Oecologia

DOI 10.1007/s00442-008-0959-4

GLOBAL CHANGE ECOLOGY - ORIGINAL PAPER

Bias in phenology assessments based on Wrst appearance data of butterXies

Arco J. van Strien · Willem F. Plantenga · Leo L. Soldaat · Chris A. M. van Swaay · Michiel F. WallisDeVries

Received: 26 May 2007 / Accepted: 19 December 2007© Springer-Verlag 2008

Abstract Data on the Wrst appearance of species in theWeld season are widely used in phenological studies. How-ever, there are probabilistic arguments for bias in estimatesof phenological change if sampling methods or populationabundances change. We examined the importance of bias inthree measures of phenological change: (1) the date of theWrst X appearances, (2) the date of the Wrst Y% of all Wrstappearances and (3) the date of the Wrst Z% of the individu-als observed during the entire Xight period. These measureswere tested by resampling the data of the Dutch ButterXyMonitoring Scheme and by simulations using artiWcial data.We compared datasets diVering in the number of samplingsites, population abundance and the start of the observationperiod. The date of the Wrst X appearances proved to be sen-sitive to the number of sampling sites. Both the date of theWrst X appearances and the date of the Wrst Y% of all Wrstappearances were sensitive to population trend. No suchbiases were found for estimates of the Wrst Z% of the Xightperiod, but all three measures were sensitive to changes inthe start of the observation period. The conclusions weresimilar for both the study on butterXy data and the simula-tion study. Bias in phenology assessments based on Wrstappearance data may be considerable and should no longerbe ignored in phenological research.

Keywords Climate change · First observation · Flight period

Introduction

Recent studies have reported changes in phenology of spe-cies caused by an increase in average temperature in the lastthree decades, ranging from several days to a few weeks(Inouye et al. 2000; Parmesan and Yohe 2003; Stefanescuet al. 2003; Primack et al. 2004; Both et al. 2005; Menzelet al. 2006). Much of the evidence for these phenologicalchanges is based on Wrst observations, such as the earliestreturn date of migratory birds or the Wrst appearance of but-terXies in the season. Such data are easily collected in theWeld, and an increasing number of volunteer observersreport Wrst sightings in the framework of phenological net-works (Bruns and van Vliet 2003).

Data on Wrst observations can be applied in two diVerentmeasures, both widely used. The Wrst type of measure isbased on a speciWc number of Wrst observations, often thedate of the very Wrst observation, but also, for example, thedate of the 25th observation. Examples are the Wrst layingdate of lapwing (Both et al. 2005), the Wrst arrival date ofspring migrants (Sparks 1999), the Wrst observation of win-tering birds (Sparks and Mason 2004) and the Wrst sightingsof butterXies (Forister and Shapiro 2003). The second phe-nological measure is based on all Wrst observations, forexample the mean date or the date of the Wrst 25% of allWrst observations. Examples are the mean date of all Wrstappearances of butterXy species (Roy and Sparks 2000;Stefanescu et al. 2003) and the median Wrst arrival date ofspring migrants (Barrett 2002).

However, when calculating phenological changes, biasmay creep in if data consist of Wrst observations. This can

Communicated by Diethart Matthies.

A. J. van Strien (&) · W. F. Plantenga · L. L. SoldaatStatistics Netherlands, P.O. Box 4000, 2270 JM Voorburg, The Netherlandse-mail: asin@cbs.nl; lnmasint@wxs.nl

C. A. M. van Swaay · M. F. WallisDeVriesDe Vlinderstichting/Dutch ButterXy Conservation, P.O. Box 506, 6700 AM Wageningen, The Netherlands

123

Oecologia

be caused by two main methodological reasons. First, if thenumber of sites at which the Wrst observations of a speciesis recorded increases during the study period, the probabil-ity of early observations also increases in time. A hypothet-ical example is shown in Fig. 1. The Wrst appearance at theWrst site is expected 2 days earlier in a study that includes100 sites than in a study with 25 sites. The Wrst appearance

at the 25th site is even expected 13 days earlier in the 100-sites study (see 25th in Fig. 1). If a Wxed percentage of allWrst observations is taken, no such bias is expected (see25% in Fig. 1). Thus, the bias will depend on the phenolog-ical measure used. Secondly, earlier appearance may beobserved as a consequence of an increase in populationabundance which may aVect both phenological measures.An additional potential source of bias is the possibility ofearlier starts of data collection in warmer springs: if Weld-work started earlier because of the recent warmer springs,earlier Wrst appearances are to be expected.

