Assessing reliability in neuroimaging research through intraclass effect decomposition (ICED)
Abstract
Magnetic resonance imaging has become an indispensable tool for studying associations of structural and functional properties of the brain with behavior in humans. However, generally recognized standards for assessing and reporting the reliability of these techniques are still lacking. Here, we introduce a new approach for assessing and reporting reliability, termed intraclass effect decomposition (ICED). ICED uses structural equation modeling of data from a repeatedmeasures design to decompose reliability into orthogonal sources of measurement error that are associated with different characteristics of the measurements, for example, session, day, or scanning site. This allows researchers to describe the magnitude of different error components, make inferences about error sources, and inform them in planning future studies. We apply ICED to published measurements of myelin content and resting state functional connectivity. These examples illustrate how longitudinal data can be leveraged separately or conjointly with crosssectional data to obtain more precise estimates of reliability.
https://doi.org/10.7554/eLife.35718.001Introduction
Neuroimaging techniques have become indispensable tools for studying associations among brain structure, brain function, and behavior in multiple contexts, including aging, child development, neuropathology and interventions, with concerted efforts increasingly focusing on comprehensive quantitative analyses across multiple imaging modalities (Lerch et al., 2017). Surprisingly, however, generally recognized standards and procedures for assessing and reporting the reliability of measurements and indices generated by noninvasive neuroimaging techniques are still lacking. This state of affairs may reflect the rapid evolution of a research field that straddles several wellestablished disciplines such as physics, biology, and psychology. Each of these fields comes with its own methodology, including conceptualization of error of measurement and reliability, and an articulation of these diverse methodologies into a coherent neuroscience framework is currently lacking. The goal of our contribution is twofold. First, we introduce a signaltonoise perspective that reconciles these seemingly disparate approaches. Second, we apply an analytic framework, based on the ideas of Generalizability Theory (GTheory; Cronbach et al., 1972) and Structural Equation Modeling (SEM) that allows us to separate and gauge various sources of measurement error associated with different characteristics of the measurement, such as run, session, day, or scanning site (in multisite studies). The proposed tool enables researchers to describe the magnitude of individual error components, make inferences about the error sources, and inform them in planning the design of future studies. We proceed without loss of generality but with an emphasis on applications to human cognitive neuroscience.
Materials and methods
Prelude: Coefficient of variation and intraclass correlation coefficient represent different but compatible conceptions of signal and noise
Request a detailed protocolPhysics and psychometrics offer two fundamentally different but equally important and compatible conceptions of reliability and error. Physicists typically inquire how reliably a given measurement instrument can detect a given quantity. To this end, they repeatedly measure a property of an object, be it a phantom or a single research participant, and for expressing the absolute precision of measurement, evaluate the dispersion of the different measurement values obtained from this object to their mean. The prototypical index produced by such approach is the coefficient of variation (CV), which is defined as the ratio of an estimate of variability, $\sigma}_{i$, and a mean, $m$, with i representing the object undergoing repeated measurements:
The interpretability of the CV depends upon the quantity having positive values and being measured on a ratio scale. When these conditions are met, the CV effectively expresses the (im)precision of measurement, with larger values meaning lesser precision. Imagine, for instance, that the same quantity is being measured in the same research participant or the same phantom on two different scanners. All other things equal, comparing the CV obtained from each of the two scanners shows which of the two provides a more reliable (in this case, precise) measurement.
Note that in this context, the scanner with the greater precision may not necessarily yield more valid data, as the mean of its measurements may be further away from the ground truth (see Figure 1). Bearing this distinction in mind, we limit our discussion to the issues of reliability (precision), rather than validity (bias). In Table 1, we list terms used in various disciplines to express the difference between precision and bias. We maintain that the confusion surrounding these concepts may to a large extent reflect terminological differences among disciplines.
In contrast to physics that deals with welldefined objects of measurement, in human neuroscience, we focus on a different meaning of reliability. Informed by psychometric theory and differential psychology, reliability here refers to the precision of assessing betweenperson differences. Researchers concerned with gauging individual differences as a meaningful objective express this form of reliability in a ratio index, termed intraclass correlation coefficient (ICC), which represents the strength of association between any pair of measurements made on the same object. However, instead of relating variance to the mean, the ICC quantifies variance within persons (or groups of persons), in relation to the total variance, which also contains variance between persons (or between groups of persons; cf. Bartko, 1966). Hence, the ICC is a dimensionless quantity bracketed between 0 and 1, and is tantamount to the ratio of variancebetween, ${\sigma}_{B}^{2}$, to the total variance that includes the variancewithin, ${\sigma}_{W}^{2}$:
In repeatedmeasures studies on human participants, the variancewithin corresponds to the variance within each person, whereas the variancebetween represents differences among persons. Thus, for interval or ratio scales, the ICC expresses the percentage of the total variance that can be attributed to differences between persons.
The similarities and differences between CV and ICC become clear when one conceives of both as expressions of signaltonoise ratio. For a physicist, the mean represents the soughtafter signal, and the variation around the mean represents the noise to be minimized. Hence the use of the CV to evaluate measurement precision normalized with the metric of the given scale. For a psychologist interested in individual differences, the betweenperson variation is the signal, and the withinperson variation is regarded as noise. Therefore, a measure that quantifies the contribution of betweenperson differences to the total variance in the data, the ICC, is chosen for this purpose (in other contexts, not discussed in this article, withinperson variability itself may be an important marker of individual differences, e.g., Garrett et al., 2013; Nesselroade, 1991).
Clearly, CV and ICC do not convey the same information. To illustrate this point, we simulated data under two conditions, which show that each measure can be manipulated independently of the other. We illustrate how CV remains unchanged, while drastic changes occur in ICC (see Figure 2). Instead of individual CV values, we report an aggregated CV computed as the squareroot of the average withinperson variance divided by the overall mean. For each condition, we simulated for each of five persons ten repeated measures of a fictitious continuous outcome variable $X$. Across conditions and persons, withinperson variability was identical and only betweenperson variability varied between conditions. In the first condition, the simulated data have identical betweenperson and withinperson variance. As a result, we obtain a low ICC and conclude that the measurement instrument fails to adequately discriminate among persons. However, critically, we also obtain a rather low CV, implying high precision to detect deviation from zero (see left panel of Figure 2). In the second instance, betweenperson standard deviation was larger than withinperson standard deviation by a factor of five. This condition yields a high ICC reflecting the fact that the measure discriminates well among persons. At the same time, CV remains low, which implies reasonable precision of detecting differences from zero. This is because the withinperson variance is still relatively low in comparison to the means (see right panel of Figure 2).
In summary, whereas the CV refers to the precision of measurement obtained from each object, the ICC expresses a partwhole relation of variance observed in the data. All other things being equal, a less precise measurement will increase the variancewithin, and hence compromise our ability to detect betweenperson differences. On the other hand, a rather imprecise measurement (as indexed by the CV obtained for each object of measurement) may nevertheless yield high reliability (as indexed by the ICC) if the betweenperson differences in means are large.
Intraclass effect decomposition (ICED)
Request a detailed protocolThe extant neuroimaging literature typically offers little justification for the choice of the reliability index. Based on the preceding considerations, this is problematic, as the various indices differ greatly in meaning. The ICC and variants thereof are appropriate for evaluating how well one can capture betweenperson differences in a measure of interest. Put differently, it is misleading to report the CV as a measure of reliability when the goal of the research is to investigate individual or group differences. Both approaches to reliability assessment are informative, but they serve different purposes, and cannot be used interchangeably. Below, we focus on individual differences as we present a general and versatile method for estimating the relative contributions of different sources of error in repeatedmeasures designs. Our approach can be seen as an extension of ANOVAbased approaches to decomposing ICCs. In this sense, it is tightly linked to GTheory, which has been used successfully before in assessing reliability of neuroimaging measures (Gee et al., 2015; Noble et al., 2017). The method, termed intraclass effect decomposition (ICED), has ICC as its core concept. The key feature of the method, however, is its ability to distinguish among multiple sources of unreliability, with the understanding that not all sources of error and their separation are important and meaningful in repeatedmeasures designs. For example, different sources of error may be due to run, session, day, site, scanner, or acquisition protocol variations. Furthermore, there may be more complex error structures to be accounted for, for example, runs nested in sessions; and multiple sessions, again, may be nested within days, and all may be nested under specific scanners in multisite investigations. Neglecting these nuances of error structures leads to biased reliability estimates. The ability to adequately model these relationships and visually represent them in path diagrams is a virtue of our approach.
