Power Analysis for mixed-effect models in R

The power of a statistical test is the probability that a null hypothesis will be rejected when the alternative hypothesis is true. In lay terms, power is your ability to refine or "prove" your expectations from the data you collect. The most frequent motivation for estimating the power of a study is to figure out what sample size will be needed to observe a treatment effect. Given a set of pilot data or some other estimate of the variation in a sample, we can use power analysis to inform how much additional data we should collect.

I recently did a power analysis on a set of pilot data for a long-term monitoring study of the US National Park Service. I thought I would share some of the things I learned and a bit of R code for others that might need to do something like this. If you aren't into power analysis, the code below may still be useful as examples of how to use the error handling functions in R (withCallingHandlers, withRestarts), parallel programming using the snow package, and linear mixed effect regression using nlme. If you have any suggestions for improvement or if I got something wrong on the analysis, I'd love to hear from you.

1 The Study

The study system was cobblebars along the Cumberland river in Big South Fork National Park (Kentucky and Tennessee, United States). Cobblebars are typically dominated by grassy vegetation that include disjunct tall-grass prairie species. It is hypothesized that woody species will encroach onto cobblebars if they are not seasonally scoured by floods. The purpose of the NPS sampling was to observe changes in woody cover through time. The study design consisted of two-stages of clustering: the first being cobblebars, and the second being transects within cobblebars. The response variable was the percentage of the transect that was woody vegetation. Because of the clustered design, the inferential model for this study design has mixed-effects. I used a simple varying intercept model:

where y is the percent of each transect in woody vegetation sampled n times within J cobblebars, each with K transects. The parameter of inference for the purpose of monitoring change in woody vegetation through time is β, the rate at which cover changes as a function of time. α, γ, σ2γ, and σ2y are hyper-parameters that describe the hierarchical variance structure inherent in the clustered sampling design.

Below is the function code used I used to regress the pilot data. It should be noted that with this function you can log or logit transform the response variable (percentage of transect that is woody). I put this in because the responses are proportions (0,1) and errors should technically follow a beta-distribution. Log and logit transforms with Gaussian errors could approximate this. I ran all the models with transformed and untransformed response, and the results did not vary at all. So, I stuck with untransformed responses:

Model <- function(x = cobblebars,
  type = c("normal","log","logit")){
  ## Transforms
  if (type[1] == "log")
    x$prop.woody <- log(x$prop.woody)
  else if (type[1] == "logit")
    x$prop.woody <- log(x$prop.woody / (1 - x$prop.woody))

  mod <- lme(prop.woody ~ year,
             data = x,
             random = ~ 1 | cobblebar/transect,
             na.action = na.omit,
             control = lmeControl(opt = "optim",
               maxIter = 800, msMaxIter = 800)
  mod$type <- type[1]


Here are the results from this regression of the pilot data:

Linear mixed-effects model fit by REML
 Data: x 
        AIC       BIC   logLik
  -134.4319 -124.1297 72.21595

Random effects:
 Formula: ~1 | cobblebar
StdDev:  0.03668416

 Formula: ~1 | transect %in% cobblebar
        (Intercept)   Residual
StdDev:  0.02625062 0.05663784

Fixed effects: prop.woody ~ year 
                  Value  Std.Error DF   t-value p-value
(Intercept)  0.12966667 0.01881983 29  6.889896  0.0000
year        -0.00704598 0.01462383 29 -0.481815  0.6336
year -0.389

Number of Observations: 60
Number of Groups: 
              cobblebar transect %in% cobblebar 
                      6                      30 

2 We don't learn about power analysis and complex models

When I decided upon the inferential model the first thing that occurred to me was that I never learned in any statistics course I had taken how to do such a power analysis on a multi-level model. I've taken more statistics courses than I'd like to count and taught my own statistics courses for undergrads and graduate students, and the only exposure to power analysis that I had was in the context of simple t-tests or ANOVA. You learn about it in your first 2 statistics courses, then it rarely if ever comes up again until you actually need it.

I was, however, able to find a great resource on power analysis from a Bayesian perspective in the excellent book "Data Analysis Using Regression and Multilevel/Hierarchical Models" by Andrew Gelman and Jennifer Hill. Andrew Gelman has thought and debated about power analysis and you can get more from his blog. The approach in the book is a simulation-based one and I have adopted it for this analysis.