Although the sensitivity of Wrst appearance data to biasis frequently acknowledged in the literature (e.g. Roy andSparks 2000; Forister and Shapiro 2003; Tryanowksi et al.2005; Tøttrup et al. 2006), few attempts have been made toassess the signiWcance of bias, and procedures to adjust Wrstappearance data for bias have hardly been addressed.

Here, we examine bias in measures of Wrst appearancesof common butterXy species by resampling the data of theDutch ButterXy Monitoring Scheme (DBMS), as well as bysimulations with artiWcial data. The aims are: (1) to test thesigniWcance of bias, and (2) to contribute to the develop-ment of procedures to reduce it. We compare datasetswhich diVer in the number of sampling sites (transects),population abundance and start of the observation period.Two diVerent measures of Wrst appearances are applied andcompared with a third measure of phenological change,which we expect to be insensitive to both sampling eVortand population abundance. This latter measure was basedon the entire appearance period and is also frequently usedin the literature. Examples are mean Xight date of butterXies(Roy and Sparks 2000; Stefanescu et al. 2003), peak Xow-ering date of woody plants (Primack et al. 2004), medianlaying date of Xycatchers (Both et al. 2005) and springtotals of mistnetted migrant birds (Tøttrup et al. 2006). Inaddition, we examine whether bias in Wrst appearance esti-mates leads to biased relations between Wrst appearance andspring temperature.

Materials and methods

Field data

The Weld method of the DBMS is based on the Britishscheme (Pollard and Yates 1993). The counts are conductedalong Wxed transects of 1 km. The Weldworkers record allbutterXies 2.5 m to their right, 2.5 m to their left, 5 m aheadof them and 5 m above them (van Swaay et al. 2002).Weekly butterXy counts are conducted from 1 April to 30September, resulting in a maximum of 26 visits a year,although the average number is about 15. Gaps in the dataoccur mainly for periods with bad weather, as visits are

Fig. 1 Expected frequency distribution of Wrst appearance dates forsamples of n = 25 sites and n = 100 sites (example data), based on thesame underlying probability distribution caused by natural variation.For n = 25 sites, the expected frequencies of sites with a Wrst observa-tion on days 1–5 are 0.03, 0.08, 0.17, 0.33 and 0.59, respectively. Thevery Wrst observation is then expected before day 6 and the 25th sight-ing (FA-25) before day 22. For n = 100 sites, the expected frequencyfor day 1 is 4 times higher, i.e. 0.13, the very Wrst observation is expect-ed before day 4 and FA-25 before day 9. The 25% arrows indicate thedate before which the Wrst 25% of appearances are observed (FA-25%)

123

Oecologia

only conducted when weather conditions meet speciWedcriteria (van Swaay et al. 2002).

Phenological measures

The two measures we used of Wrst appearance are: (1) thedate of the 25th appearance (FA-25), and (2) the date onwhich the Wrst 25% of all Wrst appearances had beenobserved (FA-25%). We compared these two measureswith: (3) the date of the Wrst 25% of all observed individu-als in the entire Xight period on all transects together (FP-25%). The latter contains all observations in the entire sea-son rather than only the Wrst appearance per transect.Yearly Xight period numbers per species were calculated intwo steps: the number per transect was obtained by sum-ming all individuals per species observed per week, inter-polating for the weeks without visits; and the numbers forall transects were added. Note that each transect can pro-duce only one Wrst appearance observation per species peryear, but all visits per transect contribute to the estimationof the Xight period numbers. The date of appearance isexpressed as the Julian calendar date (day 1 = 1 January,etc.).

The analyses included only transects in which a particu-lar species had been observed in that year, with at least oneto three visits during the Xight period (depending on thelength of this period). As a result, the number of transectsvaried between species, but each analysis used the samenumber of transects to calculate all phenological measuresper species. We used only data for the Wrst generations of31 abundant species (see Table 3). The Xight period ofhibernating adults, e.g. Aglais urticae and Gonepteryxrhamni, already starts before the onset of the DBMS on 1April. Although these butterXies are very sensitive to warmspring conditions, they were excluded from the analysis.The data in Table 3 refer to the Wrst oVspring of the hiber-nating adults.

Analysis of number of sampling sites

We started by testing the eVect of the number of samplingsites on the date of Wrst appearance. For each separate spe-cies we did 1,000 random selections of 25, 50 and 100 tran-sects with observations from the data for 2004. For eachselection we calculated FA-25, FA-25% and FP-25% andaveraged the result of all 1,000 trials. The year 2004 waschosen because data were most numerous for recent years.We repeated the test with data for 2005. As this testrequires data for a large number of transects, we used onlythe 20 most widespread species (see Table 3 for a specieslist).

