Methodological considerations in assessment of language lateralisation with fMRI: a systematic review

The involvement of the right and left hemispheres in mediating language functions has been measured in a variety of ways over the centuries since the relative dominance of the left hemisphere was first known. Functional magnetic resonance imaging (fMRI) presents a useful non-invasive method of assessing lateralisation that is being increasingly used in clinical practice and research. However, the methods used in the fMRI laterality literature currently are highly variable, making systematic comparisons across studies difficult. Here we consider the different methods of quantifying and classifying laterality that have been used in fMRI studies since 2000, with the aim of determining which give the most robust and reliable measurement. Recommendations are made with a view to informing future research to increase standardisation in fMRI laterality protocols. In particular, the findings reinforce the importance of threshold-independent methods for calculating laterality indices, and the benefits of assessing heterogeneity of language laterality across multiple regions of interest and tasks. This systematic review was registered as a protocol on Open Science Framework: https://osf.io/hyvc4/.


INTRODUCTION
A wealth of evidence has demonstrated that language is predominantly mediated by the left cerebral hemisphere in the majority of individuals, a phenomenon known as hemispheric specialisation. This has been recently defined by Tzourio-Mazoyer & Seghier (2016) as ''the hosting by a given hemisphere of specialized networks that have specific functional properties and interact interhemispherically in a way that optimizes brain processing ''. However, our understanding of the nature and correlates of such lateralisation has been hampered by the high level of variability in methods of its measurement, making integration of findings from across different studies difficult.
Non-invasive techniques for assessment of language lateralisation make it possible to probe the characteristics of language lateralisation in neurologically intact populations. Functional magnetic resonance imaging (fMRI) is a prominent non-invasive method that has been used to assess laterality. A laterality index (LI ) is calculated based on a comparison of activation measures from each hemisphere, according to the following formula: This calculates laterality as the difference between activity in each hemisphere (L and R) divided by the total activity across the hemispheres. The LI gives a single value indicating the relative strength of left and right hemisphere activation for an individual. LI measurement may be required for clinical purposes in order to establish an individual's hemispheric dominance for language prior to surgery, as in patients with intractable epilepsy. Alternatively, a study may measure an LI to assess the strength or variability in lateralisation for a given language function in order to make inferences about the neural organisation of the language system. That is, studies may vary in whether the aim of LI measurement is to classify or to quantify lateralisation. This will have important implications for which methods of LI calculation are optimal for laterality measurement.
Interpretation of fMRI lateralisation research has been problematic due to a lack of standardisation of fMRI laterality protocols. Multiple arbitrary decisions must be made when calculating the L and R terms for use in the LI equation which might affect the LI value obtained (Jansen et al., 2006;Seghier, 2008). Such variability in methodology can preclude systematic study of language lateralisation.
For example, when calculating an LI from active voxels in each hemisphere or region of interest (ROI), a decision must be made as to the threshold p value at which to view and analyse the images. Multiple studies have documented the dependence of the LI obtained on the threshold chosen (Rutten et al., 2002;Adcock et al., 2003;Seghier et al., 2004;Abbott et al., 2010;Nadkarni et al., 2015). As illustrated in Fig. 1, as the threshold value is increased, the number of voxels surviving thresholding decreases, typically leading to an increase in the LI. Ultimately, above a certain threshold, no active voxels will remain in the non-dominant hemisphere, resulting in an LI of 1; and below a certain threshold many voxels will survive across both hemispheres, resulting in an LI of 0. Indeed, there are even reports of individuals whose LI shows a switch in dominance with a change in threshold level (Jansen et al., 2006;Suarez et al., 2008;Wilke & Lidzba, 2007;Ruff et al., 2008).
This illustrates just one preliminary issue that must be addressed when considering how to quantify lateralisation from fMRI data. Further decisions have to be made as to tasks used in an activation paradigm, whether the analysis focuses on a specific region of interest (ROI) or the whole hemisphere, and whether the quantification of activation is based on magnitude or extent of activity. If the LI is used to categorise individuals as left-, bilateral or right-lateralised, a suitable cut-off for categorisation must also be determined.
The purpose of this review is to assess different protocols for fMRI measurement of language lateralisation used by studies published between 2000 and 2016. We aimed to (1) look at the methods used by different studies over this time period in order to consider whether the field is converging on common criteria for evaluating language lateralisation, and (2) consider evidence for the robustness and reliability of these different methods in order to make recommendations for future research in this field.