3 Analysis Procedure

For the current analysis we needed to know three things: effect size, sample size, and estimates of population variance. We set effect size beforehand. In this context, the parameter of interest is the rate of change in woody cover through time β, and effect size is simply how large or small a value of β you want to distinguish with a regression. Sample size is also set a priori. In the analysis we want to vary sample size by varying the number of cobblebars, the number of transects per cobblebar or the number of years the study is conducted.

The population variance cannot be known precisely, and this is where the pilot data come in. By regressing the pilot data using the model we can obtain estimates of all the different components of the variance (cobblebars, transects within cobblebars, and the residual variance). Below is the R function that will return all the hyperparameters (and β) from the regression:

  ## Get the hyperparameters from the mixed effect model
  fe <- fixef(x)
    b<-fe[2] # use the data effect size if not supplied

  mu.a <- fe[1] 

  vc <- VarCorr(x)
  sigma.y <- as.numeric(vc[5, 2]) # Residual StdDev
  sigma.a <- as.numeric(vc[2, 2]) # Cobblebar StdDev
  sigma.g <- as.numeric(vc[4, 2]) # Cobblebar:transect StdDev

  hp<-c(b, mu.a, sigma.y, sigma.a, sigma.g)
  names(hp)<-c("b", "mu.a", "sigma.y", "sigma.a", "sigma.g")

To calculate power we to regress the simulated data in the same way we did the pilot data, and check for a significant β. Since optimization is done using numeric methods there is always the chance that the optimization will not work. So, we make sure the regression on the fake data catches and recovers from all errors. The solution for error recovery is to simply try the regression on a new set of fake data. This function is a pretty good example of using the R error handling function withCallingHandlers and withRestarts.

fakeModWithRestarts <- function(m.o, n = 100,  ...){
  ## A Fake Model
    i <- 0
    mod <- NULL
    while (i < n & is.null(mod)){
      mod <- withRestarts({
        f <- fake(m.orig = m.o, transform = F, ...)
        return(update(m.o, data = f))
      rs = function(){
        i <<- i + 1
  error = function(e){
  warning = function(w){
    if(w$message == "ExceededIterations")
      cat("\n", w$message, "\n")

To calculate the power of a particular design we run fakeModWithRestarts 1000 times and look at the proportion of significant β values:

dt.power <- function (m, n.sims = 1000, alpha=0.05, ...){
  ## Calculate power for a particular sampling design
  signif<-rep(NA, n.sims)
  for(i in 1:n.sims){
    lme.power <- fakeModWithRestarts(m.o = m, ...)
      signif[i] <- summary(lme.power)$tTable[2, 5] < alpha
  power <- mean(signif, na.rm = T)

Finally, we want to perform this analysis on many different sampling designs. In my case I did all combinations of set of effect sizes, cobblebars, transects, and years. So, I generated the appropriate designs:

factoredDesign <- function(Elevs = 0.2/c(1,5,10,20),
                           Nlevs = seq(2, 10, by = 2),
                           Jlevs = seq(4, 10, by = 2),
                           Klevs = c(3, 5, 7), ...){
  ## Generates factored series of sampling designs for simulation
  ## of data that follow a particular model.
  ## Inputs:
  ##   Elevs - vector of effect sizes for the slope parameter.
  ##   Nlevs - vector of number of years to sample.
  ##   Jlevs - vector of number of cobblebars to sample.
  ##   Klevs - vector of number of transects to sample.
  ## Results:
  ##   Data frame with where columns are the factors and
  ##   rows are the designs.

  # Level lengths
  lE <- length(Elevs)
  lN <- length(Nlevs)
  lJ <- length(Jlevs)
  lK <- length(Klevs)

  # Generate repeated vectors for each factor
  E <- rep(Elevs, each = lN*lJ*lK)
  N <- rep(rep(Nlevs, each = lJ*lK), times = lE)
  J <- rep(rep(Jlevs, each = lK), times = lE*lN)
  K <- rep(Klevs, times = lE*lN*lJ)
  return(data.frame(E, N, J, K))