We then examined the eVect of an increasing number ofsampling sites on the changes in phenological measures, by

comparing datasets with an increasing number of transectswith those with a constant number of transects. The numberof transects in the standard dataset of the DBMS increasedfrom 230 in 1992 to about 400 in 2005. Therefore, we cre-ated extra datasets for species with a constant number oftransects per year by Wrst identifying the year with the few-est transects per species, and then—for each year sepa-rately—making 1,000 random selections of the samenumber of transects with observations. The number of tran-sects in the latter datasets were smaller than in the standarddataset, with averages of 74 and 113, respectively. Regres-sion coeYcients of phenological measures with year werecompared between the datasets with an increasing numberof transects and those with a constant number of transects.DiVerences were tested using a paired t-test, unless other-wise stated. In this analysis, in addition to FA-25, FA-25%and FP-25% we also used FA-10, FA-50, FA-10%, FA-50%, FP-10% and FP-50%, in order to test whether theresults depend on the exact parameters chosen (seeTable 2). For several species the number of transects withobservations was too small to calculate FA-25 and FA-50.In these cases this measure was computed for all availabletransects, which may have resulted in an underestimation ofthe eVect of the number of transects.

Furthermore, we compared the regression coeYcients ofyearly estimates of FA-25, FA-25% and FP-25% withyearly spring temperature for the two datasets. This wasdone for 11 early species only, i.e. species observed in theWrst week of April at least once in 1992–2005 (see Table 3for the species list) and was again based on 1,000 trials.Spring temperature was taken as the average of the dailymean temperatures in March and April, as recorded at themeteorological station at De Bilt in the central part of theNetherlands.

Analysis of population abundance

First we tested the eVect of population abundance by acomparison of dates of Wrst appearance between transectswith a high and with a low population abundance of a spe-cies. However, an earlier Wrst observation in high abun-dance rather than in low abundance transects would notnecessarily prove statistical bias. It may also indicate a realphenological change of a species in response to changingtemperature, associated with an increase of local popula-tions. To rule out biological explanations while focusing onpossible bias, we selected two species, Pieris napi andVanessa cardui. Neither species showed an overall popula-tion change in the Netherlands, whereas a trend might beexpected if the species were adapting to warmer springs.Moreover, V. cardui is a migrant species with yearly recol-onization from Africa, and no local populations in theNetherlands. Therefore any diVerences in phenology

123

Oecologia

between transects with low and high population densities ofthese two species are more likely to be caused by statisticalbias rather than biological responses.

Secondly, we examined the relation between populationchange and change in phenological measures. Earlier stud-ies on the eVect of population trend on Wrst appearancescompared the trends in population abundance and phenol-ogy between species (Tryanowksi et al. 2005; Tøttrup et al.2006). Our preliminary analyses showed the date of the25th appearance advanced more strongly for increasingbutterXy species than for declining species (Pearson’s cor-relation between trend in FA-25 and population trendr = ¡0.41, n = 31, P < 0.05). Again we do not regard this asunambiguous evidence for statistical bias, because speciesmight really be adapting to climate change. In order toobtain a better test for bias, we again selected P. napi andV. cardui and assessed the correlation between phenologi-cal change per transect and population trends per transect.

Population trends were computed as multiplicative over-all trend slopes using TRIM (Pannekoek and van Strien2001), a program used to analyse time series of counts withmissing counts using Poisson regression (log-linear mod-els; McCullagh and Nelder 1989). Population trends werecomputed for all transects together. For P. napi and V. car-dui population trends were also assessed for individualtransects, as were the changes in Wrst appearance, but onlyfor transects with time series longer than 7 years.

Analysis of observation period

To study the eVect of changes in the start of the observationperiod, we compiled two extra datasets to compare with thestandard dataset: (1) with all data excluded before 8 April,and (2) with all data excluded before 15 April. For eachdataset we calculated FA-25, FA-25% and FP-25%. Thisanalysis included only 11 early species (see Table 3), as noeVects of a shift in observation period in the Wrst 2 weeks inApril are likely for species with later Xight periods.

Monte Carlo simulations

We applied simple Monte Carlo simulations to explore theeVect of the number of transects and the population abun-dance on phenological measures. Each transect in the simu-lations was assumed to follow the trajectory in the numberof individuals in the season as indicated by the Wrst genera-tion of Araschnia levana in 1992–1994 (Fig. 2). But inaddition, each transect was given a random shock derivedfrom a normal probability distribution with mean 0 and SD5 days. Thus, the Xight period of each transect was ran-domly shifted to make it earlier or later, but overall therewas no change in phenology. The frequencies were thenadded for each Julian date across all transects to obtain the

Julian date for FA-25, FA-25% and FP-25%. Simulationswere done 1,000 times for 25, 50, 100 and 1,000 transectsand for 10, 100 and 1,000 individuals per transect per year.

