If you do not chose a method that is perfectly suited to the data it is usually not a big problem. If you, however, chose a method that respond to a completely different research question you are likely to present something that is very wrong and your conclusions could potentially lead to incorrect recommendations. There are e.g. random controlled trials that has shown completely different results than analyses of observational studies and this shows well that the statistical method was not feasible for the observational study even if the study used a common "conventional" method.
Purely as a statistician you are not able to foresee problems that can arise due to confounding and similar effects that are unavoidably involved in an observational study. Overall, you heavily rely on the professional researcher within the research field of the study to give crucial information to make good/high quality analyses for a paper. As a non-statistician it is necessary that you can speak a language that the statistician can understand and interpret correctly.
I do not agree with the description given by some of the responders in the topic about the use of statisticians. I think that it is too easy to just say that you should let the statistician do the job for you. It requires that you either have a statistician that got special competence in the area and can himself/herself identify confounders and other important details or that the non-statistician can well explain these important details so that the statistician understand the context well enough.
To clarify, you do not need a statistician to write a paper but you need someone with an expert knowledge about the possible analysis methods for a given problem. It is not essential that the main author has this competence but it is necessary that there is enough competence in the research team to make sure that the one doing the statistical analyses has enough knowledge to perform a good enough analysis. I think that too often a standard method is used because the researchers believes that it is good enough for their purpose as it has been widely used before in their and others field. More is required of an analysis than just to assume that it works because it has worked in the past. You need to know that it can be used to answer to your specific research question.
Many times, a good work might be not much, when we don't do the proper analysis. Researchers need to learn not only about the application or the development of statistical methods; I think it's more important to make the correct choice of test will allow us to obtain correct conclusions.
It is not necessary to do advanced statistical analysis as the statistical methods are clear and correct. Keep in mind to do the statistical tests needed to make it easy for the reader to understand.
Statistics is the science of collecting, analyzing and making inference from data. Statistics is a particularly useful branch of mathematics that is not only studied theoretically by advanced mathematicians but one that is used by researchers in many fields to organize, analyze, and summarize data. Statistical methods and analyses are often used to communicate research findings and to support hypotheses and give credibility to research methodology and conclusions. It is important for researchers and also consumers of research to understand statistics so that they can be informed, evaluate the credibility and usefulness of information, and make appropriate decisions.
I edit lots of different articles from authors that love chucking in 'statistical analysis' which 'supports' their position .... but sadly, more often than not, it fails to do so. as many have said, it depends very much on your field. I have enough knowledge to use basic SPSS, work out some basic percentage correlations etC, but I also wondered if I perhaps needed some more in-depth statistics, both for my own work and to interpret that of others ... and for me, the answer was no. I took the course, manually calculated 'pee hat' etC and found that although I had gained personal knowledge, it was unlikely to be reflected in my particular work. I do however have an insight as to where more advanced statistical enquiries may be of help, so I can always consult with someone who specializes in statistics when needed.
What I would like to see however, is a more governed view on the inclusion of statistics in journal articles (of all impact factors). Just because I have included some Greek letters, played around with my outliers until my boxes are almost equal etC, it does not always mean I am making a proven point (like posters at a conference - who actually makes sure that 2+2=4, unless they are following through the article in a test setting?). Quite often, we can clearly see that a particular argument is not robust or well supported, despite the fact that the paper is riddled with scatter plots and language from Harry Potter's spell book. Additionally, if someone simply chooses to take a more literary approach to their 'analysis', perhaps it may open up their work and its meaning for those who are not so well versed in statistical techniques .... we all work & interact with a range of colleagues and professionals, but how many of them actually have the skills or inclination to work out a mean confidence interval for whether your local GP has accessed current research before choosing this or that headache pill (& your local GP probably can't do it themselves either) ;-)
My view is that most of us need to read and understand something on the level of 'Statistics for Dummies', and that will let us write our own standard work and understand the routine work of others. Few of us though, need to be self-supporting experts unless our fields specifically call for it ... after all, that is what Statistician's do :-)
Just to back up the need for statistical depth however, check out the debate going on with another question being asked in RG: What is the use of reporting the exact p value in a research paper? Is it necessary?
