The p-value= 0.050 is considered significant or insignificant for confidence interval of 95%. or the result is inconclusive? And what about p-value = 0.053?
Amna -
This is greatly dependent upon the sample size and a sensitivity analysis such as a power analysis. The practice of using 0.05, regardless, is not appropriate.
https://www.researchgate.net/publication/262971440_Practical_Interpretation_of_Hypothesis_Tests_-_letter_to_the_editor_-_TAS
Thanks. - Jim
Article Practical Interpretation of Hypothesis Tests - letter to the...
Just adding to James commens: what you consider significant is your choice. If you say that getting a p of exactly 0.05 is significant, then it is. If you don't, then it isn't. Significance is nothing god-given. It is a decision we (you) make. You may have good reasons to consider a p of 0.7 still significant, another one, for another experiment, may have good reasons to consider a p of 0.001 still as insignificant.
The previous answers are correct.
But to answer what you are asking...
Just grazing the internet, it looks like most sources that pop up say a result is significant if the p-value is less than or equal to the alpha value.
But I've always said a result is significant if the p-value is less than the alpha value.
Maybe I'm just unconventional?
Slavatore, in most cases the p-values are on a continuous scale, and in these cases it is practically irrelevant to distinguish "p
It is a matter of 'who will judge' and what are the overall errors at the field you are searching.
If everything is tight (small variances), then somebody (a 'bad guy') could blame you for p=0.053.
But in general everything is relative...
Agreed with other scholars that statistical significance of p < 0.1, p < 0.05 or p < 0.01 etc is relative. Want to add which threshold of p to set is also depending on which area of the research is being conducted. For examples, generally in social science research, normally we use p < 0.05. For other sciences e.g. drug / medicine testing, normally we might use p < 0.01 i.e. there is a one percent chance (1 out of 100 patients) that the result was accidental - as five percent chance (i.e. 5 out of 100 patients) might not be acceptable in medical science as too many human lives are involved here.
If a research is a social science research whereby we agreed to set p < 0.050, then when p = 0.053 it is considered not significant as p > 0.050.
Han, you seem to confuse alpha with the "chance to find an 'accidental' result". But this is not the same. The chance that a ('positive') result is 'accidental' depends on the prevalence of 'positive results'. This is easier to understand by example:
Consider a diagnosic test. The test works just like a hypothesis test and has a known specificity and sensitivity. From these properties one can derive the probability of the test to give a 'positive' result when the patient is actually healthy (a false-positive result). This probability is denoted by "alpha" in hypothesis tests. A good test surely has a low "alpha". For our example, let's say alpha is 1%. This means: if 1000 healthy (!) persons are tested, we expect about 10 tests turning out to be (false) positive and 990 (correct) negative.
Given this scenario, what is the probability of the 10 persons with a positive test result to have the disease? - It is zero (actually a frequentist would say that this is undefined because the probability applies only to repeatable processes and not to fixed cases) - we know that these persons are healthy...
Now consider we test 1000 persons from which we know that they have the disease. The test will give us, say, 995 correct positive results and 5 false nagative results. What is the probability of the 995 persons with positive test results having the disease? - it is 100% - we know that these persons do have the disease.
Now in we simply use the test on some person. The test is positive. What is the chance that this person has the disease? - This depends on the prevalence of the disease. If the prevalence is tiny, the chance that this person has the disease is still very small. If the prevalence is moderate, the chance that this person has the disease is large. When we do not know the prevalence we can not say what the chance is that the tested person ihas the disease.
This is a reason why screenings for rare diseases have to be taken with care. The prevalence in the tested population is very low, and so the chance that a screend person with a positive test result does really have this disease is still very low. But the positive result can considerably worry the patient and have bad side effects. So what you do with this screening is: you can find some very few diseased people and help them - but at the cost of unneccesarily worrying many more people (and this can easily be very relevant - think of a (false!) diagnosis of HIV, Alzheimer or a brain tumor). In real life such diagnoses will have to be confirmed by a second test, but the time between the first diagnosis and the second test can be very hard challenge for the poor patient!
Thanks Jochen for the correction, agreed the chance that the result was accidental as stated in my post is dependent on prevalence of positive results & not patients.
P-value should be less that 5% for a result significant at 5% level of significance. In alternative hypothesis there is no equality.
