Видео

This was the reference I used. www.graphpad.com/guides/prism/latest/statistics/stat_sample_size_for_nonparametric_.htm

The effect size you use depends on the effect size you expect your future experiment (the one you want to run and this estimate sample size for). The main point of changing the effect size in the example is to show that in the SIMR framework, you can manually change the effect size. So maybe you think the effect size of 0.11 from the dataset is an accurate estimate of what the future experiment would produce. Then, use the 0.11. But perhaps you think that the real effect size is smaller than that, say about 1/2 that size. Then, adding the effect size of 0.05 manually would let you generate the sample size for such an effect.

@@daccotabiostatistics Clear explanation! Thank you very much!😀

I am not sure what you mean.

That is a good question. I think the sufficient power is for the overall model. Then, if the overall model is significant, you'd look at the F-values for specific variables and then compare mean values across different levels for what the significant difference was.

I don't have plans to repost without music in the short term. You can access the content in slide format using the link below: med.und.edu/research/daccota/_files/pdfs/berdc_resource_pdfs/sample_size_r_module_glmm2.pdf

med.und.edu/research/daccota/_files/pdfs/berdc_resource_pdfs/sample_size_r_module_glmm2.pdf

The mean and standard deviations for the t-test (m1=20, m2=25, std1=10, std2=10) were just used for an example. There is no underlying data. However, if you wanted to, you could use the rnorm function in R to generate sample data. rnorm(n=10, mean=20, sd=10) rnorm(n=10, mean=25, sd=10)

My understanding would be that if the data is what you actually want to analyze, yes, it would be post-hoc. But if the data is preliminary data to get an estimate for a full study, I'd consider it a priori.

Not that I've made. I haven't found a good approach to GLMs in R for power analysis.

@@daccotabiostatistics thank you. I think SIMR works on glm() objects as well.

you can estimate power in GLMs using the pwr function-very simple

Looking in WebPower manual (cran.r-project.org/web/packages/WebPower/WebPower.pdf), Cluster randomized trials and multisite randomized trials are described the same way, as shown below. "[CRT/MRT] are a type of multilevel design for the situation when the entire cluster is randomly assigned to either a treatment arm or a control arm (Liu, 2013). The data from [CRT/MRT] can be analyzed in a two-level hierachical linear model, where the indicator variable for treatment assignment is included in second level. If a study contains multiple treatments, then multiple indicators will be used." In the variable inputs between the two functions, CRT is much simpler, just needing the icc in addition to the number of clusters, alpha, power, and effect size. MRT needs three variances (tau00, tau11, sg2). I think this is because the MRT can look at three types of tests (main, site, and variance).

@@daccotabiostatistics Thank you so much for clarification.

From what I understand, linear mixed models (and other similar models) are more complex, so simulation is better than replying on specific formulas (see: stats.stackexchange.com/questions/373597/how-can-you-compute-sample-size-for-a-linear-mixed-model-gpower-only-does-repe). My default for power analysis is G*power, but it doesn't have anything for more complex linear models. Therefore, I found SIMR in R for such things. I suppose you could do something like run power analysis for linear multiple regression in G*power and just ignore the random effects. As for code not working, if you could provide more details, I might be able to help. SIMR doesn't always work smoothly for me either.

I do not have a publication on this. These were put together as video/pdf source. You could cite the PDF document. med.und.edu/daccota/_files/pdfs/berdc_resource_pdfs/sample_size_gpower_module.pdf

I am not very familiar with partial least squares (PLS) methods. However, there does appear to be a paper that goes through a tutorial on power analysis in R: www.researchgate.net/publication/281690984_Sample_Size_Determination_and_Statistical_Power_Analysis_in_PLS_Using_R_An_Annotated_Tutorial

It is hard to tell without seeing the full code. But I don't see how the arrows function and a list of numbers as characters would be used in running a power curve.

@@daccotabiostatistics if you don't mind, can I email it to you sir?.. so you could have a look and give an opinion.. its important to me

The format (1|group) represents a random group intercept. The number 1 is sort of a placeholder. So for (1|herd), herd is the random effect with random intercept. The format (X|group) represents a random slope and random intercept, that is a random slope of X within group with correlated intercept. This is a good resource I found that explains random effects models in R: bbolker.github.io/mixedmodels-misc/glmmFAQ.html

@@daccotabiostatistics Thank you so much for your kind reply and for your time. I really appreciate it 🙏🏼 And the link that you shared is sooo good. Many thanks. Wishing you a good day :)

I have not run a multinomial logistic regression for repeated data, so I can't say for sure. I would start with trying out a simple case of multinomial logistic regression. Here is a good source from UCLA for trying that out (stats.idre.ucla.edu/r/dae/multinomial-logistic-regression/). From there, you could test if the multinom function can handle repeated data. Otherwise, I did find this post that may be useful (hlplab.wordpress.com/2009/05/07/multinomial-random-effects-models-in-r/). Hopefully those sources help.

@@daccotabiostatistics thank you :-)

You'll want to make sure your model has both a fixed and a random effect, glmer(y~x+(1|z)). I tried just having a fixed effect and got the same problem. Below is R code that I got to work for a simple simulated dataset drawn from a gamma distribution. library(lme4) library(simr) #create simulated dataset mygamma1 <- rgamma(n=20, shape=2, rate=2.0) mygamma2 <- rgamma(n=20, shape=2, rate=2.5) hist(mygamma1) hist(mygamma2) gamma_df <-data.frame('treatment'=c(rep(1,20),rep(2,20)), 'block'=rep(seq(1,10,1),4), 'rate'=c(mygamma1,mygamma2)) View(gamma_df) boxplot(rate~treatment, data=gamma_df) #create model gamma_glm <-glmer(rate~treatment+(1|block), family="Gamma",data=gamma_df) summary(gamma_glm) fixef(gamma_glm)["treatment"] powerSim(gamma_glm, nsim=50) #extend number of blocks gamma_glm_ext <- extend(gamma_glm, along="block", n=60) gamma_PC <-powerCurve(gamma_glm_ext, along="block", breaks=c(20,30,40,50,60)) print(gamma_PC)

​@@daccotabiostatistics Thanks very much! I will work through the code provided in the morning (it's way past bedtime here in Australia). But just as a reference, my model looks somewhat similar to the model you created: glmer(gamma_distributed_score ~ time + age + sex + score.1 + (1 | patient.id), family = Gamma(link="identity"), data = mydata) ### time has only two levels. Thanks again for your help will let you know how I go!

Thanks

The 'test' option specifies the test to perform. Otherwise, the first fixed effect will be tested by default. So, in the example, I wanted to look at 'time' specifically and then 'treat' specifically, so I set it that way.

Thanks for the feedback. I've been working to make sure the videos don't have sudden shifts in volume.