Artificial selection on reproductive timing in hatchery salmon drives a phenological shift and potential maladaptation to climate change

Abstract The timing of breeding migration and reproduction links generations and substantially influences individual fitness. In salmonid fishes, such phenological events (seasonal return to freshwater and spawning) vary among populations but are consistent among years, indicating local adaptation in these traits to prevailing environmental conditions. Changing reproductive phenology has been observed in many populations of Atlantic and Pacific salmon and is sometimes attributed to adaptive responses to climate change. The sockeye salmon spawning in the Cedar River near Seattle, Washington, USA, have displayed dramatic changes in spawning timing over the past 50 years, trending later through the early 1990s, and becoming earlier since then. We explored the patterns and drivers of these changes using generalized linear models and mathematical simulations to identify possible environmental correlates of the changes, and test the alternative hypothesis that hatchery propagation caused inadvertent selection on timing. The trend toward later spawning prior to 1993 was partially explained by environmental changes, but the rapid advance in spawning since was not. Instead, since its initiation in 1991, the hatchery has, on average, selected for earlier spawning, and, depending on trait heritability, could have advanced spawning by 1–3 weeks over this period. We estimated heritability of spawning date to be high (h 2 ~0.8; 95% CI: 0.5–1.1), so the upper end of this range is not improbable, though at lower heritabilities a smaller effect would be expected. The lower reproductive success of early spawners and relatively low survival of early emerging juveniles observed in recent years suggest that artificial and natural selection are acting in opposite directions. The fitness costs of early spawning may be exacerbated by future warming; thus, the artificially advanced phenology could reduce the population’s productivity. Such artificial selection is known in many salmon hatcheries, so there are broad consequences for the productivity of wild populations comingled with hatchery‐produced fish.