@Jochen : I would be less stringent about the use of alpha and significance associated with p-values. In the H0-only approach, you still have the result that p-value (p) is uniformly distributed in [0, 1] if H0 is true, hence if you declare H0 false if p < alpha, the probability of rejecting H0 when it is true is alpha, and it is the Type I error by definition.
The problem is more on the lack of any precise H1, that prevent for any power analysis, hence we have no idea of the Type II error: taking decision when "rejecting H0" is OK, but when not-rejecting is not.
So you can use the p < 0.05 (or any other cut-off) to proove that an hypothesis is (said to be) false with no problem; of course the problem is that that non-conclusive results (p >= 0.05) are often interpretated as « conclusive : H0 is true » which is not at all possible without the complete framework of H0 vs H1 as you presented.
And additional problem is that when prooving that H0 is false, you don't know how it is false (for instance, in Student's test, is it because of a difference in mean, or because of different variances, or because of non-Gaussian data or...), and that usually the fact it is false is not really relevant (do I really expect that in two different conditions, a given gene has exactly the same expression level?)
"The problem is more on the lack of any precise H1, that prevent for any power analysis, hence we have no idea of the Type II error:"
I agree. But additionally we also do not have any clue about a reasonable Type I error we should select. The "conventional" value of 0.05 is surely a bad choice in 99% of the cases. Even considering such a "fixed value" or "precise cut-off" is a source of a lot of fuzz in papers.
"taking decision when "rejecting H0" is OK, but when not-rejecting is not."
I am more with Neyman here: "not rejecting H0" is a factual decision. I (or others) behave as if H0 was true, I (or others) do not engage another Ph.D. student on that topic, I (or others) turn to other topics, hypotheses, theories, models ... so I (or others) do something, and this action is related to (founded or justified by) the fact that I was not able to reject H0. This decision has consequences and these consequences should be valued in order to "not reject H0" at some sensible, resonable level of significance. But this requires the formulation of a substantive alternative.
I do not confuse "non-rejection" with "H0 is deemed 'true'". I am very clear that "non-rejection" only means the the data is "not conclusive". And this "non-conclusiveness" is the crucial point here! When is comes to the judgement of "conclusiveness" and pre-defined cut-off for p makes no sense. Conclusiveness of the data/results can only be judged in relation to the experimental setup, the sample size, the particular methodological difficulties, and underlying model (explanations, resonability, etc.). It is NOT ABOUT MAKING A WRONG DECISION, with any well-controlled rate - it is about judging the conclusiveness of a corps of data relative to a model. This is nothing where a p-value is very helpful at all, it may be used as some strange lifebelt in situations where there is no other grip on the interpretation.
This is a completely different way and not a decision problem with some underlying decision strategy at all. It is based on a completely different philosophy about inference ("learing from data" in contrast to "inductive behaviour"). The mess was started by the unfortunate illogical mixture of hypothesis tests and significance tests (to NHST).
@ Jochen : I agree with some of your arguments, but not with all. I think abuse of p-values is real, and of the p
Emmanuel,
I am not sure if I found the point where we disagree.
A "conventional cut-off" is nonsense in either regime (hsignificance tests and hypothesis tests). The decision-theoretic approach just makes the formulation or the statement of the cost-function explicit. In cases where this is achievable this is a nice thing to have. The "decisions" made with significance testing uses such cost-functions implicitely. A decision without a cost-function cannot be resonable.
In research it is next to impossible to state sensible cost-functions. This is always a open to discussion and a matter of opinion. By sticking to some "conventional cut-off" we simply try to circumvent this. A strategy that, to my opinion, is quite disadvantagous for science.
Yes, we must start being smart (instead of pointing on p-values and "significance"). This is harder work, requires more understanding, and will involve the exchange or arguments and the formulation and modification of opinions. Right now we mostly "let computers decide for us", there is not much we do in this process. We should change this and take science back in our responsibility.
I still don't see where we disagree ;)
if p value = 0,05 it means significant. but if p value=0,053 is insignificant, bacuse it more than alpha(=0,05)
@Jochen : well, may be we don't really disagree and I just misunderstood something... I though you were against p-values, and not only against the blind use of « p-value < 0.05 » as a magical formula. I personally see this rule, in most setting, just as a hint among other to interpret experimental results, and help presenting the results (I have no problem adding a magical « p < 0,05 » if it helps publication or makes co-workers comfortable, if it's not the only argument to discuss the results).
