nine is okay - you can do a meta-analysis of two studies but obviously that is just going to have some bias but nine is good.

2

Yes two studies will suffice.  I've attached a publication that will be of help to you.

Thank you very much for all your help 

Nine is reasonably good. So keep going!

Dear Usama,

I agree with other scientists, minimal number is 2. You get 9 and it is very good if they belong with a systematic search.

Actually you could meta-analyze even one single study if for some reasons it has multiple results of the outcome of your interest.

The thing you want to be sure is (1) that you have performed an appropriate systematic review before and (2) Heterogenity: Does it make sense to combine the studies together? Are they different? How much is this difference important? Could the analysis adjust for this?

Hope this can help

3 - 10 studies will avoid biases. But it should not be prescriptive.  Yes you can have 2. 

nine studies are fine, but the samples you included should be not low. If you have enough samples, two studies are also feasible.

I found a meta-analysis which included only four studies...

I have come across a study with 5 studies but it is always better to have a lot more.

Here there is no role of thumb: you should only be careful to avoid introducing selection bias in to your literature search strategy. Every study that meet selection criteria have to be analised. More study, More studies equals to shorter confidence intervals and higher statistical power, as usual.

I suspect that someone may have developed a statistic regarding sufficiency of the selected studies for a meta-analysis.  You might also check the guidelines for meta-analysis, PRISMA.  As another person indicated, the sample size of the original samples will be one parameter that determines whether the aggregate studies are sufficient.  A second parameter will be the studies themselves: are they sufficiently representative of the hypothesis being tested.  A third parameter will be the factors of interest that are extracted from the studies, and the resultant cell sizes.  Finally, means and SDs will give you information regarding sufficiency.

Assuming your meta-analysis is carried out in a context of a comprehensive systematic review of the literature, you can combine even two studies. Sample sizes also have an impact - two large randomized controlled trials can generate high quality data and clinically useful information. In the Cochrane Library you'll find plenty of SR with meta-analysis of few, small RCTs. In my opinion those are meaningless. 

Hi

Theoretically, you do not have any restriction for number of studies in your meta-analysis. we have some meta-analysis with just two studies. However , more studies means that your meta-analysis have more power and is more exact and reliable. if your meta analysis have about five and more studies your result may consider as reliable result.

you can use the REVIEW MANAGER 5.3 soft ware from Cochrane Library .

Good Luck

There is no rule about the number of studies, but it is important to select the appropriate literature

Dear Ramiro López-Pulles ,

Would you specify? What should be appropriate? Method, sample size, ...  so on and so forth.... ?

As with the majority of the comments here, nine studies is acceptable for meta-analysis. I would think that the overall sample size is an important consideration too. I too endorse getting familiar with PRISMA and MOOSE guidelines as well as PROSPERO as you embark on conducting a meta-analysis. If you have a resident statistician in your department, employ their assistance. If you are concerned about your selection criteria, trying being less stringent when searching through PubMed/Scopus etc but then screen more abstracts. Good luck.

I have seen meta-analysis with 3 primary studies. I guess 9 is more than enough.

I have just copy-pasted a paragraph from the Book "Advances in Meta-Analysis" written by Pigott, Terri.

LINK: http://www.springer.com/gp/book/9781461422778

Another common question is: How many studies do I need to conduct a meta-analysis? Though my colleagues and I have often answered “two” (Valentine et al. 2010), the more complete answer lies in understanding the power of the statistical tests in meta-analysis. I take the approach in this book that power of tests in meta-analysis like power of any statistical test needs to be computed a priori, using assumptions about the size of an important effect in a given context, and the typical sample sizes used in a given field. Again, deep substantive knowledge of a research literature is critical for a reviewer in order to make reasonable assumptions about parameters needed for power.

Good Luck ....

It should fine with nine studies as you will be examining the best available evidence concerning a specific issue

All the best

Ma'en

Nine is fine but remember rubbish in equals rubbish out when analyzinganalyzing

I believe there is no rule of thumb on this topic. I believe there are two major arguments here. First, those who defend a minimum number rely their arguments on the statistical point of view where a low number of observations may lead to inaccurate estimates. Therefore, the minimum number of studies may be debatable and depends on the model you are using. But accuracy estimates also depends on the number of subjects in each study and the variance in and between the study, as commented above in other answers. The answer is not straightforward. The second argument is in the opposite way. Meta-analysis main purpose is to increase precision of a conclusion which the single studies were not able to due to imprecision. Therefore results with two studies are more interesting than with one and so on. The conclusion of a meta-analysis with three studies which results are not that precise or not interpretable is that more studies are needed in this field to come to a definite conclusion. I have seen systematic reviews with zero studies, and a conclusion like this. 

More important than the final number of studies is the initial number of studies that your search strategy enabled us to identify .