| INTRODUC TI ON
The ability of animal populations to respond and persist in the face of myriad human-induced environmental changes is a key concern for scientists, natural resource managers, and those who rely on fish and wildlife for their livelihoods (Badjeck, Allison, Halls, & Dulvy, 2010;Dolan & Walker, 2006). In the face of rapid and often unpredictable environmental change, maintaining or increasing resilience-the capacity of individuals, populations, and ecosystems to adapt and persist-is a common goal of resource management and conservation (Allen, Cumming, Garmestani, Taylor, & Walker, 2011;Gunderson, 1999;Quinlan, Berbés-Blázquez, Haider, & Peterson, 2016;Walker, Holling, Carpenter, & Kinzig, 2004). Pathways for adaptation include phenotypic plasticity (i.e., behavioral, morphological, and physiological changes at the individual level), range shifts, and evolution (Bernhardt & Leslie, 2013;Parmesan & Yohe, 2003). However, because environmental changes and resulting biological responses are uncertain, the specific traits, populations, or species most likely to persist under future conditions are seldom known (Webster et al., 2017). Consequently, management strategies to maximize resilience often focus on genetic, phenotypic, and species diversity, and the connectivity of populations and habitats. Indeed, benefits of maintaining diversity at multiple scales are widely recognized (Schindler, Armstrong, & Reed, 2015) and include increased abundance and reduced year-to-year variability in productivity (Hilborn, Quinn, Schindler, & Rogers, 2003;Schindler et al., 2010). On the other hand, human activities including some management and conservation strategies may unintentionally erode the diversity on which resilience relies (Carlson & Satterthwaite, 2011;Webster et al., 2017).
In addition to harvest, artificial propagation of species for subsequent release to join wild populations can also exert selection and spread maladaptive traits (Derry, 2019;Frankham, 2008). If propagated individuals are later integrated with wild breeders, these trait changes may spread in the broader population, and if the trait changes are maladaptive in the wild, then conservation efforts may undermine the population's productivity or persistence (Baskett & Waples, 2013;Gering, 2019). The approach is particularly common in anadromous fishes where propagation methods are well established, and freshwater habitat loss or degradation often limits populations (Lorenzen, 2005). Indeed, billions of juvenile Pacific salmon (Oncorhynchus spp.) are released from fish hatcheries around the Pacific Rim each year (Ruggerone, Peterman, Dorner, & Myers, 2010), to increase harvest opportunities, supplement depleted wild populations, or prevent extinction . These fish experience different regimes of selection on a number of traits, including reproductive timing, a highly heritable trait in salmonids (McLean, Bentzen, & Quinn, 2005;Quinn, Peterson, Gallucci, Hershberger, & Brannon, 2002;Tipping & Busack, 2004). In some hatchery salmon populations, reproductive timing has been intentionally altered through selection to temporally separate hatchery from wild runs or to enhance the efficiency of the hatchery operations (Crawford, 1979). In other cases, hatchery managers tend to spawn early arriving fish because they are uncertain how many fish will eventually return (McLean et al., 2005), or there are practical challenges to capturing fish for breeding later in the season (e.g., high stream flows). Either or both processes can cause unintended advances in spawning timing (e.g., Flagg, Waknitz, Maynard, Milner, & Mahnken, 1995;Quinn et al., 2002). Early return may reduce fitness by exposing adults to higher water temperatures (Quinn & Adams, 1996) that can result in prespawning mortality (Bowerman, Roumasset, Keefer, Sharpe, & Caudill, 2018). Earlier spawning also advances hatching and emergence of juveniles, potentially leading to a mismatch with environmental conditions or prey resources (Quinn, 2018). Moreover, the salmon that return to spawn over the course of the season commonly differ in body size, in-stream life span, and reproductive allocation (Doctor & Quinn, 2009 Quinn, 1999); thus, timing is a very important trait with direct and indirect connections to many aspects of salmon ecology and life history.
Given the prevalence of salmon hatcheries, their potential influence on migratory and reproductive timing, and the importance of phenology in adaptation to climate change, an interaction between these two selective forces seems probable. In this study, we examine a potential case of such counteracting selection in a population of sockeye salmon (O. nerka) that has undergone marked changes in reproductive timing over the past four decades and has been supplemented with hatchery propagated fish as an increasing proportion of the total run since the early 1990s ( Figure 1). Prior to 1991, the average escapement to the Cedar River, Washington, USA, (Figure 2) was ~213,000 with no hatchery contribution. In the decades since, the average run size has declined while an increasing proportion of returning adults have been captured for spawning in the hatchery (1990s: 87,000 escapement, 8% spawned in hatchery; 2000s: 77,000 escapement, 11% spawned in hatchery; since 2010:45,000 escapement, 28% spawned in hatchery). Meanwhile, between the early 1970s and early 1990s, natural reproduction by sockeye salmon in the Cedar River appeared to be shifting later in the year, consistent with an effect of local environmental change (e.g., Warren, Robinson, Josephson, Sheldon, & Kraft, 2012) during this period (i.e., increasing water temperatures and decreasing late-summer flow). In the mid-1990s-approximately coincident with the initiation of a hatchery supplementation program-this pattern apparently reversed and spawning became progressively earlier in the year. Despite annual hatchery supplementation and an absence of targeted fishing since 2006, the population has declined in abundance in recent years.
Because naturally spawned and hatchery-produced fish interbreed in both the hatchery and the river, there is concern that unintended selection in the hatchery may have contributed to the recent trends in reproductive timing and abundance in the integrated population.
Our overall goal was to describe and explain the patterns of migration and reproductive timing in Cedar River sockeye salmon, testing two hypotheses that might explain them. First, we considered that changes might be phenotypic responses to environmental variation in marine or freshwater habitats (e.g., water temperature and river discharge: Hodgson, Quinn, Hilborn, Francis, & Rogers, 2006). To do so, we compared the patterns of salmon migration and spawning timing to a variety of data sets on conditions along the migratory corridor from the coastal ocean to the spawning stream, including a comparison to the patterns in another Puget Sound sockeye salmon population, Baker Lake. Second, we assessed the potential role of directional selection from hatchery operations. To do so, we estimated the heritability of spawning timing based on data from this system and from estimates of other salmonid species.
We then determined whether hatchery operations have been selective, relative to spawning in the river. Finally, we modeled the extent to which the observed changes in timing could be explained by the strength of selection and heritability of timing, and conclude by discussing the potential fitness consequences of artificially altered F I G U R E 2 Study region showing the Skagit River and Lake Washington watersheds with sockeye salmon enumeration locations identified reproductive phenology for this population, given the regional environmental conditions.