I am generally against the habit of painting a black-and-white picture, against the habit of dinstinguishing "significant" from "non-significant". Statistical significance (i.e. "p-values") can be a useful tool to rank results. I am generellay against formutaion (and answering) questions of the form "is there an effect?".
I would love to see the careful painting and interpretation of grayscale pictures, of using statistical significance as a gradual, continuous measure (regarding a particular set of data, not a null hypothesis), and that the questions being asked (and answered) are of the form: "how strong is the effect?", or even better: "given the data (and the model, and possiply any othe prior knowledge): what is the best we can say about the effect?".
PS: you will find many papers I co-authored containing the "p
@Jochen : well, may be we don't really disagree and I just misunderstood something... I though you were against p-values, and not only against the blind use of « p-value < 0.05 » as a magical formula. I personally see this rule, in most setting, just as a hint among other to interpret experimental results, and help presenting the results (I have no problem adding a magical « p < 0,05 » if it helps publication or makes co-workers comfortable, if it's not the only argument to discuss the results).
If p-value is exactly equal to 0.05, is that significant or insignificant?. Available from: https://www.researchgate.net/post/If_p-value_is_exactly_equal_to_005_is_that_significant_or_insignificant [accessed May 15, 2017].
@ Atif: why do copy my 2 years old reply to Jochen without additional comment?
I think this question was raised to know about p=0.050 is considered as significant or not. If we are considering 95% confidence interval, beyond this 95% region is significance, i.e. there is 5%. Hence P
It depends upon you, if you choose to test with an alpha level of 1 it will be significant but if you are testing at an alpha value of 0.05 it is better to consider it not significant because at that point it is really a thin line. And if you are interested to explore more about the P-value and its misuse check this link https://www.nature.com/news/statisticians-issue-warning-over-misuse-of-p-values-1.19503
Concerning the question of p precisely =0.05, if p alpha(0.05), you cannot reject H0.
New people to this question should read earlier responses.
"Significance" is a misnomer.
In my table results there was a 0,50 sig. value and I was really search "0,50 significant or not" then i come across with this question topic. I thank all contributors. I take the value as significant.
From a Fisherian point of view it is neither, and honestly a p-value equal to 0.05 is not that interesting: you should get more data and repeat the experiment (as per Fisher's recommendation).
From a N-P point of view, it is really up to you - depends on where you set the bar (i.e. your type I error), but I doubt you would be interested in a p-value equal to 0.53. Again from a Fisherian point of view, you might find that result of any interest and repeat the experiment.
Please use proper terminology. There is no "insignificant" finding in regards to statistics. If the findings are null then it is non-significant.
Jon, "proper" according to whom?!
That is simply not true and you answer is not really of any help to the discussion
I am personally more accustomed to significant VS. non-significant, but there is plenty of assessed literature that uses the dichotomy significant VS insignificant, especially in econometrics.
Statistical significance is not a terminology set in stone:
"significance level" sometimes is used a synonym of p-value in statistical literature. The dichotomy significant VS non-significant comes from merging together the approach of Fisher on statistical testing and the mathematical theory of Neyman-Pearsond deriving from their lemma.
Hello everyone
Please how could I interpret the t-test sample.
Intact family sample size is 360 while non intact family size is 71. from my Group statistics I got Intact Mean=72.15and Non intact = 37.66, Std. Dev. =15.71 & 12.39. my Sig=.005, Sig 2(tails)=.000, t= 17.214 & 19.848
Thanks
Please, be more conscious that the 0.05 exercise is merely nonsensical.
Start being more analytical.
This mixes two different visions of statistical inference (Fisher vs NP).
YOU as a researcher should decide if your results are relevant, there is no such thing as a dichotomy significant/non significant.
Article Mindless Statistics
Amna Khan please be aware that this dichotomy is a false one, except in quality control
https://www.tandfonline.com/doi/full/10.1080/00031305.2019.1583913
https://www.nature.com/articles/d41586-019-00857-9
I know the original question is old but with some new replies, I thought I would make a brief comment. While .05 may be statistically significant, consider a program like SPSS may not be providing decimals beyond 2. Therefore, .05 on a table may actually be .050000000001, which is not statistically significant.
Daniel G. J. Kuchinka although I understand your point, this general indication is only useful in repeated experimentation, e.g. in quality control, where 0.051 and 0.049 have different outcomes if the type I error is set to 0.05, i.e. in a Neyman-Person perspective.