Nicolas, you made an interesting comment. I also see that authors use plenty formulas and complicated descriptions just to make simple things look more complicated or more "scientific". Unfortunately, many of such authors quickly collect a lot of "followers", citing their work (because it sounds so scientific), referring to these formulas (because they look so scientific), and finally get distracted from more important, fundamental, or influential questions/insights. The main reason for this (I call it secondary) misbehaviour is a severe lack in understanding maths (logic!) and stats.
Nice example with the p-value thing you cited: the maths is clear, but the appplication or the interpretation of the results of the math-formulas depend on the degree to which the underlying assumptions are met. To my knowledge, the assumptions are never met exactly. Now, if the assumptions are not met exactly, what the hell is the resoning in calculating, presenting and interpreting "exact" p-values? The exactness should not exceed the exactness whit which the assumptions are met. But to determine quantitatively, how exactley the assumptions are met is a mission impossible. It can only be based on rough educated guesses.
But I would not abandon the need for some good basic statistical knowledge, for the reasons I discussed above (statistics being the only objective approach to quantify how much we can learn from the data we know).
Hi Jochen - I liked a comment of yours .... 'many things are presented in a complicated way just because a) the teacher wants to be adored for understanding such complicated stuff and b) the teacher did not understand the topic deeply enough to be able to present it in a simple way'
I used a pseudo-equation in a poster once, simply to get peoples attention & they thought it was very 'scientific' (by which means I proved my point ! I guess we have to conform in a way to others expectations (valid or not), or they simply will not read or publish the work & it will be no use to man nor beast. This might especially apply to those who are reasonably 'new' in their fields and perhaps not in a position to mention 'The Emperors New Statistics' ;-)
Antonio, your statement "Once you comprehend the concepts behind statistical tests and methods..." reminds me on Pierre-Simon Laplaces nice words: "Statistcs is nothing else but common sence reduced to calculus". This "common sense"-part is often lost when people think about statistics.
Laplace also put forward the principle of indifference, further stating that "Probability is relative in part to this ignorance, and in part to our knowledge."
I liked very much the Laplace quotation about 'common sense and calculus' and I often repeat (in a myriad of different ways) this concept to my statistics students. But the big crisis of nowadays science starts from a disease science was kept centuries ago but that in the last years (with the exponential fragmentation of science fields) became devastating and that exactly coincides with the elimination of common sense from the horizon of science. Scientists use common sense in many occasions in their personal life: when they decide where to go in vacation with their family, when looking at a menu in the restaurant, when they drive ..but they totally forget it when they think at scientific matters. This is the reason why even if they know very well that a plenty of biological processes arise in terms of between cell communications (then they are posited at the tissue organization level) they continue to look at causes internal to the single cell just because in the books they read at the University all the box and arrows schemes had as boundaries the cell membrane, this is why even if someone at chemistry courses told them tri-atomic encounters in a diffusion regime are so rare to be paractically impossible they continue to imagine biochemical pathways with dozens of ordered steps as happening in a diffusion condition, similarly physicists still think 'string theory' has to do with science even if it cannot be experimentaly proven and so forth...
Thus I came to the conclusion that the poor knowledge of statistics in science (for which a discussion like the one we are participating into can happen, given in a reasonable world it can be equated to a discussion about the relevance to have legs for playing soccer) is strictly related to the common sense crisis in science , thus teaching statistics should be, in my opion, mainly based to an effort to drive students to apply their common sense even when they approach scientific matters.
Perfectly fitting to your statement that "the poor knowledge of statistics in science [...] is strictly related to the common sense crisis in science": when teaching stats for graduate students I regularily find that the students only thinking of data analysis in terms of "how do I get a significant p-value and what test should I use?". It is a very hard job to take them back thinking about their scientific questions. Sometimes it comes out that some of them just don't have a specific question at all, or they miss to ask the important question(s) (and instead formulate questions that are hoped to be answerable by means of the known hypothesis tests). But, to come to their defence, this is also a consequence of (1) a bad teaching habit (frequentistic philosophy, focused on testing) and (2) a bad reviewing habit (same, mainly because reviews are done by scientists also trained this way in statistics). I saw some improvement by sending mauscripts to statisticians to review the data analysis part, but this often fails to adress the coherence of analysis and science (I mean, the analysis is reviewd as being technically ok and adequate for the properties of the data, but often it can't be judged by the reviewer if the analysis is helpful to [best] answer the scientific questions). Exactely here a sufficient statistical literacy of the scientists is required and this can't be effectively substituted by using professional statisticians as reviewers.