| Study system
Lake Washington, WA, USA, is a large, natural lake near Seattle, WA, that has been substantially altered over the past century by human development. Today, the primary input is the Cedar River, which enters from the south and contributes over 50% of the annual inflow (Arhonditsis, Brett, Degasperi, & Schindler, 2004); the lake drains to the marine waters of Puget Sound through the Lake Washington Ship Canal (Figure 2). Prior to the construction of the Ship Canal and the Hiram M. Chittenden Locks (commonly, and hereafter, called the Ballard Locks), and diversion of the lower river reaches in the early 20th century, the Cedar River did not flow into Lake Washington; rather, the lake drained through the now dry Black River (Edmondson, 1991). Because sockeye salmon typically require lakes for juvenile rearing, it is generally thought that the Cedar River had few if any sockeye salmon prior to hydrological modifications of the Lake Washington system (Darwin, 1917). Lake Washington sockeye salmon are thought to be primarily descendants of fish from Baker Lake, Washington, a tributary to the Skagit River located ~120 km north of the Cedar River that were stocked in the Cedar River and other tributaries between 1934 and 1944 ( Figure 2; Ames, 2006;Spies, Anderson, Naish, & Bentzen, 2007). Consequently, we also obtained and analyzed data from the Baker Lake system for comparison with the marine entry timing of the Lake Washington fish, as they might be affected by common oceanic factors. Although sockeye salmon spawning also occurs on Lake Washington beaches and in other tributaries, the Cedar River has produced on average more than 85% of all returns since adult counting efforts began in the 1960s. Since 1991, a portion (2%-61%; mean 14%) of returning Cedar River sockeye have been collected using a weir in the river for spawning in a hatchery and their offspring released back to the river as fry to complete the rest of their lives naturally. Hatchery fish are not externally marked, and no effort is made to separate hatchery and natural origin fish. Both hatchery and naturally spawned fish typically spend one year rearing in the lake and two or (less often) three years at sea, and thus achieve a total age of four or five years before returning to complete their life cycle.

| Environmental data
Cedar River flow data were obtained from USGS gaging stations located at river kilometer (rkm) 2.6 near Renton, and rkm 32.8 below the Landsburg diversion dam; daily average discharge was available for all study years . Water temperature data along the migratory route (Lake Washington, Lake Union (a small lake between Lake Washington and Puget Sound), and the Ship Canal) were obtained from King County (https ://green2.kingc ounty.gov/ lake-buoy/) and the University of Washington (Edmondson, 1994).
Temperatures were typically recorded monthly or biweekly, in which case observations were averaged by month. Monthly temperature data were available for the Ship Canal nearly continuously since 1975 and range from the surface to 5 m depth, while in Lake Washington observations began in 1965 and range from the surface to 60 m. Daily surface water temperatures recorded at a marine site along the migration route into Puget Sound (Race Rocks, British Columbia) were obtained from racerocks.com and are available for all study years.
We summarized flow and temperature data into potential covariates of timing based on their general influence on sockeye salmon timing (Hodgson & Quinn, 2002) or the Cedar River population in particular (Ames, 2006). For Lake Washington water temperatures,