MATERIALS AND METHODS
A protocol for this review has been registered on Open Science Framework and can be found at https://osf.io/hyvc4/. This paper addresses those objectives outlined in the protocol relating to assessment of the methods used to quantify lateralisation in fMRI studies of language lateralisation. Assessment of the impact of language task and baselining methods will be considered in a companion paper. non-European languages. The rationale for restricting the search to studies on healthy, monolingual, adult participants was to reduce heterogeneity within our study sample.

Search strategy and selection process
The search and selection process is illustrated in Fig. 2. We searched Web of Science for studies published between 2000 and 2016 using the following search terms: laterali* OR asymmetr* OR dominance; AND language OR reading; AND fMRI OR functional MRI OR functional magnetic resonance imaging OR functional MR OR function MRI; NOT schizophrenia; NOT development*; NOT child*; NOT bilingual*. This was last searched on 05/12/16. Two of the study authors (Abigail Bradshaw and Zoe Woodhead) screened the titles and abstracts of the resulting 90 papers to assess their eligibility then conducted full-text scans to determine whether the inclusion criteria were met. Selected lists were compared between reviewers and any discrepancies discussed and a mutual decision made.
This yielded a total of 34 papers selected from the original 90. To ensure thorough coverage of the literature, papers citing these 34 articles were searched to look for additional articles that met criteria. From this, 50 additional papers were selected, bringing the total to 84 papers. A final search to re-check all 84 papers against search criteria identified 7 ineligible papers. During the review, a further paper was judged to not meet criteria. A list of the final 76 selected papers can be found in Appendix S1.

Data collection and data summary
For each paper, we recorded the following parameters relating to the protocol used: the type of fMRI design used, the activity measures used for LI calculation, the threshold level chosen, the use of global or regional LI calculation, the specific regions considered, the language and baseline tasks used, the use of a single or a combined task analysis and the task difficulty. We also recorded sample size and sample handedness for each study. Information on these measures for each paper was collected and managed using REDCap electronic data capture tools (Harris et al., 2009) hosted at Oxford University. REDCap (Research Electronic Data Capture) is a secure, web-based application designed to support data capture for research studies, providing: (1) an intuitive interface for validated data entry; (2) audit trails for tracking data manipulation and export procedures; (3) automated export procedures for seamless data downloads to common statistical packages; and (4) procedures for importing data from external sources. The full database can be found in Appendix S2 . A summary table drawn from this database with the key outcomes of interest for this paper is provided in Appendix S3.
The variable nature of the methods used and measures reported by different fMRI studies of language lateralisation means the data are not suitable for a meta-analysis. Instead, this review will document the range of methods used, and provide a qualitative summary of information from these studies that is relevant for our understanding of the robustness and reliability of LI measurement.

