search

itchyshin / castration_meta-analysis

Castration meta-analysis
https://itchyshin.github.io/castration_meta-analysis/
MIT License
1 stars 0 forks source link

Problematic data (rows) #3

Closed itchyshin closed 3 years ago

itchyshin commented 3 years ago

@Mike-Garratt - these are problematic data I am excluding them at the moment (we can add them in later once we have data)

  1. Benedusi et al. 2015 - you said you emailed the authors - did you get data?
  2. Oneil et al. 2013 - CI is only provided for 2 rows but you have 4 rows (I do not really understand your comment)
  3. Bronson 1988 - no information is available

Please let me know what you will be doing for these

Mike-Garratt commented 3 years ago

Hi Shinichi,

  1. I didn't get any reply from the author. I am guessing that we probably won’t get it so we can exclude for now.

  2. For the two Oneil papers they only provide linear regression results for each sex and neuter status group against females who are intact. So they provide a parameter estimate for female neutered versus female intact (e.g. 0.6 years) but then male intact versus female intact, and also male castrated versus female intact. So all results are relative to females who are intact. Is there any way that we can use this to calculate the difference between male intact and male castrated? Ive also attached a screenshot of the table where these results are displayed in the paper to you can see what I mean.

image

We also have the same problem for Oneil 2015.

  1. Are you talking about Bronson 1981? I’ve actually extracted the data for this paper now – what’s the best way to add the data to GitHub?

Let me know if you have any other questions!

Mike

itchyshin commented 3 years ago

@Mike-Garratt - thanks for this.

  1. Got it
  2. OK - I can try to get data from the table
  3. please keep your master excel file and add the data and send to me (I will send you the version with Oneil data on. Thanks
itchyshin commented 3 years ago

@Mike-Garratt - actually I realise that we have the latest excel in our shared Dropbox folder.

So all fine. I will send you to ask convert CI, SEM etc to SD for Mean and median type data - a bit later so please stay tuned

itchyshin commented 3 years ago

@Mike-Garratt

Do you think you could create a new column for SD? (you just need to do this for Lifespan_parameter == Mean or median - not survival rate or likelihood)

So you need to turn SEM to SD by

SD = sqrt(n)*SEM

SD = 95% CI (the while width) /(2*1.96) assuming we have enough n (which is fine for all of your CIs)

For interquartile range - please use the formula from this paper

https://bmcmedresmethodol.biomedcentral.com/articles/10.1186/1471-2288-14-135

Note that we assume medians are close enough to means and also SD for rate (%) data follow binomial distributions - in such a case, I do not need SD for this type of data.

Thanks! - if not clear, do let me know (have a nice weekend!)

itchyshin commented 3 years ago

BTW @Mike-Garratt - please take time - meanwhile I will do more coding using % part of the dataset - this bit is easy to get Cohen's d

Mike-Garratt commented 3 years ago

Thanks Shinichi, that all makes sense. I'll try to get it done on Wednesday and Thursday after I finish some teaching!

Mike-Garratt commented 3 years ago

Hi Shinichi,

I have now converted everything over to SD, which is shown as new columns in the final extracted data file. There are a two studies I am stuck on. One does not present any variance parameter in the paper (Sichuk 1965) and the other does not provide the exact sample size for the castrated group (but error is provided) (Moore et al 2001). Any ideas if there is a way to create usable data? Can provide more information if required.

For calculating SD from CI, I am assuming that we need to also need to multiple by sqrt (n) after we do the calculation you outline above?

So SD = sqrt(n)(95% CI (the while width) /(21.96))

This is what it is outlined as in the chapter below and without that extra calculation the results are abnormally small.

https://handbook-5-1.cochrane.org/chapter_7/7_7_3_2_obtaining_standard_deviations_from_standard_errors_and.htm

Apart from that I think we are nearly there!

Let me know if you have any other thoughts.

Thanks

Mike

itchyshin commented 3 years ago

Hi, Mike

Thanks for this

I will look into (Sichuk 1965) and (Moore et al 2001)

I think you got the formula a bit wrong

Not - SD = sqrt(n)(95% CI (the while width) /(21.96)) SD = sqrt(n)(95% CI (the while width) /(2*1.96))

I managed to do a bit on % data - there is a huge effect (a bit more later)

Thanks

Shinichi

itchyshin commented 3 years ago

@Mike-Garratt

Just a little update - so castration does reduce the risk of death (or live longer) - regardless of sex. The effect is a 50% decrease in the odds of death so pretty substantial it seems.

I am just using the % data at the moment but will get to Mean data soon.

Best

Shinichi

Mike-Garratt commented 3 years ago

Thanks Shinichi, that sounds exciting.

Just a couple of notes regarding my conversion of all the variation parameters to SD. I'm just going to paste them here so we have them documented.

There was a typo in my last message to you, so the caculation for SD from CI was correct. Just to confirm, I used the whole width of the CI for this calculation. So if the CI was 14.03-14.16 I used 0.13 as the 95% CI whole width.