In applied research, especially if we are not talking about data collected after a controlled experiment, one has only one single set of data at hand and 0.049 or 0.051 provide the same amount of evidence against the null hypothesis, also a pvalue around 0.05 does not indicate a strong evidence against the null hypothesis.
http://www.dcscience.net/Sellke-Bayarri-Berger-calibration-of-P-2001.pdf
Pvalues should not be declared significant because they exceed a threshold that has arbitrarily set after the collection of data, but inductive inference should be left to the researcher avoiding methods that automatic and eliminate any sort of judgment.
Results should never be indicated as “significant” or “not significant
Article Statistical Rituals: The Replication Delusion and How We Got There
@ Mojtaba: if p-value = 0,05 exactly, then the 95 % confidence interval will have the critical value (typically 0) just on its border; so the problem is exactly the same.
Emmanuel Curis
when P-value= 0.05, regularly the 95 % confidence interval will not have the critical value ( typically 0) exactly on its border, (with 2 or 3 rounded it is exactly on borders.)
@ Mojtaba : it *does* have. Confidence intervals and hypothesis test are equivalent, if made using the same assumptions. There are only two possibilities for not being the case, and both of them are related to other issues:
1) either p-value or confidence interval bound were rounded. In that case, either the p-value is not exactly 0.05 or the confidence interval not exactly 95 % (or both), so the comparison is unfair. And there is no reason to trust more the confidence interval to conclude than the p-value.
2) or p-value and confidence interval were computed using different assumptions (like, for instance, p-value is using a Wald test and confidence interval by profiling/using a score test/...). In that case, they cannot be compared...
Emmanuel Curis
You are true.
But regularly we employ some software's ( such as SPSS, SAS ,...) and so we have only 3 digits of P-value. But in CI we have bigger than 3 digits and then we can have a decision.
It means that under H0 there is a 5% probability of the Wilk-statistic (W) being grater than the one of your sample. Note that the p-value is exact only for n=3 (as only in this case the sampling distribution of W is known). For other sample sizes, monte-carlo methods or approximations are used. So it's rather silly to be concerned with with question whether or not p is exactly 0.05.
Technically p=0.053 is not significant. But for your analysis of results, you may take p=0.05 - 0.08 as a TREND and mark those, further you can test for those to your objective parameters by correlation and/regression. In that case, you may find some significant or high R2 value for positive confirmation as 'p' is near to the significance level.
Nope Rajib. I see that often, but that's nonsense. As soon as you say that you do see a "trend", you consider the data significant, because you do interpret the direction of the effect as estimated from your sample. This is what the term "significant" means - that you believe your data is sufficient to interpret a trend (up vs. down, positive vs. negative). Saying that a result is not significant but at the same time interpret the trend (or sign, or direction - call it like you want) is a contradiction in itself.
The ritual of looking at the p-values should just STOP us from interpreting trends where the data is not sufficient to confidently do so. It makes no difference if you call your data significant because the data looks "trendy" or because the p-value is small. But because we tend to see trends everywhere (especially if we worked hard on the data and desperately need to publish great stuff), it's saver to look at p-values, and to take the sign of the sample estimate (i.e. a "trend" in a particular direction) serious only of p is very small (there still one can cheat, but that's a different topic).
@Dr Debojyoti Moulick , that would amount to bias. At that stage, you don't force your outcome to be fine-tuned to your personal decision or alter it. Generally and technically speaking, 0.05 is insignificant, however here we are again racking brain on what seems to die a natural death; the use of p value in judging an outcome of event should riddle off the terrain in scientific reporting using statistical evidence to make a decision.
Sir, please have a look it once again I put my opinion with a condition, less than or equal to 0.05!
Now in the question p value is equal to 0.053. If the last digit is either 5 or more than 5. Then it would be of insignificant type.
Any value of p which is 0.05 or less is significant. Clearly since 0.053>0.05, in your case it is not significant.
That depends on your null hypothesis, when the test include hypothesis at P≤0.05 and all assumptions of the test are available, then 0.05 represent significant effect, that mean 0.05 and less included in the significant effect.
It depends on your alpha value! If you've set your alpha value to the standard 0.05, then 0.053 is not significant (as any value equal to or above 0.051 is greater than alpha and thus not significant). If your alpha value is a less conservation 0.1, then a p value of 0.053 would be considered significant. In general though, a p value of 0.05 or less is considered significant.