Methods for calculating LI
As shown in Fig. 3, a range of methods have been used to compute a laterality index in the set of studies that we considered. These will be briefly described before moving to consider the relative advantages and disadvantages of each. Note that the majority of studies used the standard LI ratio approach using the LI formula as previously outlined (55 within our search), but use of bootstrapping approaches started to be seen in 2010 and has gained in popularity since that time. Use of different methods likely reflects their availability as well as their robustness; for example, the development of LI -tool (Wilke & Lidzba, 2007), a tool-box for LI calculation within MATLAB software (Mathworks, Natick, MA, USA) has made bootstrapping readily available and easily implementable. In contrast, other approaches such as the flip method and t -weighting may be less widely used due to the lack of dedicated software for their implementation. Relatively few studies (3) explicitly compared different methods for calculating the LI : we will cover those that did so in more detail in the following sections.  Figure 4 illustrates the different thresholding approaches used by studies, and how these have changed over time. As can be seen, the majority of studies (33) used a single fixed threshold approach for determining the LI in which a single threshold level is chosen at which either extent of activation, or magnitude of activation in a given region is measured for left and right, and entered into the LI formula. As noted above however, use of a single threshold when calculating an LI is likely to yield an unreliable and inadequate measure of an individual's pattern of laterality. Awareness of this has led to a decline of the single fixed threshold approach in more recent years, in favour of approaches that aim to address the problem of threshold dependence. Each of these will be described and evaluated in turn in the following sections.

Multiple thresholds and threshold dependent laterality curves
One way to address the problem of threshold dependence is to calculate the LI across multiple thresholds. One can then produce a plot of LI as a function of threshold (see   (Seghier et al., 2004;Jansen et al., 2006;Ruff et al., 2008;Suarez et al., 2008;Abbott et al., 2010). Such curves can allow one to decipher an individual's general tendency towards one pattern of dominance, often showing a transition point at which the increase in laterality plateaus at a particular laterality level. However, such curves are not always informative, since in some cases they may fail to reach a plateau, or are not reproducible within a subject (Rutten et al., 2002;Jansen et al., 2006).

Variable thresholds
A second approach uses a variable or adaptive threshold, in which the threshold is set according to subject-specific parameters. One such method involves choosing that threshold which yields a fixed number of active voxels for each individual participant (Knecht et al., 2003;Jansen et al., 2006;Abbott et al., 2010;Fesl et al., 2010). Using simulated data, Abbott et al. (2010) demonstrated that thresholding at a fixed number of voxels was more robust against variability in signal strength than the standard thresholding method. They advocated plotting the LI as a function of the number of active voxels, similar to threshold dependent curves; these curves are however tighter and more stable. Furthermore, Fesl et al. (2010) reported improved reliability of LI measurement when using this variable threshold method as opposed to a single fixed threshold. However, this approach does not remove the need for arbitrary decisions, since a 'reasonable' fixed number of active voxels must be decided on. Interestingly, when using this method Jansen et al. (2006) set the criterion number of activated voxels at a different level for each language task; this can thus enable one to take into account the fact that different tasks may require different threshold levels.
Other adaptive thresholding methods are based on setting the threshold in proportion to the maximum or mean intensity of voxels in an image. Methods include identifying the highest 5% of voxels with the highest t -values, and setting the threshold at half of their mean value (Fernandez et al., 2001;Van Veelen et al., 2011); alternatively, the mean intensity of voxels within an area of interest can be used (Stippich et al., 2003;Wilke & Lidzba, 2007;Partovi et al., 2012a;Partovi et al., 2012b;Allendorfer et al., 2016). Using this latter method, Wilke & Lidzba (2007) reported more stable LI s using variable threshold methods as compared to fixed thresholding, suggesting that this may make LI s more robust. This study also demonstrated a flattening of laterality curves when only those voxels that formed a significant cluster or that had a sufficiently low level of variability were included in the LI calculation. These clustering and variance weighting methods thus allow calculation of LI s to become more stable across threshold levels.