Beyond reliability per se, researchers may often be interested in the specific sources of error variance and measurement characteristics that contribute to it. For example, in applying MRI to studying longterm withinperson changes in the course of aging, child development, disease progression, or treatment, one may wish to determine first what effect repositioning of a person in the scanner between sessions has on reliability of measured quantities (e.g., Arshad et al., 2017). Similarly, it may be important to determine how much variation is associated with scanning on a different day relative to conducting two scanning sessions on the same day (e.g., Morey et al., 2010). These types of questions are of utmost importance in longitudinal studies, in which researchers collect data on the same person using an ostensibly identical instrument (e.g., MRI scanner ) under an identical protocol (sequence), but inevitably under slightly different measurement characteristics, including position of the participant within the scanner, body and air temperature, or time of day. From a design perspective, knowing the distinct components of measurement error and their relative magnitudes may enhance future study designs and boost their generalizability.
In the proposed SEM framework, observed variance is partitioned into several orthogonal error variance components that capture unreliability attributable to specific measurement characteristics, with the number of components depending on identification constraints based on the study design. Figure 3 shows a minimal, or optimally efficient, repeatedmeasures study design for estimating the contributions of the main effects of day, session, and residual variance to measurement error. The design consists of four measurements (scans) performed over two days and three sessions. In this design, unique contributions of each error source are identified as depicted in the path diagram in Figure 4. In the diagram, observed variables correspond to image acquisitions and are depicted as rectangles; latent variables are depicted as circles and represent the unobservable sources of variance, that is, the true score variance (T) and the error variance components of day (D), session (S), and residual (E). Doubleheaded arrows represent variances of a latent variable. Singleheaded arrows represent regressions with fixed unit loadings.
In this example, total observed variability in an outcome across measurements and persons is partitioned into truescore variance and three error variance terms: the dayspecific error variance, the sessionspecific error variance (here capturing the effect of repositioning a person between scans), and the residual error variance. The full measurement model is depicted as a path diagram in the left panel of Figure 4. The structural equation model specifies four observed variables representing the repeated measurements of the outcome of interest. One of the latent variables represents the true values of the construct of interest. Its variance, ${\sigma}_{T}^{2}$, denotes the betweenperson variance. Fixed regressions of each measurement occasion on the latent construct express the assumption that we are measuring the given construct with each of the four repeated measures on the same scale. There are four orthogonal error variance sources with identical residual variance ${\sigma}_{E}^{2}$, that is, residual errors that are not correlated with any other type of error or among themselves over time. In classical test theory, this is referred to as a parallel model, in which the construct is measured on the same scale with identical precision at each occasion. Typically, there is no explicit assumption of uncorrelated error terms even though many measures derived from this theory assume (and are only valid under) uncorrelated error terms (Raykov et al., 2015). Here, we focus on a parallel model while accounting for the correlated error structure implied by the greater similarity of multiple runs within the same session compared to runs across different sessions. Note that, in the SEM framework, we also can extend the parallel model to more complex types of measurement models (e.g., congeneric or tauequivalent models) that allow for different residual error variances or different factor loadings. To account for the nested structure in our design, we introduce two dayspecific error variance sources with variance ${\sigma}_{D}^{2}$ that represent dayspecific disturbances and imply a closer similarity of measurements on the same day. Finally, there are three sessionspecific variance sources (depicted in blue) representing the session effect (including, for example, the effect of repositioning a person between sessions). The modelimplied covariance matrix has the total variances for each observed variable in the diagonal. It can be analytically or numerically derived using matrix algebra (McArdle, 1980) or pathtracing rules (Boker et al., 2002), and is typically available in SEM computer programs (e.g., von Oertzen et al., 2015). The full modelimplied covariance matrix is given in Table 2. For the given study design, each variance source is uniquely identifiable, as there is a unique solution for all parameters in the model. From the covariance matrix, it is apparent that the inclusion of the session variance term differentially affects the similarity of measurements between days 1 and 2. The correlation between first and second scan is $\frac{{\sigma}_{T}^{2}+{\sigma}_{D}^{2}+{\sigma}_{S}^{2}}{{\sigma}_{T}^{2}+{\sigma}_{E}^{2}+{\sigma}_{D}^{2}+{\sigma}_{S}^{2}}$ whereas the correlation between measurements 3 and 4 is $\frac{{\sigma}_{T}^{2}+{\sigma}_{D}^{2}}{{\sigma}_{T}^{2}+{\sigma}_{E}^{2}+{\sigma}_{D}^{2}+{\sigma}_{S}^{2}}$. Thus the similarity of the two measurements on the first day is greater than the similarity of measurements on the second day. In other words, the difference in correlation is the proportion of variance that the sessionspecific variance accounts for in total variance.
For this model (see Figure 4), we define ICC equivalently to the common ICC formula as ratio of betweenperson variance to total variance at the level of observed variables:
We estimate the components using the full information maximum likelihood procedure for SEM (Finkbeiner, 1979), which allows estimating all components under the assumption of the data missing at random. This maximumlikelihoodbased ICC is similar to the analytical procedure based on relating the ANOVAderived within and between residualsumsofsquares. The main difference is that the maximum likelihood estimator cannot attain negative values when we allow only positive variance estimates (Pannunzi et al., 2018).
In many cognitive neuroscience studies, one may be interested in constructlevel reliability, and not only in reliability of indicators (i.e., observed variables). This construct reliability is captured by ICC_{2} (Bliese, 2000). Based on the above SEMbased effect decomposition, we use power equivalence theory (von Oertzen, 2010) to derive the effective error of measuring the latent construct of interest. The effective error can be regarded as the residual error that would emerge from a direct measurement of a latent construct of interest. Here, it is an index of the precision with which a given study design is able to capture stable individual differences in the outcome of interest. The effective error is a function of all error components and its specific composition depends on the specific design in question. Effective error is the single residual error term that arises from all variances components other than the construct that is to be measured. As such, it represents the combined influence of all error variance components that determine construct reliability:
Effective error can be computed using the algorithm provided by von Oertzen (2010) and for some models, analytic expressions are available (see the multiindicator theorem in von Oertzen, 2010). For the study design in our example, effective error is:
Relating true score variance to total variance yields $IC{C}_{2}$ – a measure of reliability on the construct level. For our model, $IC{C}_{2}$ is then:
As a check, when assuming no dayspecific and sessionspecific effects by inserting ${\sigma}_{D}^{2}=0$ and ${\sigma}_{S}^{2}=0$, we obtain the classical definition of $IC{C}_{2}$ that scales residual error variance with the number of measurement occasions (here, four occasions):
In sum, ICC is a coefficient describing testretest reliability of a measure (also referred to as shortterm reliability or intrasession reliability by Noble et al., 2017) whereas ICC_{2} is a coefficient describing testretest reliability of an underlying construct (an average score in parallel models) in a repeatedmeasures design (longterm reliability or intersession reliability according to Noble et al., 2017).
For our hypothesized measurement model that includes multiple measurements and multiple variance sources, the analytic solution of $IC{C}_{2}$ allows, for instance, to analytically trace reliability curves depending on properties of a design, such as the number of sessions, number of runs per sessions, number of sessions per day, or varying magnitudes of the error component. Of note, this corresponds to a Dstudy in GTheory that can demonstrate, for example, how total session duration and number of sessions influence resting state functional connectivity reliability (see Noble et al., 2017).
A virtue of the proposed SEM approach is the possibility of applying likelihoodratio tests to efficiently test simple and complex hypotheses about the design. For example, we can assess whether individual variance components significantly differ from zero or from particular values, or whether variance components have identical contributions (corresponding to Ftests on variance components in classical GTheory). Such likelihoodratio tests represent statistical model comparisons between a full model, in which each of the hypothesized error components are freely estimated from the data, and a restricted model, in which the variance of a target error component is set to zero. Both models are nested, and under the null hypothesis, the difference in negativetwo loglikelihoods of the models will be ${\chi}^{2}$distributed with 1 degree of freedom. This allows the derivation of $p$ values for the null hypotheses of each individual error component being zero. Moreover, the generality of SEM allows testing complex hypotheses with hierarchically nested error structures or multigroup models. It also allows inference under missing data or by evaluating informative hypotheses (de Schoot et al., 2011) whereas ANOVAbased approaches become progressively invalid with increasing design complexity.
Results
An empirical example: Myelin water fraction data from Arshad et al. (2017)
To demonstrate how the proposed approach separates and quantifies sources of unreliability, we reanalyzed data from a study of the brain regional myelin content by Arshad et al. (2017). In human aging, changes of myelin structure and quantity have been proposed as neuroanatomical substrates of cognitive decline, which makes it particularly interesting to obtain a highly reliable estimate of regional myelin content, here, represented by myelin water fraction (MWF) derived from multicomponent T_{2} relaxation curves. The data in this demonstration were collected in 20 healthy adults (mean age ± SD = 45.9 ± 17.1 years, range of 24.4–69.5 years; no significant difference between men and women: t(18)=–0.81, p=0.43) and are freely available (Arshad et al., 2018); for detailed sample description see Arshad et al. (2017). The study protocol stipulated three acquisitions for each participant in a single session. In the first part, T_{1}weighted and T_{2}weighted MRI images were acquired, followed by a backtoback acquisition of the MET_{2} relaxation images without repositioning the participant in the scanner. At the end of the first part, participants were removed from the scanner and, after a short break, placed back in. In the second part, T_{1}weighted, T_{2}weighted and MET_{2} multiecho sequences were acquired once. All further details relating to the study design, MR acquisition protocol, and preprocessing can be found in the original publication by Arshad et al. (2017). In the following, we focus on the MWF derived from a multiecho gradient recall and spinecho (GRASE) sequence. The study design allows separating the influences of repositioning expressed as sessionspecific variance from true score variance (defined as the shared variance over all three repetitions) and individual error variance (the orthogonal residual error structure). Figure 5 presents a diagram of the hypothesized contributions of the individual variance components. Parameters in the SEM correspond to estimates of true score variation (T), sessionspecific error variance component (S), and a residual error variance component (E). Model specification and estimation was both performed in Ωnyx (von Oertzen et al., 2015) and lavaan (Rosseel, 2012) via full information maximum likelihood. We provide the Ωnyx models and lavaan syntax in the Supplementary material.
