Although there is no universal definition of a marine regime shift, they are typically described as an abrupt change in the ecosystem from one state to another, which is detectable in multiple physical and biological components of the system (Lees et al., 2006; Daskalov et al., 2007; deYoung et al., 2008; Andersen et al., 2009; Schwing, 2009). Generally, the magnitude of the regime shift is large and it occurs rapidly relative to the time spent in the different states (e.g. a shift from one year to the next that persists on decadal or longer timescales). The regime shift can be a linear response to an abrupt change in forcing (e.g. climate shift), a nonlinear response to a small change in forcing or driven by the internal dynamics of the system (Andersen et al., 2009; Bestelmeyer et al., 2011), but the exact mechanisms are often unknown.

Key drivers of marine regime shifts include changes in ecosystem habitat, biotic processes such as dynamics of the foodweb and abiotic processes such as changes in physical and chemical conditions (deYoung et al., 2008). These drivers can be natural or anthropogenic, or a combined influence, which can increase the vulnerability of ecosystems (e.g. an ecosystem which has less resilience due to increasing human pressure tends to respond differently to an ecosystem subject only to natural disturbances; Folke et al., 2004). Excessive fishing is an example of an anthropogenic biotic driver where a decrease in top predators (top-down control) can cause a trophic cascade, resulting in a new bottom-up controlled state (Daskalov et al., 2007). Abiotic factors such as climate change or ocean and atmosphere oscillations may initiate bottom-up regime shifts in the food web via changes affecting the abundance of phytoplankton or zooplankton (Cury and Shannon, 2004). Typically, bottom-up driven shifts in biological components of the ecosystem generated by climate shifts manifested in changes in sea surface temperature or mixed-layer depth are considered the most easily identified (deYoung et al., 2008) and are the focus of this study.

Temporal and spatial scales of regime shifts may also affect their detectability (e.g. from a small scale coral reef regime shift occurring within a single year to a North Pacific-wide ecosystem regime shift taking a few years to transition; Drinkwater, 2006; deYoung et al., 2008). Hence, detection of a shift in a large complex marine ecosystem such as the North Pacific or North Atlantic, in which there may be lags between the expression of the shift in the abiotic and biotic components of the system, may be more difficult than detecting a regime shift in a small coral reef (deYoung et al., 2008).

Regime shifts associated with changes in physical conditions have been previously reported in the North Atlantic (Drinkwater, 2006; Beaugrand et al., 2009; Alheit et al., 2014), North Sea (Reid et al., 2001; Beaugrand, 2004; McQuatters-Gollop et al., 2007) and North Pacific (Polovina et al., 1995; Mantua et al., 1997; Hare and Mantua, 2000; Litzow and Mueter, 2014), among others. The late 1970s North Pacific regime shift has been comprehensively studied (Mantua et al., 1997; McGowan et al., 1998; Francis et al., 1998; Hare and Mantua, 2000; Yatsu et al., 2008). It was observed in a composite time series of 100 physical and biological variables, which revealed an abrupt and sustained change during 1976/77 (Hare and Mantua, 2000). At that time, there was a deepening of the Aleutian low-pressure system which doubled the eastward wind stress and brought cooler winds over the central North Pacific, causing a drop in sea surface temperature (SST) and a deepening of the mixed-layer depth (MLD). This resulted in moister and warmer air settling over the California Current region and the Gulf of Alaska, which caused an increase in SST in these two regions (Mantua et al., 1997). This mechanism has been described as the Pacific Decadal Oscillation (PDO), which switched from a negative to a positive state in 1976/77 (Mantua et al., 1997). The late 1970s shift is thereby implicitly related to El Niño Southern Oscillation (ENSO) variability, whose shorter timescale fluctuations combined with random atmospheric forcing enforce decadal variability in the PDO (Newman et al., 2003). Alternatively, other large-scale climate patterns such as the North Pacific Gyre Oscillation (NPGO) may impact on marine ecosystem dynamics. Concurrent with the switch in the PDO state, an increase in zooplankton biomass was observed in the Gulf of Alaska between the periods of 1956–1962 and 1980–1989 (Brodeur and Ware, 1992). In upper trophic levels, abrupt increases in groundfish recruitment and salmon catches were observed, while some forage fish populations collapsed with consequences for piscivorous sea birds and marine mammal populations (Anderson and Piatt, 1999). Overall the yield of fish stocks in the Gulf of Alaska increased from the 1970s to the 1990s (McGowan et al., 1998).