T-weighting and threshold independent methods
Alternatively, the issue of threshold dependence can be avoided by the use of a 'thresholdindependent' method. One such widely used threshold-independent method is t -weighting (Branco et al., 2006;Suarez et al., 2008;Propper et al., 2010;Zaca, Jarso & Pillai, 2013), illustrated in Fig. 5. This involves plotting a histogram for each hemisphere of the number of active voxels against t -score threshold, and then multiplying this distribution by a weighting function that assigns weight in a way directly proportional to t -score. The integrated areas under each hemisphere's curve can then be used as the input for the standard LI equation. Suarez et al. (2008) reported that such a method yielded reduced within-subject and between-subject variability in LI compared to fixed thresholding, resulting in clear left lateralisation across subjects.
Other threshold-independent methods have been developed, but these have been less widely used. Harrington, Buonocore & Farias (2006) reported that taking the average signal magnitude within activated voxels across multiple thresholds yielded higher and more reproducible LI s than a single threshold approach. Seghier et al. (2011b) used a method developed by Nagata et al. (2001) in which the L and R terms are calculated by taking the regression of the curve obtained by plotting the number of activated voxels against threshold for each hemisphere separately. This provides a fixed term for each hemisphere that is independent of threshold for use in the LI calculation, providing a more robust measure.

Bootstrapping
A further method developed to remove the issue of threshold dependence is bootstrapping (Wilke & Schmithorst, 2006). This involves iterative resampling and calculation of LI s across multiple threshold levels, illustrated in Fig. 6. At each threshold, a vector containing all voxel values is created from an image (b), one for each hemisphere. Multiple random samples of values (e.g., 100 samples) from these vectors are then taken (c) and an LI calculated for all possible right/left sample combinations (d). All LI s are then plotted in a histogram (e) and a trimmed mean is taken by selecting the central 50% of the data in order to reduce the effect of outliers. A weighted overall mean is then calculated from this resulting data by assigning a higher weight to higher thresholds. This method has been widely adopted in recent research on measuring language lateralisation (Häberling, Badzakova-Trajkov & Corballis, 2011;Van der Haegen et al., 2011;Van der Haegen, Cai & Brysbaert, 2012;Perlaki et al., 2013;Berl et al., 2014;Mazoyer et al., 2014;Miro et al., 2014;Tzourio-Mazoyer et al., 2015;Häberling, Steinemann & Corballis, 2016;Sepeta et al., 2016). As well as being threshold-independent, its key strengths include greater resistance to outliers and built-in markers for detecting the presence of outliers within the process of LI calculation (Wilke & Schmithorst, 2006), making it a robust method for assessing language laterality from fMRI data.

Activity measure
A key decision in calculation of a laterality index from fMRI data concerns which activity measure to use; signal extent (i.e., the number of suprathreshold voxels in each hemisphere) or signal magnitude (i.e., the average intensity of suprathreshold voxels in each hemisphere). Figure 7 documents the different activity measures used by studies within our search. It can be seen that the majority of studies opt for an extent measure, although in recent years there has been an increased use of magnitude measures. Of the studies that compare both methods, many have reported finding similar laterality indices and curves (Jansen et al., 2006;Bethmann et al., 2007;Wilke & Lidzba, 2007;Ocklenburg, Hugdahl & Westerhausen, 2013). Others have reported differences in LI strength, with reports of both higher LI s for magnitude measures (Harrington, Buonocore & Farias, 2006) and higher LI s for extent measures (Jensen-Kondering et al., 2012). Further still, Jansen et al. (2006) reported that differences in the activity measure used for calculating laterality with a picture naming task could yield different dominance classifications for a given participant. This was not the case however for verbal fluency and semantic decision tasks, suggesting that this may reflect something particular about the activity patterns induced by naming.
There is evidence that these measures can yield the same high levels of reproducibility for LI measurement: Morrison et al. (2016) reported 100% reproducibility for classification of language dominance and almost identical test-retest LI correlations using both activity measures for a rhyming task. However, the majority of studies report higher reproducibility for signal magnitude measures compared to signal extent measures (Adcock et al., 2003;Jansen et al., 2006;Harrington, Buonocore & Farias, 2006;Morrison et al., 2016). Importantly, Jansen et al. (2006) reported that a magnitude measure determined dominance reproducibly only when just those voxels that exceeded a criterion activation level were included; an extent measure was not reproducible. Magnitude measures have also been reported to be less sensitive to noise than a thresholded extent measure (Adcock et al., 2003). One possible explanation for higher reliability with magnitude than extent LI s relates to intersessional differences in the distribution of t -values. Gorgolewski et al. (2012) discuss the phenomenon in which shifts in t -value histograms can occur between sessions due to global effects which have a systematic effect across the whole brain. This can result in the distribution of t -values in a second session being shifted relative to the first, whilst the pattern or shape of that distribution is still preserved. In such cases, use of the same fixed threshold across both testing sessions could result in greatly different activity maps, leading to a low estimation of reliability, despite that fact that the overall pattern has not changed. Extent measures would be affected by such shifts to a greater extent than magnitude measures, thus resulting in their reduced reliability.
Of further note is Jansen et al. 's (2006) finding that LI s based on signal extent lacked meaningful variation, often yielding LI values of 1; in contrast, LI magnitude measures gave greater between-subject variation in LI values. Which constitutes better laterality data depends on one's view of lateralisation measurement. That is, in cases when one wishes to classify individuals' language dominance, having LI values close to 1 or −1 Figure 8 The flip method. By contrasting a right-side contrast image with a mirror image (flipped so that the right hemisphere is on the left), a new contrast image is generated with significant voxels indicating regions in which left activity is statistically significantly greater than right homologue activity.
would be useful to allow decisions to be clear cut. Conversely, when one is interested in quantifying individual variability in the degree of language lateralisation beyond the binary typical/atypical distinction, an LI measure that reveals greater between-subject variation would be more useful.