For illustration, we only report estimates of the first of the six regions of interest reported in the original study, the anterior limb of the internal capsule (ALIC). The estimates of the individual variance components explaining the observed variance are shown in the diagram in Figure 5. To assess the significance of these components’ magnitudes, we used likelihood ratio tests against null models, in which each component’s variance was set to zero. For testing the residual error variance component, we used a Wald test because the null model without an orthogonal error structure cannot be estimated. Of the total betweenperson variance in measurements of MWF, we found that 86% were due to true score variance (est = 6.97; χ^{2} = 27.759; df = 1; p<0.001), 8%  to sessionspecific variance (est = 0.59, χ^{2} = 3.951; df = 1; p=0.047), and 6%  to residual error variance (est = 0.52; Z = 9.64; p=0.002). Testing whether the variance contribution of the sessionspecific variance and the residual error variance were equal yields a nonsignificant result (χ^{2 }= 20.02; df = 1; p=0.89) and, thus, cannot be decided.
As shown before, we can obtain ICC as the ratio of systematic (t) and all variance components, which by means of standardization of the observed variables sums up to unity, resulting in:
To compute ICC2 as a standardized estimate of the precision with which the repeatedmeasures study design can measure individual differences in MWF in ALIC, we equate the dayspecific variance with zero since it is not identified in this design but rather subsumed under the estimate of the truescore variance component, yielding:
The fact that dayspecific variance and true score variance are inseparable in this design (both are shared variance components of all three measurement occasions) leads to an inflation of the true score variance estimate if nonzero dayspecific variance is assumed and, thus, to an overly optimistic estimate of reliability. To be able to separate the individual variance contributions, one would have to rely on an augmented design that includes additional scanner acquisitions on at least one different day, such as the design shown in Figure 4.
Arshad et al. (2017) only reported pairwise ICCs, based either only on the two backtoback sessions of a single day, or on a single session of each day (again omitting a third of the available data). In the following, we derive the corresponding pairwise ICCs using the full data set. Our estimates are similar even though not identical to the results obtained by Arshad et al. (2017) because our results were jointly estimated from three measurements. First, the authors report an estimate of ICC based on one measurement from the second session of the first day and one measurement from the single session of the second day, resulting in ICC = 0.83, which is close to our estimate of ICC = 0.86. Second, they reported an estimated reliability (ICC) derived only from the two backtoback sessions on the first day as ICC = 0.94, Similarly, we can derive the reliability of a single measurement, had we measured only the two backtoback sessions, achieving the identical result:
The estimates of constructlevel reliability obtained imply that individual differences in MWF can be measured quite well. As expected, the reliability estimate is higher for the backtoback session than for the complete design because one error variance component, sessionspecific error variance, is not apportioned to the total error variance. Such a simple design commingles true score variance and the sessionspecific variance, and reliability studies should thus, by design, take into account potential different error sources, such as sessionspecific error variance.
A comprehensive SEM approach to assessing reliability allows for using the complete dataset in a single model to estimate reliability as either itemlevel reliability (ICC) or a constructlevel reliability (ICC_{2}). A particular benefit of the proposed approach is its ability to tease apart individual error components as far as the study design permits this, that is, as far as these components are identified. Future studies may very well increase study design complexity to test for additional error variance components. To compare the effect of repositioning a participant versus scanning a participant backtoback, Arshad et al. (2017) compared pairwise ICCs of either the two backtoback acquisitions or the second of the backtoback acquisition with the repositioned acquisition. Using the SEMbased approach described above, we can directly estimate a variance component that quantifies the contribution of the session to the total error variance. We can also formally test whether this contribution is nonzero, or, if necessary, whether it is greater than some value or some other error variance component in the model. Furthermore, our estimates are always based on the complete dataset and there is no need to select certain pairs of runs for computing subset ICCs and potentially disregarding important dependencies in the data – a limitation that Arshad et al. (2017) explicitly mentioned in their report.
Linkwise reliability of resting state functionalconnectivity indices
Restingstate functional connectivity was proposed as a promising index of agerelated or pathologyinduced changes in the brain, and has been used to predict brain maturation (Dosenbach et al., 2010) or disease state (Craddock et al., 2009). These applications can only prove practically useful if reliability is sufficiently high, so that differences between persons can be reliably detected in the first place, as a methodological precondition for prediction. Thus, there has been increasing interest in examining reliability of methods for assessing resting state connectivity (Gordon et al., 2017; Noble et al., 2017; Pannunzi et al., 2018). Here, we demonstrate how ICED can be used to evaluate reliability of pairwise functional indices obtained from restingstate functional connectivity analyses.
To illustrate such a model, we obtained the resting state functional connectivity (rsFC) dataset from Pannunzi et al., 2018, which is based on the publicly available raw data from the Day2day study (Filevich et al., 2017). In that study, six participants were scanned at least 40 (and some up to 50) times over the course of approximately seven months, and another sample of 50 participants (data from 42 participants of them available) were each scanned only once. In the following, we show how both datasets can be jointly investigated to estimate linkwise reliability of resting state functional connectivity (rsFC). We present a reliability analysis of the linkwise connectivity indices of brain regionsofinterest based on 5 min of measurement. For each measurement, as our main outcome, we obtained a 16 × 16 correlation matrix of rsFC indices, for pairs of regions including prefrontal, sensormotor, parietal, temporal, limbic, occipital cortices, cerebellum and subcortical structures. In our model, we assume independence of the measurement occasions. Thus, we decompose the covariance structure of the repeated measurements into one betweenperson variance and one withinperson variance component. For simplicity, we illustrate this model by using the first ten observations. Figure 6 shows a path diagram of this model. We estimated this model using Ωnyx and lavaan and significance tests were performed using Wald tests. For example, we first estimated our model only for the link between left prefrontal cortex and right prefrontal cortex. The true score variance was estimated to account for 49% of the total variance (est = 0.013; W = 2.46; df = 1; p=0.117) and the error variance contributed 51% of the total variance (est = 0.014; W = 27.00; df = 1; p<0.0001), thus, ICC was 0.49.
With up to fifty measurement occasions, we can expect to get sufficiently precise measures of withinperson fluctuations but since only eight participants contributed, we augment this dataset with crosssectional data from additional 42 persons treating them as quasilongitudinal data with the majority of data missing. This more precise measurement of betweenperson differences yields a somewhat different pattern of results. The true score variance was 39% of the total variance (est = 0.008; W = 6.31; df = 1; p=0.012) and the error variance was 61% of the total variance (est = 0.013; W = 33.53; df = 1; p<0.0001). Thus, our estimate dropped from 0.49 to 0.39. Due to a small sample size in the first analysis, we likely had overestimated the betweenperson differences in rsFC and had obtained an exceedingly overoptimistic ICC. By augmenting the initial analysis with a second dataset, we have obtained more precise and, here, even more pessimistic estimates of rsFC reliability.
Figure 7 shows a reliability matrix of all links between the investigated brain regions with estimates based on the joint model. Pannunzi et al., 2018 reported that ICCs range from 0.0 to 0.7 with an average ICC of 0.22, which is typically considered an unacceptably low reliability (i.e., signal is outweighed by noise by a factor of about 4). The average ICC in our analysis is 0.28 and, thus, very much in line with the original analysis. Compared to Pannunzi et al., 2018, we find a compressed range of ICCs from 0.0 to 0.55 and second the claim that rsFC obtained from 5 min scans performs poorly as a marker for individual subjects (also see Gordon et al., 2017).
Discussion
When the true scores are changing: Extending ICED to growth curve modeling
So far, we have assumed that the construct of interest does not change over time. Thus, any change between repeated measures was assumed due to unsystematic variability, that is, noise. But what if the construct of interest varies over time? For example, had we modeled all fifty measurements from the day2day study that spanned roughly six month, we would have confounded reliability and lack of stability. it is very likely that the difference between repeated measures in the beginning and at the end of the study represent a mixture of measurement error and true withinperson shortterm variability, longterm change, or both (also see Nesselroade, 1991). When assessing reliability over repeated measures in practice, one seeks avoiding this problem by reducing the interval between measurements. At the same time, one is interested in independent measurements, and the degree of dependence may increase with shorter time spans as the chance of itemspecific or constructgeneral temporal effects that may affect multiple measurements may artificially increase the reliability estimate. If, however, measurements are numerous or if the reliability estimate must be obtained from an existing study with a considerable time lag between measurements, it is likely that true change in the construct is present, and that persons differ regarding its magnitude, direction, or both. If substantive change is not accounted for, reliability estimates are biased towards lower values (Brandmaier et al., 2018). The resulting biased measure may still be useful when interpreted as a stability coefficient, while keeping in mind that instability may be caused by change as well as imprecise measurement. What is, however, the best strategy when we wish to know whether true scores have changed?