Results

EVects of number of sampling sites

As expected, a higher number of transects sampled from2004 data resulted in considerably earlier dates for FA-25(Fig. 3). FA-25% and FP-25% were unrelated to the numberof transects. The results were similar for the 2005 data. Thesimulations conWrmed these results: the number of transects

Fig. 3 Julian date (§SE) for three phenological measures (FA-25,FA-25%, FP-25%) in relation to the number of transects sampled. Re-sults are for data from 2004 and are averaged for 20 widespread butter-Xy species. DiVerent letters refer to signiWcant diVerences (one-tailedpaired t-test, P < 0.05). For abbreviations, see Figs. 1 and 2

Fig. 2 Mean Xight period of the Wrst generation of Araschnia levanain 1992–1994 and 2003–2005. Arrows indicate the Julian date of 25%of the Xight periods (FP-25%)

123

Oecologia

aVected FA-25, but not FA-25% nor FP-25% (Table 1). As aconsequence, an increasing number of transects led to stron-ger negative regression coeYcients of FA-n with year, butnot for FA-% and FP-% (Table 2). Regression coeYcientsof FA-n averaged over species were 1.5–2 times higher forthe dataset with an increasing number of transects than forthe dataset with a constant number of transects.

The regression coeYcients of FA-n and FA-% did notdepend on the number of transects or percentages in pheno-logical measures (Table 2). This is not surprising because by

deWnition Wrst appearances cover the same phenophase of thebutterXies. By contrast, the regression coeYcients were morenegative at lower percentages of FP-% (Table 2), indicating astronger phenological change earlier in the season. Fig. 2shows an example of the shift in Xight period for A. levana.The Wrst 25% of its Xight period fell 8 days earlier, the Wrst50% fell 5 days earlier and the Wrst 75% only 3 days earlier.

The regression coeYcients of FA-25 with temperaturediVered between the datasets with an increasing and thosewith a constant number of transects (Fig. 4). The relations

Table 1 Results of Monte Carlo simulations using artiWcial datasets with diVerent numbers of transects and abundance per transect. All estimatesare based on 1,000 trials

a FA-25 diVered signiWcantly between datasets with diVerent number of transects. Both FA-25 and FA-25% diVered between datasets with diVer-ent yearly abundances (t-test; P < 0.05). No signiWcant diVerences were found for FP-25%. Note that for 100 transects, FA-25 equals FA-25%.Also note that FP-25% values correspond with the FP-25% values indicated in Fig. 2

Phenological measurea Number of transects

Yearly abundance

10 100 1,000

Date of 25th appearance (FA-25) § SE

25 132.81 § 0.49 122.84 § 0.50 116.77 § 0.51

50 122.89 § 0.13 112.92 § 0.14 106.83 § 0.13

100 119.53 § 0.07 109.53 § 0.07 103.55 § 0.07

1,000 113.11 § 0.02 103.11 § 0.02 97.13 § 0.02

Date of Wrst 25% of all Wrst appearances (FA-25%) § SE

25 119.80 § 0.28 109.82 § 0.29 103.78 § 0.28

50 119.65 § 0.14 109.67 § 0.15 103.60 § 0.14

100 119.53 § 0.07 109.53 § 0.07 103.55 § 0.07

1,000 119.70 § 0.01 109.72 § 0.01 103.70 § 0.01

Date of Wrst 25% of Xight period (FP-25%) § SE

25 128.40 § 0.21 128.56 § 0.22 128.52 § 0.21

50 128.55 § 0.11 128.55 § 0.12 128.50 § 0.11

100 128.52 § 0.06 128.54 § 0.06 128.53 § 0.06

1,000 128.57 § 0.02 128.57 § 0.02 128.57 § 0.02

Table 2 Mean regression coeYcients (§SE) of phenological mea-sures with year (1992–2005) for datasets with increasing number oftransects [standard dataset Dutch ButterXy Monitoring Scheme(DBMS)] and constant number of transects, averaged for 31 butterXy

species. The same number of transects was used to calculate all mea-sures per dataset. DiVerent letters within measures and datasets refer tosigniWcant diVerences (one-tailed paired t-test, P < 0.05). NS not sig-niWcant; for other abbreviations, see Table 1

a Mean regression for dataset with constant number of transects was based on 1,000 trialsb Compares the results between datasets for an increasing and constant number of transects

Phenological measure

Mean regression (§SE) with year for increasing number of transects

Mean regression (§SE) with year for constant number of transectsa

Paired t-testb

Date of Wrst X appearances (FA-n)

