Conducting a meta-analysis with R
Вставка
- Опубліковано 12 лип 2024
- UPDATE: The audio quality isn't great for this video, so I re-redid this tutorial in a new video with better audio quality: • How to perform a meta-...
Meta-analysis synthesizes a body of research investigating a common research question. This video provides a practical and non-technical guide showing you how to perform a meta-analysis of correlational datasets. I use a supplementary R script to demonstrate each analytical step described in the paper, which is readily adaptable for people to use for their analyses. While the worked example is the analysis of a correlational dataset, the general meta-analytic process described in this paper is applicable for all types of effect sizes.
The paper - journal.frontiersin.org/articl...
The script and datasets - github.com/dsquintana/corr_meta - Наука та технологія
Daniel! Thanks for so much for posting; this is going to be a lifesaver to putting the final touches to my dissertation. Now the hard part of collating all the data from wide and varied sources. I might have some questions for you later if I get stuck on the code.
Great to hear! You may also be interested in a recent podcast episode where I cover some wider meta-analysis issues soundcloud.com/everything-hertz/4-meta-analysis-or-mega-silliness
Fantastic presentation!! Thank you for sharing.
+Ingrid Oliveira Thanks, glad to hear you got something out of it
Thank you Daniel, your presentation helped me a lot. But I have a simple question in the funnel plot. Xlab says, correlation coefficient, but as to me, these values are rather the Fisher's z-transformed values (res
Excellent! Thanks for the video,
Thank you so much for this video. You saved my masters thesis.
Thanks for the comment, very happy to hear this! 😃
Inspiring teaching, thank you!!!
nice concise explanation, thanks!
HI, Thank you so much for this video! DO i not need to calculate heterogeneity when i am running the RVE: hierarchical effects model?
Thanks!
Great tutorials! Thanks for your explanation of MA. 7:30 in the first video, I wonder how to explain the size of the square, is it representing the weight of each study?
Hello, thank you for tutorial. What about if you have groups of variables like rate,blood,pressure and you want to measure which has greater impact on cardiovascular diseases but for the data you have only mean and SD ,you do not have control and treatment group, should i use only metamean in R or is any better solution suggested? Thanks!
Amazing ! Thanx
Hey Dan. It is a really helpful video for meta analysis. I have a small question. Which version or R are you using for analysis because I am using 3.6.1 and in that I am not able to access the libraries which you have mentioned.
Thanks for this video, really helpful. I am adapting your script for my data and had a question what the numbers for the forest plot in the last two lines relate to? (the text (-1.6,18) As I am unsure if I need to adapt them to my data. Also I wondered, can you do the meta-analysis by subgroup? Including have the forest plot organised by subgroup? Thanks so much!
forest(res, xlim=c(-1.6,1.6), atransf=transf.ztor,
at=transf.rtoz(c(-.4,-.2,0,.2,.4,.6)), digits=c(2,1), cex=.8)
text(-1.6, 18, "Author(s), Year", pos=4, cex=.8)
text( 1.6, 18, "Correlation [95% CI]", pos=2, cex=.8)
Amazing! really helpful
johan heemskerk Glad to heat that!
Great paper and video thaks for providing the code.
excellent tutorial
Sekhar Dutta thanks, appreciate your comment!
Thank you Daniel for your great script and effort in simplifying meta-analysis pipieline. I'm doing my first steps in this analysis and wondering if you could recommend some source on how to prepare data for your pipeline including different types of effect (like OR etc.)?
+Konrad Stawiski Thanks for the kind words. Can you provide a little more information on what you would like to do in relation to preparing your data?
I saw your blog in medium if i think...I have a basic question how do you aggregate data from multiple studies?
Nice video , thanks :))
HI Dan, I am conducting a meta analyses using beta and se (from linear regressions). Could you please give me some suggestions on editing your code? Thanks so much!
Hey Dan! I am having trouble understanding the the vi and yi mean and how those values are calculated?
Thanks and great video! Using your research to help with my thesis!
Abuzar T “vi” is a measure of effect size and “vi” is a measure of variance. These are calculated using the “escalc” function, by entering the correlation coefficient and sample size. Glad you found the video useful!
I have a question: the evaluation of study methodological quality
using a 4-point assessment tool developed for this meta-analysis by two coders. ; is there other tools to help me to define the study methodological quality?
A really helpful video, thanks!