Elsewhere, we have applied the logic presented here to linear latent growth curve models (Brandmaier et al., 2015; Brandmaier et al., 2018; von Oertzen and Brandmaier, 2013). Effective error of the change component (or, slope) in a latent growth curve model reflects the precision with which a growth curve model can measure betweenperson differences in change. By scaling the magnitude of individual differences in change (i.e., betweenperson variance in slope) with effective error, we obtain effective curve reliability (ECR; Brandmaier et al., 2015). Major components of effective error for individual differences in change are the number of measurement occasions, the temporal arrangement of measurement occasions, the total study time span, and instrument reliability. We have shown that effective error, reliability, and statistical power are all potentially useful measures that quantify the sensitivity of a longitudinal design, or any repeated measures design (for example, multiple sessions within a day), to measure individual differences in change (Brandmaier et al., 2015; Brandmaier et al., 2018). All these measures may be used for a priori design optimization. Such optimization entails either tradingoff multiple design factors against each other, while keeping power constant, or changing power as a function of the various design factors and treating them as important measures to communicate reliability of change beyond crosssectional reliability.
Intraclass effect decomposition of group differences and interactions among error sources
In Section 4, we have discussed a research design that is optimal in factorizing the total error variance into three orthogonal error components of dayspecific, sessionspecific, and unspecific residual variance. Optimality referred to a design that comprises the smallests number of measurements necessary to identify the soughtafter error components. However, the ICED framework easily generalizes to more complex designs. For example, with a greater number of sessions, it would be possible to identify additional sources of error, such as experimenterspecific or sitespecific errors. We can think of this framework as a variance decomposition approach just as in regular analysis of variance (see Noble et al., 2017), with the only difference that we are not interested in the sources of true score variance with the residuals set aside but rather in the decomposing the error score variance.
In the example reported in Section 4, we only examined main effects of day and session. Note, however, that the ICED method also can handle interaction effects. For example, we may be interested if there is an interaction of day and session, that is, if it matters on which day repositioning happened. To test this interaction, one can easily add a second group to the design presented earlier, with two sessions at Day one and one session at Day two (i.e., the mirror image of the current design, in which there is one session at Day one and two sessions at Day two). In this model, we could estimate the ICED components separately for each group. To test a potential interaction, we would state a null hypothesis of no differences in error variance across groups for the session effect. This is a null model of no interaction between session and orderofday. Explicitly testing the initial model against the restricted null model yields a χ^{2} significance test of the interaction. Now, imposing this equality constraint on the session effect across groups would effectively test for the presence of reliable session by day interaction (e.g., does it make a difference whether repositioning within a day takes place at Day one or Day two). One could also conduct the same study with different groups, such as children, older adults, or patients with a particular disease or condition to evaluate group differences in day and session error contributions.
Summary
In this paper, we have discussed the distinction and complementarity of ICC and CV in gauging reliability of brain imaging measures, a topic that thus far has received only limited attention. Considering the increasing demand for longitudinal and multicenter studies, there is a dire need for properly evaluating reliability and identifying components that contribute to measurement error. ICC and CV, as measures of (relative) precision, or reliability, fundamentally relate information about lasting properties of the participants to the precision with which we can measure this information over repeated assessments under the assumption of no change in the underlying construct. We have shown how the generality of the SEM approach (cf. McArdle, 1994) may be leveraged to identify components of error sources and estimate their magnitude in more complex designs in more comprehensive and general ways than achievable with standard ANOVAbased ICC decompositions. The underlying framework for deriving the individual error components as factors of reliability is closely related to Cronbach’s generalizability theory (or GTheory; Cronbach et al., 1972), which was recently expressed in a SEM framework (Vispoel et al., 2018). Our approach is similar to those approaches but was derived using the power equivalence logic (von Oertzen, 2010) to analytically derive effective error and reliability scores in a SEM context. This means that our approach easily generalizes to complex measurement designs beyond standard ANOVA, and that effective error, ICC, ICC_{2} can automatically be derived using von Oertzen (2010) algorithm from any study design rendered as a path diagram or in matrixbased SEM notation.
As noted at the beginning of our article, ICC and CV represent two perspectives on reliability that correspond to a fundamental divide of approaches to the understanding of human behavior: the experimental and the correlational (individual differences), each coming with its own notion of reliability (Cronbach, 1957; Hedge et al., 2018). In experimental settings, reliable effects are usually those that are observed on average, that is, assumed to exist in most individuals. To facilitate detection of such effects, the withinperson variability must be low in relation to the average effect. The experimental approach is therefore compatible with the CV perspective. In individual difference approaches, reliable effects distinguish well between persons, which is only true if the withinperson variability is low in relation to the betweenperson variability. The two notions of reliability are associated with competing goals; hence, it is not surprising that robust experimental effects often do not translate into reliable individual differences (Hedge et al., 2018).
In addition to ICC and CV, other reliability indices have been reported. When researchers compare the similarity of sets, as in gauging the overlap of voxels identified in two repeated analyses of the same subject, the Sørensen–Dice similarity coefficient (or, Dice coefficient; Dice, 1945; Sørensen, 1948) is often used. Since we are focusing on the reliability of derived continuous indices (e.g., total gray matter volume, fractional anisotropy or indices of myelin water fraction in a region of interest, or linkwise resting state functional connectivity), we did not consider the Dice coefficient here. Others have used the Pearson product moment correlation coefficient, $r$, to quantify the consistency of test scores across repeated assessments. The linear correlation is a poor choice for reliability assessment because due to its invariance to linear transformation, it is insensitive to mean changes (Bartko, 1966). Moreover, it is limited to twooccasion data. Therefore, we have also not considered Pearson’s $r$ here.
Outlook
Effective error variance partitioning as described above can be useful for communicating absolute precision of measurement, on its own and complimentarily with reliability. Importantly, one needs to specify what kind of reliability is being sought: reliability with respect to an anchoring point (e.g., the scale’s zero) or with respect to the heterogeneity in the population. It needs to be emphasized that ICC can only be large if there are individual differences across persons in the measure of interest. Critics of ICCbased approaches to estimating reliability have argued that this method confounds group heterogeneity in the outcome of interest and measurement precision, and therefore must ‘be perceived as an extremely misleading criterion for judging the measurement qualities of an instrument.’ (Willett, 1989, p. 595). We strongly disagree with this narrow view of measurement quality. In the proverbial sense, ‘one man’s trash is another man’s treasure,’ and what some may view as a ‘confound,’ is for others a virtue of the measure in as much as it determines the capability of detecting heterogeneity in the population. However, the ICC may reveal nothing about the trialtotrial differences expressed as deviations in the actual unit of measurement; those are better represented by the withinperson standard deviation or standardized versions of it. We maintain that ICC is the appropriate measure of reliability when assessing diagnostic instruments and especially while focusing on individual differences.
In this article we introduced ICED as a variancepartitioning framework to quantify the contributions of various measurement context characteristics to unreliability. ICED allows researchers to (1) identify error components; (2) draw inferences about their statistical significance and effect size; and (3) inform the design of future studies.
Given the remarkable pace of progress in human brain imaging, researchers often will be interested in the (yet unknown) reliability of a new neuroimaging measure. Whether this reliability is sufficient can roughly be decided using thresholds, which essentially are a matter of consensus and conventions. For example, reliability larger than 0.9 is often regarded as excellent, as it implies a signal to noise ratio of 10:1. However, there may be good reasons to adopt less conservative thresholds (e.g., Cicchetti and Sparrow, 1981). In addition, using ICED, researchers can go beyond a summary index of ICC and instead report the magnitudes of individual variance components that contribute to lowering the overall ICC. These different components may differ in their methodological and practical implications. Often, researchers will be interested in using inferential statistics to test whether each of the individual variance components differs from zero and, maybe, whether the components differ from each other. Finally, the results of these analyses can guide researchers in their subsequent attempts to improve measurement reliability. For instance, using ICED, researchers may discover that a hitherto overlooked but remediable source of error greatly contributes to unreliability, and work on improving the measurement properties influencing this component. Also, researchers may ask what combinations of measurements are needed to attain a target reliability (Noble et al., 2017) while optimizing an external criterion such as minimizing costs or participant burden (Brandmaier et al., 2015).
To conclude, we hope that the tools summarized under ICED will be applied in human brain imaging studies to index overall reliability, and to identify and quantify multisource contributions to measurement error. We are confident that the use of ICED will help researcher to develop more reliable measures, which are a prerequisite for more valid studies.
Data availability
The dataset on myelin water fraction measurements is freely available at https://osf.io/t68my/ and the linkwise resting state functional connectivity data are available at https://osf.io/8n24x/.