Once we know our effect sizes, the different sample sizes we want, and the estimates of population variance we can generate simulated dataset that are similar to the pilot data. To calculate power we simply simulate a large number of dataset and calculate the proportion of slopes, β that are significantly different from zero (p-value < 0.05). This procedure is repeated for all the effect sizes and sample sizes of interest. Here is the code for generating a simulated dataset. It also does the work of doing the inverse transform of the response variables if necessary.

fake <- function(N = 2, J = 6, K = 5, b = NULL, m.orig = mod,
                 transform = TRUE, ...){
  ## Simulated Data for power analysis
  ## N = Number of years
  ## J = Number of cobblebars
  ## K = Number of transects within cobblebars
  year <- rep(0:(N-1), each = J*K)
  cobblebar <- factor(rep(rep(1:J, each = K), times = N))
  transect <- factor(rep(1:K, times = N*J))

  ## Simulated parameters
    b <- hp['b']
  g <- rnorm(J*K, 0, hp['sigma.g'])
  a <- rnorm(J*K, hp['mu.a'] + g, hp['sigma.a'])
  ## Simulated responses
  eta <- rnorm(J*K*N, a + b * year, hp['sigma.y'])
  if (transform){
    if (m.orig$type == "normal"){
      y <- eta
      y[y > 1] <- 1 # Fix any boundary problems.
      y[y < 0] <- 0
    else if (m.orig$type == "log"){
      y <- exp(eta)
      y[y > 1] <- 1
    else if (m.orig$type == "logit")
      y <- exp(eta) / (1 + exp(eta))
    y <- eta
  return(data.frame(prop.woody = y, year, transect, cobblebar))

Then I performed the power calculations on each of these designs. This could take a long time, so I set this procedure to use parallel processing if needed. Note that I had to re-~source~ the file with all the necessary functions for each processor.

powerAnalysis <- function(parallel = T, ...){
  ## Full Power Analysis
  ## Parallel
    cl <- makeCluster(7, type = "SOCK")
    clusterEvalQ(cl, source("cobblebars2.r"))
  ## The simulations
  dat <- factoredDesign(...)

  if (parallel){
    dat$power <- parRapply(cl, dat, function(x,...){
      dt.power(N = x[2], J = x[3], K = x[4], b = x[1], ...)
    }, ...)
  } else {
    dat$power <- apply(dat, 1, function(x, ...){
      dt.power(N = x[2], J = x[3], K = x[4], b = x[1], ...)
    }, ...)


The output of the powerAnalysis function is a data frame with columns for the power and all the sample design settings. So, I wrote a custom plotting function for this data frame:

plotPower <- function(dt){
  xyplot(power~N|J*K, data = dt, groups = E,
         panel = function(...){panel.xyplot(...)
                               panel.abline(h = 0.8, lty = 2)},
         type = c("p", "l"),
         xlab = "sampling years",
         ylab = "power",
         strip = strip.custom(var.name = c("C", "T"),
           strip.levels = c(T, T)),
         auto.key = T

Below is the figure for the cobblebar power analysis. I won't go into detail on what the results mean since I am concerned here with illustrating the technique and the R code. Obviously, as the number of cobblebars and transects per year increase, so does power. And, as the effect size increases, observing it with a test is easier.

Author: Todd Jobe <toddjobe@unc.edu>

Date: 2009-09-18 Fri

HTML generated by org-mode 6.30trans in emacs 22

Labels: , , , , ,


Blogger Dan said...

This is awesome Todd! I'm trying to adapt your parallel processing code for a randomization procedure I'd like to speed up.
Dan McGlinn

October 26, 2009 at 12:25 PM  
Blogger Integrated said...

this is really complex

Online Tutoring

August 15, 2010 at 12:02 AM  
Blogger Integrated said...

This is remarkable work

Online Tutoring,Home Tutor,Private Tutor

August 23, 2010 at 12:02 PM  
Blogger martha said...

This is really very complicated and comlex, I don't how people can solve this.

bioinformatics training india

September 22, 2010 at 5:36 AM  
Blogger Shagor said...

This blog is nice and amazing. I love your post! It's also nice to see someone who does a lot of research and has a great knack for ting, which is pretty rare from bloggers these days.
Thanks a lot!
Pilot license

January 29, 2012 at 1:54 PM  
OpenID Muhammad said...