If you are investigating a new topic and found few studies, it may not be justifiable you conduct a meta- analytic treatment.

In addition , we must evaluate whether the small number of studies found is due to the novelty of the topic or whether it is due to lack of interest of the scientific community to investigate this issue .

Best Regards

Poton

I agree with Dr Pedro Avarenga. Sometimes, even with a small number of trials, a meta-analysis will be providing the best evidence at that time and must be updated as new trials become available.

Keep in mind: 

When you are performing a Random Effects Meta-analysis, your software take into account the within studies variance plus the estimate of the between studies variance. 

 τ2 (Tau-square) is the between studies variance and, if the number of studies is small, then the estimate of the τ2 will have poor precision. 

In other words, during a random effects meta analysis the weight you give to each study is (generally) still equal to the inverse of the variance, just as in the fixed-effect model.

But instead, in the random-effect model, the variance has two components.

One is the within study variance. One is the between study variance.

So that's the difference. Statically, the only difference between fixed and random effect model is how you're going to weight each study.

The weight equals the inverse of the variance, but the variance is modified by the between study variance. So, we use the tau squared to modify the weights used to calculate the summary estimate. That's the difference between the random effects model and the fixed effect model when you're doing it.

I know that you didn't ask about the difference between fixed or random effect models, but the decrease in the precision of the Tau-square was one of the "problems" that came to my mind when you asked about the number of trials you should include in a meta-analysis. Nonetheless, this fact  must not prohibit you to do your meta-analysis with few studies. 

I have F value, df, sum of squares, mean squares but neither sample size of control group and experimental group nor mean for control and experimental group. Could you please advice me how to insert data in comprehensive meta-analysis?

yes, there should be a minimum of 2 studies to perform a meta-analysis.

When the coefficient of variation for data in a study  is less than 15%. why we exclude the study ?

I mean, Is there any criteria for coefficient of variation during data collection?

According to Sterne et al. (2011),  the number of studies required to test metabias by funnel plot is ten or larger (argument k.min=10). 

Source: Sterne, JAC et al. (2011), Recommendations for Examining and Interpreting Funnel Plot Asymmetry in Meta-Analyses of Randomised Controlled Trials. BMJ (Clinical research ed.), 343, 1, doi: 10.1136/bmj.d4002

It depends on what precision we are looking for I think with at least two studies we can do meta analysis.

Simulation studies suggest very poor error estimation when less than 5 studies - see SAS simulation file at the link below

https://www.researchgate.net/post/Is_the_random_effects_estimator_worse_than_the_fixed_effect_estimator_under_heterogeneity

At least 2 or more RCTs are needed for meta-analysis

At leat two studies should be there for meta-analysis. But, the studies should be alike. Never pool two different studies i.e apple and oranges...

Question:

Is there any minimum number for studies that should be included in meta-analysis, when I want to confirm relationships beetween constructs?

Minimum two article but new studies say that we have to calculate the power of meta analysis. The power can be calculated with number of observations.

When I do not have ECR, my area is not health, but Engineering / Applied Social Sciences. I not consider the fixed-effect models nor random. In addition to the meta regression, which other statistical models can be used?

Two, see intro into Geoff Cummings

"Understanding The New Statistics: Effect Sizes, Confidence Intervals, and Meta-Analysis"

also see "meta essentials" package on web, it will save you a lot of efforts

Elaine you can use IVhet and Quality effects models. See MetaXL at www.epigear.com

I have seen systematic review studies published that have only included two studies in their metanalysis

Even 2 good publications can make a meta-analysis.

Theoretically, only 2 studies are needed for a meta-analysis, but it may be hard to get published.

There's a balance between the novelty and the generality of your findings. If you're doing something very very specific, it may not be applicable to a wide group of readers.

More technically, if you want to do have further analysis such as detecting for publication bias, at least ten studies are needed as recommended by Cochrane handbook.

Ref:

Cochrane Collaboration. Cochrane handbook for systematic reviews

of interventions, version 5.1.0. London, UK: Cochrane Collaboration;

2011.

Hi All,

The first steps of your systematic review are to outline your specific aim and refine your search strategy. We tell our students that, in many cases, the systematic review is used to identify gaps in the literature. As such, in the event that not much has been published on a given topic, your systematic review could be very valuable in identifying potential directions for future research.

On the other hand, I believe that failing to include all relevant publications on a given topic could potentially bias your results. Traditionally we recommend an electronic database search to identify potentially relevant articles. The references below provide a little more detail. In addition, I would also recommend contact authors or leading experts in the field in case they may know of other articles not returned during the database search. Finally, I would recommend manually searching the reference lists of previous systematic reviews, meta-analyses, and original research papers for additional articles.