Restingstate fMRI correlations: from linkwise unreliability to whole brain stabilityPublicly available at Open Science Framework.

Reliability of Myelin Water Fraction in ALICPublicly available at Open Science Framework.
References

The intraclass correlation coefficient as a measure of reliabilityPsychological Reports 19:3–11.https://doi.org/10.2466/pr0.1966.19.1.3

Multilevel Theory, Research, Andmethods in Organizations: Foundations, Extensions, and New Directions349–381, Withingroup agreement, nonindependence, and reliability:Implications for data aggregation and analysis, Multilevel Theory, Research, Andmethods in Organizations: Foundations, Extensions, and New Directions, San Francisco, CA, JosseyBass.

An algorithm for the hierarchical organization of path diagrams and calculation of components of expected covarianceStructural Equation Modeling: A Multidisciplinary Journal 9:174–194.https://doi.org/10.1207/S15328007SEM0902_2

Developing criteria for establishing interrater reliability of specific items: applications to assessment of adaptive behaviorAmerican Journal of Mental Deficiency 86:127–137.

Disease state prediction from resting state functional connectivityMagnetic Resonance in Medicine 62:1619–1628.https://doi.org/10.1002/mrm.22159

BookThe Dependability of Behavioral Measurements: Theory of Generalizability for Scores and ProfilesNew York: John Wiley.

The two disciplines of scientific psychologyAmerican Psychologist 12:671–684.https://doi.org/10.1037/h0043943

Estimation for the multiple factor model when data are missingPsychometrika 44:409–420.https://doi.org/10.1007/BF02296204

Momenttomoment brain signal variability: a next frontier in human brain mapping?Neuroscience & Biobehavioral Reviews 37:610–624.https://doi.org/10.1016/j.neubiorev.2013.02.015

Reliability of an fMRI paradigm for emotional processing in a multisite longitudinal studyHuman Brain Mapping 36:2558–2579.https://doi.org/10.1002/hbm.22791

The reliability paradox: why robust cognitive tasks do not produce reliable individual differencesBehavior Research Methods 50:1166–1186.https://doi.org/10.3758/s1342801709351

Causal modeling applied to psychonomic systems simulationBehavior Research Methods & Instrumentation 12:193–209.https://doi.org/10.3758/BF03201598

Structural factor analysis experiments with incomplete dataMultivariate Behavioral Research 29:409–454.https://doi.org/10.1207/s15327906mbr2904_5

Scanrescan reliability of subcortical brain volumes derived from automated segmentationHuman Brain Mapping 31:1751–1762.https://doi.org/10.1002/hbm.20973

BookThe Warp and Woof of the Developmental Fabric HillsdaleHillsdale, NJ: Lawrence Erlbaum.

The importance of the assumption of uncorrelated errors in psychometric theoryEducational and Psychological Measurement 75:634–647.https://doi.org/10.1177/0013164414548217

lavaan : an R package for structural equation modelingJournal of Statistical Software 48:1–36.https://doi.org/10.18637/jss.v048.i02

A method of establishing groups of equal amplitude in plant sociology based on similarity of species and its application to analyses of the vegetation on danish commonsBiologiske Skrifter 5:1–34.

Structural equation modeling with ωnyxStructural Equation Modeling: A Multidisciplinary Journal 22:148–161.https://doi.org/10.1080/10705511.2014.935842

The effect of multiple indicators on the power to detect interindividual differences in changeBritish Journal of Mathematical and Statistical Psychology 63:627–646.https://doi.org/10.1348/000711010X486633

Power equivalence in structural equation modellingBritish Journal of Mathematical and Statistical Psychology 63:257–272.https://doi.org/10.1348/000711009X441021

Some results on reliability for the longitudinal measurement of change: implications for the design of studies of individual growthEducational and Psychological Measurement 49:587–602.https://doi.org/10.1177/001316448904900309
Decision letter