BTW, could you please advise what is the % of credible interval in the estimate? if I need to set a % of credible interval, say 90% Cr, where I should put the argument?
Hi Eleanor,
This is actually given from the "predict(res, digits=3, transf=transf.ztor)" command. The values under "cr.lb" and "cr.ub" are the bounds of the 95% credibility intervals. However, I'm unsure how to specify other limits, like 90%. Hope this helps!
thanks Daniel, that helps a lot!
Thanks again! I tried it with my data, and on the column "authors, year", the list of authors is not displaying on the forest plot. On that column, study 1, study 2,... lists are displayed instead of the name of the authors. What do you think for this?
Hello NG. Asefa. How did you solve this issue? I am facing the same
can you explain codes for making risk of bias table, and how to change colors in risk of bias table
Hi there, first of all, thanks for writing that paper. It's been extremely useful in a meta-analysis that I'm conducting at the moment. I was wondering though if, you could tell me how to reference the tests that the R package performs, as per in your script (I could not find any of this in the documentation). An example from the PRISMA: "we tested for heterogeneity with the Brestlow-day test, and used the method proposed by Higgins et al. to measure inconsistency... (termed I-squared)"... So i guess my point is, could you tell me what/who's "method" is being used when calculating the Q-statistic and the I-squared?
Hi Santiago, good question! Sometimes this information can be tricky to find in R package reference manuals, particularly comprehensive ones like metafor cran.r-project.org/web/packages/metafor/metafor.pdf
The reference on the Q-statistic (Cochran, 1954) can be found on page 225. The reference for I-squared (Higgins & Thompson, 2002) can be found on page 174.
Cochran, W. G. (1954). Some methods for strengthening the common χ2 tests. Biometrics, 10, 417-451.
Higgins, J. P. T., & Thompson, S. G. (2002). Quantifying heterogeneity in a meta-analysis. Statis- tics in Medicine, 21, 1539-1558.
Thank you so much Daniel for this, you're right on how tricky finding this can be. This is something I'm still struggling to get use to in R. (I'm sure this info would be useful for others too).
By the way, on that note, when plotting and using the baujat() funtion, it might be useful to pass along the argument (symbol = "slab") in this way the numbers indexing the studies, would actually correspond to the study_id tags (rather than the R generated row-names). This might not make much of a difference to most people, but in my case I had to split my sample and wanted to plot these two sub-samples, conserving the study_ids assigned to each of the studies when together (e.g., some would only belong to the second plot and vice versa). Anyway, I hope that is clear.
dear Daniel,
l can access both the video and script and i found it very impressive. Great job
just one question: revman can do the rest of the jobs but i can't do meta regression. Can you help me out if there is a possibility with latest versions?
Hi, if you are interested in performing meta-regression in R using the metafor package, I would recommending working your way through the examples on the metafor webpage. Many of these examples include meta-regression www.metafor-project.org/doku.php/analyses
Thank you very much for this helpful video.
Now I am conducting a meta-analysis, using effect size instead of Pearson r. When I get to the step of normalizing individual study effect size and corresponding size, I don't know how to convert it. Is there any other variable I need? Could you give me some suggestion if you know that. Thank you very much.
Hi Frank,
What specific effect size are you using? The 'escalc' function in the metafor package can be used for the calculation of effect sizes and their corresponding variances. Can you give me a little more information about what kind of meta-analysis you are performing?
Cheers,
Dan
Hi Dan, any chance you'll be making a similar video for diagnostic accuracy meta-analysis in R using the mada package? :)
Hi Joyeeta, I'm not familiar with the 'mada' package so I don't think I'll be making a video on that topic in the near future, but I'll check out the package.
Cheers,
Dan
Hi! Your video and paper is really awesome and very helpful. I am almost done with my master thesis. But now I have a question concerning the moderators. I have (among others) a categorial moderator wich is the instrument used in the study. Each of the instrument has its coding numer (1-8). Can I now just take these numbers or does it make more sense to write it in letters (how you did) ? You know any scientific paper which says it is ok to do so? (as I have to proof what i am doing is ok in my paper;)). How do categorial moderators differ from continous? Thank you!!
Hi Dominique, thanks for your comment. With the information you've provided, I think it would be best to write out the names of each instrument as they are their own categories. The script to assess categorical and continuous moderators are a little different, have a look at the script to perform the "mean age" moderator (continuous) and the "controlled variables" moderator (categorical) github.com/dsquintana/corr_meta/blob/master/script Hope this helps!
