A Study of Clustered Data and Approaches to Its Analysis

Abstract

Editor's Note: Toolboxes are intended to describe and evaluate methods that are becoming widely relevant to the neuroscience community or to provide a critical analysis of established techniques. For more information, see http://www.jneurosci.org/misc/ ifa_minireviews.dtl. Statistical analysis is critical in the interpretation of experimental data across the life sciences, including neuroscience. The nature of the data collected has a critical role in determining the best statistical approach to take. One particularly prevalent type of data is referred to as "clustered data." Clustered data are characterized as data that can be classified into a number of distinct groups or "clusters" within a particular study. Clustered data arise most commonly in neuroscience when data are compiled across multiple experiments, for example in electrophysiological or optical recordings taken from synaptic terminals, with each experiment providing a distinct cluster of data. However, there are many other types of experimental design that can yield clustered data. Here, we provide a statistical model for intracluster correlation and systematically investigate a range of methods for analyzing clustered data. Our analysis reveals that it is critical to take data clustering into account and suggests appropriate statistical approaches that can be used to account for data clustering. A Study of Clustered Data and Approaches to Its Analysis

A solution to dependency: using multilevel analysis to accommodate nested data

“…The problem of nesting Nested designs are designs in which multiple observations or mea surements are collected in each research object (for example, animal, tissue sample or neuron/cell) 7 . Consider the following fictive, yet representative, research results.…”

“…Alternatively, multilevel analysis can be circumvented by conduct ing conventional analyses on cluster based summary statistics, for example, by performing a t test on the means or medians calculated in each cluster. Although this strategy is statistically valid, informa tion contributed by the individual observations is lost, and, relative to multilevel analysis, statistical power to detect the experimental effect of interest decreases 7,9,10 . Conducting t tests on cluster based means instead of multilevel analysis on all observations results in up to a 40% loss of statistical power, depending on the number of clusters in the study and the ICC (Fig.…”

Feed-forward motor control of ultrafast, ballistic movements

“…5). Linear mixed effects models for clustered data were used to account for inter-animal variability (Galbraith et al, 2010). Data clusters were defined by each individual's strikes.…”

“…Then, we examined whether strike kinematics and spring compression could be predicted by the duration of particular motor phases, the number of motor spikes during particular motor phases and the timing of spike generation across motor phases. We performed alternative hypothesis testing through the use of Akaike's information criterion (AIC) and linear mixed models (Bolker et al, 2008;Galbraith et al, 2010), thus allowing a robust examination of the key control variables (Gordon et al, 2015;Lu et al, 2015). If mantis shrimp vary their strikes and this variation can be explained by prior variation in motor activity, then these animals exhibit feed-forward control (Dickinson et al, 2000;Kubow and Full, 1999;Nishikawa and Gans, 1996).…”

Sensitivity analyses for clustered data: An illustration from a large-scale clustered randomized controlled trial in education

“…Both Peters et al (2003) and Galbraith et al (2010) (3) Finally, as Peters et al (2003) note, the number of clusters as well as the distributional assumptions of the data within clusters (i.e., the extent to which data are normally distributed or skewed) can influence the overall findings.…”