So i'm doing a meta analysis, and i have some question about a study that i have.
This is the data for one group(lower intensity group):
pre intervention mean(SD): 274(70.5)
post intervention mean(SD): 286.7(73.9)
mean difference pre and post (95%CI): 12.7 (-27.1, 40.0)
This is the data for other group(higher intensity group)
pre intervention mean(SD): 267.2(61.3)
post intervention mean(SD): 291.3(63.7)
Mean difference pre and post(95%CI): 24.2(-12,63.7)
As you can see the 95%CI between pre and post is asymmetrical around the mean. What causes this?
i'm comparing the mean differences (between pre and post) of higher intensity group compared to lower intensity group using revman, how do i use this data for meta analysis?
Here's the link to the study: (table 2, 6mwd)Article Benefits of different intensities of pulmonary rehabilitatio...
It doesn't mean anything. Not all methods of calculating confidence intervals yield symmetrical results - for example if means were log transformed or converted to some kinds of effect sizes during calculation or if they were bootstrapped, there will often be an asymmetrical CI. I can't tell you the exact reason in this case given that you haven't said how the CI was derived, but it doesn't have any implicit meaning. You can still use this data for the analysis without any complications.
For RevMan, you can just input the mean and SD for each study arm and the MD will be calculated automatically.
I can think of three possible reasons, among many others that more experienced contributors might share later on:
1. Modifications of paired t-test that consider/assume substantial sample skewness/asymmetry based on an important variable were used rather than straightforward paired t-test. Please check the following articles that may help:
1.1. https://www.jstor.org/stable/2286597?seq=1
1.2. Article A Confidence Interval and Test for the Mean of an Asymmetric...
2. Clerical error that was not spotted during peer review
3. Computational error that was not spotted during peer review
Regardless of the cause, it will be prudent for you to get in touch with the corresponding author so he/she may provide clarification. If possible, you can even ask for the raw data so you can do the computations yourself.
Jack Henry thanks for your answer, i have edited the question to make it clearer, yeah i have tried to put the mean and SD in the revman but I get different CI(MD: 12.7; 95%CI: -11.67, 37.07) compared to the one shown in the study(MD:12.7; 95% CI: -27.1, 40.0), is it okay if i use the CI from the revman?
here's the link to the study(table 2, 6mwd)
Article Benefits of different intensities of pulmonary rehabilitatio...
Aedrian Abrilla thank you so much i will read it! yeah I previously thought of contacting the author but i thought that asking here first is a good idea, and if i dont get the answer i'll contact the author
Christian James Ibrahim Apologies, I didn't realise these were calculated by RevMan. RevMan calculates SMD via Hedge's g which adjusts for sample bias and is probably the cause of the asymmetry. There is no error here and you can continue your analysis as normal.
Jack Henry No sir, the asymmetric CI data isn't from revman, it is from the study data i linked in my previous answer. (MD:12.7; 95% CI: -27.1, 40.0) -> asymmetrical CI
The CI that i calculated manually using pre and post intervention mean is symmetrical (MD: 12.7; 95%CI: -11.67, 37.07)
sorry to confuse you
Christian James Ibrahim Apologies - is the problem that you are unsure whether to pass the asymmetrical CI to RevMan? There is no issue or likelihood of error on behalf of the study authors given the asymmetrical CI as this is likely to simply be related to the way the CIs are calculated.
Instead of calculating the MD and SE/CIs yourself, you should use the in-built calculator functionality in RevMan. You can input the mean and SD for each group, and RevMan will automatically calculate the SE and SMD for you, I believe. If RevMan cannot do this, you should use a software package that can - meta for R is certainly capable of this. Calculating these values by hand is very prone to error.
Jack Henry i've tried to use the calculator to convert the 95%CI to SD but i think the calculator doesn't work if the 95%CI isn't symmetrical.
Christian James Ibrahim That is not true. You do not need to put the CI into the calculator. You should put the mean and SD for each group into the calculator.
The asymmetrical CI does not mean that you cannot use the SD reported by the study, as the CI is derived from the SE (and by extension the SD) and the fact that it is adjusted does not mean the SD is not usable.
Input the mean for each group and SD for each group into RevMan and let it calculate everything else. The data is absolutely fine.
Jack Henry But there's no reported SD for the change score (only MD and 95%CI)
the only data available is :
pre-intervention: mean &SD,
post intervention: mean&SD
but there's no data for the change score SD(only MD and 95%CI). change score is difference between post intervention and preintervention
Because what i need to compare is the change score from the higher intensity group, compared to change score from the lower intensity group
Christian James Ibrahim Apologies, now I see what you are asking. This is a common issue. There is no robust way of calculating the standard deviation for the change manually from the data available.
It is perfectly reasonable to use the asymmetric confidence interval for your analysis in this case - while it is clearly adjusted, SDs are often imputed in far less accurate ways. If you are still worried about the impact of this, consider performing a sensitivity analysis in the form of a leave-one-out analysis of influence to investigate if this one study affects your overall result. Assuming it doesn't, this is absolutely not an issue.
If you really don't want to include this asymmetric CI, you can impute it by estimating a correlation coefficient from a study with full data and imputing the SD from this, but there is really no need as this will probably be far less accurate than whatever minor adjustment was applied to the CI of this individual study.
- Badges
- Science topic