Thank you very much! This is awesome:)
Question: If I want to work with the Hunter and Schmidt approach; where and when do I type in (method.tau="HS") ?
Hi Dominique,
To use the Hunter and Schmidt approach, you can use this script:
dat
Thank you so much Dan for the quick response! Your are basically just saving my master thesis:)
Concerning the approach: I assume in this tutorial you are using Hedges/Borenstein? Z-Transformation and weigthing according to sample size?
Glad I can help! In the tutorial and talking the Pearson's r values and sample size, performing a Fisher's r to Z transform (along with corresponding sample variances). I use the default Restricted maximum-likelihood estimator (no special command here, metafor uses this by default unless you use another command). I then back transform Z to R for easier interpretation. This paper is a good resource www.jstatsoft.org/article/view/v036i03
Hi Daniel. Any quick code for forest plots with subgroup analysis? Metafor tutorial kills me on that one. Cheers
Check out the metaviz package for subgroup forest plots cran.r-project.org/web/packages/metaviz/vignettes/metaviz.html#subgroup-analysis
Thank you so much for publishing this video. I am doing the meta-analysis in uni, that helps me a lot. However, I have one question which is related to my project: In order to calculate the effect size, there are the correlation, and the standardized mean difference, like d or g. But what if in my database, some of papers did the correlation study, while some of them did the t-test? Then which element in effect size should I use in R?
Jasmin Zhang What you’ll need to do is calculate a common effect size measure and associated measure of variance. I would suggest carefully working through this tutorial, which demonstrates how you can do this www.metafor-project.org/doku.php/tips:assembling_data_smd Once you have these measures you can perform your meta-analysis
Thanks for replying me. I will check that video out.
hi daniel,
i want to ask what is 'ri' and 'ni' in ur data?
plz reply soon..
Hi Anjali,
'ri' represents the Pearson's r value and 'ni' represents the number of participants from the study. Hope this helps!
Dear Daniel, thank you for posting how to make meta-analysis with r. I am involved in the meta-analysis of heritability, can you please help how to meta-analyze heritability estimation? The effect size is a bit different from OR,RR, MD. It is basically similar to proportion, but it may still be different.
Hi Nigus, I'm not familiar with the meta-analysis of heritability estimation. However, these two papers have used the metafor package for heritability estimation meta-analysis may assist psycnet.apa.org/record/2015-20360-001 www.ncbi.nlm.nih.gov/pubmed/25263742
Thank you for your replay! I will have a look on these articles.
Dear Daniel, the quantitative bias and moderation testing methods are impressive, yes the funnel plot is very subjective. I have learnt that heritability can be estimated using this package, too. It showed me almost similar results to those analyzed by other softwares. But, I have still one more question, if I am not bothering you. I was trying to meta-analyze subgroups, can you give me a clue about this?
Solved, with "byvar". Thanks
do you manage differently a meta-analysis on longitudinal studies?
Hi everyone, I'm doing a monograph and my teacher asked me to do a meta-analysis article instead of the traditional literature review. I've never had contact before with the meta-analysis and I have many doubts. Well, I need to do the meta-analysis on the data I got from published articles by other authors, but I'm evaluating which are the major bacteria that cause inflammation of the mammary gland of cows. So in the case I created a file in Excel with the data of the work with the columns containing: Author, Number of samples evaluated in that work, Bacterial species 1, Bacterial species 2, Bacterial species 3, Bacterial species 4, Other species. In this case, how could a metanalysis be performed on the data? Thanks to whoever can help me!
thank you so much, please can you put the link of the script
hi Daniel, I am getting this error while trying to open Molloy2014 dataset. Please suggest what is wrong
dat
Solved. It requires loading library Matrix prior to Metafor
> library("Matrix")
> library("metafor")
> dat View(dat)
Let´s say if I have data for 10 RCTs, where I want to compare A to placebo. In 3 of the RCTs they look at A vs placebo+B. And in 2 of the analysis they have A+B vs placebo.
Would it be incorrect to include from the three studies: A
and from the 2 studies: placebo
and then the remaining 5 studies: A vs placebo
into a funnel plot? (lets pretend I'm using SMD and SE continuous data) .
And brilliant work.
Hi, I would like to help if I can but would you be able to rephrase the second half of your question so I can better understand the scenario?