| Salmon timing data
Sockeye salmon returning to Lake Washington enter freshwater at the Ballard Locks which separates the Lake Washington Ship Canal from the marine waters of Puget Sound. Since 1972, state and tribal fisheries personnel have conducted daily counts of sockeye salmon passing through the locks and/or the fish ladder during an index (12 June-31 July) period that captures on average >95% of the run (i.e., salmon returning to the basin as a whole), with the great majority destined for the Cedar River. Expansion factors are used to account for proportions of the run that occur at night or during uncounted locking cycles (Ames, 2006). Although imperfect, these daily visual counts were very similar to estimates from hydroacoustic methods (Thorne, 1979) and are considered unbiased with regard to the hypotheses addressed here. Counts have ranged from 22,159 to 530,063 per year and have been consistently low since 2007 (Pre-2007 mean: 261,440;-2015. This stage, when the salmon transition from marine to freshwater environments, we hereafter refer to as arrival which is separated in space and time from entry onto the spawning grounds and breeding. In addition to the timing of arrival in freshwater, the other life history event of primary interest was the timing of spawning in the Cedar River. This could not be directly observed at the population level and so was estimated from a two-step process, detailed below. Briefly, surveys of live salmon in the river yielded an annual estimate of occupancy, which was the basis of an estimate of when those salmon entered the Cedar River, and from that we estimated the timing of spawning. Surveys of live adult salmon in the Cedar River have been conducted by Washington Department of Fish and Wildlife (WDFW) or other agencies using standardized methods since 1969 (Ames, 2006). An index reach extending from rkm 6.8 upstream to the Landsburg Diversion Dam (rkm 35.1), the upper limit of sockeye spawning, has been surveyed since 1969. Occasional surveys of the lower section of the river (below rkm 6.8) documented few fish and little spawning activity prior to 1987, but the use of this reach by sockeye salmon increased dramatically in recent decades (Timm & Wissmar, 2014).
Thus, since 1987, survey counts have been reported for all available habitat (rkm. 0.0-35.1) in addition to the traditional index reach. The inclusion of the lower reaches in more recent years was not expected to bias our analysis because few fish were present in this reach prior to the late 1980s, and because fish spawning in the lower river tend to enter later, which should lead to a delay in observed timing-the opposite of the recent trend. To summarize timing of entry into the Cedar River, survey counts were converted to estimated daily counts through linear interpolation (Barnett, Simmons, & Peterson, 2013). Interpolated daily counts of sockeye salmon for 1969-2012 were obtained from prior reports on counts of live Cedar River sockeye salmon (Cascade Environmental Services Inc. 1995, Barnett et al., 2013) and updated through 2015 with field data from WDFW. As in previous reports, five years (1970, 1973, 1974, 1975, and 1990) were excluded from analysis because too few surveys were conducted to be reliable (Barnett et al., 2013).
Although survey counts of live salmon allow for interannual comparison of timing of occupancy, they do not allow for evaluation of selection on spawning timing within a year, because individual fish may be observed multiple times throughout the spawning season.
We therefore used a simple accounting model based on interpolated daily in-river abundance and estimated in-stream life span to calculate the number of fish dying each day. Then, the estimated number of deaths was added to the change in live fish between days to yield the number of new fish entering the river. Mathematically, and Because of observation error (i.e., error estimating live fish from stream surveys) and process error (i.e., variability in true stream-life), the model can produce negative estimates of entering salmon which result in oscillations as the estimates are used in the calculation of future deaths. To accurately calculate descriptive statistics for the distribution of entries, we stabilized such oscillations by fitting a cumulative normal distribution to the cumulative distribution of modeled entries using ordinary least squares; these fitted distributions were used for further analysis of entry timing. The temporal distribution of spawning in the river was then estimated by offsetting entry timing by the average number of days that fish are in the river before completing spawning.
Cedar River sockeye mature while holding in Lake Washington for several months prior to river entry (Newell & Quinn, 2005;Newell, Fresh, & Quinn, 2007) and then typically initiate spawning relatively soon after entering the Cedar River. Tagging studies in both the river and broodstock held at the hatchery suggest this delay averages slightly more than one week (Ames, 2006;WDFW, unpublished data).
The arrival timing of Baker Lake sockeye salmon, the ancestral population of the Lake Washington run, was used for comparison with the Ballard Locks data on arrival of Cedar River salmon to indicate possible shared marine influences on timing. The Baker Lake fish are captured in a trap approximately 88 river km upstream from marine waters and are counted and transported daily above two hydroelectric dams. Similar to the Lake Washington populations, Baker Lake sockeye arrive primarily during June and July, but do not spawn until the fall. Daily counts of transported sockeye were obtained from the WDFW and the Skagit River System Cooperative for the years 1965-2016. Prior to 1992, Baker Lake sockeye salmon were severely depleted, directed fishing was negligible, and so the counts at the collection facility represented essentially the entire run. The population expanded in recent years and was exposed to some fishing, so from 1992 to 2016, daily counts were corrected for commercial and sport catches in terminal areas (i.e., Skagit Bay and the Skagit River). For this period, daily catches were assigned trap arrival dates by combining the fishing location with a estimate No attempt was made to correct for any harvest of Lake Washington or Baker Lake sockeye salmon in distant fisheries; interception by coastal fisheries is minimal because the Puget Sound runs are earlier than the much more abundant runs to the Fraser River and not sufficiently numerous to merit dedicated fisheries (Starr & Hilborn, 1988). Data were not collected on the spawning timing of this population so they were only used for comparison of arrival timing.