For Kirkman and Yau no variance information was provided in the results, just mean lifespan per group. However, they did provide various graphs that showed how many animals died in 100 day brackets. I used this to estimate the interquartile range, which I then used to estimate SD. It's a fairly rough approximation. I will paste a pic of how I did this from the figures.

Page_01.png

For data were we have probable error, I used probable error/0.6745 to get standard deviation. This is according to wikipedia for how to calculate probable error.

Thanks for your work on this! Let me know what else I can do.

Mike

itchyshin commented 3 years ago

@Mike-Garratt

All good - I will have look at these values and see.

Have a good Easter break

Best

Shinichi

itchyshin commented 3 years ago

@Mike-Garratt (hope you had a good Easter break)

I have changed a few things in the "Final extracted data.xlsx" file in our dropbox.

  1. I make column names a string - e.g. "Lifespan unit" to "Lifespan_unit"
  2. Mooe et al (2001) study - you have the total N of 927 and out of these, 641 got castrated. I divided the rest to the control and opposite-sex groups (also you put SE but it was SD so fixed for this paper)
  3. Sichuk did not have SD but we replaced with SD values of Kirkman and Yau as both studies from Hamsters and the same species should have very close SDs

Some assumptions I will make (I will do this within R) - if you are against them, please let me know and probably provide an alternative.

  1. We have a lot of missing Ns for opposite-sex groups - we replace NA with the smallest of either Control or Treatment N
  2. We have a lot of missing SD for opposite-sex groups - we replace NA with the mean of control and treatment SDs

By doing the above, we deal with most of the missing value issues for getting effect size = Hedges' g.

More later and best

Shinichi

Mike-Garratt commented 3 years ago

@itchyshin (Happy Easter too, hope you managed some family time ;))

Thanks for changing the column names to string, sorry for my initial laziness.

For Moore et al 2001, I am happy with the way you have estimated the N for castrated males. However we do not have any opposite sex data for the Moore paper. They only had intact males, castrated males and spayed (ovariectomized) females in their population, and the remaining animals in that study are spayed females. Because we do not have any intact females to compare to we cannot use the spayed female data in this study.

For the assumptions you will make in R:

Where we are missing opposite sex N and SD, this is usually because we do not have any data on the lifespan/surivival for the opposite sex. In these studies they only compared the effect of sterilization on lifespan within the one sex.

I was thinking about using the opposite sex data for a secondary analysis to test whether sterilization reduces the sex differences in aging, but for the first analysis we will not use this column of data.

So if there are studies were we have a survival value for the opposite sex but no N or SD, I am happy with your assumption, but I do not think that there are many of these.

Let me know if that makes sense or we need to discuss further!

Thanks for all your work on this!

Mike

itchyshin commented 3 years ago

@Mike-Garratt

OK - this makes sense (sorry I should have been more careful with this). OK, it seems like we are ready to go. I will get some initial results by Monday.

Shinichi

itchyshin commented 3 years ago

@Mike-Garratt

I will soon send some preliminary results. Actually, I found some mistake in my initial code. Anyway, when we account for phylo and shared controls etc, overall and sex effects are not significant.

Anyway, there is a lot of variation to be explained.

Do you think you could get the timing of castration - here is my suggestion and interested to hear what @Jeff-Lemaitre thinks too

  1. If you could turn all the timing data into months
  2. For all 14 species, can you get the age of sexual maturity - @Jeff-Lemaitre do you have this data? - I think you could probably get at AnAge or scrape from the web or Wiki
  3. Then, you can get the ratio between this timing data and maturation age - so we can see whether this explains the huge variation we see in the data

Also, there are no other moderators to test? apart from this and Sex.

Mike-Garratt commented 3 years ago

Hi Shinichi,

Thanks for making so much progress with the analysis! The other main moderators are environment, type of sterilization and the sex difference in lifespan.

I would predict that that in wild conditions we will see stronger effects of sterilization, particularly in females. I have left the human data blank for this parameter, maybe we should just call this human.

Type of sterilization is also an important parameter. This is only really relevant in females, where several different approaches have been used, and only ovariectomy influences sex hormone levels.

I've attached a word document that is in our shared dropbox, that outlines some of the important tests I have been thinking of.

We may also want to make another moderator that accounts for the type of experiment - some of these studies randomly allocated to control and treatment groups, while others like in humans are less controlled and just make some historical/prospective analyses.

Good ideas about the age at castration. Let's see what Jeff says about the sexual maturity data - it may also be possible to change this moderator into a categorical variable of "before sexual maturity" "at sexual maturity" and "after sexual maturity", although I like your idea better.

Thanks again for all your work!

Mike

castration meta analysis_Questions and proposal for analysis.docx

Jeff-Lemaitre commented 3 years ago

Hi Mike & Shinichi,

Thanks for all the work you have made so far. Just a few comments:

All best wishes, Jeff

itchyshin commented 3 years ago

@Mike-Garratt - I have some ideas on how to get effect size out of ONeill papers - discuss at our meeting.

We need to assume a few things but we can discuss this

assumption about N assumption about SD