Although a climate shift occurred over the entire North Pacific, the ecological response varied between regions depending on their respective dominant processes (Schwing, 2009). For example, in the California Current region the ecological changes associated with the 1977 climate shift were different from those that occurred in the Gulf of Alaska with lower salmon catches after 1977 (Mantua et al., 1997). Investigation of the magnitude and extent of the regime shift and the proposed mechanism is challenged by the paucity of data covering adequate time and space scales in the Gulf of Alaska. Most support for the observed biological changes comes from fisheries stock assessments, which are not designed to study how climate shifts are affecting marine ecosystems (McGowan et al., 1998). A few modelling studies have attempted to simulate the chain of events for the late 1970s shift, but the direction of changes in the simulations of the physical and biological parameters are sometimes opposite, and also vary according to the space/timescale of the study (e.g. Polovina et al., 1995; Haigh et al., 2001; Capotondi et al., 2005; Alexander et al., 2008). By using the late 1970s regime shifts in the Gulf of Alaska as a case study we aim to assess the ability of five global ocean biogeochemical models to simulate this shift. The models were part of the UK Integrated Global Biogeochemical Modelling Network (iMarNet) intercomparison, which aimed to evaluate the models' ability to simulate global-scale bulk biogeochemical properties using the same ocean general circulation model and atmospheric forcing (Kwiatkowski et al., 2014). These physically identical hindcast simulations allow any model differences to be ascribed only to their representation of biogeochemical processes, thereby providing insight into the mechanisms leading to marine regime shifts.

A substantial part of the literature on regime shifts uses principal component analysis to compress a large number of time series representing the state of the ecosystem to a smaller number of uncorrelated ones, which indicates to what extent the different components of the system are responding coherently. For example, Hare and Mantua (2000) reduced a total of 100 time series of physical and biological variables representing the state of the North Pacific to two leading modes of variability. The presence of regime shifts in the reduced set of time series may render the presence of shifts more evident to visual inspection, but this is often done without further significance testing (Andersen et al., 2009). In order to objectively identify the timing of a shift and distinguish it from a random fluctuation, change-point techniques can be used, especially methods designed to detect multiple shifts in the mean of a time series (e.g. Andersen et al., 2009). For example, the shift detection methodology proposed by Rodionov (2004) consists of applying a t-test successively to compare the means of two segments of a time series, considering all possible timings for a shift, and repeats this until all shifts have been detected. This method has been applied widely in the marine regime shift literature (e.g. Daskalov et al., 2007; DeYoung et al., 2008; Overland et al., 2008, 2010; Yatsu et al., 2008; Möllmann et al., 2009). However, it is not designed to distinguish a shift from a trend, which may lead to the detection of a series of spurious shifts in the presence of a background long-term trend (e.g. Spencer et al., 2011). Furthermore, it may lead to the detection of spurious shifts in the presence of red noise, which creates patterns that may be interpreted as shifts, but which are purely random (e.g. Wunsch, 1999; Rudnick and Davis, 2003). Red noise is often present in biological time series such as chlorophyll (e.g. Beaulieu et al., 2013) or plankton abundance (e.g. Di Lorenzo and Ohman, 2013), and manifests through a slow integrated response to random weather forcings (Di Lorenzo and Ohman, 2013). Therefore, we opt for a methodology capable of separating a long-term trend from an abrupt change signal (e.g. which occurs from one year to the next) and distinguishing these signals from red noise (Beaulieu et al., 2012). In order to provide further insights as to whether the shifts detected are a linear response to a shift in the forcing itself (e.g. climate shift) from shifts generated through a nonlinear response of some change in the forcing, also called thresholds or “tipping points” (Scheffer et al., 2009), the relationship between the forcing and the response was explored using regression models (Bestelmeyer et al., 2011).

