Temperature change effects on marine fish range shifts: A meta‐analysis of ecological and methodological predictors

The current effects of global warming on marine ecosystems are predicted to increase, with species responding by changing their spatial distributions. Marine ectotherms such as fish experience elevated distribution shifts, as temperature plays a key role in physiological functions and delineating population ranges through thermal constraints. Distributional response predictions necessary for population management have been complicated by high heterogeneity in magnitude and direction of movements, which may be explained by both biological as well as methodological study differences. To date, however, there has been no comprehensive synthesis of the interacting ecological factors influencing fish distributions in response to climate change and the confounding methodological factors that can affect their estimation. In this study we analyzed published studies meeting criteria of reporting range shift responses to global warming in 115 taxa spanning all major oceanic regions, totaling 595 three‐dimensional population responses (latitudinal, longitudinal, and depth), with temperature identified as a significant driver. We found that latitudinal shifts were the fastest in non‐exploited, tropical populations, and inversely correlated with depth shifts which, in turn, dominated at the trailing edges of population ranges. While poleward responses increased with rate of temperature change and latitude, niche was a key factor in predicting both depth (18% of variation) and latitudinal responses (13%), with methodological predictors explaining between 10% and 28% of the observed variance in marine fish responses to temperature change. Finally, we found strong geographical publication bias and limited taxonomical scope, highlighting the need for more representative and standardized research in order to address heterogeneity in distribution responses and improve predictions in face of changing climate.