Heidi JohansenBergReviewing Editor; University of Oxford, United Kingdom

Sabine KastnerSenior Editor; Princeton University, United States
In the interests of transparency, eLife includes the editorial decision letter and accompanying author responses. A lightly edited version of the letter sent to the authors after peer review is shown, indicating the most substantive concerns; minor comments are not usually included.
Thank you for submitting your article "Assessing Reliability in Neuroimaging Research Through IntraClass Effect Decomposition (ICED)" for consideration by eLife. Your article has been reviewed by two peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Sabine Kastner as the Senior Editor. The following individual involved in review of your submission has agreed to reveal his identity: Nico Dosenbach (Reviewer #3).
The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.
Summary:
In this article Brandmaier et al. discuss the differences between coefficient of variation (CV) and intraclass correlation coefficient (ICC) and then introduce a novel measure called the intraclass effect decomposition (ICED). The ICED estimates sources of error by explicitly modeling the contributions of each latent source of variance (e.g., person, session, day) to each scan via confirmatory SEM. These sources are then combined into a power equivalence form of the ICC, and used to estimate ICC under different conditions. The analyses of Arshad et al. (2017) are replicated using this method in order to demonstrate that different forms of the ICC (using different error components) can be estimated from the full dataset.
Essential revisions:
1) The motivation for the approach, and what it adds over existing methods, needs to be clarified. The authors state that the "key feature of this method is its ability to distinguish among multiple sources of unreliability, with understanding that not all errors are equally important and meaningful in repeatedmeasures design", and again highlight this in the subsection “An Empirical Example: Myelin Water Fraction Data from Arshad et al. (2017)”, last paragraph. However, these benefits are already present in ANOVAbased ICC. Generalizability Theory (GTheory; Webb and Shavelson, 2005) has been used in neuroimaging to decompose error into constituent sources and reestimate ICCs (e.g., Gee et al., 2015, Noble et al., 2017). Despite this crucial point, the authors only mention GTheory in the third paragraph of the subsection “When the true scores are changing: Extending ICED to growth curve modeling”. From a practical standpoint, the estimates of the ICC2 are very similar to the estimates of Arshad et al. (2017). Although they use a subset of the data to estimate backtoback versus repositioned ICCs, this can be estimated with the full data and an ANOVA using GTheory.
While the authors state another motivating virtue – the ability to use "likelihood ratio tests to efficiently assess whether individual variance components significantly differ from zero" – they do not acknowledge that this is also possible in a traditional ANOVA framework, e.g., via simple Ftest.
The value of this approach, which should be at the heart of its motivation, is that it provides a theoretically more valid way of estimating error from multiple sources, particularly in complex and timedependent designs. The ANOVA and repeatedmeasures ANOVA become increasingly invalid with the complexity of the design. In addition, the SEM framework allows the user to test different assumptions about the structure of the model (e.g., tauequivalent vs. congeneric). The aforementioned subsection starts to hint at this, though this should be the core feature throughout. Note however, that this is a theoretical argument; it is difficult to demonstrate the "utility" of this method over the simpler ANOVA, especially for simpler designs where the gains may be small.
2) Subsection “Intraclass effect decomposition (ICED)”, fifth paragraph: Significantly, this form of the ICC, with error variance divided by k (here, 4), represents an average measure over multiple scans. Reliability of averages are not relevant to most purposes, where a single scan is of interest. Arshad et al. get similar values (ICC=0.83) as the value in the third paragraph of the subsection “An Empirical Example: Myelin Water Fraction Data from Arshad et al. (2017)” in ALIC with repositioning, though it is difficult to tell whether they are using average measures from their methods (the exact form of the ICC is not given).
3) The extensive discussion in the Introduction/subsection “Comparing CV and ICC: Different but compatible conceptions of signal and noise” of the relative pros and cons of ICC and CV, alongside language about reconciling disparate approaches (Abstract and Introduction, first paragraph), leads the reader to anticipate a measure that reconciles CV and ICC. However, this is not the case. Therefore, the indepth discussion of CV therefore seems out of place. The authors also return to a confusing/imprecise discussion about this in the fourth paragraph of the subsection “When the true scores are changing: Extending ICED to growth curve modeling”.
4) Subsection “Intraclass effect decomposition (ICED)”, fourth paragraph: More details are needed for the SEM estimation procedure. For instance, does "identical and fixed" mean tauequivalent same weight across paths, but that weight is freely estimated? Since each residual is estimated separately for each scan, how are these residuals then combined into a single residual error term for the ICC?
5) Subsection “Intraclass effect decomposition (ICED)”, fifth paragraph: Do the reliability curves mentioned here refer to explicitly varying the magnitude of the error terms? If so, what is the utility of this? Or does this refer to vary the number of measurements in a Decision Study, as in Noble et al. (2017)?
https://doi.org/10.7554/eLife.35718.016Author response
Essential revisions:
1) The motivation for the approach, and what it adds over existing methods, needs to be clarified. The authors state that the "key feature of this method is its ability to distinguish among multiple sources of unreliability, with understanding that not all errors are equally important and meaningful in repeatedmeasures design", and again highlight this in the subsection “An Empirical Example: Myelin Water Fraction Data from Arshad et al. (2017)”, last paragraph. However, these benefits are already present in ANOVAbased ICC. Generalizability Theory (GTheory; Webb and Shavelson, 2005) has been used in neuroimaging to decompose error into constituent sources and reestimate ICCs (e.g., Gee et al., 2015, Noble et al., 2017). Despite this crucial point, the authors only mention GTheory in the third paragraph of the subsection “When the true scores are changing: Extending ICED to growth curve modeling”.
We are grateful for this comment. We were not aware of the fact that previous work had applied GTheory to neuroimaging data and we are glad that this was brought to our attention. We now acknowledge these previous applications at various places throughout the manuscript. Furthermore, we draw a stronger link to GTheory much earlier in the manuscript and also throughout the manuscript. For example, in the Introduction:
“Below, we focus on individual differences as we present a general and versatile method for estimating the relative contributions of different sources of error in repeatedmeasures designs. […] In this sense, it is tightly linked to GTheory, which has been used successfully before in assessing reliability of neuroimaging measures (Gee et al., 2015; Noble et al., 2017).”
Or, in the Discussion:
“The underlying framework for deriving the individual error components as factors of reliability is closely related to Cronbach’s generalizability theory (or GTheory; Cronbach et al., 1972), which was recently expressed in a SEM framework (Vispoel, Morris, and Kilinc, 2017). […] This means that our approach easily generalizes to complex measurement designs beyond standard ANOVA, and that effective error, ICC, ICC_{2} can automatically be derived using von Oertzen’s (2010) algorithm from any study design rendered as a path diagram or in matrixbased SEM notation.”
From a practical standpoint, the estimates of the ICC2 are very similar to the estimates of Arshad et al. (2017). Although they use a subset of the data to estimate backtoback versus repositioned ICCs, this can be estimated with the full data and an ANOVA using GTheory.
While the authors state another motivating virtue – the ability to use "likelihood ratio tests to efficiently assess whether individual variance components significantly differ from zero" – they do not acknowledge that this is also possible in a traditional ANOVA framework, e.g., via simple Ftest.
This observation is correct. We now state that our results can equivalently be derived using ANOVAbased ICCs following GTheory logic and we acknowledge that the Ftest is applicable in the first example using Arshad’s data. We further motivate our approach by stating that we are able to test more complex hypotheses using our framework, involving hierarchically nested structures, multigroup models, or informative hypothesis testing. For example:
“The key feature of the method, however, is its ability to distinguish among multiple sources of unreliability, with the understanding that not all sources of error and their separation are important and meaningful in repeatedmeasures designs. […] The ability to adequately model these relationships and visually represent them in path diagrams is a virtue of our approach.”
and
“Beyond, the generality of SEM allows us to test complex hypotheses involving hierarchically nested error structures or multigroup models and allows inference under missing data or by evaluating informative hypotheses (de Schoot, Hoijtink, and JanWillem, 2011) whereas ANOVAbased approaches increasingly become invalid with the complexity of the design.”
The value of this approach, which should be at the heart of its motivation, is that it provides a theoretically more valid way of estimating error from multiple sources, particularly in complex and timedependent designs. The ANOVA and repeatedmeasures ANOVA become increasingly invalid with the complexity of the design. In addition, the SEM framework allows the user to test different assumptions about the structure of the model (e.g., tauequivalent vs. congeneric). The aforementioned subsection starts to hint at this, though this should be the core feature throughout. Note however, that this is a theoretical argument; it is difficult to demonstrate the "utility" of this method over the simpler ANOVA, especially for simpler designs where the gains may be small.
