I'm doing meta-analysis in which a paper is reporting outcomes as a mean, range. Instead, I need to calculate the SD to be able to pool the effect in the meta-analysis model.
Yes agree with Klaus. I also cant see how you can go from range to SD, except if you have multiple samples with ranges reported then maybe you can construct an SD, but with one mean and range reported you cant. SD is variance including all the data so from three points (range and the mean) I cant see how you can approximate SD. Cheers.
Definitely you can't directly calculate SD from data you mentioned. You have 3 options:
1) Exclude the study from meta analysis since it does not present all required data,
2) Ask the authors for more precise information,
3) You can impute SD by 'borrowing' it from similar studies and perform a meta-analysis (please refer to Chochrane handbook: http://handbook.cochrane.org/chapter_16/16_1_3_1imputing_standard_deviations.htm)
If you choose #3 please remember that you receive result with limited credibility as you used imputed SD. Therefore, usually such analyses are presented as additional sensitivity analysis.
I totally agree with Piotr Wojciechowski and I will just add the following from the Cochrane Handbook
7.7.3.6 Ranges
Ranges are very unstable and, unlike other measures of variation, increase when the sample size increases. They describe the extremes of observed outcomes rather than the average variation. Ranges should not be used to estimate standard deviations. One common approach has been to make use of the fact that, with normally distributed data, 95% of values will lie within 2×SD either side of the mean. The SD may therefore be estimated to be approximately one quarter of the typical range of data values. This method is not robust and we recommend that it should not be used
Thanks everyone for your response to my question. I had asked this question when I was a 2nd year medical student and since this question has been read by multiple researchers till today, I'm adding a clear answer now:
There is no direct transformation from mean and range to SD
Possibility (1): I can obtain the SD from the means of multiple samples (only in case of multiple samples). In this condition, the variance between sample means will yield the standard error (SE)
Possibility (2): If the data were normally distributed, we can assume that the mean and median are similar and therefore, we apply the equations provided by Hozo et al. 2005 at BMC Research Methodology. These calculation will estimate the variance and SD from median, range and N.
Article Hozo SP, Djulbegovic B, Hozo I. Estimating the mean and vari...
Got a few questions on whether this is published. Most likely because - I imagine (?) - you will be able to reference it.
Thanks for the messages; privately and here.
The answer is that it is accepted as part of one publication.
“There is no Face like Home”: Ratings for Cultural Familiarity to own and other Facial Dialects of Emotion with and without Conscious Awareness; A British Population Study. Perception.
It is pending with more elaboration as to how the metric works in another publication.
“Look not at what is Contrary to Propriety”: Religiosity, Disgust Sensitivity and the uncharted Framework of their Association .
The problem you will have in quoting either of these is that the metric is used as part of the analysis and is not the dedicated focus of either paper.
Since the metric works there are a couple of ways to go from here given the demand, I could either upload the metric with examples in a dedicated site, or submit a small monograph to a mathematical journal. Latter will take longer than the former. Tell me what you think.
Sorry, was not expecting to have such interest but I am glad to help out :)
Myron Tsikandilakis in my case I prefer a brief report about how the metric works, but if it possible to upload the metric with examples in a dedicated site it will be useful too. Thanks in advance.
Thanks Iratxe, let me contact Mathematical Psychology that is the obvious outlet for this metric and see if the metric warrants publication according to the chief editor, or whether it is a minor addition even for a brief report, and I will be back to you. In case it does I will give you the title and you can quote it as "Publication Pending", in case it doesn't I will upload the metric and examples on a server and give you a way to reference that.
OK, the subject qualifies for consideration as a brief report in Mathematical Psychology. I will most likely start writing it mid this month. If you have any specific asks concerning what the brief report could include I am very glad to put them in text.
I will most likely code a simple open-source interactive software that will present the result given that the mean and range are imported and add a link in the paper.
I prepared the paper, was ready to submit and I was forwarded the document below. The paper essentially presents the algorithm I was suggesting. Its a shame I spent so much time preparing what was already in existence but the good news for you is that you can use and quote it now :)
It is also a very accessible and an excellent paper.
If you struggle with it, I would be glad to help you through it.
Ce la vie :) and well done professors Cox and Ramirez, beautiful work.
Weir, C. J., Butcher, I., Assi, V., Lewis, S. C., Murray, G. D., Langhorne, P., & Brady, M. C. (2018). Dealing with missing standard deviation and mean values in meta-analysis of continuous outcomes: a systematic review. BMC Medical Research Methodology, 18(1), 25. https://doi.org/10.1186/s12874-018-0483-0
For useful methods to estimate SD from range and sample size:
Wan X, Wang W, Liu J, Tong T. Estimating the sample mean and standard deviation from the sample size, median, range and/or interquartile range. BMC Med Res Methodol. 2014;14:135.
Walter, S. D., & Yao, X. (2007). Effect sizes can be calculated for studies reporting ranges for outcome variables in systematic reviews. Journal of Clinical Epidemiology, 60(8), 849–852. https://doi.org/10.1016/j.jclinepi.2006.11.003