| INTRODUC TI ON
Over the last century, global warming has had substantial impacts on marine ecosystems, with species locally extirpating (Pinsky et al., 2019), changing distributions in depth and latitude (Brown et al., 2016;Chen et al., 2011;Kortsch et al., 2012;Lenoir et al., 2020;Poloczanska et al., 2013), or in some cases shifting phenotypes in response to climatic pressures (Manhard et al., 2017;Ryu et al., 2020). In marine ectotherms such as fish, population distributional limits are influenced by physiological thermal constraints, as temperature affects critical functions such as metabolism, growth, and reproduction (Addo-Bediako et al., 2000;Angilletta et al., 2002;Roessig et al., 2004), and are restricted by narrower thermal safety margins . Accordingly, species' range changes in response to climate change have been up to sevenfold faster in the ocean as compared to on land (Poloczanska et al., 2013). As marine temperatures are forecasted to continue rising (Pörtner et al., 2019), the ability to predict fish redistributions will be vital to protect ecosystem functions, maintain food security, and other contributors to human well-being (Bonebrake et al., 2018;Pecl et al., 2017). A central challenge in predictive species range modeling has been the observation that, although many ranges have displayed anticipated poleward shifts in response to warming (Chen et al., 2011), a substantial number of range shifts have not followed projections and show significant variation in rate and direction of movements (Poloczanska et al., 2013;Urban, 2015), complicating population response predictions and conservation management. A key development in addressing this variation has been the acknowledgment that a suite of other non-temperature associated biotic factors, including species interactions (Ellingsen et al., 2020;Louthan et al., 2015), ecological and life history traits (MacLean & Beissinger, 2017), and ecoevolutionary dynamics (Cacciapaglia & van Woesik, 2018;Fredston et al., 2021;Nadeau & Urban, 2019), can also affect a population's ability to colonize and establish in novel environments, and should thus be incorporated into forecasts. However, an often overlooked factor in predicting and synthesizing climate change responses are differences in methodological approaches to measuring population distribution changes over time (Brown et al., 2016;Wolkovich et al., 2012), which might explain part of the observed variation in direction and velocity of responses to temperature, even within the same geographical and taxonomic context. For example, for some marine fish species within the same geographic regions seemingly contradictory responses are being reported. In the North Atlantic, for example, some studies suggest rapid environmental tracking at a rate corresponding to the local climate velocity (the pace and direction of climate shift across landscape; Frainer et al., 2017;, while other multidecadal studies on range shifts suggest that only few are completely keeping pace with changing climate  and report significantly slower distribution responses (Campana et al., 2020). Addressing this variation will be key to improved response predictions informing conservation management, particularly as the magnitude of range shifts is likely to increase under climate change forecasts. Some syntheses have indeed highlighted the complexity of interacting functional and taxonomic predictors of climate responses in marine taxa (Lenoir et al., 2020), with Brown et al. (2016) demonstrating higher importance of methodological biases in marine range shift estimates than previously thought. To date, however, no recent synthesis with a focus on marine fish exists. As such, there is a need to build upon this initial work and to summarize the most recent literature to test an extended scope of interacting ecological factors influencing both the latitudinal and depth changes of marine fish species in response to climate change, and the confounding methodological factors that can affect their estimation.
The scarcity of analyses of methodological biases in marine range shift research is surprising considering the wide range of methods for data acquisition, processing, and modeling, resulting in high heterogeneity of research quality and results. While some methodological details need to be tailored to be suitable for specific taxa, ecosystems, and geographical conditions, large heterogeneity in other variables potentially affecting accuracy such as population sampling effort, temporal resolution, and statistical approaches remains. For example, redistribution inferences may be affected by sampling methods including choice of proxy for distribution measurement Wernberg et al., 2012), including the "center of distribution" (COD) which constitutes the mean latitude of the spatial extent (e.g., , or a population's most extreme boundaries of longitude, latitude, or depth, inferred, for instance, by presence-absence data (e.g., . How these distribution indices are obtained also affects the predictions that are produced (Brown et al., 2016): common data sources include abundance data from survey trawls by long-term fisheries or research programs , tagging-recapture data , historical records , or genetic molecular methods (Knutsen et al., 2013;Spies et al., 2020). Each of these methods has various costs and benefits, such as tradeoffs associated with monetary expense, sampling effort, and feasibility in contrast to the likelihood of observing specific species or species types, achieving adequate sample sizes, and spatial-temporal resolution. Variation also exists in the data analysis stage, including the decision of whether to report movement estimates for a single species or cumulative inferences for whole assemblages reflecting changes in community traits and composition (e.g., Frainer et al., 2017). Response estimates in marine taxa were also shown to be affected when climatic predictors, other than temperature, such as salinity (Champion et al., 2021), oscillation indexes , bathymetry , or non-climatic drivers, such as food availability  or exploitation by fishing , were included (Brown et al., 2016). Nevertheless, robust data from wild marine fish populations incorporating both biotic and abiotic drivers of climate responses remain scarce (but see Adams et al., 2018), with potential differential effects on response estimates between single and multi-predictor models remaining unexplored. Overall, while this methodological variation is known to exist, it remains unclear whether it has generated any systematic biases in the existing literature which may distort estimates of geographical shifts across fish species.
This review aims to summarize the current state and remaining gaps of knowledge on ecological and methodological factors influencing latitudinal and depth shifts in response to ocean warming in marine fish. First, we carried out a systematic literature review to gather data from existing original articles meeting criteria of measuring range shifts in response to temperature change. The aim was to investigate trends between rate of temperature change and range shifts across different niches, habitats, and other ecological factors such as life stage and marine exclusivity. Second, we summarized the current state of methodology prevalent across these studies, such as data acquisition and analysis methods, temporal and spatial resolution, and estimated the effects of study methods on population redistribution inferences.

| Literature search
The methodology of this review and meta-analysis was guided by the Preferred Reporting Items for Systematic reviews and Metaanalyses (PRISMA; Page et al., 2021).  (Table 1). Additionally, suitable articles were identified further by scanning reference lists and review articles on related topics. Authors of four studies were contacted via email to obtain missing information on results and methodology. Of these, Dr. Maria Fossheim and Dr. Raul Primicerio provided species-wise raw data of latitudinal changes in distribution from the paper by . The three remaining studies, for which no data were received, were dropped from analyses.