Our analysis is organized as follows. First, we investigate whether shifts are present in the Gulf of Alaska as predicted in a multiple model intercomparison hindcast experiment, iMarNet (Kwiatkowski et al., 2014; www.imarnet.org). Specifically, we analyse model physical and biological variables for regime shifts and verify whether these shifts are internally coherent. Then, we investigate the contribution of the different physical and biological variables to the observed late 1970s and late 1980s shifts in the Gulf of Alaska and the type of forcing–response relationship that led to abrupt changes.

Figure 2 presents the results of the change-point analysis on gridded SST observations for the North Pacific. This reveals a predominant shift in 1977 over the Gulf of Alaska region, which also extends as a coastal band towards the California Current region and the Bering Sea. A late 1980s shift is detected in a smaller area in the middle of the gyre. It must be noted that spatial homogeneity of the shift is to be expected here given that the ERSST dataset is produced using empirical orthogonal functions (Smith and Reynolds, 2003).

In the observed SST time series averaged over the Gulf of Alaska, a statistically significant shift is detected and manifests as a rapid increase in the mean of ∼ 1 ∘C after a decreasing trend (Fig. 3a). In the model physical time series (which are identical in all five OBGC models), SST exhibits the same signal as the observations: a shift in the intercept and gradient occurring in 1976, while the MLD is best represented by a linear trend. However, the model MLD time series show strong decadal variability with large changes occurring in the mid-1970s and at the beginning of the 1990s (Fig. 3b–c). Results of change-point analysis on large-scale oscillations characterizing the climate over the region are also presented in Fig. 3d–f and show a significant shift in the PDO in the late 1970s while the NPGO and MEI annual time series do not indicate a shift.

The change-point analysis was performed on PC1 for each model (Fig. 4, Table 3), which explains most of the variance for each model (except MEDUSA, 36 % of variance explained; Table 4). HadOCC exhibits a shift in 1977 in PC1 (Table 3), for which all variables except MLD have large relative contributions (> 10 % relative contribution, Table 4). The first principal component in Diat-HadOCC exhibits a shift in 1976 and explains 63 % of the total variance. The variable offering the smallest relative contribution is again the MLD (Table 4). In MEDUSA, a shift is also detected in the late 1970s in the first component, which explains only 36 % of the variance. The SST, CHL and nutrients are the most important variables with relative contributions larger than 10 % (Table 4). The MLD has a relative contribution of 0.94 % to PC1 (Table 4). The relative contributions of the nutrients in the HadOCC, Diat-HadOCC and MEDUSA late 1970s shift detected in the first principal component suggests the controlling factor is nutrient limitation (i.e. bottom-up control) in these models. In ERSEM and PlankTOM10, there are no shifts detected in the first principal component. Similar results are obtained when excluding SST and MLD from PC1 (Appendix A).

The results of the change-point analysis on all observational and model individual time series are presented in Appendix B (Table B1). The fit of the most appropriate statistical models for the biological variables for each OBGC model are also presented in Appendix B (Figs. B1–B5). Statistically significant shifts are found more often in the simpler OBGC models (HadOCC, Diat-HadOCC and MEDUSA) than the complex ones (Table B1), which is consistent with the results obtained on the first principal component for each model. Of the statistically significant shifts identified in these models, the majority occurred in the late 1970s. In HadOCC, the late 1970s shift corresponds to a decrease in DIN, CHL, PHY and ZOO, while a large increase in PP is detected in 1991. Nevertheless, PP is decreasing over the period 1957–1990 (Fig. B1). In Diat-HadOCC, all parameters exhibit a shift in the late 1970s, although it is not significant in PHY and ZOO. The significant shifts in the late 1970s manifest as a decrease in SI, FE, CHL and PP. In MEDUSA, shifts in DIN and FE (although not significant) are identified in the late 1970s. ERSEM exhibits a significant shift in CHL in the late 1970s, while PlankTOM10 does not have any significant shifts for that period.