It seem surprising to me that there is not a clear consensus on what is an arbitrary criterion that statisticians and scientists have established and so often use. In my lab, we treat an exact value of p = .05 as significant and interpret the criterion as p < or = to .05. If SPSS has by default given us .05, rounding from three places, I sometimes check out to 4 places to see how it has rounded and then decide accordingly.
Dear D. Alan Bensley , there is no consensus because this procedure is not supported by statisticians whatsoever, because it is very ritualistic and makes little sense
Article Mindless Statistics
"No scientific worker has a fixed level of significance at which from year to year, and in all circumstances, he rejects hypotheses; he rather gives his mind to each particular case in the light of his evidence and his ideas. (Ronald A. Fisher, 1956, p. 42)"
0.053 is still not statistically significant, and data analysts should not try to pretend otherwise. A p-value is not a negotiation: if p = 0.05, the results of p = 0.53 are not significant. Period.
In China, you would a firing squad for allowing it to be significant (just to show how serious it is).
Null Hypothesis Statistical Testing, a mashup of Fisher with Neyman-Pearson approaches, leads to these conundrums, in my interpretation of Perezgonzalez (2015). Article Fisher, Neyman-Pearson or NHST? A Tutorial for Teaching Data Testing
It seems to me that we tend to approach our own work from the perspective of Fisher (exploratory, simplistically speaking), yet expect others to approach theirs from the perspective of Neyman-Pearson (confirmatory, simplistically speaking).
So, for me, it depends on what you are genuinely doing in the research project: Is this new ground? Then you are in exploratory mode (Fisher): Don't get hung up on if the p-value crosses a magical barrier, use phrases such as 'marginally significant', 'significant', 'highly significant', etc. to describe the degree of evidence.
If you are confirming an effect (Neyman-Pearson) in a new area or extending findings, then you need to do many more things: Get a high power (80%), declare an alpha and stick to it (keep in mind how it affects power). Then, if you get p=.051, you can't reject. Your evidence for confirmation is insufficient.
Ugh... all that work to get no career advancement... which leads to the other problem, retention and promotion is so tightly tied to publication, which is tightly tied to p
I agree to your practice, Paul :)
Nothing to fight you, just a comment: If you really follow Neyman/Pearson, it's not about "rejecting H0". It's about accepting one of two substantially different alterantives (A vs. B). It's only in the math that one of these two hypotheses is treated just the way H0 is treated in Fisher's procedure. But the philosophy and interpretation is entirely different. The confidence levels to accept alternative A (=1-β) or B (=1-ɑ) are chosen based on a loss or cost function. The statistical rule ensures the minimum expected loss or cost. It's not about thinking that one of these hypotheses is true and the other is false. So it's even worse than you say, that NP-testing can be used only for confirmatory studies: it can only be used rationally when you can state a loss function! Without having a cleary stated, sensible loss function, the decision rule can be arbitrarily unreasonable. I find it perplexing that here so often the advice is given that empirical scientists should behave unreasonable, or at least irgnore the degree of reasonability (or the lack thereof) of their procedures.
The paper you cited says that NP-testing ia "acceptance testing", so better not use the word "rejection" when talking about NP-tests. However, the authors also fail to acknowledge the fact that NP-testing is a reasonable procedure only when there is a reasonable cost function associated with the chosen hypotheses (B-A may not be the "expected effect size" (page 4, after "Step I") - it should be an effect that would call for a particular action, if it was the case - an action that comes at a quantifiable cost under A and quantifiable benefit under B). Apart from that it's a good read, I think.
Hello dear colleague
thank you for question , p-value until 0.05 is significant and p-value = 0.05 is significant too.
When the p-value of a relationship exceeded 0.05 (for the confidence interval level of 95%), then the relationship is not significant. In the contrary, the p-value of less than 0.05 (for the confidence interval level of 95%) suggest that the relationship between variables is significant.
Of course, the cut-off level for the p-value will change or alter depends on which level of confidence interval that you're applying for your research at the moment (90%; 95%; 99%)
Hope it could help you in solving your problem
Dear Colleague
hello
How for confidence interval of 90% or 99%?
the p-value is significant?
Another alternative may be to evaluate the minimum effects, equivalence, inferiority test based on the results presented by the results using two-tailed tests (TOST). This allows to determine if an observed effect is really small since there is a true effect greater than the minimum effect (SESOI). The free jamovi program has the TOSTER option that allows you to evaluate this analysis in a simple way.