The flip method
The standard LI ratio does not indicate whether the difference in activity between the hemispheres is significant or not, but simply quantifies the bias in activity towards one hemisphere. The flip method (Baciu et al., 2005) was developed to provide a direct statistical comparison of activity between the hemispheres. As illustrated in Fig. 8, this involves contrasting two sets of functional images created for the contrast of interest (i.e., task versus control); a right side images set, in which the left hemisphere is on the left and a mirror images set, in which the image is flipped such that the left hemisphere is on the right. By contrasting these two images, one can identify those homotopic voxels that show a significant difference in activity across the hemispheres. The resulting significant voxels in each hemisphere can then be used as input for the standard LI equation.
This method has been used in a small number of studies (see Fig. 3) as a means of measuring laterality (Cousin et al., 2007;Seghier et al., 2011b;Hernandez et al., 2013). Baciu et al. (2005) compared the flip method to the standard LI ratio by comparing the correlations of each method's LI s with handedness lateralisation indices. While both methods yielded language LI s that were poorly correlated with handedness LI s for a verb generation task, when a rhyming task was used the flip method yielded higher correlations than the standard LI method. Such a finding is difficult to interpret, especially given the inconsistent relationship between handedness and language dominance (e.g., Mazoyer et al., 2014). It should be noted that the flip method cannot be used to describe bilaterality, but only non-significantly lateralised activity. Further, since significant voxels from this contrast typically reside exclusively in the dominant hemisphere, using these voxels as input for the LI equation will often yield an LI of −/+1, rendering it an inappropriate method to provide a meaningful description of degree of lateralisation.