| Hatchery operations data
Comparisons between reproductive timing of hatchery-spawned and naturally spawning fish were used to assess artificial selection on timing by hatchery practices. In the hatchery, all females were checked every few days and those with eggs free from the connective tissue were euthanized and the eggs removed. At this stage, they would be expected to spawn had they been in the river, and so the dates of hatchery spawning were compared to the estimated dates of spawning in the river without any adjustment. The average time from capture to spawning in the hatchery is around eight days; given the similar water temperatures experienced by hatchery and naturally spwaning fish during this period, we believe it unlikely that the delay between river entry and spawning varies substantially based on location (i.e., hatchery vs. in-river). Daily counts of eggs taken were obtained from WDFW for all years of hatchery operation TILLOTSON eT aL. . All hatchery origin fish embryos were exposed to controlled thermal shifts during embryonic development that induced a permanent set of marks on their otoliths, unique to each group, that can be examined in adult salmon after death (Volk, Schroder, & Grimm, 1994). These marks not only indicated that the fish was produced in the hatchery, but in many years specific marks were also applied based on the timing when their parents were spawned (early, middle, or late in each season). For fish returning between 2005 and 2012, a sample of otoliths was collected during each hatchery spawning event (N = 9,571 over all years), and for these years, we compared the timing of spawning in the parental and offspring generations to estimate the genetic control over timing. It should be noted that salmon spawning takes place more or less throughout the season, and thus, the emergence of fry is similarly broad. It is impractical to separately mark the embryos from each day of spawning and then keep separately the associated lots of emerging fry. So, the embryos of fish spawned in the early, middle, and late parts of the run were combined and given the same mark. Even within these groups, they did not all emerge on the same day and so were held briefly until the rest of that group emerged and could be transported to the river for release. Thus, the returning adults were assigned a specific date of spawning but it could only be related to their parents' timing group (early, middle, or late).

| Analysis of long-term trends in phenology
We calculated median arrival timing from daily counts at the Baker Lake outlet trap, and at the Ballard Locks, river entry timing from interpolated daily counts of sockeye salmon in the Cedar River (because counts were not made daily), and egg take dates (i.e., spawning of females) in the Cedar River hatchery by calculating daily cumulative counts, dividing these by the annual total at each site, and taking the first date that exceeded 50% of the total as the median. We then employed an information theoretic model selection approach to evaluate the shape, magnitude, and influence of environmental covariates on trends in median timing for Lake Washington entry (i.e., passage of the Ballard Locks), entry into the Cedar River, and arrival at the Baker River trap.
Because migration and reproductive timing in salmonids have ostensibly evolved to maximize fitness given prevailing environmental conditions, long-term change in such conditions may also explain contemporary evolution or plastic changes in these traits (Crozier, Scheuerell, & Zabel, 2011;Quinn & Adams, 1996;Warren et al., 2012). Water temperature and stream flow influence reproductive phenology in many salmonid species, including sockeye salmon throughout their geographic range (Hodgson et al., 2006). Changes in temperature and flow regimes have also been implicated in altered migration timing of other sockeye salmon populations (Crozier

| Heritability of spawning timing in cedar river sockeye
All the progeny of adult sockeye salmon spawned in the hatchery from 2004 through 2011 had their otoliths permanently marked by exposure of the embryos to a series of thermal shifts during development.
These marks were specific to the periods in the fall when the parents had been spawned (early, middle, and late). These juveniles entered Lake Washington in the years 2005 through 2012, and the great majority returned three years later (i.e., 2008 through 2015). When these returning adults were spawned at the hatchery, otoliths were collected after the fish were killed for spawning from a consistent subsample, revealing whether the fish's parents had been spawned early, middle, or late season four years earlier, as well as its age. These data allowed us to calculate, for each timing group, in each year, the average spawning date for the parental and offspring generations.
Because younger Pacific salmon often spawn later in the season than do older salmon (Quinn, 2018), we analyzed these ages separately. We standardized both generations to zero mean and unit variance, and performed a linear regression of offspring mean timing against parental timing for each age at return. The slope coefficients of these regressions can be interpreted as the realized heritability in spawning timing (Hard, Bradshaw, & Holzapfel, 1993).

