Increasing phenological asynchrony between spring green-up and arrival of migratory birds

Consistent with a warming climate, birds are shifting the timing of their migrations, but it remains unclear to what extent these shifts have kept pace with the changing environment. Because bird migration is primarily cued by annually consistent physiological responses to photoperiod, but conditions at their breeding grounds depend on annually variable climate, bird arrival and climate-driven spring events would diverge. We combined satellite and citizen science data to estimate rates of change in phenological interval between spring green-up and migratory arrival for 48 breeding passerine species across North America. Both arrival and green-up changed over time, usually in the same direction (earlier or later). Although birds adjusted their arrival dates, 9 of 48 species did not keep pace with rapidly changing green-up and across all species the interval between arrival and green-up increased by over half a day per year. As green-up became earlier in the east, arrival of eastern breeding species increasingly lagged behind green-up, whereas in the west—where green-up typically became later—birds arrived increasingly earlier relative to green-up. Our results highlight that phenologies of species and trophic levels can shift at different rates, potentially leading to phenological mismatches with negative fitness consequences.

. Significance 'b' is given when the raw p value is less than the Bonferroni critical value a / m , where α = 0.05 and m, the number of tests = 48. The Benjamin-Hochberg (BH) test controls the false discovery rate . Significance 'c' and 'd' are given for false discovery rates (FDR) of 0.05 and 0.25 respectively when the raw p value is less than its BH critical value (i/m)Q where i is the rank of the raw p, and Q is the FDR selected.

Supplementary Note S1. Analysis of effects of passing migrants on arrival date estimates
Observations of migratory birds may consist of two groups of individuals. Individuals who arrive and remain for the breeding season will be termed "breeders", while individuals who arrive in a cell but quickly pass through and depart to another cell will be termed "passers". Our interest is primarily in breeders as these individuals may be more reliably associated with broad environmental conditions and their synchrony with the environment may be more important for demography. However, we were unable to differentiate among breeders and passers. Here we offer several evaluations of the impact of this limitation on our results, an important step in leveraging the exceptional richness of eBird or similar data to facilitate our understanding of phenological mismatch.
There are several possible reasons why passers may be less of a problem than one may at first imagine. First, phenological interval between arrival and environment may occur and be relevant for both passers and breeders, even if the case for demographic impacts of high phenological intervals is greater for breeders. Second, passers will only bias studies of phenological interval if they arrive at substantially different times than individuals remaining to breed. Third, phenological interval trends were more strongly driven by green-up than by arrival (Fig. 1).
Although we could not exclude passers from the analysis, we investigated the potential sensitivity of our results to passers in three ways. First, passers are expected to decrease with latitude within a species' range, so we re-estimated phenological interval trends while controlling for latitude (see note 1a). We found that more species showed significant trends in phenological interval when controlling for latitude, indicating both the sensitivity to latitude and that our estimates of the number of species with significant phenological interval trends may have been conservative. Second, if passers influence the proportion of presences observed over time (for example, by arriving later than breeders), then the estimation of arrival dates may be sensitive to the range in dates (window size) over which logistic curves were estimated. This effect may vary with latitude if passers are influential. We investigated this sensitivity and found that arrival was sensitive to window size but that this effect did not vary with latitude (see note 1b). Third, if passers arrive later than breeders, our estimates of arrival dates may be biased late. We re-estimated the arrival dates at several points on the logistic curves to simulate earlier arrival and found that more species displayed significantly increasing phenological interval trends and that the mean phenological interval trend was stronger, indicating that our estimates may have been conservative (see note 1c).

1a. Controlling for latitudinal effect on phenological interval.
We expected that the proportion of passers of any given species would decrease with latitude. Individuals on the northernmost edge of a breeding range would, for example, likely consist only of breeders. Although exceptions may exist to this pattern, it may serve as a general assumption to test the impacts of passers on phenological interval trends. If this assumption is valid, and if passers decrease our ability to detect phenological interval trends by biasing arrival date estimation, we expected that controlling for latitude would reduce the number of species for which significant trends in phenological interval would be observed.
To test if latitude was a strong predictor of phenological interval, we modeled phenological interval as follows, implemented with package lme4 and lmerTest in R: where * denotes random intercept, and Year was fit as a categorical factor.
We found latitude was a strong predictor of phenological interval, as shown in the following Several factors may account for the impact of latitude on phenological interval, including: fewer passage migrants at higher latitudes (within a species), or increased fitness pressure on phenological synchrony with green-up in more seasonal environments typically at higher latitudes.
We re-modelled trends in phenological interval for each species using the method described in the main text, except that we controlled for latitude with the following model specification: where * denotes random intercept.
We found that when controlling for latitude, phenological interval increased significantly in 35 of 48 species (73%). This contrasts our finding that when not controlling for latitude (see main text), phenological interval increased significantly in 9 of 48 species (19%).
We conclude that if latitude is a good proxy for the proportion of passers, passers are unlikely to have resulted in "false positives", that is, falsely identifying increasing phenological interval in species. Our results suggest phenological interval trends may be more widespread across species than our analyses ignoring latitude revealed.