Whiting P, Westwood M, Burke M, Sterne J, Glanville J. Systematic reviews of test accuracy should search a range of databases to identify primary studies. Journal of clinical epidemiology. 2008;61(4):357-64.

Papaioannou D, Sutton A, Carroll C, Booth A, Wong R. Literature searching for social science systematic reviews: consideration of a range of search techniques. Health information and libraries journal. 2010;27(2):114-22.

Also discussed here in a little more detail:

https://www.researchgate.net/post/What_is_the_minimum_number_of_papers_that_one_can_use_for_a_systematic_review#view=5a60005ecbd5c2d1e73918b9

Good Luck!

MVF

Hi Usama,

Statistically, the results of a meta-analysis may not be valid if the number of studies are fewer than 5, given the lower power in the overall estimation of effect. However, this is often not an impediment to pooling studies, the limitations just need to be acknowledged.

Regards,

Hi Godfrey, do you happen to have any reference to support the 5 studies minimum rule?

Thanks

Hi Usama, Theoretically two but if others are available you need to ensure that your data are not leading to selection bias. It is not a requirement, however, to include all possible available sources. A strong justification would need to be made for exclusion, though. If this is due to an overpowered model fit an "expensive" way of dealing with claims of selection bias is to fit your model to several mixed permutations of the available data and compare the results!

In theory, the bare minimum is 2, but the highest the number of the included studies in your meta-analysis the strongest is the effect size estimate. If you would like to undertake publication bias tests, you should include at least 10 data sets as it is the minimum number to run such tests.

The general answer still is "two studies", however, it depends on your objective, field, plan to publish... you may refer more in Valentine et al., 2010 (How Many Studies Do You Need?: A Primer on Statistical Power for Meta-Analysis)

The advantage of meta-analysis is answer to question(s) with high power. we know number of study and sample size within studies is more effective on the power. So if the number of study ( number of sample) is low, our answer to question have not enough credit.I think it is better no perform a meta-analysis at least 5 studies unless we know 2, 3 or 4 studies are very similar in methodology. specially in CRT studies.Also all simulation study run at least five study.

Theoretically a minimum of 2. But the standard error if using a random effect model is very large. in addition, the small number of studies can lead to publication bias, namely the results of meta-analysis do not reflect the actual conditions. In other words the results of the meta-analysis are influenced by subjective findings. So Boronstein et al. Suggest large samples.

A would consider the size of the studies, it would make better sense to have two large well conducted trials than 5 small studies

Carlos Ayán - it took me a while to find a good reference, but this article explains why you should have at least 5 studies for a random effects meta-analysis: Article Power analysis for random-effects meta-analysis

Thanks a lot, I am sure it will be very helful

Kimberley L Way - Very timely and important reference for me, thank you!

If this is because the data is not available in the original studies, the ideal is to always contact the authors to request the raw data. The ideal number of studies in meta-analyzes is in the range of 20 to 30 effect sizes, although it may contain many more. For more details see chapter 14 of the book that is attached.

I would also recommend checking how to calculate the power for meta-analysis:

https://jtiebel.shinyapps.io/MetaPowerCalculator/

https://www.jtiebel.com/2018/08/26/how-to-calculate-statistical-power-of-a-meta-analysis/

The power should be seen related to effect size, number of participants and studies rather than just the number of studies included :D

Cheers

more than 2 studies.

What if on a topic there are no more than 10 from which we should select

Dear Atoosa,

Select at random and include, sequentially, until saturation is achieved or all 10 have been used. If saturation is achieved prior to exhausting all cases, select at least one remaining and compare to ensure that the convergence is robust.

Good luck with your research.

James

Hi Usama,

Cases are sufficient because you are saying that the population equals the sample. Here is the reason why it is sufficient. Imagine, you want to study a personality, such as a leader, and the focus of the study is graduation speech, then you know how many speeches any president makes. It is not the number of studies, but it is the question of what you do with it.

Here is an example of cases where meta-analyses interest at the question level in the interview. Still, it produces a quantitative study. You can ignore the topic, but the methodology might be insightful for your purpose.

https://www.nomos-elibrary.de/10.5771/0935-9915-2020-1-55/contextual-bricolage-and-the-founder-s-apprehension-in-the-inter-generational-succession-of-small-family-businesses-volume-31-2020-issue-1?page=1

NO !: simply analyze, in such a meta-analysis, the maximum number of studies and publications on the subject in question ... although, by statistical criteria it would be desirable that they were more than 30 (which is considered a sufficient "n" for some tests statistical significance); but, if there aren't, go ahead!

Some simulation studies showed that at list five studies.But in practice at least 10 studies in random effect model is need.

two is a minimum

Most research suggested a minimum of 20-30 to calculate its effect size. But there is no black and white decision.

Two studies, per Cochrane