We are thankful for this comment. We have added specific references to the ANOVA framework and have made clearer when the SEM framework is preferable. We have added a complete new analysis to illustrate a complex design in which we benefit from the flexibility that the SEM framework has to offer. Please see the newly added text describing our reliability analysis of the data from the day2day study.
Concerning the assumptions about the measurement model, we clarified that we are assuming a parallel measurement model. We acknowledge that a specific virtue of our framework is the generality to allow for modeling and testing other forms of measurement models (congeneric, tauequivalent) but we believe that most usecases as well as our demonstrations will assume a parallel model. We have revised the whole section introducing the model. Here is a revised excerpt from the manuscript:
“In classical test theory, this is referred to as a parallel model, in which the construct is measured on the same scale with identical precision at each occasion. […] Note that, in the SEM framework, we also can extend the parallel model to more complex types of measurement models (e.g., congeneric or tauequivalent models) that allow for different residual error variances or different factor loadings.
2) Subsection “Intraclass effect decomposition (ICED)”, fifth paragraph: Significantly, this form of the ICC, with error variance divided by k (here, 4), represents an average measure over multiple scans. Reliability of averages are not relevant to most purposes, where a single scan is of interest. Arshad et al. get similar values (ICC=0.83) as the value in the third paragraph of the subsection “An Empirical Example: Myelin Water Fraction Data from Arshad et al. (2017)” in ALIC with repositioning, though it is difficult to tell whether they are using average measures from their methods (the exact form of the ICC is not given).
It is correct to note that the ICC2 relates to the reliability of an average obtained from multiple scans. However, we disagree that ICC2 is not important. In fact, it is the primary value of interest when our framework is used to inform a future study design, in which one can decide how many scans will be performed and under what conditions the measurements will take place (e.g., two runs in a single session or two runs in two sessions, etc.). For example, Noble et al. (2017) use a ICC2type measure to investigate the effects of total scan time and number of sessions on (average) reliability. We have strengthened this point in response to another comment below. Newly added excerpts to the manuscript are:
“In many cognitive neuroscience studies, one may be interested in constructlevel reliability, and not only in reliability of indicators (i.e., observed variables). […] This construct reliability is captured by ICC_{2} (Bliese, 2000).”
and
“In sum, ICC is a coefficient describing testretest reliability of a measure (this was also referred to as shortterm reliability or intrasession reliability by Noble, 2017) whereas ICC_{2} is a coefficient describing testretest reliability of an underlying construct (an average score in parallel models) in a repeatedmeasures design (this was also referred to as longterm reliability or intersession reliability by Noble, 2017).”
We have also clarified the connection between Arshad et al.’s original results and our results. Revised excerpt from the manuscript:
“Arshad et al. (2017) only reported pairwise ICCs, based either only on the two backtoback sessions of a single day, or on a single session of each day (again omitting a third of the available data). […] Similarly, we can derive the reliability of a single measurement, had we measured only the two backtoback sessions, achieving the identical result:
ICC=0.94”.
3) The extensive discussion in the Introduction/subsection “Comparing CV and ICC: Different but compatible conceptions of signal and noise” of the relative pros and cons of ICC and CV, alongside language about reconciling disparate approaches (Abstract and Introduction, first paragraph), leads the reader to anticipate a measure that reconciles CV and ICC. However, this is not the case. Therefore, the indepth discussion of CV therefore seems out of place. The authors also return to a confusing/imprecise discussion about this in the fourth paragraph of the subsection “When the true scores are changing: Extending ICED to growth curve modeling”.
We thank you for this comment. From our discussions with various researchers both within our own group and with others, it became clear that there is some confusion about when (or why at all) one should use CV or ICC to assess reliability. We believe such a thorough discussion is much needed and indispensable before one can fully acknowledge the importance of our framework.
Still, we agree that we may have built up some wrong expectations in the Abstract and the Introduction, so we have changed the manuscript in several places to clarify that we are addressing two related goals, here. First, we motivate the use of ICC while acknowledging its relation to CV, and, second, we introduce a general variance partitioning framework related to Gtheory that allows us to better understand sources of unreliability in our neuroimaging measurements.
We have clarified the Discussion in the following way:
“As noted at the beginning of our article, ICC and CV represent two perspectives on reliability that correspond to a fundamental divide of approaches to the understanding of human behavior: the experimental and the correlational (individual differences), each coming with its own notion of reliability (Cronbach, 1957; Hedge, Powell, & Sumner, 2017). […] The two notions of reliability are associated with competing goals; hence, it is not surprising that robust experimental effects often do not translate into reliable individual differences (Hedge et al., 2017).
4) Subsection “Intraclass effect decomposition (ICED)”, fourth paragraph: More details are needed for the SEM estimation procedure. For instance, does "identical and fixed" mean tauequivalent same weight across paths, but that weight is freely estimated? Since each residual is estimated separately for each scan, how are these residuals then combined into a single residual error term for the ICC?
We thank you for this comment. We have clarified details of the estimation procedure (see our response to item 1, for example, regarding the parallel model and potential extensions to tauequivalent models). In fact, residuals are not necessarily separately estimated for each scan. The residual errors are sums of the individual variance components and their composition depends on the path diagram which again represents a given study design. They can be directly read off from the modelimplied covariance matrix and the actual residual error components can be either be computed or manually derived using pathtracing rules. We have clarified the entire related section and added, for example:
“The modelimplied covariance matrix has the total variances for each observed variable in the diagonal. It can be analytically or numerically derived using matrix algebra (McArdle, 1980) or pathtracing rules (Boker, McArdle, and Neale, 2002), and is typically available in SEM computer programs (e.g., von Oertzen, Brandmaier, and Tsang, 2015).”
We also added further clarifications on how to compute the residual error (“effective error”) for ICC_{2}, for example:
“The effective error is a function of all error components and the specific composition depends on a specific study design. […] Effective error can be computed using the algorithm provided by von Oertzen (2010) and for some models, analytic expressions are available (see the multiindicator theorem in von Oertzen, Hertzog, Lindenberger, and Ghisletta, 2010).”
5) Subsection “Intraclass effect decomposition (ICED)”, fifth paragraph: Do the reliability curves mentioned here refer to explicitly varying the magnitude of the error terms? If so, what is the utility of this? Or does this refer to vary the number of measurements in a Decision Study, as in Noble et al. (2017)?
Correct. We have made explicit the link to Noble’s very convincing application of this idea. We also explicitly link to the idea of decision studies in general in a new paragraph in the Discussion. Here is an excerpt from the revised manuscript:
“For our hypothesized measurement model that includes multiple measurements and multiple variance sources, the analytic solution of ICC_{2} allows us, for instance, to analytically trace reliability curves depending on properties of a design, such as the number of sessions, number of runs per sessions, number of sessions per day, or varying magnitudes of the error component. Of note, this corresponds to a Dstudy in GTheory that can demonstrate, for example, how total session duration and number of sessions influence resting state functional connectivity reliability (see Noble et al., 2017).”
https://doi.org/10.7554/eLife.35718.017Article and author information
Author details
Funding
Horizon 2020 Framework Programme (732592)
 Andreas M Brandmaier
 Simone Kühn
 Ulman Lindenberger
MaxPlanckGesellschaft
 Andreas M Brandmaier
 Elisabeth Wenger
 Nils C Bodammer
 Naftali Raz
 Ulman Lindenberger
 Simone Kühn
National Institutes of Health (R01AG011230)
 Naftali Raz
Openaccess funding. The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication
Acknowledgements
We thank Muzamil Arshad and Jeffrey A Stanley from the Department of Psychiatry and Behavioral Neuroscience, School of Medicine, Wayne State University, Detroit, Michigan, for providing the raw data on Myelin Water Fraction measurements. This work was supported by European Union‘s Horizon 2020 research and innovation programme under grant agreement No. 732592: ‘Healthy minds from 0–100 years: optimizing the use of European brain imaging cohorts ('Lifebrain')’ to AB, SK and UL, and by NIH grant R01AG011230 to NR.
Senior Editor
 Sabine Kastner, Princeton University, United States
Reviewing Editor
 Heidi JohansenBerg, University of Oxford, United Kingdom
Publication history
 Received: February 6, 2018
 Accepted: July 1, 2018
 Accepted Manuscript published: July 2, 2018 (version 1)
 Version of Record published: July 13, 2018 (version 2)
Copyright
© 2018, Brandmaier et al.
This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.
Metrics