FA-10 ¡0.64 § 0.13 ¡0.37 § 0.09 P < 0.01

FA-25 ¡0.74 § 0.14 ¡0.37 § 0.10 P < 0.01

FA-50 ¡0.66 § 0.15 ¡0.43 § 0.14 P < 0.05

Date of Wrst Y% of all Wrst appearances (FA-%)

FA-10% ¡0.46 § 0.09 ¡0.50 § 0.09 NS

FA-25% ¡0.39 § 0.10 ¡0.47 § 0.10 NS

FA-50% ¡0.40 § 0.10 ¡0.42 § 0.10 NS

Date of Wrst Z% of Xight period (FP-%)

FP-10% ¡0.56 § 0.10 a ¡0.55 § 0.10 d NS

FP-25% ¡0.47 § 0.10 b ¡0.46 § 0.10 e NS

FP-50% ¡0.34 § 0.11 c ¡0.34 § 0.11 f NS

123

Oecologia

of both FA-25% and FP-25% with temperature did notdiVer between the two datasets.

EVects of population abundance

First observations of P. napi and V. cardui were lateralong transects with low population abundance than alongthose with high abundance (see FA-25 in Table 4). ThesediVerences conWrm the expected statistical bias. FA-25%also diVered between the sets of transects with low andhigh abundance (Table 4), as did the mean date of Wrstobservation of P. napi (123.4 § 1.1 and 113.2 § 0.8respectively; t-test, P < 0.05) and V. cardui (211.5 § 2.5and 193.2 § 3.7 respectively; t-test, P < 0.05). By con-trast, FP-25% did not diVer between low and high abun-dance transects for either species. These results wereconWrmed in the simulation study: the yearly abundanceper transect aVected both FA-25 and FA-25%, but not FP-25% (Table 1).

If a larger population leads to earlier Wrst observations,then we can expect population increase to result in advanc-ing Wrst appearances. Indeed, for both P. napi and V. carduiwe found a signiWcant negative correlation between trendsin Wrst appearances per transect and population trends pertransect (P. napi Pearson’s r = ¡0.44, n = 172, P < 0.05;V. cardui r = ¡0.40, n = 145, P < 0.05).

EVects of observation period

The later the start of the observation period, the later the WrstbutterXies were observed (see FA-25 and FA-25% inTable 5). The same was true for the shift in the Xight period(FP-25%). On average, the diVerence in the three measuresbetween datasets starting from 1, 8 and 15 April respectively

was 1.5–2 days. Later starts of the season also led to smallerregression coeYcients of phenological measures with year.These trends were smallest for the dataset starting from 15April, presumably because of the later phenophases covered.

Phenological changes

As FP-% is insensitive to both sampling eVort and popula-tion change, it shows a real shift towards an earlier date in1992–2005 (Table 2). For FP-25%, a total 24 of the 31 spe-cies advanced, 12 of them signiWcantly (Table 3). EarlybutterXy species showed the strongest changes in time:seven out of 11 early species had a signiWcantly earlierXight period, compared with only Wve of the 20 non-earlyspecies (�2-test, P < 0.05). All 11 early species had nega-tive linear relations of FP-25% with temperature, eight ofthem signiWcantly (Table 3). Thus, their appearance peryear was earlier when the temperature was higher, withshifts of 3–9 days per °C.

Discussion

The results show that the theoretically predicted biases inFA-n caused by sampling eVort may be considerable. Forinstance, the average regression coeYcient of FA-25 withyear was ¡0.74 in the standard dataset, which correspondsto FA-25 being brought forward 9.6 days in the period1992–2005 (Table 2). Adjusting for the increase in thenumber of transects, however, produced a regression coeY-cient of only ¡0.37 (Table 2), corresponding to an advanceof 4.8 days in this period. For FA-%, on the other hand, noeVects were found of the number of transects, so the sensi-tivity to sampling eVort depends on the Wrst appearancemeasure applied. Furthermore, the results indicated largediVerences in both FA-n and FA-% caused by diVerences inpopulation abundance, as demonstrated for P. napi andV. cardui (Table 4) and by the simulations (Table 1). Inaddition, a relatively small shift of 1–2 weeks of the start ofthe observation period caused a detectable bias in bothFA-n and FA-% (Table 5).

The eVect of the number of transects on FA-25 presentedin Fig. 3 was stronger than in the example data of Fig. 1.This is because in the DBMS data the SDs of Wrst appear-ances were considerably larger than applied in the example,as not all observers start Weldwork immediately on 1 April.