In my opinion and limited experience: The purpose of most of statistical tests is to extrapolate the findings in the "study sample" to "general population" or find what is described as "universal truth".
Statistics is a tool to achieve the same as also a part of the yard stick. Hence to justify and properly employ statistical test to advantage, a researcher need to have basic knowledge of statistics, take the help of statistician in the initial stages of designing itself or for performing & tabulating the results concisely. But the extrapolation of the result to research question need to be the domain of the researcher. A well trained statistician will also be able to steer the manuscript, but as the captain, the PI or CoPI need to take up extrapolation responsibility, unless the statistician is the CoPI or PI. Emphasis in research shall rest on "generalizing" the results of the experiment rather than relying on numbers or stressing on p values.
Just like previously mentioned is my reflection from my own masters in Mathematical statistics that the scientific problem easily becomes a parenthesis in the learning process in statistics courses. The consequence from this is that the statistician knows how the methods works in theory but are not able to apply them on real data. The researcher who applies statistics usually are from the other extreme, i.e. knows the application very well but doesn't understand the output from statistical methods. You do not need a bachelor in statistics to be able to apply statistics as practitioner and you do not need a bachelor in the expertise field to understand the problem as a statistician. The problem is that you need to find a way where either you are able to deal with both processes, i.e. the statistical method and the research question, or you are able to in cooperation with others mix the two processes so that the statistical method correctly answers to the research question.
I think that too much focus in this thread has been about the inclusion of a statistician. I think that to believe that purely the inclusion of a statistician will solve things are wrong. I think that it also is common that people misjudge the role of the statistician in solving the problem. If a statistician that is not well enough informed about the problem he/she will not give you the solution. It is the capacity in mixing the knowledge about the research question and the appropriate statistical method that is the key. It is generally enough to have a statistician to support the researcher. Important to remember though is that a statistician without prior knowledge of the research field will have big difficulties in securing that the research question is properly answered. It is therefore crucial that others in the research team got capacity to secure that the research question is dealt with well enough.
When there are confounders that disturbs that is important to consider to get appropriate conclusions from a study it is even more important that the person responsible for the research question (usually first author) can secure that the analysis can be performed in a suitable way, with or without a "professional" statistician.
My response to the question in the start of the thread is therefore that the key author must make sure that the statistical analysis in the paper are of high quality. If there is capacity in the research team that can compensate for the lack of a statistician then you do not need a statistician. It must not be the key author's task to verify that the research question has been dealt with correctly through the use of statistical methods but the main author must be sure that things are correctly performed. I use the word key author instead of first author because the process is often owned by the last author and then it is his/hers responsibility to secure the process.
I think it becomes more importance for clinicians to master an analysis tool - mathematics (statistics). This is because medicine science will be more precise and sophisticated. I am interested in the courses of Epidemology and Statistics...Hope atteched (a simple biostatistics book) can help you...
So, to quote an anonymous source: "A statistician is one who comes to the rescue of figures that cannot lie for themselves."
In the case of a young researcher he can get help from a statistician or he should first learn statistics?
Islande:
My response is very much in line with Antonio's excellent response.
You will, as Antonio mentions, suffer in the understanding of the problem if you lacks methodological skills. The research design are not necessarily something that people relates to statistics but it is at least directly related to the statistical analysis. To make it possible to make a good statistical analysis the research team must be well aware about how to do it and it is reasonable to assume that also a young researcher are heavily involved in this and understands how and why it is performed. If it is your paper and you lack skills in this it will be very hard to draw good conclusions and that must certainly be expected from the first author.
The discussion on how to go from research design to statistical analysis is something that you might be able to have help from e.g. your supervisor. However, in the long run you will struggle to make a career within academics if you do not learn statistics well enough and have to rely on others to do it for you.
To not learn statistics well enough and use someone else as the link between yourself and the statistician is for me comparable with using a language interpreter when you discuss your research with your expert colleague. The risk of failure and misunderstandings is pretty obvious in that case.
To extend a response to your question, as an individual researcher are you not expected to decide on the statistical analysis method but you need to understand what the method does with your data. The statistician is supposed to find the analysis method for you. It is too much to ask from you to be an expert both within your field and in statistics, but without pure knowledge about statistics you will be very limited as it will be difficult to recognize if the statistician has misunderstood the problem, including missed out to take care of confounders, based on the description he/she has been given. The statistician can of course be replaced by someone who has the same expertise level within statistics but has another work description.