Hi Jeeva! Thank you for your elaborate instructions. I will request you at a later date if the 'do' file is needed, because I have to read up the Stata instructions first ! 

I had this doubt about calculation of pooled prevalence using RevMan because I read a paper that used RevMan to calculate pooled prevalence, and I know you need two comparative groups to calculate pooled data in RevMan. 

Sumedha

pooled prevalence can be calculated with forest plot.

chs

Just to update you, yes we can calculate pooled prevalence in RevMan (ref. Allagh et al, 2015, http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0118961). The principle is the same as explained by Jeeva. The Standard error needs to be calculated separately. You have to use the Inverse variance option, enter the prevalence and standard error, and the software calculates the pooled prevalence, depending on the preference of your model (random or fixed effects).

I'm having a quite similar question, as I'm writing a systematic review about a global performed action. Included studies show results of prevalences per country or region. I'd like to calculate a global prevalence in RevMan. The instructions by Sumedha Dharmarajan were too brief for me to understand. Is there a more comprehensive elaboration or reference? 

Hi, there is no appropriate reference, because I myself have written to authors of other articles and figured out the method. But for a detailed explanation, if you know RevMan, these are the following steps

1) Enter your studies in RevMan

2) Go to Data and analysis

3) Click on add outcome

4) Click on the Generic Inverse Variance option

5) Enter the name of your outcome. Click next

6) The statistical method should be chosen as Inverse Variance

7) You can choose the model you wish to use. Click on next 

8) Click on the option for which CI you wish to use (90 or 95%)

9) Customize your graph labels, since I am analysing for prevalence data, I prefer that the graph labels do not read anything.

10) Please adjust the scale of the graph at this point. Click next

11) Choose the option to enter the study data. Click finish

12) Add in the study data. For this you will need your study references ready, Prevalence data with the standard error of mean calculated separately for prevalence data. The formula to be used is SQRT(p*(1-p)/n).

I hope these steps help you all out !!! 

Thank you Sumedha for you elaborate yet clarifying answer. I am still wondering though which effect measure to use. All options seem to suppose that there is an exposed and non-exposed population. However, the studies I'm pooling just report on an incidence rate of an adverse event in one type of patients (that underwent a particular intervention).

No not always. You can select the option "name of effect measure, and rename it as prevalence. Ïn that way, the software automatically asks you to enter the Standard error of mean (SE). That way you can plot the forest plot on RevMan.

I am planing to do similar work and reading each comments.I got the comments of a great importance for my work.

I do not think that Revman is doing this properly. If we follow this approach Revman thinks it is summarising continuous numbers with an associated precision, not proportions. We are not providing the information that a prevalence is bounded below by zero. That Plos One paper linked to illustrates this, with the forest plot being (incorrectly i believe) symmetric about zero (i.e. the graphs x axis ranges into negative prevalences, which is meaningless).  

When we don't treat prevalences as prevalences we will get our 95% confidence interval coverage wrong. For a start, this method in RevMan produces symmetric CIs, which is typically not what happens when they are calculated appropriately. The RevMan approach is most wrong when the prevalences are close to 0 or 1 (which is the case in the PlosOne paper). 

I found this while searching for the solution to a similar issue.

A better approach uses the metaprop function from the meta package in R. This conducts meta-analysis of a proportion invoking lots of Robert Newcombe's work. Happy to share code on my example if anyone wants a walkthrough

Mark Kelson

What about using metaprop command in Stata? does it have any problem?

Not a Stata user, but have just read this, which indicates it is accounting for the binomial data properly (bolding mine). (from here https://archpublichealth.biomedcentral.com/articles/10.1186/2049-3258-72-39)

"Metaprop implements procedures which are specific to binomial data and allows computation of exact binomial and score test-based confidence intervals. It provides appropriate methods for dealing with proportions close to or at the margins where the normal approximation procedures often break down, by use of the binomial distribution to model the within-study variability or by allowing Freeman-Tukey double arcsine transformation to stabilize the variances. "

I'm not sure whether RevMan has the option to calculate pooled propotion/prevalence. Use Stata. If not you can use MetaXL. it is free and add on to Excel.