Sure, thanks Daniel.
I´m looking at A vs Control
Study 1 has: A vs Control
Study 2: A vs control + B
Study 3: A+B vs control
Would it be incorrect to pool data as
Study 1 A + Study 2 A
Study 1 Control + Study 3 Control
Then compare the pooled estimates with inferential statistics? I should mention that these are all pre- post- RCTs.
As these are pre-post RCTs the effect sizes depend on the correlations between pre and post means (check out this paper www.ncbi.nlm.nih.gov/labs/articles/27790968/). So if you were to pool data in this way you'd be making a lot of assumptions regarding how these effects would be correlated pre-and post. The goal for these types of meta-analysis is to pool together standardized mean differences for each individual study but you wouldn't be doing this in this case, only comparing the pre-mean and the post-mean indendent of the within-study differences.
Thank you, this is very helpful. I'm trying to run a meta-analysis in R but every time I type my code in it says Error: could not find function "escalc". Don't suppose you have any idea why this happens?
Glad to hear this was helpful. You may be receiving that error because the 'metafor' package isn't loaded. Type in "library("metafor")" and try again. Also, make sure you're running the latest version of R.
Hi thank you, I tried doing that again but it still isn't working (this isn't with your dataset btw). I'm also using the latest version. Have you got any other tips for why it might not be working?
Hard to say, I'd try using one of the datasets that comes with metafor. If they work then there might be something odd with how your dataset is structured. You might need to restructure your dataset to match how the examples are set up. If you can post your dataset online (e.g., dropbox, github) and send me a link I could also take a quick look.
Thankfully it is working now, I'm not sure what I did differently! Thanks very much for getting back to me!
What about meta-analysis in for diagnostic test accuracy?
simple question what is the simple way to aggregate the studies ?i have list of PMIDs which i have filtered based on my criteria . Your first object where you have studies ,authors designs etc etc.So all these if I have to run for my set of analysis I have to collect and collate the data in a single table form manually or there is any programmatic way of doing it?
Krushna Chnadra Murmu you will have to manually extract the relevant information from each article
@@dsquintana thank you
@@dsquintana i see the metafor in built as dataframe ..so I can simply store the data in a standard table format which i can read it into the library .?
Hey how can I find the podcast?
It seemed really interesting
Sebawe Syaj this link should work, enjoy! everythinghertz.com/4
I really appreciate your response!
Thank you 😍
Subscribed
Hello,
I am not comparing studies, I have to combine them... Can R help me out with that?
Hi Adil, can you provide a little more information?
Study 1 has a 95% Confidence Interval Lower Bound of -0.5, and a 95% Confidence Interval Upper Bound of +6.5. Study 2 also has a 95% Confidence Interval Lower Bound of -0.5, and a 95% Confidence Interval Upper Bound of +6.5.Study 1 and Study 2 also have the same p-value, which was .10. What is the R code that produces the lower and upper bound for a meta-analysis combination 95% confidence interval based on those two studies?
Hi Daniel,
Not sure if you still follow questions/comments on this video, but I wanted to chance it and ask my question anyways. It seems that any time I try to transform r to Z and calculate the corresponding sample variances, with the "escalc" function, it never adds the two new variables to the end of my file, but rather it seems to completely overwrite it and leave only those two columns in it.
I tried running the analysis in all the possible ways, even downloading your original file and running the script exactly as you have it, with no luck.
Many, many thanks for posting the video though, it is tremendously useful.
All the best
I don’t have access to my computer right now, but if you were to run the file included here this should add the two columns to the end of the file github.com/dsquintana/corr_meta If this doesn’t work, I would suggest double-checking you have the latest version of R and ‘metafor’ installed
Hi Daniel Quintana
Great vidieo.
I would like to ask how i can get the plot of Egger’s linear regression and Begg’s funnel plot using R?
Hi Mubarak,
If you follow along with the video, the paper (journal.frontiersin.org/article/10.3389/fpsyg.2015.01549/full) and the R script (github.com/dsquintana/corr_meta) you'll be able to calculated Egger's regression test and construct a funnel plot. In short, using the 'metafor' package you first fit a random-effects meta-analysis model using the 'rma' function. Once you have fit this model, then you can use the 'funnel' and 'regtest' functions. By recreating the script from start to finish you'll get a good understanding on how to perform a meta-analysis and associated tests.