I recommend reading:
Equivalence Testing for Psychological Research: A Tutorial
Daniel Lakens, Anne M. Scheel, & Peder M. Isager
Two points are essential:
1- In Manuscript, the value of "the significance threshold" should be specified.
2- It is wrong to decide that 0.05 is significant or not automatically. Besides, effect size and power of association should be reported. For example, the t value, the F value, correlation coefficients, unstandardized and standardized regression coefficients, eta square value should be reported.
The sample size is very important for p-value decision.
Bonferroni correction can be used to compare p-values that show group differences in case of a large number of tests.
Otherwise, .050 will be considered significant and .051 will be considered insignificant.
Best.
Mithat
0.05 and less than 0.0.5 is significant other all are insignificant
Thanks to everyone who has contributed to this thread. Yes, there are limitations to significance testing and setting a criterion, such as .05 that corresponds to a critical value to reach significance. The issue seems to be a problem, in part, of the precision of the test statistic value that is measured. However, common practice seems to be to call a value corresponding to .05 as significant.
When p=< 0.05 then there is a significant , so reject the null hypothesis
Amna -
In my first response, the first one in this thread on Oct 27, 2015, I noted that "The practice of using 0.05, regardless, is not appropriate." Even if you happen to have decided that 0.05 was a good threshold to use, in a particular case, considering sample size, and, say effect size, it would have just been an approximation. So the whole argument of the question is rendered arbitrary.
This brings up my objection to the use of a cut-off point for a somewhat arbitrary decision. One should consider the evidence, and estimation is often more useful.
Consider this example:
One can test for the presence of heteroscedasticity in the estimated residuals for a regression. There are hypothesis tests to say Do we or don't we have heteroscedasticity? But this will be a matter of degree, which is somewhat vague for the user. You really want to know if you can assume homoscedasticity, because that may be easier. But if instead you estimate the coefficient of heteroscedasticity, you can see the results, and judge on the basis of each application, with helpful information, not an automatic response that relieves one of responsibility, but leaves you with the product of a vague, somewhat arbitrary decision.
I'd say this question illustrates the flaw in using a threshold to decide everything when one does not know the consequences. Hypothesis tests are often overused and misused, and an isolated p-value and threshold is not very meaningful. This question shows that to be a concern. Don't just stop with a p-value.
Cheers - Jim
For interpretation of results in research, we would need to specify a degree of accuracy( the commonest being 0.05). Any value equal or less than this is regarded as statistically significant.
The p-value is the probability of an observed difference having occurred by chance.
p-value of 0.05 means the probability of the observed difference having occurred by chance is 5%
p-value of 0.01 means the probability of the observed difference having occurred by chance is 1%
It means it is unlikely to have occurred by chance, therefore statistically significant.
p-value of 0.001 means the probability of the observed difference having occurred by chance is 0.1%(highly significant)
However, the p-value is always reported with confidence interval.
Dear sir,
Your observed p value (0.053) is great than 0.05, in this situation the results are statistically non-significant.
Dear colleague
the p-value =0.05 consider significant but p-value =0.053, statistically non significant.
Dear Ghahraman Mahmoudi
I have p-value = 0.051, is it also statistically non-significant?
Kind Regards
Irfan Ullah : 1) Since you do not mention your Type I error, it is not possible to answer.
2) If you assume the usual Type I error of 5%, technically and strictly speaking, yes
3) Please take the time to read all the previous discussion to understand how these questions are in fact not very relevant, but the call for a more in-depth understanding of your usage of statistic, and not using a blind rule.
Dear Colleague
the p-value =.05 is significant but p-value = .053 is statistically non significant.
Ghahraman Mahmoudi -
See #3 in Emmanuel Curis's response just above.
Using a threshold of 0.05 everywhere is not a good idea, as you may see from above. Besides standard deviation, consider the extremes of small sample applications versus "Big Data."
In addition, making a decision based on a threshold, even if it weren't arbitrary, does not excuse one from including other considerations.
Jim
hello dear colleague
in base on statistical references the p-value =0.053 was considered in significant .
Based on your hypothesis. Normally, less or equal 0.05 will be considered significant.
P-value of equal or less than 0.054 is considered significant. 0.05 is considered non-significant because it can be approximated to 0.06. However, scientific conclusion or policy decisions should not be based solely on achieving a specific p-value.
@ Murumba : your answer is non-sensical, 0.054 is not significant, and 0.05 can not be approximated to 0.06, especially in this context.