Dominance classification
If a categorical dominance classification is required, some form of standardised procedure is needed once an LI has been calculated. The range of methods used by different studies within our search is illustrated in Fig. 9. The most standard method of dominance classification uses cut-offs at −0.2 and 0.2, to divide left dominance (LI > 0.2) from bilaterality (−0.2 ≤ LI ≤ 0.2) and right dominance (LI < −0.2). However, such cut-offs are arbitrary, and we found multiple studies within our search that chose their own cut-offs (see 'other cut-off values used'), including 0.1, 0.33, 0.4, 0.5 and 0.6. Thus, it can be seen that there is a high level of heterogeneity in the methods of dominance classification used by different studies. This makes it very difficult to draw conclusions and comparisons between the proportions of typically and atypically lateralised individuals reported by different studies, and thus impedes progress in understanding the distribution of such lateralisation profiles across different populations.
Other researchers have investigated data-driven methods of defining dominance categories, in which the typicality/atypicality of cases is assessed in relation to norms within the data itself or a separate 'normative' dataset (Adcock et al., 2003;Seghier et al., 2004;Abbott et al., 2010;Berl et al., 2014;Mazoyer et al., 2014;Tzourio-Mazoyer et al., 2016). One simple way of deriving cut-offs is to use 2 standard deviations below the mean as a threshold to divide typical from atypical laterality (Adcock et al., 2003;Seghier et al., 2004). Abbott et al. (2010) used a similar approach in which an individual's LI distribution (LI as a function of voxel count) was compared to a normative distribution based on a sample of controls with 'typical' lateralisation. As illustrated in Fig. 10, the individual's threshold-dependent laterality curve is established as being either within or below the lower 95% confidence interval for the control group (represented by the shaded area in Fig. 10). If an individual's laterality fell below such an interval, they were classed as atypical, on the basis of a low probability of their laterality data having come from the 'population' of the normative group. This method thus provides an objective means of establishing an individual as either atypical or typical in relation to a normative group.
Other data-driven approaches to dominance classification suggest the existence of more than two dominance categories. Berl et al. (2014) used a hierarchical clustering method which gradually and iteratively combined cases into clusters and indicated at what level in the hierarchy the optimal cluster solution was obtained. They discussed both a three cluster solution and a two cluster solution for a large sample of right handers; the former divided subjects into left dominant, crossed dominance (one outlying case) or bilateral, whereas the latter divided typically and atypically lateralised subjects. This latter solution was found to divide at an LI of 0.5, which was therefore argued to represent a meaningful cut-off for dominance classification.
A larger-scale study by Mazoyer et al. (2014) looked at dominance categories in both left and right handed participant groups. They used Gaussian mixture modelling to extract dominance categories from laterality data (consisting of LI values between −100 and +100), which involves determining the optimal number of Gaussian functions that can be fitted to the data. This found different model solutions for right and left handed groups. For the right handed group (Fig. 11A), a three function solution was optimal, consisting of two overlapping 'typical' (left dominant) functions (both with LI values above +18) and a third 'ambilateral' function (consisting of LI values between −50 and +18). This agrees with Berl et al.'s (2014) study, which found no evidence for right hemisphere dominance in a right handed group. However, in the left handed group (Fig. 11B), an additional 'strongly atypical' function was found with strongly negative LI values (below −50). Only 10 left handers from this large sample (297 right and left handers) were strongly atypical, indicating that it is very rare. They thus argued for the need to treat atypical laterality as a heterogeneous group consisting of the subgroups of 'ambilateral' and 'strongly atypical'; conversely, it was argued that the overlapping typical distributions (from both left and right handed participants) could be combined into a single homogenous typical group. These data-driven methods of deriving dominance categories suggest that the traditional −0.2 and 0.2 cut-offs do not reflect the true distribution of LI s across individuals, and that revision of standards for dominance classification is needed in the field, with ramifications for both clinical practice and research. It should be further noted that classification of bilaterality from fMRI data is typically not reproducible across repeated testing sessions or even across different methods of LI calculation (Jansen et al., 2006), suggesting the need to interpret bilateral classification with caution. Further work is needed to establish the reproducibility of classification into the 'ambilateral' group identified by Mazoyer et al. (2014).