| Artificial selection on spawning timing
To  (Ames, 2006), studies from many other salmon populations (Perrin & Irvine, 1990) and anecdotal observations in the Cedar River (Ames, 2006) suggest that early arriving salmon live longer than those arriving later.
There is also some evidence that average Cedar River stream-life may be declining in response to spatial shifts in spawning activity or increased rates of prespawning mortality (Ames, 2006).
The sensitivity analysis, therefore, explored the effect of average stream-life as well as the influence of declining stream-life within years (i.e., early arrivals live longer than late arrivals) and between years (i.e., stream-life is declining over time). The Monte Carlo procedure was conducted three times, with high, moderate, and low levels of spawning timing heritability (h 2 ). Table 2 shows the range of values used in the sensitivity analysis, and the point estimates considered to be best estimates for each model parameter.

| Analysis of long-term trends in phenology
The most parsimonious models indicated that the temporal trends in timing varied between the Ballard Locks, Cedar River, and Baker River. The timing of sockeye salmon arrival at the Ballard Locks and Baker River changed little between ~1970 and 2016 (Figure 3d,e).
For each of these time series, the selection procedure did not TA B L E 3 Summary of model comparisons for temporal trends in arrival of sockeye salmon to the Baker Lake system, reflecting entry into freshwater there; the Ballard Locks, reflecting entry into the Lake Washington system; and the Cedar River, where most Lake Washington salmon are produced TA B L E 4 Pairwise comparisons between sockeye salmon run timing metrics at three locations: Ballard Locks (the entry into the Lake Washington basin), the Cedar River (the primary Lake Washington basin spawning area), and Baker Lake (entry into that basin by a geographically proximate and genetically similar population) Year Cedar River flow during September (Figure 3a). Model 4b (Table 3) received over 80% of the AIC weight, while model 4d which also included average autumn Lake Washington surface temperature received the remaining weight. In both of these models and the other two including multiple linear trends, the break year was consistently identified as 1993 (Table 3). Examination of coefficients from model 4b showed that from 1969 through 1993, higher September flow increases were associated with earlier entry timing, but since 1994, this influence has been greatly reduced (Figure 3b). Conversely, after accounting for environmental change (i.e., a decreasing pattern in September flow; Figure 5), no significant temporal trend was apparent prior to 1994, but since then, median timing has become earlier at a rate of 1.26 days per year (Figure 3c; Table 5). Timing was not highly correlated between the Ballard Locks and Cedar River (r = 0.30; Table 4).

| Heritability of spawning timing in cedar river sockeye
Regression analysis of standardized offspring and parent spawning timing resulted in separate estimates of realized heritability of this trait in Cedar River sockeye salmon returning at age-4 and age-5.
Consequently, we considered a broad range of h 2 in our sensitivity analysis.