As a visual support for the change-point analysis, cumulative sums of the z scores of each time series within each model are presented in Fig. 5. A shift in a time series is revealed by a change of slope of the cumulative sums. The change of slope in SST is sharp, as one would expect given the significant shift detected. Even though our analysis suggests a long-term deepening of the MLD rather than an abrupt change, subtle decreases are suggested by smooth change of slope in the cumulative sum in the late 1970s and 1980s. These changes are clearly propagated to the other parameters in HadOCC, DiatHadOCC and MEDUSA with a sharp change of slope, but smoother change in ERSEM and PlankTOM10.

We further investigate the forcing–response relationship between SST and the biological variables (CHL, PP, PHY, ZOO) in HadOCC, DiatHadOCC and MEDUSA (Fig. 6) before and after 1977, as a significant shift is present in PC1 in these models. The slopes of the linear relationships between SST and the biological variables are mostly similar before and after 1977 (Table 5). This is consistent with a linear, rather than nonlinear, response to changes in SST forcing. There is one exception for ZOO for which the difference in slopes is significant with a stronger relationship after 1977 in HadOCC and DiatHadOCC (Table 5, Fig. 6), which could suggest an amplified nonlinear response.

Using the Gulf of Alaska as a case study, our results demonstrate the usefulness of OBGC models to infer the chain of events responsible for regime shifts, especially in regions where observations are scarce. Although there are many definitions of regime shifts in the literature, they can be generally described as an abrupt change (e.g. from one year to the next) that occurs across both physical and biological parts of the ecosystem. Therefore, to determine if a regime shift has occurred in the five OBGC models tested here the shift has to be traceable from physical parameters through to biological parameters. With the change-point detection method used here, we found statistically significant shifts in the late 1970s in the Gulf of Alaska simulated in five OBGC models. A shift in model SST occurred in 1976 and matched a shift in observed SST. This abrupt change in SST was followed by an overall decrease in nutrients and productivity. The three OBGC models simulating an abrupt change in 1977 in PC1 (i.e. HadOCC, DiatHadOCC and MEDUSA) are consistent in the direction of change (Fig. 4). The decrease in nutrients after 1977 seems to be the dominant driver in the reduction in productivity and outweighs changes due to fluctuations in light availability. The dominance of declining nutrients in explaining the variability in the principal components of HadOCC, Diat-HadOCC and MEDUSA, supports this hypothesis.

Previous studies have linked the late 1970s shift in the North Pacific with the PDO, which switched from a negative to a positive state in 1976/77 (Mantua et al., 1997). The PDO fluctuations have been suggested to exhibit a red noise response to atmospheric noise and ENSO events (Newman et al., 2003), thereby raising the possibility of a link between ENSO and the North Pacific shift in the late 1970s. Nevertheless, the PDO (and implicitly ENSO) alone is not enough to characterize the North Pacific climate (Bond et al., 2003). Alternatively, the North Pacific Gyre Oscillation (NPGO) has been suggested as a global-scale mode of variability that plays an important role in decadal changes in marine ecosystems (DiLorenzo et al., 2008). For example, in the California Current, the PDO correlates with SST while NPGO is more closely related to variability in salinity, nutrient and primary production (DiLorenzo et al., 2008). Thus, if both the PDO and NPGO fluctuations drive changes in the North Pacific climate and ecosystem functioning, the question arises whether either or both of these indices exhibit a shift at a similar time. Underscoring some of the conclusions of the prior work discussed above, the shift in 1976/77 manifests in the PDO index, but notably we find no significant shifts in the multivariate ENSO index or the NPGO index. Clearly, by detecting a shift in the late 1970s in PDO only we cannot conclusively tie the PDO and untie the NPGO and ENSO to the shift in climate and ecosystem dynamics of the Gulf of Alaska. However, these corresponding changes are an important piece of information to future work aimed at determining causal mechanisms, mode of teleconnection and coupled physical/biogeochemical dynamics that link global climate patterns to ocean fertility of the Gulf of Alaska.