The eVects of both number of transects and abundance onFA-n and FA-% can be explained by probabilistic arguments.Flight periods diVer between transects as a result of diVer-ences in, for example, microclimate and availability of hostplants. If the number of transects observed increases, onemay expect a higher absolute number of transects with earlyXight periods and thus an earlier FA-25. The relative number

Fig. 4 Mean regression coeYcients (§SE) of three phenological mea-sures (FA-25, FA-25%, FP-25%) with average temperature in Marchand April for datasets with increasing (standard dataset) and constantnumber of transects, averaged for 11 early butterXy species. DiVerentletters within measures refer to signiWcant diVerences (sign-testP < 0.05). For abbreviations, see Figs. 1 and 2

123

Oecologia

of transects with early Xight periods does not change with anincrease in number of transects, so FA-25% remainsunaVected. If the local abundance increases, however, ahigher number of early individuals can be expected to appearper transect, as a result of the natural variation in develop-ment rate between individuals. On average, this would leadto an earlier appearance per transect, thereby aVecting bothFA-25 and FA-25% when transects are combined. By con-trast, FP-% is not aVected by local population change,because the shift in Wrst observations is compensated by ahigher number of observations during the entire Xight period.

Figure 3 and Table 1 show that the eVect of the numberof transects on FA-25 gradually diminishes as the numberof transects increases. Also the eVect of yearly abundancediminishes with higher abundance levels, for both FA-25and FA-25%. But Table 1 suggests a detectable bias in FA-25 and FA-25% even with a large number of transects andhigh abundance levels (Table 1).

Both the butterXy study and the simulations conWrm thecaution expressed in the literature concerning the use ofWrst appearance data. The eVects were even stronger thanexpected: we found bias in Wrst appearance measures forcommon and conspicuous species, not only for rare orsecretive species as suggested by Roy and Sparks (2000)and Tryanowksi et al. (2005). The question is, then, how todeal with these biases.

One option is to take bias into account when using Wrstappearance data. We then recommend the following: (1)when using FA-n, randomly select an equal number of Wrstobservations per year, or use FA-%; (2) if this is not feasi-

Table 3 Regression coeYcients of FP-25% with year (n = 31 species)and with temperature (n = 11 early species only) and population trends,using the standard dataset of the DBMS (1992–2005)

* P < 0.05a The 20 most widespread species

Species Regression coeYcient Population trend

FP-25% with year

25% with temperature

Thymelicus sylvestris ¡0.22 ¡0.09*

Thymelicus lineolaa 0.36 ¡0.08*

Ochlodes faunusa ¡0.50 ¡0.06*

Pyrgus malvae ¡0.60* ¡0.06*

Gonepteryx rhamnia 0.02 ¡0.09*

Pieris brassicaea ¡1.10* ¡5.26* ¡0.04*

Pieris rapaea ¡0.63* ¡3.34* ¡0.02*

Pieris napia ¡0.76* ¡3.43* ¡0.01

Anthocharis cardaminesa ¡0.62* ¡3.77* 0.02

Callophrys rubi ¡1.33* 0.03*

Neozephyrus quercus 0.11 ¡0.06*

Lycaena phlaeasa ¡0.56 ¡4.89* 0.06*

Plebeius argus ¡0.84* ¡0.04*

Aricia agestis ¡0.02 ¡0.03*

Polyommatus icarusa ¡0.46 0.03*

Celastrina argiolus ¡0.62* ¡4.33* 0.17*

Vanessa atalantaa ¡0.29 ¡5.03 0.01*

Vanessa carduia ¡0.49 ¡8.02 0.01

Aglais urticaea ¡2.38* ¡0.08*

Inachis ioa 0.06 ¡0.11*

Polygonia c-albuma ¡0.58 0.13*

Araschnia levanaa ¡0.61* ¡4.12* 0.02*

Issoria lathonia ¡0.90 ¡9.27* ¡0.06*

Argynnis niobe 0.00 ¡0.16*

Pararge aegeriaa ¡0.54* ¡4.33 0.09*

Lasiommata megera ¡0.53 ¡0.08*

Aphantopus hyperantusa ¡0.19 ¡0.01

Coenonympha pamphilusa ¡0.84* 0.04*

Pyronia tithonusa 0.08 ¡0.03*

Maniola jurtinaa ¡0.22 0.01*

Hipparchia semele 0.75 ¡0.12*

Table 4 Phenological measures in 2004 for transects with low andhigh abundance of P. napi (<10 and ¸10 individuals per transect peryear) and of V. cardui (<5 and ¸5 individuals per transect per year).DiVerent letters within measures and species refer to signiWcant diVer-ences (t-test, P < 0.01). For abbreviations, see Table 1

Phenological measure

P. napi V. cardui

Low abundance (n = 143)