 1,911
 Page views

 226
 Downloads

 27
 Citations
Article citation count generated by polling the highest count across the following sources: Crossref, Scopus, PubMed Central.
Download links
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Further reading

 Neuroscience
Developmental language disorder (DLD) is a common neurodevelopmental disorder characterised by receptive or expressive language difficulties or both. While theoretical frameworks and empirical studies support the idea that there may be neural correlates of DLD in frontostriatal loops, findings are inconsistent across studies. Here, we use a novel semiquantitative imaging protocol – multiparameter mapping (MPM) – to investigate microstructural neural differences in children with DLD. The MPM protocol allows us to reproducibly map specific indices of tissue microstructure. In 56 typically developing children and 33 children with DLD, we derived maps of (1) longitudinal relaxation rate R1 (1/T1), (2) transverse relaxation rate R2* (1/T2*), and (3) Magnetization Transfer saturation (MTsat). R1 and MTsat predominantly index myelin, while R2* is sensitive to iron content. Children with DLD showed reductions in MTsat values in the caudate nucleus bilaterally, as well as in the left ventral sensorimotor cortex and Heschl’s gyrus. They also had globally lower R1 values. No group differences were noted in R2* maps. Differences in MTsat and R1 were coincident in the caudate nucleus bilaterally. These findings support our hypothesis of corticostriatal abnormalities in DLD and indicate abnormal levels of myelin in the dorsal striatum in children with DLD.

 Neuroscience
Many studies have investigated the contributions of vision, touch, and proprioception to body ownership, i.e., the multisensory perception of limbs and body parts as our own. However, the computational processes and principles that determine subjectively experienced body ownership remain unclear. To address this issue, we developed a detectionlike psychophysics task based on the classic rubber hand illusion paradigm where participants were asked to report whether the rubber hand felt like their own (the illusion) or not. We manipulated the asynchrony of visual and tactile stimuli delivered to the rubber hand and the hidden real hand under different levels of visual noise. We found that (1) the probability of the emergence of the rubber hand illusion increased with visual noise and was well predicted by a causal inference model involving the observer computing the probability of the visual and tactile signals coming from a common source; (2) the causal inference model outperformed a nonBayesian model involving the observer not taking into account sensory uncertainty; (3) by comparing body ownership and visuotactile synchrony detection, we found that the prior probability of inferring a common cause for the two types of multisensory percept was correlated but greater for ownership, which suggests that individual differences in rubber hand illusion can be explained at the computational level as differences in how priors are used in the multisensory integration process. These results imply that the same statistical principles determine the perception of the bodily self and the external world.