In conclusion, the 1977 regime shift in the Gulf of Alaska was observed in sea surface temperature and in the abundance of a range of commercial fish species (McGowan et al., 1998). Here, we infer the behaviour of the nutrients and lower trophic levels using OBGC models, and the relationship of these changes to physical variables that are plausible drivers. Our novel approach based on change-point detection offers a helpful framework to evaluate previous modelling studies that have attempted to reproduce the extent of changes from physics to biology for the late 1970s shift in the Gulf of Alaska (e.g. Polovina et al., 1995; Haigh et al., 2001; Capotondi et al., 2005; Alexander et al., 2008). For example, Haigh et al. (2001) used the Miami isopycnic coordinate ocean model combined with an ecosystem model of four compartments (Denman and Peña, 1999) to show that a year-round deepening of the mixed-layer depth after 1976 led to a slight decrease in nutrients and phytoplankton as well as zooplankton biomass. These findings are broadly consistent with the model simulations analysed here. Other studies instead suggest that the MLD shoaled after 1977 resulting in increased plankton production in the region. This is the case in the Polovina et al. (1995) study, which suggested that shoaling in the spring/winter MLD led to increased productivity in a plankton population dynamics model. More recently, Alexander et al. (2008) used the National Center for Atmospheric Research Climate System Model Ocean Model (NCOM) combined with a biological model that contains 10 compartments (Chai et al., 2002) to simulate the chain of events in the region. In that study, an increase in SST simulated in the late 1970s is accompanied by a shoaling in the winter mixed-layer depth, giving rise to an early spring increase in primary production, phytoplankton and zooplankton biomass followed by a late spring decline in both phytoplankton and zooplankton biomass. Despite the caveat that we are analysing annual mean time series it is important to point out the contradictory direction of change in mixed-layer depth. Possibly reconciling this discrepancy, Capotondi et al. (2005) suggest, based on NCOM model simulations, a deepening trend in MLD in a broad band along the coast and shoaling in the central part of the Gulf of Alaska. Thus, the comparison of the various attempts to simulate the late 1970s regime shift of the Gulf of Alaska raises the possibility that the observed abrupt and spatially coherent ecosystem change was actually caused by a previously unappreciated heterogeneous set of environmental changes with distinct spatial pattern and timing in the annual cycle. If so, the inherent assumption underpinning our own and previous work to understand the Gulf of Alaska ecosystem shift as a single mechanistic causal sequence may be overly simplistic. Consequently, future analysis aimed at spatial and temporal heterogeneity of abrupt regional ecosystem change has the potential to greatly improve our understanding of the underlying dynamics and the vulnerability of marine ecosystems to abrupt future changes.

A second major outcome of this study involves the role of model complexity in determining a system's propensity for abrupt ecosystem change. All the OBGC models used in this study have the same underlying physical model, and were run with the same initial conditions and forcing fields. Their performance in terms of a fit to observations has been assessed globally in a previous study by Kwiatkowski et al. (2014), showing that all models have skills in simulating some variables, but simpler models were broadly closer to observations overall. In the Gulf of Alaska, the five models systematically differ in nutrient and biological responses as a function of model ecosystem complexity. Simple to intermediate complexity models such as HadOCC, Diat-HadOCC, and MEDUSA simulate a shift in the late 1970s, which manifests as an abrupt change in SST and many nutrients and biological parameters. As the model complexity increases to PlankTOM10 and ERSEM, these changes are mostly in the same direction but become less abrupt. The simpler models have fewer plankton groups responding to environmental changes (both HadOCC and Diat-HadOCC have one zooplankton group, and Diat-HadOCC has two phytoplankton groups), which might explain a more direct response than a model with a larger number of plankton groups interacting with each other. More complex models could potentially unveil shifts in the community structure (i.e. increase of a certain type of plankton and decrease of another one), as regime shifts can affect different species in opposite ways (Benson and Trites, 2002). Feedbacks and interactions between groups in the models are in need of thorough exploration to determine how they affect the simulation of observed regime shifts. Such differences between model results raise the question as to what degree of model complexity is needed to appropriately simulate the complexity of regime shifts in the real world. Extremely simple models are easy to interpret but may not be able to reproduce realistic behaviour, while too much complexity will lead to uncertainty and problems in interpretation of the model (Allen et al., 2010). Given the observed differences between models, our results suggest caution on relying on a single “ultimate” model for understanding regime shifts behaviour and rather favour multiple lower to intermediate complexity models, as recommended by Fulton et al. (2003). However, our results should not be generalized too easily, as we focused uniquely on the Gulf of Alaska region here. More complex models could outperform simple models in different ecosystems. For example, higher complexity models have been suggested to be more portable (i.e. ability to perform well in diverse regions and physical settings) in a comparative study focusing on the equatorial Pacific and Arabian Sea (Friedrichs et al., 2007). Future work should involve a regime shift analysis in several ecosystems using models with traceable complexity. Furthermore, an ensemble approach to quantify the effects of model and internal variability uncertainty in regime shift detection would be beneficial.