High abundance (n = 118)

Low abundance (n = 129)

High abundance (n = 66)

FA-25 114 a 107 b 172 c 168 d

FA-25% 115 e 107 f 187 g 163 h

FP-25% 118 118 203 204

Table 5 Mean date and regression coeYcient with year (1992–2005)for datasets with diVerent observation period. Values are averages of11 early butterXy species. DiVerent letters within measures refer to sig-niWcant diVerences (sign-test, P < 0.05). For abbreviations, see Table 1

a The dataset starting on April 1 is the standard dataset of the DBMS

Phenological measure

Date of start of observationperiod

Mean date (§SE)

Mean regression with year (§SE)

FA-25 91 (=April 1)a 127.8 § 6.4 a ¡1.30 § 0.19 d

98 (=April 8) 129.0 § 6.5 b ¡1.20 § 0.22

105 (=April 15) 129.8 § 6.4 c ¡1.15 § 0.22 e

FA-25% 91 (=April 1) 127.2 § 6.3 f ¡0.88 § 0.14 i

98 (=April 8) 128.1 § 6.6 g ¡0.82 § 0.14 j

105 (=April 15) 128.6 § 6.6 h ¡0.80 § 0.14 j

FP-25% 91 (=April 1) 140.9 § 9.8 k ¡0.65 § 0.06 n

98 (=April 8) 141.4 § 9.7 l ¡0.59 § 0.07 o

105 (=April 15) 142.3 § 9.6 m ¡0.52 § 0.06 p

123

Oecologia

ble, for example if only one Wrst observation per year isavailable (FA-1), check whether sampling eVorts are suY-ciently similar across years, or be very careful when usingthe magnitude of phenological change as well as anyparameters associated with temperature; (3) take intoaccount bias caused by changes in population abundance,as it is unlikely for a population of a species to remain sta-ble during the usually long-term studies on phenologicalchanges. If no monitoring scheme is available, data on pop-ulation abundance can be collected simultaneously withWrst appearance data, for example by recording the numberof butterXies observed during the visits. Another possibilityis to record not only the Wrst observation per site, but to col-lect information on several subsequent observations aswell, and to use the diVerence in dates between these obser-vations to adjust for any local population changes.

A further option would be to avoid using Wrst appearancedata and instead apply metrics free from bias. We foundFP% to be insensitive to the number of sites and the popula-tion abundance (Table 6). Such a measure based on dataencompassing the entire Xight period may, however, bemuch more demanding for observers. A less time-consum-ing alternative would be to collect data on butterXy appear-ances regardless of Wrst appearances, and then to calculatethe mean date of all data recorded by multiple observers peryear. This measure can be expected to be free from bias andmight also be compared with the mean recording date ofmuseum specimens (see Primack et al. 2004).

Whichever measure is used, it remains important to stan-dardize the observation period and to check whether onaverage observers indeed did not start Weld work earlier inwarm springs. In addition, the representativeness of studylocations should be taken into account. In many phenologi-cal studies, observation locations are not randomlyselected. Estimates will be biased if locations wherechanges are most apparent—for example urban areas—areover-represented in the data. If a stratiWed random design isnot feasible, data from diVerent locations could be weightedin the same way as in methods applied in population moni-toring. In the DBMS, for instance, weighting proceduresare applied to adjust for oversampling of particular areas(van Swaay et al. 2002).

The unbiased FP-% clearly shows earlier Xight periodsof butterXies, especially for early phenophases and earlyspecies (Tables 2 3). This result conWrms Wndings else-where (Roy and Sparks 2000; Stefanescu et al. 2003; Men-zel et al. 2006; Tøttrup et al. 2006). These changes in FP-25% were correlated with FA-25 and FA-25% (Pearson’sr = 0.44 and 0.48, respectively, using the standard DBMSdataset; n = 31; P < 0.05). Thus, Wrst appearance data canstill be useful to detect phenological change, even thoughthey are biased. The estimated magnitude of changeremains questionable, however. Biased estimates of Wrstappearances such as FA-25 also result in bias in the assess-ment of relations with temperature, as a higher spring tem-perature coincides with expansion of the number oftransects in the DBMS.

Current phenological research focuses on detectingchanges in phenology in relation to climatic change. Littlepriority is given to developing procedures to obtain unbi-ased estimates of the extent of phenological changes. Unbi-ased estimates of phenological change are a prerequisite formeaningful comparisons of changes between species andregions, and to draw conclusions on mismatches betweenparts of the food chain, such as predators and prey species(Visser and Both 2005). Therefore, bias in phenologicalestimates should no longer be ignored in phenologicalresearch. Studies with insuYcient information on the qual-ity of estimates of phenological changes should be treatedwith caution.