| Study selection
Records retrieved from the database were screened for duplicates, and for the first round of eligibility abstracts were manually checked to confirm the study focus included marine fish and distributional range changes in response to temperature (Figure 1). Four further rounds of filtering were performed according to inclusion and exclusion criteria ( Table 2). This process was performed independently by one reviewer, while the second reviewer randomly selected a sample of five studies in every stage to assess, with disagreements between reviewers being resolved by consensus. Articles extracted from references were simultaneously screened for eligibility in the same manner.
Only original research papers documenting latitudinal or depth responses to temperature in marine fish were considered ( Table 2).
The terms range and distribution shifts are used in this study interchangeably and refer to, based on definitions used by Parmesan et al. (2005) and Sorte et al. (2010), a change in the distribution of native species' boundaries from their historical boundaries, including relocations, expansions, contractions along range edges.
For a study to be included in the analysis, it had to discuss temperature as a likely driver of distributional range changes (preferably by statistical association) and have a span of at least 5 years, as fewer temporal sampling points may increase bias of short-term responses to climate fluctuations rather than long-term redistribution trends (Poloczanska et al., 2013). Studies looking at seasonal distribution responses or being only concerned with response predictions were excluded as this review is focused on historical long-term range changes. This review was limited to studies reporting quantified measurements of spatial change in mean latitude, either of centers of distribution (COD), or range edges (mean maximum and minimum latitudes, or lower and upper 5th latitudinal percentile), or estimates of depth changes (in meters) over a defined time span. The final step (  (Fogarty et al., 2017), such measurements should be treated with particular caution due to increased chance F I G U R E 1 Flow Chart representing stages of the study selection process. From the original 896 records found in the bibliographic database Web of Science with search terms shown in Table 1, studies were scanned first by title and abstract for eligibility, and further filtered by criteria concerning methodology (see Table 2). 23 additional studies meeting criteria were retrieved from relevant references, totaling a final of 39 articles included in this analysis. of detection bias and representing outliers (Brown et al., 2016).

Records identified from
Accordingly, due to their reduced spatial and temporal resolution, many studies based on sightings failed to meet the minimum selection criteria. To avoid biases due to local population abundance changes, we excluded estimates based on changes in relative community composition and species richness or stemming solely from abundance data with sparse time points (less than 5 years).

| Data collection process
After the filtering process, an extraction sheet with variables of interest (described under Data Items) was created (Supporting Data S1). We pilot-tested five records and refined the sheet accordingly. In cases where variables were provided only in graphical rather than numerical representations (either not provided or authors were unresponsive to requests), numeric data were extracted manually from graphs using WebPlotDigitizer version 4.5 (Rohatgi, 2021).

| Data items
Information from each study was extracted covering the following:  (Rayner et al., 2003); 5. Whether other significant predictors of distribution changes, such as chlorophyll-a concentrations, ocean currents, pH, and oxygen concentrations in addition to water temperature, were identified by the study; 6. Methods of measurement which were classified into three categories: observations based on abundance data (A), presenceabsence data (P), or a combination of both (AP), where we expect abundance-only data to bias toward lower range shift estimates as it is less influenced by potential outliers as in occurrence data. For the study by , of which raw data were obtained for 29 species, two sets of LRS estimates were included in this study-one set of COD estimates weighted by abundance data and another weighted by presence-absence data, totaling 58 entries included in further analyses. We also considered which portion of a species' range was measuredthe center (usually as the mean latitude or abundance weighed centroid), the leading or trailing edge (the upper and lower percentiles of a species distribution range), as we expect faster response rates at the leading front and center compared to the trailing edge; 7. If and how the overall size of shift (OSS) in depth and longitude/ latitude was provided in a quantifiable form (e.g., °latitudes, km year −1 , or km °C −1 ). Some studies provided only combined averages for grouped species (such as by habitat affinities) either for latitudinal and depth or latitudinal changes only and were marked appropriately (OSS C , OSS C* , respectively), which may reduce accuracy and statistical power in further meta-analyses.
Distinguishing between different OSS reporting approaches (single or multi-population averages) allowed to test for their potential effect on reported distribution responses, as averages from multiple taxa are expected to be less accurate. Studies were further divided into three categories according to sampling frequency: those which measured distributional and temperature changes annually, irregularly (e.g., excluding some years during the study period), or between two points in time, such as studies which divided the study period into cold and warm years according to yearly temperature anomaly estimates and based further analyses on the comparison between cold and warm years; reports based on long-term sighting records (e.g.,  or using population genetic techniques (e.g., Knutsen et al., 2013); and the yearly average of stations fished, such as in studies relying on abundance data from trawling surveys, were extracted. All data are provided in Supporting Dataset S1.