Thanks for sharing very nice info. your post is very informative and awesome and

its very helpful for the reader.I like your post.Keep it


June 18, 2013 at 10:39 AM  
Blogger Jenis Broad said...

You have some really good ideas in this article . I am glad I read this. I agree with much of what you state in this article. Your information is thought-provoking, interesting and well-written. Thank you.

Term Papers Essay Services

July 10, 2013 at 3:16 AM  
Blogger Arun Kumar said...

I would like to thank for the efforts you have put in writing this blog. I am hoping the same high-grade blog post from you in the upcoming as well. In fact your creative writing abilities has inspired me to get my own blog now. Really the blogging is spreading its wings quickly. Your write up is a good example of it.
Best security system for home in Bangalore

August 6, 2013 at 9:06 AM  
Blogger Some Man said...

great blog man, really helped with some of my research.

See secret documents about Nikola Tesla

August 17, 2013 at 3:37 PM  
Blogger ali raza said...

I love the way you are teaching us. Waiting to read more from you. I love to design websites.

August 30, 2013 at 4:06 AM  
OpenID electricfireplaceheaters said...

http://electricfireplaceheater.org/component/k2/item/66-10-best-electric-fireplace-heaters-by-user-reviews.html - find best electric fireplace heaters ...

October 29, 2013 at 2:49 PM  
Blogger romunov said...

This is not spam. :) Unfortunately links to figures don't work any more. Can you look at this?

November 6, 2013 at 8:38 AM  
Blogger laurawkelly said...

The Apache® Hadoop project is a framework enabling distributed processing of Hadoop Online Training large data sets through a network of computers using simple programming models.
It is born to scale single servers up to thousands of machines, each that can offer local computation and storage. Hadoop Online Training To deliver high availability and uptime of 99% and more rather than relying on hardware, the library can itself detect and handle failures at the application layer.
So delivering a value based service on top of a network of computers Hadoop Online Training which may be prone to failures is the objective that is attained with the Hadoop project.

February 21, 2014 at 11:53 PM  
Blogger Richard C. Lambert said...

Collect and learn them". William Penn, founder of the State of Pennsylvania.makita impact driver

November 3, 2015 at 1:34 AM  
Blogger Richard C. Lambert said...

Given a set of pilot
data or some other estimate of the variation in a sample, we can use
power analysis to inform how much additional data we should collectc10fce2 review

November 29, 2015 at 6:02 AM  
Blogger henris said...

Thank you for posting it will be helpful by knowing more about GDAL. Thank you and please keep update like this with this site. Definitely it will be useful for all.

SQL DBA Training in Chennai

March 2, 2016 at 7:50 AM  
Blogger Unknown said...

Hi Todd,

Thanks for this post, it's very helpful. A couple of the images are currently unavailable though. I was wondering if you could possibly fix them?

March 5, 2016 at 3:01 PM  
Blogger Trong Trong said...

Excellent info here, I am currently doing some research and found exactly what I was looing for. my review here

March 10, 2016 at 6:54 PM  
Blogger Trân Huyền said...

Really good job on the site, Thanks for guide! best 14 bandsaw

March 12, 2016 at 3:33 AM  
Blogger Aamala Ahona said...

All are saying the same thing repeatedly, but in your blog I had a chance to get some useful and unique information, I love your writing style very much, I would like to suggest your blog in my dude circle, so keep on updates.

Peridot Systems Adyar Contact Number

March 22, 2016 at 8:14 AM  
Blogger Jubaer said...

Really Nice post Todd, you just made my problem solved. Thanks again. web design

April 13, 2016 at 3:35 PM  
Blogger Karis said...

The image with your model structure doesn't load in my browser. I am not sure if this is a problem with my firewall or a broken link.

May 24, 2016 at 4:18 PM  
Blogger Training Coimbatore said...

great post

Hadoop training in coimbatore
Java training in coimbatore
Oracle training in coimbatore
Informatica training in coimbatore
Oracle training in coimbatore
Informatica training in coimbatore

June 22, 2016 at 11:33 AM  
Blogger Prashanthi said...

Your post is really great. Thank you for taking time to provide us some of the useful and exclusive information with us. Keep on blogging!!
Best SharePoint Training Institute in Chennai

August 16, 2016 at 6:07 AM  
Blogger poori konidala said...

