Long‐term population studies can identify changes in population dynamics over time. However, to realize meaningful conclusions, these studies rely on accurate measurements of individual traits and population characteristics. Here, we evaluate the accuracy of the observational methods used to measure reproductive traits in individually marked black‐tailed godwits (Limosa limosa limosa). By comparing estimates from traditional methods with data obtained from light‐level geolocators, we provide an accurate estimate of the likelihood of renesting in godwits and the repeatability of the lay dates of first clutches. From 2012 – 2018, we used periods of shading recorded on the light‐level geolocators carried by 68 individual godwits to document their nesting behaviour. We then compared these estimates to those simultaneously obtained by our long‐term observational study. We found that among recaptured geolocator‐carrying godwits, all birds renested after a failed first clutch, regardless of the date of nest loss or the number of days already spent incubating. We also found that 43% of these godwits laid a second replacement clutch after a failed first replacement, and that 21% of these godwits renested after a hatched first clutch. However, the observational study correctly identified only 3% of the replacement clutches produced by geolocator‐carrying individuals and designated as first clutches a number of nests that were actually replacement clutches. Additionally, on the basis of the observational study, the repeatability of lay date was 0.24 (95% CI 0.17 – 0.31), whereas it was 0.54 (95% CI 0.28 – 0.75) using geolocator‐carrying individuals. We use examples from our own and other godwit studies to illustrate how the biases in our observational study discovered here may have affected the outcome of demographic estimates, individual‐level comparisons, and the design, implementation, and evaluation of conservation practices. These examples emphasize the importance of improving and validating field methodologies and show how the addition of new tools can be transformational.

Materials and Methods

Fieldwork
Fieldwork occurred from March through June 2012 – 2018, in our 12,000 ha long-term study area in southwest Fryslân, The Netherlands (52.9643°N, 5.5042°E; Senner et al. 2015b). Starting on 15 March, we checked every field within the study area at least once every week for six weeks. During this period, godwits arrive from the non-breeding areas, form pairs and establish territories. We consequently had a good sense of where in the study area godwits were present, and used that knowledge to find nests when the godwits started laying in April. We used the egg flotation method to estimate the lay date of each nest and, consequently, their expected hatch dates (Liebezeit et al. 2007). We visited each nest three days before the estimated hatch date and, if it was still active, returned 1 – 3 days later to band the chicks. We also caught a portion of incubating godwits using walk-in-traps, automated drop cages, or mist nets placed over the nest. After capturing an adult, we individually marked it with colour rings and took a blood sample for molecular sexing. In the years after capture, we linked marked individuals to specific nests through observations of incubating birds or by recapturing them coincidentally.
               Each breeding season we outfitted 42 – 69 adult godwits with geolocators (i.e., 26 – 61% of the adults caught annually). We used geolocators from Migrate Technology, Ltd: the 0.65g Intigeo W65A9 model from 2012 – 2013 and the 1g Intigeo C65 model thereafter. These geolocators were attached to a coloured flag and placed on the tibia. The total weight of the attachment was ~3.3g from 2012 – 2013 and ~3.7g from 2014 – 2017, representing 1 – 1.5% of an individual’s body mass at capture. The return rate of geolocator-carrying individuals to the breeding grounds in the year following deployment was 0.90, which is similar to their apparent annual survival rate (0.85, Kentie et al. 2016).
               From 2013 onward, these geolocators were programmed to log the ambient light level for up to 26 months (i.e. up to two consecutive breeding seasons). In the years following deployment, we put considerable effort into recapturing godwits carrying geolocators. We retrieved light-level data from 129 geolocators. Of these, 22 logged for 23 months or more, while most logged only 11 – 22 months either because the battery ran out or because we recaptured the bird within 22 months. We also retrieved 32 geolocators that logged for less than 11 months and which thus failed to log the start of the next breeding season. We retrieved geolocators from both live and dead birds; after retrieving a geolocator from a live bird, we re-deployed a new geolocator on the same bird in all but 6 cases (5%).
              
Inferring incubation duration and hatching success from geolocator data
The geolocators were programmed to log ambient light level every five minutes and, because they were mounted on the leg, recorded those periods of time when the geolocator was shaded during incubation (see also Bulla et al. 2016). To inspect the daily light patterns (Figure 1), we used the function “preprocessLight” from package “BAStag” (Wotherspoon et al. 2016) in Program R (R Core Team 2018). We manually identified the beginning and end of an individual’s incubation period, as well as the number of times each individual nested within a breeding season (Figure 1). In 111 of 151 cases, we observed an egg-laying phase denoted by 20 or more minutes of shading for 1 – 3 days, immediately followed by an incubation phase denoted by long shaded periods lasting 1 – 10 hours. This pattern is consistent with known godwit nesting behaviour, as most godwits lay 3 – 4 eggs (Haverschmidt 1963, Verhoeven et al. 2019), both females and males spend short periods sitting on the nest during the egg-laying phase, and incubation begins after the penultimate or ultimate egg is laid (Haverschmidt 1963). In the remaining 40 cases, we did not observe an egg-laying phase but did observe a clear incubation phase. Observing egg-laying phases shorter than two days or no egg-laying phase at all could be the result of females laying fewer than four eggs, birds starting to incubate earlier then the penultimate egg, males that did not sit on the nest during the laying phase, or because we were unable to accurately identify a complete egg-laying phase. Because of these uncertainties, the estimated lay date in these cases might be 1 – 3 days later than the actual lay date. This, in turn, might have caused us to overestimate an individual’s renesting interval or to underestimate the repeatability of an individual’s lay date across years. However, we do not believe these possible sources of error affected our conclusions, because (1) we use the average renesting interval across years and (2) despite being a potential underestimate, the geolocator-based estimate of repeatability was already substantially higher than the observational-based estimate.         