| Summary measures
The aim was to estimate standardized responses of latitudinal and depth shifts in marine fish distributions over time from studies that used a diversity of measurement methods. The meta-analyses were performed by selecting multivariate models with random effects, with the best models chosen according to likelihood ratio tests. We included 'Study' as a random effect to account for multiple estimates derived from the same paper. The primary analytical unit was the estimate for a given species or group of species of distance in latitude moved per year (LRS; km year −1 ) in response to temperature. As the dataset to which the full model was fitted was reduced to 179 data For the best fitting model, marginal and conditional effect sizes (R 2 ) for mixed-effect models were calculated in the MuMIn package (v.1.46.0, Barton & Barton, 2015) according to Equations (1) and (2), respectively. The marginal R 2 represents variance explained by fixed predictors, while the conditional statistic shows the variance explained by both fixed and random effects, f representing the variance of fixed effects, α the variance of random effects, and ε the observation-level variance (Nakagawa & Schielzeth, 2013). Relative contributions of predictors to explained variation in range shift rates were compared by calculating partial marginal R 2 estimates (Nakagawa & Schielzeth, 2013).

| Effect size estimation by correlation coefficients
The relationship between temperature change and LRS was quantified by extracting correlation coefficient (r) values from retrieved studies.   To assess potential publication bias, a funnel plot and regression of the effect sizes (reported LRS estimates) on sample sizes (n) was computed. Depending on the study method, n was either the total number of fish sampled per population or average number of stations per year trawled. Symmetry of the funnel shape was inspected visually and tested with a regression of effect sizes (y) on 1/√n , where p-values below significance threshold (α = 0.05) suggest potential publication bias ( Figure S2a,b). While for funnel plot regression analyses the weighted standard error of effect sizes is most commonly used (e.g., Egger's test; Egger et al., 1997), this measure was not available for most studies and was replaced by sample size in Tang and Liu's test (2000), which addresses the inflated false positive rates associated with the former regression test (Jin et al., 2015).

| Assessment of methodology
(1) Average LRS estimates were expressed in medians and respec-

| Study selection
A total of 39 studies were identified for inclusion in the review (Table S1;

| Geography
The average sample size across the 11 major locations was 54.1

| Methodology
The mean study duration was 41 (±49) years, at a sampling area size of 356,628 (±358,127) km 2 on average. Strong geographical bias of study location was observed-half of the studies originated from North America, with almost a third (31%) performed in Europe.
Australia, Asia, and Africa had less representation with 11%, 6%, and 3% of the identified research articles, respectively. No eligible reports from South America and Antarctica were identified.
Most frequently investigated taxa were classified as tropical (n = 65), followed by temperate (n = 52), deep-water (n = 37), and 19 polar populations ( Figure S1)  From the filtered dataset for outliers, from which the highest likelihood model was fitted, 179 individual population-wise LRS estimates were retained, while 92 entries provided LRS estimates for grouped populations, such as species combined into assemblages according to niche or temperature affinity (e.g., . LRS was on average higher among studies which reported range shift sizes for individual populations as compared to those that grouped populations ( Figure 4c); and lower when based on occurrence data compared to those derived from abundance data or a combination of the two (Figure 4d). Moreover, estimates tended to be lower if studies started in earlier years ( Figure 4h). As sea temperature significantly increased over the years (ANOVA test: F 1,340 = 9.81, p = .002) and was positively correlated with LRS, this effect might be rather due to methodological biases, driven by significantly earlier study start among the fastest shifting temperate and tropical species, as study timing differed significantly among niche affinities (ANOVA: F 3,38 = 34.8, p < .001).