Effect of region of interest on laterality
Another key consideration in calculating an LI concerns whether to include all voxels across a hemisphere (to calculate a global LI ) or whether to define a region or several regions of interest (a regional LI ). Figure 12 shows changes in approach to choice of region over time across the studies within our search. It can be seen that the majority  of studies use a regional LI, either picking a single ROI or calculating multiple LI s from different regions. A number of studies have compared global and regional approaches to LI calculation, particularly in earlier years within our search period. This evidence on the relative advantages and disadvantages of the global and regional approaches is discussed below. It can also be seen that a substantial proportion of studies use a single ROI in LI calculation; however, a review of evidence on regional variability in laterality questions the adequacy of such an approach (see Regional heterogeneity in laterality).

Global and regional LIs
A potential issue with global LI s is that voxels outside the areas most relevant to the language paradigm can have a strong influence on the LI obtained. Whilst many studies have shown that LI s from pre-specified language ROIs are stronger and more reliable than global LI measurements (Fernandez et al., 2001;Rutten et al., 2002;Suarez et al., 2008;Pravata et al., 2011), others have found no difference (Hund-Georgiadis et al., 2002) or the opposite (Rutten et al., 2002;Wilke & Lidzba, 2007). Further studies suggest that whether global or regional LI measures are optimal may depend on other methodological decisions. For example, Jansen et al. (2006) reported that global LI s were more reliable than regional LI s when a voxel count approach was used. Similarly, Rutten et al. (2002) reported greater reliability for global LI s over regional LI s during an antonym generation task, but that the reverse was true for a verb generation task.
In general however, cases of crossed dominance or regional heterogeneity in lateralisation have been used to argue for the need for regional rather than global laterality indices, which fail to capture these finer grained individual patterns in regional laterality (e.g., Seghier et al., 2011a;Seghier et al., 2011b). Such evidence will be discussed in the following section, Regional heterogeneity in laterality.

Regional approaches to LI calculation
When using a regional rather than a global LI, careful thought must be given to deciding which areas to choose as ROIs. Typically, areas within frontal and temporoparietal cortex are chosen for measuring language laterality, such as the inferior frontal cortex or posterior superior temporal gyrus. It is not the case that any single ROI or combination of ROIs will always be optimal for assessing laterality; instead, the choice of ROI(s) must be guided by other factors such as the language function being studied or the purpose of laterality measurement.

Regional heterogeneity in laterality
Research comparing laterality across regions has reported cases of crossed or dissociated dominance across different cortical language areas, at both an individual and a group level (Vikingstad et al., 2000;Thivard et al., 2005;Jansen et al., 2006;Bethmann et al., 2007;Propper et al., 2010;Seghier et al., 2011a;Seghier et al., 2011b;Van der Haegen, Cai & Brysbaert, 2012;Vingerhoets et al., 2013;Berl et al., 2014;Häberling, Steinemann & Corballis, 2016). Bethmann et al. (2007) reported four subjects with crossed frontal and temporal dominance for a semantic decision task; in particular, one subject was classified as bilateral when ROIs were combined, whereas classification based on either only a frontal or a temporal ROI yielded contradicting dominance categories. Other studies report within-subject dissociations between the dominance measured from different regions using different language tasks. Cases of crossed dominance have been reported between frontal LI s for expressive language and temporal or occipitotemporal LI s for receptive language (Van der Haegen, Cai & Brysbaert, 2012;Häberling, Steinemann & Corballis, 2016), and between temporal and frontal LI s for two different expressive language tasks (Vikingstad et al., 2000). In Van der Haegen, Cai & Brysbaert's (2012) study, the majority of left handed subjects showed colateralisation of inferior frontal cortex for verb generation and ventral occipitotemporal cortex for lexical decision, but a small minority (3 out of 57 participants) showed crossed dominance. Such crossed dominance is not limited to left handers however; Häberling, Steinemann & Corballis (2016) reported cases of crossed frontal expressive and temporal receptive dominance among both left and right handers.
Such cases highlight the inadequacy of relying on a single global or regional LI, and how this practice in previous research may have led to an underappreciation of such crossed dominance in the literature. Indeed, such regional variation is not only found at the individual level, but also at a group level. In a comparison of lateralisation for a semantic word matching task across 50 different ROIs, Seghier et al. (2011a) andSeghier et al. (2011b) reported a negative correlation between LI s obtained from the angular gyrus and the ventral precentral gyrus. This suggests that regional heterogeneity can be a normal part of typical profiles of hemispheric lateralisation for language.