TS DSC District wise Vacancies list 2017
TS Forest Recruitment 2017
TSTRANSCO AE, AEE Recruitment 2017
TSGENCO AE, AEE Recruitment 2017

May 10, 2017 at 1:14 AM  
Blogger sri harsha said...

TS DSC 2017
Telangana DSC 2017
TS DSC Notification 2017
TS DSC Model Papers 2017
Sakshi TS DSC Model Papers 2017
Namasthe Telangana TS DSC Model Papers 2017
TS DSC District wise Vacancies list 2017

June 4, 2017 at 10:58 AM  
Blogger sri harsha said...

Telangana DSC Recruitment 2017 Notification is announced in last week of December with District wise Vacancies lists TS DSC 2017 , Telangana DSC 2017, TS DSC Notification 2017, TS DSC Model Papers 2017, Sakshi TS DSC Model Papers 2017, Namasthe Telangana TS DSC Model Papers 2017, TS DSC District wise Vacancies list 2017 Government Primary and Secondary Schools for Secondary Grade Teacher (SGT),Language Pandit (LP),School Assistant(SA)and Physical Education Teacher (PET) Posts with the TS DSC 2017.

June 26, 2017 at 1:45 PM  
Blogger sandhosh said...

Really Good blog post.provided a helpful information about power analysis of mixed model.keep updating...
Digital marketing company in Chennai

July 3, 2017 at 6:30 AM  
Blogger sri harsha said...

All the board students can Download Subject wise new Syllabus for all state and central board schools with model papers 2018 for all board of secondary education boards and students can apply scholarship 2018 for ekalyan epass to online application and status check

July 3, 2017 at 11:35 AM  
Blogger kingrani said...

SSC JHT Paper II Exam Results
UPTET Recruitment Notification

August 8, 2017 at 6:35 AM  
Blogger ghulam ali said...

Best Dyson Vacuums

August 15, 2017 at 9:37 AM  
Blogger Philips Huges said...

Its a wonderful post and very helpful, thanks for all this information. You are including better information regarding this topic in an effective way.Thank you so much

Personal Installment Loans
Title Car loan
Cash Advance Loan

August 28, 2017 at 12:22 PM  
Blogger Unknown said...

Direct Link to fill the application form of DRDO Recruitment 2017

Apply Online Punjab and Haryana High Court Clerk Recruitment 2017

September 5, 2017 at 8:25 AM  
Blogger kingrani said...

Kolkata Police Recruitment Notification - Apply Online here
Click here to Download Public Service Commission Recruitment Notification

September 5, 2017 at 8:41 AM  
Blogger Alastair Jones said...

Find best vacuum for pet hairs.

October 29, 2017 at 5:35 PM  
Blogger sophia sophia said...

Thanks for this wonderful post. The information in this article is very helpful to me. Thanks a lot for sharing. Keep blogging.
Truth or Dare Questions
Best questions to ask your girlfriend over text

November 9, 2017 at 12:27 AM  
Blogger zeeshan farhad said...

amazing !!! good work. keep it up. find shredder

November 11, 2017 at 1:24 AM  
Blogger David Smith said...

Hi, Great.. Tutorial is just awesome..It is really helpful for a newbie like me.. I am a regular follower of your blog. Really very informative post you shared here. Kindly keep blogging. If anyone wants to become a Java developer learn from Java Training in Chennai. or learn thru Java Online Training in India . Nowadays Java has tons of job opportunities on various vertical industry.

December 7, 2017 at 11:48 PM  
Blogger Anthony Paul said...

Check this
Check this
Check this
Check this
Check this
Check this
Check this
Check this
Check this
Check this

December 20, 2017 at 12:33 AM  
Blogger kingrani said...

I have read your article, it is very informative and helpful for me.I admire the valuable information you offer in your articles. Thanks for posting it.

Tatacliq Cashback

January 11, 2018 at 7:20 AM  
Blogger sophia sophia said...

Thanks for this wonderful post. Keep blogging.
If you want to know latest technical news click link below.

January 25, 2018 at 8:13 AM  
Blogger sophia sophia said...

Thanks for this wonderful post. Keep blogging.
If you want to know latest technical news click link below.

January 25, 2018 at 8:13 AM  

Post a Comment

Subscribe to Post Comments [Atom]

<< Home