| Factors affecting range shift estimation
Range shift estimates were lower in studies which found significant effects of other non-temperature predictors (Figure 4e).
Besides temperature, the most common explanatory variable for changes in marine fish ranges included oceanic oscillation indexes such as from the Atlantic and Pacific oceans, which was reported nine times across reviewed studies ( Table 3). Other factors included abiotic marine factors such as ocean currents, salinity, depth and chlorophyll-a concentration (n = 9), and exploitation by fishing (n = 7).
Density dependence was mentioned five times, which in some cases had larger effect sizes than temperature.
For individual estimates, the rate of latitudinal shifts was greater in populations which did not change mean depth (β = 9.68 km year −1 ; 95% CI (6.54-12.82); p < .001), compared to populations which were reported to shift their depth distribution (β = 6.58 km year −1 ; 95% CI Studies estimating depth changes based on abundance data found overall decreasing depth responses, while abundanceoccurrence data tended to suggest increasing depths (Figure 6c).

Individuals at the trailing edge of population distributions were
showing the largest move toward deeper waters (Figure 6a), particularly among deep-water species. Tropical taxa showed the slowest depth responses, with shallowing trends at the center and leading edge (Figure 5a). Out of 104 estimated depth shift responses, the majority (73%) shifted in the direction as expected from temperature changes (i.e., to cooler waters).

| DISCUSS ION
We found that the majority of fish populations have responded to thermal warming with a poleward change in their geographical F I G U R E 3 Frequency of methodological aspects across studies. Colors indicate counts for the complete dataset (dark) from 39 retained studies yielding 595 range shift responses to temperature change, or the data included in multivariate models (light) to test latitudinal range shift responses (12 studies with 179 observations). The reduction from 342 latitudinal shift estimates was due to only 179 observations investigating depth changes (e). Plots show (a) data acquisition method; (b) sampling frequency: whether sampling every year, sporadically or comparing two time points; (c) data type: based on abundance (a), presence-absence data (P) or a combination of the two (AP); (d) whether the study performed statistical analyses to confirm temperature effects on range shifts; (f) mean current latitude of sampling area; (g) first year of study period; number of sampled years (h) and taxa (i).  Schickele et al., 2020). Importantly, however, we also found substantial heterogeneity in degree and direction of biogeographical shifts (Champion et al., 2021), which was influenced by both ecological factors such as niche and depth changes, and methodological factors associated with data collection and reporting ( Figure S3).

| Ecological factors influencing distribution responses
We found a significant positive correlation between rate of LRS and latitude which bolsters previous findings by Lenoir et al., 2020, confirming the expectation of faster poleward movements in the Northern Hemisphere where oceans have been warming at faster rates than in the South (Friedman et al., 2013). However, mean current latitude explained only 2% of the variance in LRS, while niche affinity was a more important predictor of latitudinal and depth shifts globally. Results also show that tropical species shift latitudinally more rapidly (Chaudhary et al., 2021;McLean et al., 2021) in response to warming than other marine fishes ( Figure 4a), with disproportionate poleward movements (Figure 2). This is consistent with high sensitivity to temperature change in stenothermic species with narrow thermal tolerance limits and restricted spatial ranges, such as tropical species inhabiting shallow waters close to their tolerance limits (Storch et al., 2014). Indeed, we found that reef-associated fish tended to display the most rapid latitudinal shifts compared to other habitat affinities, although this trend was not significant (Figure 5b). Other studies have shown that, in comparison to temperate fish, tropical species may have increased sensitivity and lower adaptability to thermal increase (Comte &

F I G U R E 4
Latitudinal range shift predictors. Partial effects of fixed predictors included in the final mixed-effect model (∆BIC full-final model = 53.4) explaining latitudinal range shift (LRS; km year −1 ) in response to temperature in marine fish. Points indicate predicted means, and bars and grey shading the 95% confidence intervals. Positive LRS estimates indicate poleward shifts, while negative estimates represent equatorward movements. OSS reporting method had two categories for studies reporting LRS either for taxa individually (single) or the mean of multiple taxa (grouped). Data type was either abundance (A), presence-absence data (P) or a combination of the two (AP). According to temperature change estimates from included studies, tropical populations experienced the slowest yearly temperature increase (0.02 ± 0.02°C year −1 ), followed by deep-water (0.03 ± 0.03°C year −1 ), temperate (0.04 ± 0.03°C year −1 ) and polar taxa (0.05 ± 0.1°C year −1 ). First study year

(h)
Other non−temperature predictors 0.10 TA B L E 3 Frequency of other significant predictors of range changes. Predictors other than temperature with significant effects on redistribution in the 39 reviewed studies were summarized into the shown categories, and counted as unique occurrences across studies (n).