SUMMARY AND CONCLUSIONS
This review has highlighted the many different ways in which calculation of an LI from fMRI data presents a methodological challenge, and how the use of different methods has changed over time. Of course, the decisions one makes when designing an fMRI laterality experiment will depend on the question being investigated; however, here we have highlighted some key principles that emerge from the literature that should be considered in order to generate increased standardisation in fMRI laterality protocols across future studies. Increased homogeny in the methods used by different studies will enable better integration of research findings in order to draw conclusions as to the nature and correlates of language lateralisation. fMRI LI calculation must address the problem of threshold dependence. Bootstrapping represents a promising method for calculating a robust, threshold-independent LI, making it a widely used method in recent research. The general pattern of evidence suggests that signal magnitude may provide a more robust and reliable measure than signal extent, and that regional LI s calculated from pre-specified ROIs are stronger and more reliable than global LI s. However, such decisions need to be considered in light of other methodological parameters (e.g., the activity measure used) in order to optimise the fMRI analysis. A useful tool for implementing such analysis methods is LI -tool, a tool-box within MATLAB software (Mathworks, Natick, MA, USA), developed by Wilke & Lidzba (2007). This includes options for different thresholding techniques and activity measures, and can implement the bootstrapping method.
Data-driven methods can provide a less arbitrary means of classifying language dominance and support the validity of a three category model of language dominance within a mixed handedness sample, consisting of typical (left dominant), ambilateral, and atypical (strongly right dominant) groups; conversely in right handed samples a two-category model (typical versus ambilateral) may be sufficient (Mazoyer et al., 2014). No subsequent studies have implemented the thresholds for dominance classification suggested by Mazoyer et al.'s (2014) large scale study, except a paper reporting on the same sample of right and left handed individuals (Tzourio-Mazoyer et al., 2016). Further work is needed to implement and validate these cut-offs, to see if these generalise to other samples and thus could be used as standard practise.
The choice of which regions of interest to use for LI calculation again depends on the question being asked. If one wishes to classify laterality for a particular language function, one must consider which ROI yields the highest and most reliable LI s for that function. Frontal ROIs typically yield the strongest and most reliable laterality for expressive tasks, whereas a temporoparietal ROI may be more appropriate for receptive tasks. However, measurement of laterality from a single regional or global ROI can be misleading and does not capture potential regional heterogeneity. This was highlighted recently by Tailby, Abbott & Jackson (2017) in relation to the need to appreciate such regional variability in presurgical planning with epilepsy patients, and the consequent inadequacy of a single metric to quantify an individual's 'language dominance'. Therefore, in fMRI laterality protocols, lateralisation across frontal and temporoparietal ROIs for at least one expressive and one receptive task should be measured, to obtain a comprehensive picture of any individual's pattern of hemispheric dominance for language. This will enable further work to investigate the significance of such regional heterogeneity in dominance; for example, are there any functional consequences of having crossed frontal-temporal language laterality? In this way, fMRI as a method of laterality measurement can provide unique insights into lateralisation at a regional level; this should be fully exploited in future research.
Lastly, it should be noted that fMRI data provides opportunities for investigating the nature of lateralisation for a function that goes beyond simple calculation of laterality indices. For example, fMRI can be used to investigate more mechanistic questions such as potential relationships between inter-and intrahemispheric connectivity and lateralisation (e.g., Seghier et al., 2011a;Frässle et al., 2016). In order to pursue more advanced analyses, it will be important for the field to establish robust protocols for laterality measurement as standard across studies.