Acknowledgments Lodewijk van Duuren and Adriaan GmeligMeyling helped with part of the analyses. Jeroen Pannekoek, Willy vanStrien, Arnold van Vliet and two anonymous referees gave helpfulcomments on earlier versions of the manuscript. Lieneke Hoeksma im-proved the English. The DBMS is a joint scheme of the Dutch ButterXyConservation and Statistics Netherlands in the framework of the DutchNetwork Ecological Monitoring programme. The DBMS is Wnancedby the Ministry of Agriculture, Nature and Food quality. This workwould not have been possible without the help of many voluntary Weldworkers.

References

Barrett RT (2002) The phenology of spring bird migration to northNorway. Bird Stud 49:270–277

Table 6 Sources of bias in phenological measures covered in this study

Phenological measure Bias due to change in number of sampling sites

Bias due to change in populationabundance

Bias due to change in observation period

Date of Wrst X appearances (FA-n)

Yes Yes Yes

Date of Wrst Y% of all Wrst appearances (FA-%)

No Yes Yes

Date of Wrst Z% of Xight period (FP-%)

No No Yes

123

Oecologia

Both C, Piersma T, Roodbergen SP (2005) Climatic change ex-plains much of the 20th century advance in laying date ofnorthern lapwing Vanellus vanellus in the Netherlands. Ardea93:79–88

Bruns E, van Vliet AJH (2003) Standardisation of phenological moni-toring in Europe. Report of the European Phenology Network,Wageningen University, Deutscher Wetterdienst, Wageningen,OVenbach

Forister ML, Shapiro AM (2003) Climatic trends and advancing springXight of butterXies in lowland California. Glob Change Biol9:1130–1135. doi:10.1046/j.1365-2486.2003.00643.x

Inouye DW, Barr B, Armitage KB, Inouye BD (2000) Climate changeis aVecting altitudinal migrants and hibernating species. Proc NatlAcad Sci USA 97:1630–1633. doi:10.1073/pnas.97.4.1630

McCullagh P, Nelder JA (1989) Generalized linear models, 2nd edn.Chapman & Hall, London

Menzel A, Sparks TH, Estrella N et al (2006) European phenologicalresponse to climate change matches the warming pattern. GlobChange Biol 12:1969–1976. doi:10.1111/j.1365-2486.2006.01193.x

Pannekoek J, van Strien AJ (2001) TRIM 3 manual. Trends and indicesfor monitoring data. Research paper no. 0102. Statistics Nether-lands, Voorburg

Parmesan C, Yohe G (2003) A globally coherent Wngerprint of climatechange impacts across natural systems. Nature 421:37–42.doi:10.1038/nature01286

Pollard E, Yates TJ (1993) Monitoring butterXies for ecology and con-servation. Chapman & Hall, London

Primack D, Imbres C, Primack RB, Miller-Rushing AJ, DelTredici P(2004) Herbarium species demonstrate earlier Xowering times inresponse to warming in Boston. Am J Bot 91:1260–1264

Roy DB, Sparks TH (2000) Phenology of British butterXies and cli-mate change. Glob Change Biol 6:407–416. doi:10.1046/j.1365-2486.2000.00322.x

Sparks TH (1999) Phenology and the changing pattern of bird migra-tion in Britain. Int J Biometeorol 42:134–138. doi:10.1007/s004840050096

Sparks TH, Mason CF (2004) Can we detect change in the phenologyof winter migrant birds in the UK? Ibis 146:57–60. doi:10.1111/j.1474-919X.2004.00328.x

Stefanescu C, Penuelas J, Filella I (2003) EVects of climate change onthe phenology of butterXies in the northwest Mediterranean Ba-sin. Glob Change Biol 9:1494–1506. doi:10.1046/j.1365-2486.2003.00682.x

Tøttrup AP, Thorup K, Rahbek C (2006) Patterns of change in timingof spring migration in North European songbird populations.J Avian Biol 37:84–92. doi:10.1111/j.2006.0908-8857.03391.x

Tryanowksi P, Kuzniak S, Sparks TH (2005) What aVects the magni-tude of change in Wrst arrival dates of migrant birds? J Ornithol146:200–205. doi:10.1007/s10336-005-0079-4

van Swaay CAM, Plate CL, van Strien AJ (2002) Monitoring butter-Xies in the Netherlands: how to get unbiased indices. Proc ExpAppl Entomol 13:21–27

Visser ME, Both C (2005) Shifts in phenology due to global climatechange: the need for a yardstick. Proc R Soc Lond B Biol Sci272:2561–2569. doi:10.1098/rspb.2005.3356

123