Other identified predictors n
Oceanic oscillation indexes 9 Other oceanic variables (currents, salinity, depth, chlorophyll-a concentration)

9
Fishing pressure 7 Population abundance/density dependence 5 Reproductive (recruitment, spawning stock biomass, buoyancy) 3  Figure 6a). In line with predicted narrow temperature tolerance limits of stenotherms (Storch et al., 2014), we found polar species to experience some of the fastest increases in depth of occurrence. It is well established that polar fish communities can experience rapid and disruptive community structure changes due to arrivals of poleward shifting boreal species Frainer et al., 2017). Experiencing the fastest temperature increase (Stocker, 2014), but being limited in poleward expansion due to the edge of the sea shelf (Wassmann et al., 2006), arctic fish species might depend on moving to deeper waters as a last resort to avoid extirpation . Although leading edges showed faster poleward LRS rates compared to the trailing edge and center (Figure 5a), this difference was not significant, which is in line with previous findings suggesting similar warming sensitivities at opposite distribution fronts (Brown et al., 2016;Lenoir et al., 2020;Sunday et al., 2012), but in contrast to other reports (Poloczanska et al., 2013). Interestingly, faster depth increases were observed at the trailing edge across all niches (Figure 6a), despite similar rates of warming at the trailing and leading edges (mean ∆temperature = 0.03°C year −1 ), suggesting that depth responses at contracting range fronts may be a response to other drivers.
While additional drivers such as habitat and prey availability or resource competition for these responses were not investigated by studies, we found that commercially exploited species changed their mean depths at lower rates than non-target counterparts ( Figure 6d). Restricted responsiveness to climate change in exploited populations might be due to reduced ability to establish in new areas due to localized effects of fishing pressure on abundance and age structure (Rindorf & Lewy, 2006), which has been observed in fish stocks globally .

F I G U R E 5
Non-significant latitudinal range shift predictors. Effects of excluded predictors from simple mixed-effect models with Study as a random effect. Rate of depth change (e, 72 estimates), sapling method (k) and location (l) were tested separately and were not included in the models to avoid overfitting due to limited data. Depth change showed a significantly negative correlation with LRS (p < .001). Positive LRS estimates indicate poleward shifts, while negative values indicate equatorward movements. (e) y = 2.8-3.9x, p < .001 y = 6.2-8.4x10 -8 x, p = .18 y = 10.1-0.14x, p = .06 y = 6.5-0.02x, p = .53 Other factors, such as life stage and taxonomy, were not found to significantly affect latitudinal range shift response, even though sensitivity to warming is thought to be partly dictated by thermal tolerances changing throughout the marine fish life cycle (Killen et al., 2007;Pörtner & Farrell, 2008;Whitney et al., 2013). Early life stages, embryos in particular, are most sensitive with their thermal limit being on average 8°C lower than in other stages (Dahlke et al., 2020), and are likely a major predictor of population responses to warming (Dahlke et al., 2020). ant hypotheses on non-adult temperature response outside laboratory settings (but see Barbeaux & Hollowed, 2018) or inferences of potential range shift limitations in diadromous fish species due to affinity to natal homing grounds .

| The effects of variable study methods
Both LRS and depth responses were greater when estimated from both abundance and presence-absence data together than from abundance data alone (Figures 4d and 6c). Abundance data, mostly obtained from fishery or research trawling data, such as from the Nansen Survey Program in Namibia and Angola , has been widely used across population distribution literature as it is thought to represent the whole population F I G U R E 6 Depth shift predictors. Term plots of fixed predictors included in the final model (selected according to BIC) explaining changes in average depth (m year −1 ) in response to temperature change. Points indicate predicted marginal means, and bars and grey shading the 95% confidence intervals. Data type was either abundance (A), presence-absence data (P) or a combination of the two (AP). Positive depth change values represent deepening, while negative values indicate distribution changes to shallower waters.