               Although our individually-specific, manual approach to analysing the geolocator data could have introduced some biases in determining the timing of laying and duration of incubation, we believe that our method was the most accurate one possible. For example, the amount of time that geolocators were shaded during egg-laying and incubation varied considerably among individuals: some individuals incubated mostly at night with only 1 – 2 hours in the morning or evening, whereas others incubated mostly during the day, either in one long bout or multiple bouts of varying lengths. This considerable inter-individual variation meant that we were unable to quantitatively determine the onset of incubation, such as by using a threshold value for the number of daylight hours during which a geolocator was shaded.  

               For 43 of the nests of geolocator-carrying godwits, we know that chicks hatched successfully because we observed the newly hatched chicks in the nest; the geolocator data we retrieved for these nests showed that incubation lasted from 23 – 30 days. This corresponds with the known incubation duration of godwits (24.5 days, range 22 – 27 days; Haverschmidt 1963). Because not all nesting attempts were identified by our observational study (see Results), we lacked observational data on nest fate for some of the nests analysed in this study; we considered such nests failed if the geolocator data indicated they were incubated for 22 days or less. In most cases, it was also possible to infer chick brooding from the light-level data (see Figure 1). However, this was not failsafe, and we therefore did not use it as a measure of hatching success.
               In our data we distinguish between: (1) first clutches, (2) renesting after the failure or hatching of a first clutch (“first replacement”), and (3) renesting after the failure of a first replacement (“second replacement”). Replacement clutches do not include clutches laid by a godwit pair after it has successfully fledged chicks (also called “double-brooding”); this is a behaviour we and others have never observed among godwits (see Senner et al. 2015a). For all clutches we know the start of incubation; for successful clutches we know the date of hatching; for unsuccessful clutches we know the date of failure. We also had some incomplete incubation histories resulting from geolocators that stopped logging partway through the breeding season; this was the result of either (1) battery failure during the breeding season or (2) recapture of an individual during one breeding season (with one geolocator), but not in a subsequent breeding season (with a second geolocator). For this study, we collected a total of 103 incubation histories, both complete and incomplete, from 68 individuals: 39 females and 29 males. This included two males that likely each skipped a breeding season altogether, so our analyses include 101 complete and incomplete incubation histories from which we know the fate of the first clutch in a breeding season (Figure 2).
               Of these 101 first clutches with known fates, there were two cases in which it was not clear whether the bird renested or not, even though the geolocator remained operational. One female likely laid a first replacement clutch, and another female who lost her first replacement clutch likely laid a second replacement, but we cannot be certain (see Supplementary Material). We have therefore excluded these two cases from the analyses that estimated renesting propensity and probability; for these analyses we also excluded one case in which the parent was killed at the same time the first clutch was depredated (Figure 2).
               Renesting propensity and probability depend on whether the female produces a replacement clutch or not. However, since godwits are socially monogamous and share parental care (Cramp & Simmons 1983, Beintema et al. 1995), we can also infer renesting propensity and probability on the basis of males — except in those cases in which the female dies. In such cases, male geolocator data would show only that the female did not renest, not whether she was alive or not. In the cases where we retrieved geolocators from live birds, female geolocator data does not include this uncertainty. The calculated renesting propensity and probability would therefore be underestimated if the geolocator-based sample includes males whose partners died after laying their first clutch. Our results show that this scenario did not happen after failed first clutches, but it may have occurred after hatched first clutches or second replacement clutches.

Analysis
Observer bias in renesting propensity.
First, we calculated renesting propensity on the basis of geolocator-carrying godwits — how many individuals laid a replacement clutch after their first clutch failed, how many laid a replacement clutch after their first nest hatched, and how many renested again after their first replacement failed. The individuals carrying geolocators were part of our long-term observational study, which enabled us to compare the found renesting propensities between the two different study methods: geolocator-based and observational.