How can we do sensitivity analysis using 'metaprop' command in stata?

Hi Mark,

I do not concur with you. Revman is not calculating the standard error and CIs. You calculate SEs and CIs before inputting the data in RevMan and the Software calculates the pooled value that you have input in the software. So if your previously entered data has negative values, I think it will reflect in the pooled prevalence.

Secondly, I have no access to R, so using the functions you have described becomes a challenge for me and cannot comment on how accurate the function is. But, as far as I know, RevMan is the best available software for me to calculate prevalence data.

Hi Sumedha. I didn't mean to imply that Revman is calculating the standard error. I understand that this is user-input.

The generic inverse variance method you recommend though only allows symmetric CIs (which you are calculating using "SQRT(p*(1-p)/n)"). This is incorrect (for any prevalence that isn't 0.5) and can lead to confidence limits outside the range 0-1.

You can convince yourself of this by using Revman to enter a prevalence of 0.5 and an SE of 0.5. You will see the graph update to give a very wide confidence interval that lies outside 0-1. This is impossible for a prevalence. This generic inverse variance method does not reflect the fact that prevalences are bounded by 0 and 1 and is not a good method to use for pooling prevalences.

Hi Mark! Thank you for your inputs. I experimented by adding five studies and entered the prevalence and SE as you suggested.But, the confidence interval of the pooled estimate was 1.06, 2.56. In whichever manner I tweaked the numbers, I got confidence intervals above zero and never zero or below zero.

But, again if you are reporting prevalence studies, one rarely reports prevalence as 0.05% (for example). It will instead be reported as 5/10000. So again, not able to agree with you.

Dear Sumedha,

I have explored presenting Binomial estimates for large samples of prevalances less than 0.1 (or greater than 0.9) and found RevMan, and Stata both sub-optimal- at least for generating graphics.

My advice is to use the Binomial exact method, and present these in R. See Figure 1 from the below study as an example of what may be achieved

Article Differences in the Recurrence and Mortality Outcomes Rates o...

https://doi.org/10.1210/jc.2013-2118

Hi Dr. Mark Kelson,

We are planning for a meta-analysis which involves pooling of prevalence data. Can you please share your example code for this purpose using the "metaprop" function in R? It would be of great help for us.

Regards

Anand Srinivasan

Hi! Thank you Mark and for Ben the discussion and the inputs. Yes, planning to do further analysis in R. So that I can confirm, for myself, whether pooled prevalence estimates obtained in RevMan is really valid.

Hi Mark, Sumedha

I am doing a similar analysis of prevalance, and there are several proportions that are 0 or 1 and as mark pointed out, using simply SE in RevMan gives values that are outside the 0-1 range. I dont know how to overcome this problem. I want to use the Freeman Tukey double arcsine transformation, but Revman only asks for either prportion or its log transformation. Also, MetaXl, though it mentions that it uses the double sarcsine transformation, is giving me wierd CIs for the 0 or 1 prevalence data. Does anyone know how to solve this?

I have no experience in R/Stata

Hi everyone. I have a similar situation. I have single group studies in which a prevalence of an event is reported. I did everything I could but RevMan does not draw a forest plot. What am I missing? RevMan asks for comparison but there is no in my situation????

For Example:

Study 1: NofParticipants=200 NoofEvents=75

Study 2: NofParticipants=170 NoofEvents=46

...and so on

Hi Alp Özgün Börcek,

The steps outline previously by Sumedha worked perfectly for me. I copied it from above for you

Sumedha Dharmarajan

Hi, there is no appropriate reference, because I myself have written to authors of other articles and figured out the method. But for a detailed explanation, if you know RevMan, these are the following steps

1) Enter your studies in RevMan

2) Go to Data and analysis

3) Click on add outcome

4) Click on the Generic Inverse Variance option

5) Enter the name of your outcome. Click next

6) The statistical method should be chosen as Inverse Variance

7) You can choose the model you wish to use. Click on next 

8) Click on the option for which CI you wish to use (90 or 95%)

9) Customize your graph labels, since I am analysing for prevalence data, I prefer that the graph labels do not read anything.