In my view, p = to or < .05 is statistically significant. I treat it like the criterion that it is; and, therefore, if p = .051, I do not consider it significant. I do not round. I agree that there are other kinds of information about your data that are useful to provide, such as effect sizes and confidence intervals.
In all of this discussion, I think it's important to realise that p values such as .05, .01, and .001 were set arbitrarily (I forget by whom - was it Ronald Fisher about 100 years ago?) and have, since then, acquired a status that is probably unjustified - particularly when devotees of null-hypothesis statistical testing (NHST) rule the roost and argue until they're blue in the face concerning issues such as whether p = .055 is statistically significant or not.
When I was teaching stats, I used to enjoy us obtaining a p value like that in class analyses and asking the students to decide what to do. The debates were often quite hot, with ethics and all kinds of issues being introduced.
Indeed, a lot of us are accustomed to seeing p values and relying on them for providing some sense of the importance of results, but many journals, and organisations such as the American Psychological Society, now downplay p values and promote such things as effect sizes and confidence intervals instead. I think it's a good idea.
For me it is. There is only 5% probabilty that your result is due to chance, and 95% due to what you tested.
less than 0.05 consider significant (example 0.049), but 0.05 or more ( 0.051) are non significant.
regard
Mihai Cosmin Sandulescu : nope, your interpretation of the 5% is false. The p-value is the probability of the data *assuming H0 is TRUE*, it has nothing to do with the probability that "the data are due to chance", whatever that could be.
Mihai Cosmin Sandulescu: I agree with Emmanuel Curis. In classical statistical inference, the data and results are always obtained “by chance” derived from measurement uncertainty, random samples, random assignment, etcetera. The goodness of applying well-designed inference procedures is that if there is not a real effect, then there are “significant” (false positive) results in a proportion less than or equal to alpha. The presence of real effects increases the probability of observing “significant” (true positive) results, but a significant result doesn't necessarily correspond to a real effect.
This is the price of incomplete theories.
"Significant" obeys a definition: p-value less than or equal to alpha. In my opinion, alpha is not a probability but a threshold to control the probability of making the type I error. In Neyman-Pearson strategies, once established, we should not change its value even in cases of close proximities.
hello Dear colleague and thanks for your question, the p-value= .05 or .53
from statistical view not significant and should be little than .05.
I fully agree with Jochen,Generalising the result based only on p value has lot of limitations ,the main one being the prevalence rate and the total sample size.But, I can say more than 90% students and researchers do not bother about the importance of prevalence rate in interpreting the statistical significance. More over,5% and 95% are all arbitrarily recommended and being used by almost everybody since it was developed.Yes,the researchers have to be very careful in interpreting the results and if not,it could give a result which can harm the research and the subsequent reality
Ghahraman Mahmoudi
"from statistical view not significant and should be little than 0.05" is not a good recommendation. Excluding 0.05 as a rejection p-value may be irrelevant for continuous distributions, but if you work with discrete distributions, including 0.05 to reject the null hypothesis is very important to be consistent with the significance level. The general decision rule is
"reject H (null) if the p-value is less than or equal to the significance level (alpha)".
Jorge Ortiz Pinilla, I join you in wanting to comment on the post a couple above here - except that I couldn't understand it, so I didn't bother.
However, may I check with you please: In my experience, the criterion for rejecting the null hypothesis has been < .05, not equal to or less than .05. I come from a background in psychology and the health sciences, so might you come from a background in other disciplines?
That aside, I think that an obsessiveness with p values is misguided anyway . . .
Dear Robert Trevethan, following authors such as Lehman (Testing Statistical Hypotheses, Wiley, 1959, p. 61), Bickel & Doksum (Mathematical Statistics, Holden-Day, 1977, p. 171), Lehman (Nonparametrics: Statistical Methods Based on Ranks, Holden-Day, 1975), Navidi (Estadística para Ingenieros y Científicos, McGraw Hill, 2006, p. 443), Conover (Practical Nonparametric Statistics, Wiley, 1971, p. 81), and many other recognized statisticians, the p-value is the smallest significance level at which the experimenter using a statistic would reject Ho based on the observed outcome. Then p-value = alpha should reject Ho. It doesn't depend on any specific discipline. If we work with discrete distributions, not including equality may lead us to wrong decisions. I can illustrate this with some examples, but it will take a little more time. If you want, I can do it in a later answer.
Dear colleague ,hank you for question ,
I opinion the p-value = .05 or .053 , the statistical view is insignificant .