(d)
range, and to be less sensitive to search effort and misleading outliers (Brown et al., 2016). While fishery survey data can provide temporally and spatially high-resolution data, and decade-long records can be conveniently retrieved for new analyses, its frequent usage has created publication bias toward commercially important fish species in the northern hemisphere ( Figure S1). Alternatively, recent studies measuring changes in range limits, such as by , use only presence-absence data to infer changes in leading and trailing edges in the Northwest Atlantic, arguing that abundance data do not truly reflect potential changes of species ranges, but is rather confounded by density dependence effects through abundance changes caused by non-climatic factors such as fishing (Quinn & McCall, 1991). However, abundance and climate driven distribution shifts should be possible to distinguish by direction of shift: the former should be unselective in direction while the latter is expected to move along the temperature gradient.
In line with findings by Brown et al. (2016), we observed that studies incorporating occurrence-based data had substantially higher range shift estimates than those using abundance data only, suggesting that presence-absence data may be more sensitive to outliers.
Although response estimates from presence-absence data only were lower than estimates derived from a combination of occurrence and presence-absence data, all former observations originated from one single study and should thus be interpreted with caution.
Previous climate response syntheses have argued that singlespecies studies confirming range shifts consistent with warming may be more likely to be published and thus bias meta-analyses (Parmesan, 2007). While we did not identify publication bias due to low numbers of investigated species or sampled years, we found that range shift estimates reported as the average of multiple taxa were lower than those derived from individual species estimates.
This could be due to random bias due to lower sample size in the four studies from which all group-wise estimates were obtained  or indeed indicate that single-taxa studies are over-represented Contrary to expectations, we found only a weak negative effect of study area size on LRS values (Figure 5f). A plausible source of distributional response variation is the geographical scope of each study, with spatial sampling extents varying widely, and often spanning across whole oceans (e.g.,   . A promising tool to investigate heterogeneity in range shift responses is genetic molecular techniques which help delineate cryptic diversity  and estimate dispersal velocity of locally adapted genotypes (Jonsson et al., 2018). These techniques may improve response predictions and infer historic range changes and migration routes for both ancient and contemporary distribution responses (Knutsen et al., 2013;Robalo et al., 2020;Spies et al., 2020), although such genetic applications to climate range shift research are still scarce.
The variation in species' responses to climate change has been addressed through various predictors such as local adaptation (Jonsson et al., 2018), phenotypic plasticity (Donelson et al., 2019;Reusch, 2014), species interactions Torres et al., 2008), food availability , and even social behavior (Smith et al., 2018). In some marine fishes, the likelihood of successful range expansions and colonization of new habitats was explained by species-level traits such as dispersal ability and being a generalist (Sunday et al., 2015), although traitbased range shift forecasts seem to have generally little explanatory power (Angert et al., 2011). While the majority of reviewed studies investigated (but not always statistically tested) temperature as the sole predictor, a significant proportion of climate response variation is likely explained by a multitude of climatic and biotic factors instead of temperature alone (McHenry et al., 2019).
For example, some studies suggest that range shifts may be driven by abundance changes, as density dependence may lead to range expansions during high abundance and vice versa . Our results suggest that marine range shift estimates from single-predictor studies focusing solely on temperature were higher than those originating from studies which identified at least one additional driver to temperature (Figure 4e), possibly due to the confounding effects of additional variables explaining part of the LRS variation. While some studies found effects of fishing pressure , recruitment level  and spawning stock biomass , marine studies including multiple climatic and non-climatic effects into climate response models are generally scarce. The multi-factor approach was shown to have elevated phenology response estimates in marine organisms when compared to inferences from studies including temperature only (Brown et al., 2016). Thus, further research is needed to explore interactions between climatic and other ecological factors, and to test how these compare to singlepredictor response estimates.

