Overview

ACLUSTER statistical software estimates the significance of the treatment effect when the unit of randomization is a group or cluster. A common type of cluster randomization occurs when a lab scientist assigns an entire litter to a given treatment. Standard statistical procedures that estimate significant treatment effect assume that individuals instead of clusters were randomized to the various treatments. These procedures do not account for correlation in response to treatment among members of a cluster, and thus underestimate the variance and likewise overstate the significance of treatment differences.

ACLUSTER estimates the within-group correlation in treatment response and uses this correlation to estimate variances. Thus, it avoids overstating the significance of treatment differences. ACLUSTER provides analyses of continuous, binomial, and survival (time-to-event) data. Relative to other statistical software packages designed for the analysis of clustered data, ACLUSTER is inexpensive and easy to use if one is familiar with ANOVA analyses. If one has some estimate of inter- and intracluster variance, ACLUSTER provides methods for estimating sample sizes for many different study designs. I generated a data set with responses, containing both intracluster and intercluster variance, randomized to two groups and tested ACLUSTER on the simulated data set. The output estimates of means and intracluster correlation were reasonable. One would have to run many simulations to have a good estimate of the program's accuracy. As the theory underlying the program would predict (given the simulated variance was both among and within clusters), the ACLUSTER estimates did produce a more conservative estimate of the significance of group differences than the standard ANOVA estimates from the SAS GLM procedure.

How Long Did It Take to Learn to Use It Productively?

As a statistician with over 15 years of experience with various software packages, I needed just a few hours to install the software and run some examples on the sample data provided. However, to really understand ACLUSTER'S approach to cluster analyses, I spent several days reading the 70 pages of online documentation and a book titled Design and Analysis of Cluster Randomization Trials in Health Research by Allan Donner and Neil Klar. The documentation for ACLUSTER states that the program is based on the simpler analyses in this book. The book is well written with a good mix of theory and example.

Product Quality

Customizability

ACLUSTER is a menu-driven program. One can stipulate the pathways for the sample-size parameter and input and output data. Under Options, one can also set the level of significance for analysis and the power for sample-size calculations. ACLUSTER creates a report file with a .LST extension that can easily be opened and edited in Microsoft Word. The output for each type of analysis is fixed.

Ability to Program in Scripts, Add Extension Modules, etc.

None.

Ability to Import and Export in Different File Formats

ACLUSTER can import numerical data from a flat text file (no tabs or commas). One must type in a data dictionary according to an ACLUSTER specified format in order to read in the data. The format allows the user to provide text labels for various numerical values of a variable.

Useful or Unusual Features

ACLUSTER calculates sample size and analyzes treatment effects for trials with continuous, binomial, and time-to-event data. It will accommodate clusters of different sizes and variances. The program also allows for matched-pair and stratified designs. ACLUSTER can calculate the intracluster correlation coefficient (ICC) using either aggregate data or individual records. The ICC is useful in determining the effect of clustering on the treatment response and for calculating sample sizes in future trials using the same type of clustering. The underlying assumptions for the various statistical tests are documented in the output. The output provides various parametric and nonparametric statistical tests and confidence intervals, along with design information, without being overly verbose. Among the newer nonparametric computer-intensive statistics available with ACLUSTER, Fisher's two-sample permutation test requires no distributional assumptions about the data.

Limitations

In order for ACLUSTER to analyze the data set, it must be aggregated by cluster. Instead of containing individual records, the data set should contain parameters such as mean, number of responses, and standard deviations for each cluster. The producers of ACLUSTER assume that the user can easily obtain these parameters from other standard statistical packages. However, ACLUSTER cannot import data directly from other packages. One must export data to a flat text file containing only numerical data and then read the file into ACLUSTER using a carefully formatted dictionary located in a separate file for the data. Creating the dictionary could be tedious if there are many variables and variable labels needed. The input format is also quite rigid and literal. When I left a blank space at the bottom of my file, it was read as a new group record.

Although the program does allow for stratification, covariates cannot be entered in the model. Since ACLUSTER does not utilize command files, the program is not suitable for production work with large numbers of analyses. File names are restricted to eight letters.

Only online support is readily available. I called the only telephone number listed with the software to ask if ACLUSTER would read in character data. That number reached a person at a company that handles ACLUSTER sales, who agreed to ask around but who did not call back. Subsequently, when I tested the program on simulated data, the program would not read in data with minus signs. I received an error message stating that I could not read in character variables. In order to read in the data, I ran a new simulation and eliminated negative numbers by adding 20 units to all values. This is always one way to work around such a problem. The addition won't affect variance estimates. Just remember to subtract the same number of units from the mean estimates once the analysis is finished.

Comparisons with Similar Software

Statistical software such as S-Plus, SAS (Proc Mixed), and SUDAAN also provide procedures that allow for clustered data in the statistical analysis model. These packages provide even greater flexibility in trial design and output features than ACLUSTER. They also have more prompt and extensive (off- and online) technical support than ACLUSTER; however, they are much more complex and expensive to use than ACLUSTER. If your trial design is among those that ACLUSTER includes, the program could be a practical and cost-effective alternative.

Technical Support and Documentation

ACLUSTER includes thorough and clearly outlined online documentation. The mathematical models for each analysis are carefully annotated. There is an example with sample data to be tested for each type of analysis. The available Web site and the telephone number provided with the program provide information only on how to purchase the program. When ordering the program, I highly recommend also ordering the book upon which the program is based: Design and Analysis of Cluster Randomization Trials in Health Research by Allan Donner and Neil Klar. The book is available from Arnold Publishers for £35 (about $50) and from Amazon for $60.

Target Users

Researchers working on design and analyses of randomized trials.