Our analysis suggests that the Gulf of Alaska regime shift is consistent with a linear response to physical forcings on lower trophic levels, showing a bottom-up response due to changes in the physical environment controlled via nutrient limitation, with a potential amplified response from ZOO (only in HadOCC and Diat-HadOCC). This result is in agreement with the linear tracking window hypothesis (Hsieh and Ohman, 2006), which suggests that some populations can respond linearly to abrupt changes in physical forcing, as opposed to an amplified nonlinear response to small changes in forcing (e.g. Scheffer et al., 2009). However, it must be noted that our analysis is lacking top-down controls from upper trophic levels beyond zooplankton, and thus only partly resolves possible explanations for the observed regime shifts in the Gulf of Alaska. Many drivers (and their synergistic effects) may combine to fully explain regime shifts (Lindegren et al., 2012; Litzow et al., 2014). Models including upper trophic levels able to simulate regime shifts would also be beneficial to better understand the mechanisms leading to the shift and estimate critical thresholds.

Finally, beyond model complexity and the spatial and temporal resolution at which the output is analysed, the state-of-the-art in statistical techniques for regime shift detection is an active area of research. Here we employ an approach to detect shifts and distinguish them from a long-term trend and background red noise, i.e. evaluate if the shift is unusually large given the fluctuations that would be expected in the presence of autocorrelation and/or a trend (Beaulieu et al., 2012), which is an improvement over previous methodologies. A main current limitation of this methodology is the ability to detect at most one shift and in one time series at a time (univariate), but work to extend the methodology to detect multiple shifts in a multivariate setting is under way. Furthermore, we distinguish against a background of red noise, which is assumed constant through the time series, but the presence of changes in the red noise through time could affect the results. For example, a recent study suggests a “reddening” of the PDO and North Pacific SST as an explanation for occurrences of abrupt changes in the North Pacific ecosystem (Boulton and Lenton, 2015). However, this is unlikely to affect our results given the timescale (annual means) and length of the time series (51 years) used in this study. Further, we suggest here that analysis of the forcing–response relationship helps to distinguish between a regime shift with a linear response to a shift in forcing, and a nonlinear response after crossing a forcing threshold, as originally proposed by Bestelmeyer et al. (2011). Here we used a test that is based on a quantitative comparison of the forcing–response relationship before and after the shift. This approach can be used to detect other marine or terrestrial regime shifts and distinguish between a linear and a nonlinear response to external forcing. For management purposes, distinguishing between these two types of forcing–response relationship producing regime shifts is critical, as they will lead to different management and policy incentives (Kelly et al., 2015). For example, a routine monitoring of threshold-based systems leads to better management outcomes than “threshold-blind” management, i.e. when ignoring the possibility of a threshold and assuming a linear forcing–response relationship (Kelly et al., 2015).

The outputs from five ocean biogeochemical models averaged over the Gulf of Alaska are available as a Supplement to this article. Observed sea surface temperature data were extracted from the Extended Reconstructed Sea Surface Temperature (ERSST) dataset (version 3b; Smith and Reynolds, 2003; Smith et al., 2008) downloaded from https://www.ncdc.noaa.gov/ersst/. The Pacific Decadal Oscillation index (Mantua et al., 1997) was downloaded from http://www.atmos.washington.edu/~mantua/abst.PDO.html. The Multivariate El Niño Southern Oscillation Index (Wolter and Timlin, 1998) was downloaded from http://www.esrl.noaa.gov/psd/enso/mei/ and the North Pacific Gyre Oscillation index (DiLorenzo et al., 2008) was downloaded from http://www.o3d.org/npgo/npgo.php.