| Opportunities for future improvement
Our conclusions might have been affected by multiple statistical issues and biases associated with meta-analysis (Gurevitch & Hedges, 1999). First, the identified studies mostly originate in the northern hemisphere, particularly Northern Europe and North America with a limited number of fish species (n species = 345) of the estimated ~30,000 fish species present globally (Froese & Pauly, 2022). This suggests a significant research bias and limited taxonomic scope in marine fish climactic research. A common paradox in ecological research is observed whereas taxonomically rich ecoregions, such as the tropics, are strongly underrepresented (Hansen & Cramer, 2015). Very few or no studies could be identified Large variation in publication of LRS and temperature estimates across studies also complicated our interpretations. For example, very few studies presented supporting numeric data of both yearly population center or range edge estimates and high-resolution water temperature data. While some estimates for either of these measures were not possible to extract, others were derived from figures within published papers, which could have affected the accuracy of estimates. Improved temporal and spatial resolution of water temperature estimates, including lagged effects, or implementing tags storing individually experienced water conditions (e.g.,  would likely improve response predictions to climate changes.

| Implications and recommendations
While no single formula for inferring marine fish distribution responses to warming exists, the local ecological factors as well as the extent of current methodological variation biases highlighted here will be key to improving the accuracy and usefulness of research comparing historical distribution data, creating new time series in the future, and synthesizing literature findings. To facilitate future climate impact research, increased standardization and robustness of range shift measurement methods could be achieved by identifying population structure shaped by relevant ecological variables, such as separate spawning grounds or timing (Oomen & Hutchings, 2015;Petrou et al., 2021) and larval retention (Sinclair & Power, 2015), as well as abiotic barriers due to bathymetry, geology, oceanography (Morgan et al., 2009), and genetic factors, such as cryptic diversity and shared local adaptations (DuBois et al., 2022). For underrepresented habitats such as deep-water or tropical niches, improved spatial and temporal resolution (i.e., robust sample sizes of sampled individuals and spatial and temporal sampling frequency in long-term studies), with measurement in all three dimensions (i.e., depth, latitude, and longitude) will be needed to identify vulnerable species and populations. Bias in LRS comparisons over time could be reduced by controlling for locally relevant confounding factors, including phenomena such as the Southern Oscillation affecting temperature trends in the tropics (Jakovlev et al., 2021), or density dependencies, such as in  who accounted for biases by temporal biomass trends. There is an urgent need to expand geographical and taxonomic representation of marine fish range shift responses to climate change. In particular, expansion is needed in the highly biodiverse tropics and global south where marine taxa have been identified as the most vulnerable to warming (Comte & Olden, 2017). As in these regions marine research and long-term fisheries monitoring programs are less established than in the northern hemisphere, robust accounts of whether and how marine fish populations track their temperature niche are lacking.
Addressing the observed variation in marine range shifts will be fundamental for improving response predictions crucial to inform effective fisheries and conservation management strategies, particularly as the magnitude of distribution responses and extinction risk are likely to increase under climate change forecasts (Penn & Deutsch, 2022). In some of the most vulnerable marine ecosystems, such as the arctic, where species have limited thermal tolerance, food web structure and native biodiversity are already rapidly changing due to arrivals of invasive species from lower latitudes (Bartley et al., 2019;Kortsch et al., 2015). Globally, more frequent invasions and resulting novel community structures and interspecific interactions in temperate and arctic latitudes will have likely ecosystem-wide ramifications of yet unknown magnitude (Kortsch et al., 2015;Nadeau & Urban, 2019;Sorte et al., 2010).
Therefore, addressing the natural complexity of distributional responses should rely on innovative and robust methods to allow assessment and comparison of findings.

CO N FLI C T O F I NTE R E S T S TATE M E NT
All authors declare that they have no conflict of interest.

DATA AVA I L A B I L I T Y S TAT E M E N T
The data that support the findings of this study are openly available