10) Please adjust the scale of the graph at this point. Click next

11) Choose the option to enter the study data. Click finish

12) Add in the study data. For this you will need your study references ready, Prevalence data with the standard error of mean calculated separately for prevalence data. The formula to be used is SQRT(p*(1-p)/n).I hope these steps help you all out !!! 

I am having a similar issue working on combining studies with one group, for example looking at awareness of medical risk factors. I have several studies with the % awareness an N imported into RevMan. I also have access to SPSS. Thanks.

MetaXL is a good easy to use and free software that does this very well.

Thanks for the info. the only addition to the excellent Revman advice above is that under effect size, you should not select any ratios, but select others and then delete and type Prevalence

Is there a step by step guide for doing meta-analysis in STATA?

Very good information thank you all

I tried this in a paper about a year ago and got a figure with prominent errors. I ended up learning R and, while it takes time to learn to use R, you get more flexibility and control over every detail in your figures and can include subgroups in the same figure.

You can use Metaprop routine available to download for STATA if you only want to give details of the prevalence of a problem in a population-based on multiple studies. If you want to compare this prevalence between 2 different groups of patients within each study you can use Metan routine available to download for STATA. The last update of STATA 16.0 may have other implementation (but I am not sure). I hope that helps. Good luck!!

Alp Özgün Börcek ,

I figured out that Sumedha Dharmarajan answer is correct but she forgot one step that was mentioned by Aethele Khunda. Basically in Revman when you select the Generic Inverse variance model, it then asks you to select the effect measure (here you have things like odds ratios and relative risk ratios). Don't select any of these instead click on "Name of Effect Measure" and in this box type in "Prevalence" or whatever you want to call the measure. Then after you have set up your model, just type in the prevalence for each study manually and the S.E using the formula provided by Sumedha and it should work

Thank you for your answers

How can we enter the data if the prevalence is zero

Thank for your guidelines

I have a question about that. Could you explain me what type of data of prevalence we should put in RevMan? That is

1. Proportion % eg; 18 %

2. Another meaning of proportion eg; 0.18

3. Number of events such as count

I think your calculation in RevMan is totally ''wrong'', Sumedha Dharmarajan.

First of all, in RevMan, I-squared statistic will stand at stable value for example 100% in all analysis.

Second, CI confidence interval for each and every value will be very narrow such as 0.01 to 0.03 reguardless of sample size. This is the main point of the wrong calculation. CI is very sensitive to sample size.

Another difficult point is CI maybe beyond the null value in forest plot that is -10%.

How can we interpret it?

I can explain the way that I understand.

First of all, Inverse Variance Method in RevMan is calculated not depending on sample prevalence. So CI (Confidence Interval) is not affected by sample size.

As a result, logit and double arcsine transformation are needed in such case.

In MetaXL mentioned above , double arcsine is available which controls over Confidence Interval not going beyond to 0 or 1as in IV method. This is more conservative.

The article below maybe useful for you.

https://www.google.com/url?sa=t&source=web&rct=j&url=https://www.researchgate.net/publication/256072092_Meta-analysis_of_prevalence&ved=2ahUKEwjzhqTxhoPzAhWEwjgGHWT7Di8QFnoECAQQAQ&usg=AOvVaw29RvcBcBgIo674rHNJhamr

The tranformation is not necessary when the proportion being around 50 percent. On the other hand, Freeman-Tukey or logit transformation is needed in case of extreme value 0 or 1. The below article is very informative.

https://doi.org/10.1186/s12874-021-01381-z

I am sorry to ask but how to use this formula SQRT(p*(1-p)/n)?