| Artificial selection on spawning timing
Examination of the yearly dates when the weir was in place in the river to collect sockeye salmon for spawning in the hatchery revealed considerable variation, but in general, the very earliest returning salmon tended to arrive before the weir was in place, after which the weir was operating to trap salmon (but by no means all), and toward the end of the season high river flows precluded further trapping so the late arriving fish were not trapped. From the dates when salmon were spawned in the hatchery, we calculated the median spawning date in each year (i.e., the date by which 50% of the annual total number of eggs had been fertilized in the hatchery).
These dates advanced by ~1.26 days per year over the hatchery program's operation .
The selectivity of hatchery operations on spawning timing varied in strength and direction between years, but in most years favored earlier spawning fish (Figure 4). Using the point estimates in Table 1 for each parameter, hatchery spawning occurred on average 3.9 days earlier than natural spawning each year. Accounting for the egg-to-fry survival advantage of hatchery-spawned adults, the average selection differential was −3.2 days, and the average The Monte Carlo sensitivity analysis allowed us to explore variability in these estimates given uncertainties in model parameters. the shift toward earlier spawning that has occurred since is not well explained by environmental conditions. However, the observed shift toward earlier spawning began approximately coincident with initiation of hatchery supplementation. Analysis of hatchery spawn timing relative to the phenology of natural spawners indicated that artificial selection for earlier spawning has occurred in the past, and, depending on the heritability of the trait, can explain a substantial proportion of the observed advance in spawning timing since the mid-1990s. This was not a case of deliberate selection; rather, the operation of the weir needed to collect the salmon for spawning was not operable late in the fall, owing to high flows. Thus, the latest arriving salmon were unlikely to be spawned in the hatchery, where they would enjoy much higher survival of their embryos than would occur in the river. However, some salmon ascended at the very beginning of the season, prior to the weir's installation, and they too were unspawned. Thus, the hatchery selected, on average, for early spawning but was also to some extent exerting disruptive selection.
These findings collectively provide strong evidence that even in an integrated hatchery population, inadvertent selection on timing can induce population-level changes in phenology.
Although there is strong evidence that artificial selection on reproductive phenology has occurred, it is not clear that this ogy. Furthermore, both theory (Quinn, 2018) and studies of sockeye salmon (e.g., Kovach, Ellison, Pyare, & Tallmon, 2015) and other salmonids (e.g., Warren et al., 2012) indicate that spawning should occur later in response to warming waters; the opposite of the pattern observed since 1994 in the Cedar River. Thus, while we cannot rule the effect of other environmental factors for which we had no data, our findings suggest that inadvertent selection in the hatchery is likely occurring.
The spawning timing phenotype that is ultimately observable (i.e., median spawning date in the population) emerges from a series of complex and interrelated evolutionary, behavioral, and physiological processes. In our study population, spawning timing appears largely independent of ocean processes because, although marine environmental factors do influence the timing of arrival in freshwater, a long delay before spawning combined with a lack of relationship between lake and river entry timing appears to decouple migratory and reproductive processes (Newell et al., 2007). Thus, Cedar River sockeye salmon spawning timing appears to be determined by an underlying, genetic predisposition combined with plastic response to flow, and possibly temperature conditions in the lake and river. Spawning timing can also vary with fish size, age, and spawning location, so longterm changes in these traits may influence spawning timing (Carlson, Rich, & Quinn, 2004;McPhee & Quinn, 1998). Finally, because sockeye salmon cease feeding prior to freshwater entry, a fixed energy budget is available once the fish are in Lake Washington. It has been suggested that early upstream migration in sockeye salmon may occur if energy reserves are atypically low (Lapointe et al., 2003) or high (Katinić, Patterson, & Ydenberg, 2017 lines of evidence suggest that artificial selection for early spawning may be maladaptive in this population. First, juvenile sockeye salmon that enter Lake Washington from the Cedar River later in the spring on average experience a survival advantage over the earlier migrants of their cohort (Hovel et al., 2019). This is presumably because later migrants are more likely to experience a "match" with favorable growth conditions during the vulnerable period shortly after lake entry (Cushing, 1990). Because sockeye salmon spawn as stream temperatures are decreasing, the effect of earlier spawning is amplified in emerging juveniles. Embryonic development is strongly influenced by temperature, and earlier spawning exposes embryos to warmer water (Murray & McPhail, 1988). Given the average temperature regime of the Cedar River, a one-week shift in median spawning from mid-to early October would advance median emergence timing by over two weeks. The impact of artificial selection on spawning timing is therefore amplified in juvenile phenology. Combined with the relative survival advantage for later lake entry by juveniles, disruption of offspring trophic dynamics is a plausible fitness cost of artificial selection for earlier spawning in this population. Spring lake processes-and therefore optimal juvenile entry timing-are also advancing in response to climate change (Winder & Schindler, 2004), but the change in spawning timing since the 1990s has far exceeded the environmental change.