1b. Sensitivity analysis of logistic curve window size on arrival dates.
It is important that the logistic regression models employed appropriately estimate mean arrival dates of bird populations. However, logistic curve fits can be impacted by the behavior of occurrences following the arrival of a population. For example, if following their arrival to a cell, a large proportion of migrants continue north to other cells, these passers may generate a downturn in the proportion of occurrences that are presences.
We tested the impact of temporal window size on arrival date by refitting logistic curves for all cell-species-years at for windows of Julian days 80-160, 80-180, and 80-200. We then fit a linear mixed model using R's 'lme4' package: Arrival date = window size + year + species + cell*, where * denotes random intercept.
We determined that estimated arrival date is significantly impacted by window size (p < 0.001).
However, what is important is not whether this effect exists, but whether this effect of window size on arrival date varies systematically with geography, because that could bias our geographically based results. We therefore tested for the impact of latitude, longitude, and elevation on the effect of window size on arrival date. To do so, we first determined the effect of the window size on arrival date for each individual cell. We then fit a linear regression (window on arrival date effect = centroid latitude + centroid longitude + mean elevation) with cell as the sample unit.
We determined that the effect of window size on arrival date was independent of latitude, longitude, and elevation, as shown in the These results mirror those of Hurlbert & Liang (2012, PLoS One 7:e31662), who in a similar analysis altered the window sizes relative to original arrival date estimates and found that although window size marginally impacted arrival date, this impact did not vary significantly with latitude or migration distance.

1c. Sensitivity analysis of arrival date estimation on phenological interval trends.
The estimated arrival dates used in the main study indicate the inflection point of the logistic curve fit to the proportion of presences over time. They represent the mean arrival date of the population to the given cell by a given species in a given year, and correspond to the date at which 0.5 of the asymptotic (maximum) proportion of presences was reached.
However, if passers both a) make up a substantial proportion of observed individuals in a species, and b) systematically arrive earlier or later than breeders, then our estimates of arrival may be biased. For example, if passers tend to arrive later than breeders, our estimated arrival dates may be biased late. Passers may be more likely to arrive later than breeders because earlier arriving migrants often breed at lower latitudes (Conklin et al., 2010) and early arrival to a breeding location can be beneficial to fitness (Smith & Moore, 2005).
To determine the sensitivity of phenological interval trends to arrival date estimation, we reestimated arrival dates. Using the same fitted logistic regression curves, we determined the dates on the curves when x proportion of the asymptotic proportion of presences was reached.
Where the original model fit in R was: and where PropP is the proportion of presences, Asym is the asymptote, xmid is the date at which 0.5 of the asymptote is reached, julian is the date of year, and scal is the scale parameter.
we set PropP = x(Asym), where x is the proportion of the asymptotic proportion of presences, here x was set as 0.25, 0.333, and 0.666. used the parameters fit by the model, and solved for julian as follows: julianx = xmid -( scal * (log (( Asym / PropP ) -1 ))).
We then re-calculated phenological interval with these new arrival dates, and re-estimated trends in phenological interval over time, all using the same method as described in the main text.
We found that the number of species for which significantly increasing phenological interval (α = 0.05) was observed decreased with the proportion of the asymptotic proportion of presences (see table below).
Arrival dates x proportion of asymptote We also found that the strength of the phenological interval trend increased with the proportion of the asymptotic proportion of presences (see These observations suggest that if arrival dates were influenced by passers, then the number of species with significantly increasing phenological interval, and the overall strength in phenological interval trend, were likely sensitive to the proportion of passers in the data and the discrepancy in arrival dates between passers and breeders. The results also suggest that if passers arrived later than breeders such that earlier arrival dates are more representative of breeders, then we may have underestimated the number of species with increasing phenological interval trends and the strength of phenological interval trends. If instead passers arrived earlier than breeders such that later arrival dates are more representative of breeders, we may have overestimated the number of species with increasing phenological interval trends and the strength of phenological interval trends. Since if passers and breeders arrive at different times, passers are likely to arrive later than breeders, we conclude that our estimates of the number of species and of the strength of phenological interval trends in the main study (where x = 0.5) are likely conservative.