Hope this helps!
How did you set the value of methodological quality?
This value was generated using a custom tool. There are various tools available to calculate this. See this paper for more details www.frontiersin.org/articles/10.3389/fpsyg.2015.01549/full
@@dsquintana Thank you, I will read. I am doing a systematic literature review and part of my sample is composed with behavioural studies. I will try to categorize their methodologies based on what I saw on your video and that i will read on the article
Very nice. Thank you. You should improve resolution of your videos, 480p is too blurry this days.
Marco Antonio Díaz Thanks for your comment, all my newer videos have better resolution and video quality. This one was recorded almost five years ago.
I suggest you to kindly cover up all the points n functions used in the analysis.
Anjali Patni Hi, thanks for your comment. Can you please clarify?
I think you finally swayed me away from CMA *and* gave me a good reason to start learning R...
Robin Kok it's worth it for metafor alone
Why when i run mess_ss
As we discussed over messages (for the benefit of others who may have the same issue), make sure if you’re using your own data that it has the same properties (e.g., that ‘studynum’ is a character) of the example dataset if you would like to use the exact same script
Hi Dr Quintana, I am getting this error. Please suggest how to troubleshoot it:
Library dplyr is showing following error:
> library("dplyr")
Error in loadNamespace(j install.packages("pillar")
Installing package into ‘C:/Users/yagga/OneDrive/Documents/R/win-library/3.3’
(as ‘lib’ is unspecified)
also installing the dependencies ‘fansi’, ‘rlang’
There are binary versions available but the source versions are later:
binary source needs_compilation
fansi 0.2.2 0.4.0 TRUE
rlang 0.2.0 0.3.1 TRUE
pillar 1.2.1 1.3.1 FALSE
Binaries will be installed
trying URL 'cran.rstudio.com/bin/windows/contrib/3.3/fansi_0.2.2.zip'
Content type 'application/zip' length 141745 bytes (138 KB)
downloaded 138 KB
trying URL 'cran.rstudio.com/bin/windows/contrib/3.3/rlang_0.2.0.zip'
Content type 'application/zip' length 754726 bytes (737 KB)
downloaded 737 KB
package ‘fansi’ successfully unpacked and MD5 sums checked
package ‘rlang’ successfully unpacked and MD5 sums checked
The downloaded binary packages are in
C:\Users\yagga\AppData\Local\Temp\RtmpKUVgSF\downloaded_packages
installing the source package ‘pillar’
trying URL 'cran.rstudio.com/src/contrib/pillar_1.3.1.tar.gz'
Content type 'application/x-gzip' length 103972 bytes (101 KB)
downloaded 101 KB
Hi,
Hard to say what the issue is here, but I'd recommend restarting in a clean R session and trying to load again. Also might be an issue with where your library is installed. Also try re-installing dplyr and checking all your other packages are up to date. If that doesn't help, perhaps search for a similar issue on stackoverflow or you can post this question there.
Cheers,
Dan
@@dsquintana Hi Daniel case in my case library dplyr is not installing..dplyr is asking for Matrix and pillar libraries which are not installing..I dont know the issue
@@dsquintana Hi, Based on this thread on R Community Page community.rstudio.com/t/specific-packages-seem-to-have-crashed-my-rstudio/21222 I have updated to R 3.5 and it has solved the issue. I wanted to use R 3.3.3 with SPSS 25 but that doesnt seem working. I am just posting it here so that it may help to someone facing similar issue in future.
Reagrds
datsme888 glad to hear you solved this, thanks for posting the solution!
Egger and trim and fill did not worked here :s
Jorge Teixeira can you provide some more details?
8:58 Dat bias, tho.
In all seriousness though: Great video, great article, great everything!
Thanks, hope to make some more soon with a better mic!
I'm usually easily irritated when the sound production value is low, but in this case, the content was intriguing enough, so I didn't really care.
Oh well, off to listen to the podcast episode.
Ch vagkxk mapa kkkk
Study 1 has a 95% Confidence Interval Lower Bound of -0.5, and a 95% Confidence Interval Upper Bound of +6.5. Study 2 also has a 95% Confidence Interval Lower Bound of -0.5, and a 95% Confidence Interval Upper Bound of +6.5.Study 1 and Study 2 also have the same p-value, which was .10. What is the R code that produces the lower and upper bound for a meta-analysis combination 95% confidence interval based on those two studies?