Artificial selection for earlier spawning may also reduce fitness through increased exposure of adult sockeye salmon to temperatures warm enough to elevate prespawning mortality (PSM) rates.
Cedar River sockeye salmon enter Lake Washington prior to peak summer temperatures, hold in the lake's hypolimnion through the summer, and then enter the river and spawn in the cooler fall months (Newell and Quinn 2005). Thermal conditions in the Ship Canal and in Lake Washington have been warming and, in recent years, unexplained mortality has been documented in sockeye salmon spawning in the river (22%-34% annually) and in the hatchery (31%-41% annually: Barnett, Peterson, Yates, & Drobny, 2018). PSM such as this occurs in many salmon populations (Bowerman, Keefer, & Caudill, 2016) but has been rare in the Cedar River (Ames, 2006). The causes of PSM vary, but elevated temperatures can be a contributing factor (Bowerman et al., 2018). Earlier migration into the Cedar River, combined with a general pattern of warming in Lake Washington, may be driving an increase in the rate of PSM. Within-season patterns of PSM support this idea; the earliest fish have experienced the highest rates of failed spawning. Carcass surveys showed that PSM rates decreased from an average 49% in September, 26% in October, 24% in November, to 13% in December for years 2014-2016 (Barnett et al., 2018). If earlier spawning is associated with lower fitness from either reduced survival of progeny or inability of adults to breed successfully, then artificial selection on this trait will cause a larger proportion of the population to spawn during less favorable times.
While the abundance of Cedar River sockeye salmon has declined markedly during the past decade, it is impossible given the available data to attribute any specific portion of declining abundance to artificial selection on phenology. Nevertheless, our results suggest that altered phenology is a plausible contributing factor of population declines.
Artificial propagation may be warranted if there are demographic risks to a population's viability, and in the case of Pacific salmon, has sustained fisheries in cases where habitat loss reduced capacity for natural production (Flagg, 2015;Naish et al., 2007). However, our findings add to a growing body of evidence documenting the challenges in sustainable supplementation of fish populations (Amoroso, Tillotson, & Hilborn, 2017;Araki 2007Araki , 2008Baskett & Waples, 2013). Hatcheries can reduce genetic diversity (Berejikian & Van Doornik, 2018;Waters et al., 2015), impose artificial selection (McLean et al., 2005;Quinn et al., 2002), and may disrupt processes of natural selection and adaptive evolution (Waples, Beechie, & Pess, 2009). Efforts are underway to improve the genetic management of hatcheries in some regions (Flagg, 2015;McGarvey & Johnston, 2011;Mobrand et al., 2005) but impacts on supplemented populations may occur if wild and hatchery fish cannot be fully segregated (e.g., Seamons, Hauser, Naish, & Quinn, 2012). Our findings further highlight that phenology can be particularly sensitive to inadvertent selection in hatcheries (see also Quinn et al., 2002;McLean et al., 2005;Ford et al., 2006). As with selection on timing in fisheries (Tillotson & Quinn, 2018), selection on timing in hatcheries can result from a number of processes including the natural tendency to spawn the first arriving fish, leaving those at the end to be released into the river if the capacity of the hatchery is filled. In addition, and more germane in the present case, river conditions favor the trapping of early arriving salmon more than those coming later, and thus, the progeny of the early fish is given the benefit of the higher survival rate during incubation in the hatchery. Because the timing of migration and reproduction is a key component of salmon diversity, the management of fisheries and hatcheries should strive to maintain the natural variation in these traits (Tillotson & Quinn, 2018).
Furthermore, consideration should be given to how warming waters and changing flow regimes (e.g., Arismendi, Safeeq, Johnson, Dunham, & Haggerty, 2013) might interact with hatchery operations to ultimately shape patterns of selection on supplemented populations.

DATA A R C H I V I N G S TAT E M E N T
Data available from the Dryad Digital Repository: https ://doi. org/10.5061/dryad.3fc8m7c.

ACK N OWLED G EM ENTS
We thank the Cedar River Hatchery Technical Working Group, especially Kurt Fresh and Andy Applebee for their support of this research, and Seattle Public Utilities and the IGERT Program on Ocean Change at the University of Washington for funding. The data used in this study were obtained from many sources, and we thank all those who collect, maintain, and share these important biological and environmental data, especially Daniel Schindler for Lake Washington temperature records and Pete Kairis for sharing harvest-corrected timing data from the Baker River which are collected by Washington Department of Fish and Wildlife and the Skagit River System Cooperative, and the Washington Department of Fish and Wildlife staff working at the Cedar River Hatchery and counting salmon in the river, and the tribal biologists counting salmon as well. We also thank Jeff Hard for his input on an early version of this manuscript and three anonymous reviewers for their additional suggestions.

CO N FLI C T O F I NTE R E S T
None declared.