Observer bias in linking an adult to a nest
Our study set-up also enabled us to evaluate our observational study’s performance in linking marked adults to nests. However, of the 101 first clutches that were laid by geolocator-carrying godwits and had known fates, eight were linked to individuals that were caught for the first time while incubating that nest. Because these individuals were unmarked prior to being caught, it was not possible to evaluate the performance of our observational study for these cases. Therefore, we could only use 93 of the 101 first clutches in our evaluation.
               We used a generalised linear model with a binomial error distributand a logistic link function to test whether the chance of linking a geolocator-carrying individual to a nest on the basis of field observations (categorized as linked or not linked) depended on whether or not the nest hatched (included as a two-level factor) or when in the season the nest was laid (included as a continuous covariate). However, there are two potential caveats to these comparisons between study methods: (1) Within our observational study, we very rarely obtained data suggesting godwits were renesting. During the proofing process of our observational study, we therefore frequently disregarded the possibility of a bird renesting. Especially in cases where an adult was linked to two nests that were close to each other in time and space, the less likely nest was sometimes permanently “unlinked” from the adult in the database. At the time, we thought these cases resulted from mistakes made in the field, with single adults erroneously linked to two simultaneous nests. In light of our results here, however, it is likely that some of these adults were correctly linked to a replacement clutch laid soon after the previous failure. This means that the performance of our observational methods was actually slightly better than is shown by our comparison here. (2) Retrieving geolocators is of great value to our project and we therefore sometimes focused on geolocator-carrying individuals more than other marked individuals. The calculated performance of our observational study on the basis of geolocator-carrying individuals may thus be slightly higher than for all marked individuals.

Observer bias in the timing of laying
Some nests of geolocator-carrying individuals found in the field during our observational study and designated as first clutches were actually second or third clutches (see Results). Incorrectly assigning first and second replacement clutches as first clutches in some, but not all cases has consequences for how consistent our observational study estimates individuals to be in their timing of laying. Therefore, the individual repeatability of the lay date of first clutches estimated by Lourenço et al. (2011) on the basis of our observational study is likely an underestimate. To get a better estimate, we calculated the repeatability of lay date on the basis of the first clutches of geolocator-carrying birds. For this, we included individual as a random effect in the linear mixed model method of the function “rpt” in the R package “rptR” (Stoffel et al. 2017). The estimate made by Lourenço et al. (2011) was based on data collected in different years and with a different statistical method from our present geolocator-based study; we therefore estimated the repeatability of lay date based on our observational data collected during the same years as our geolocator data (2012 – 2018) using the same statistical method described above for our geolocator-based estimate. For this analysis we used only female lay dates because including both sexes would introduce considerable pseudo-replication from pairs comprising two marked individuals. We excluded from this analysis all nests known to be a replacement clutch on the basis of the observational study. We assessed the uncertainty of these repeatabilities with 1000 parametric bootstraps and their statistical significance with likelihood ratio tests.
 

Renesting probability
We also examined the chance of producing a replacement clutch, i.e. the renesting probability, as a function of the date of nest loss. This analysis yielded a “complete separation,” in which the explanatory variable (date) yielded a perfect prediction of the dependent variable (renesting probability). Further statistical estimates were therefore not required to assess or account for between-year and within-individual variation. Finally, we examined whether the renesting probability after the first clutch hatched depended on the date of hatch. For this we used a generalised linear mixed model from the R package “lme4” (Bates et al. 2015), with a binomial error distribution, logistic link function, and individual and year as random effects. Finally, we calculated the number of days between renests and plotted this interval against the date on which the earlier clutch was lost to investigate whether the renesting interval changed seasonally (Supplementary Material Fig A1). We also used linear mixed models to test whether this renesting interval depended on either the number of days the previous nest had been incubated or the date of nest loss. We included individual as a random effect in these models.

Comparison with van Balen
In 1954, van Balen (1959) conducted experimental research on renesting in godwits in a 100-ha area 69 km due south of our study area (52.2366°N, 5.4184°E). After van Balen marked individual incubating godwits, he collected their eggs and studied their renesting behaviour. Following the removal of eggs, he searched the area for these marked individuals and collected their subsequent nesting attempts. He thus obtained data on the renesting propensity of godwits, the interval between replacement clutches, the distance between nests, and the initiation dates of replacement clutches. We compared his findings with our own using general linear models with a Gaussian error distribution. We obtained F values and Chi-squared values for the significance of the fixed effect “study” (a two-level factor with groups “ours” and “van Balen”) of nested models with and without this fixed effect. We visually inspected the residuals to validate the model assumptions.
               From the light-level data, we obtained data on renesting propensity, the interval between replacement clutches and the initiation dates of replacement clutches. We also investigated the geographic distance between an individual’s first clutch and replacement clutches by taking the coordinates of both nests and calculating the distance between them with the function “pointDistance” from the R Package “raster” (Hijmans 2017). We used all the replacement clutches that were identified by linking a colour-marked individual to a nest as part of our long-term observational study; these include the replacement clutches of geolocator-carrying birds that were noted during the field season, but not the replacement clutches of geolocator-carrying birds that were missed by the observational study (see Results). For this analysis, we log-transformed renesting distance to achieve normality.

B3RLLL is not part of the repeatability analysis for first lay dates since it was also part of a GPS-tracking study; B3BYYY is included in this analysis because we know the laydate, but is excluded from further analyses because we don't know the fate of its first clutch. See uploaded files.