The effect of non-target species in presence-absence distribution surveys: A case study with hair-tubes

Abstract

Detection of non-target species during distribution surveys may influence the detection of the focal species, due to bait being consumed, or trapping devices inactivated. The aim of this work was to evaluate the effect of non-target species (field mice, Apodemus sp.) on the detection and occupancy estimates of a target species (the red squirrel, Sciurus vulgaris) during hair-tubes surveys. Following a modelling approach that accounted for imperfect detection of both target and non-target species, we tested the hypothesis that detection probability of the red squirrel is affected by detection of field mice. We also investigated the level of bias that occurred in estimation of key-parameters such as probability of presence and detection probability. Our results show that detection of red squirrels and field mice using hair-tubes surveys is not independent. Detection probability of the red squirrel was higher when field mice did not visit the hair-tubes. Nevertheless, there was no bias in parameter estimates, therefore relatively accurate estimates would have been obtained despite this interference.

Keywords:

• Sciurus vulgaris
• Apodemus
• detection probability
• occupancy models
• Italy

Introduction

Presence/absence data of target species are often the focus of distribution and conservation studies (Scott et al. 2002). However, if the trapping device is not selective, non-target species may be detected. While in some cases detection of non-target species may not affect results (e.g. when using camera traps or scent-stations), in others the detection probability of the target species could be lowered, as in the case of bait being consumed or trapping devices inactivated.

A thorough consideration of this problem is mandatory, as it may cause bias in parameter estimates, such as underestimation or overestimation of presence probability and detection probability (MacKenzie et al. 2004), key parameters in conservation and distribution studies.

Selective trapping devices are often used to reduce the number of species detected – for example, mouse-excluders from Longworth traps (Gurnell & Flowerdew 2006), species-specific lures (Zielinski & Kucera 1995) and selective baiting (King et al. 2007) – however, in several circumstances, exclusion of non-target species may not be achievable (Barnett & Dutton 1995).

Several analytical methods may be used to account for the effect of non-target species, such as specific covariates (e.g. presence/absence of non-target species at a site). However, these do not account for imperfect detection of both target and non-target species, resulting in false absences in the analyses, resulting in false absences in the analyses, with the associated bias (Gu & Swihart 2004; MacKenzie et al. 2006). However, only MacKenzie et al. (2004) have developed a modelling approach to deal with detection and interaction of multiple species in situations where species may be detected imperfectly. This model allows interactions to be in terms of occupancy probability (e.g. species avoidance), or in terms of detectability, whereby one species affects the probability of detection of the other species (MacKenzie et al. 2006).

Hair-tubes are often used to monitor red squirrel (Sciurus vulgaris) abundance and distribution, since they are a relatively cost-effective methodology (Gurnell et al. 2004; Finnegan et al. 2007; Bertolino et al. 2009). However, hair-tubes are a non-selective device as they may detect other species (e.g. field mouse Apodemus sylvaticus, yellow-necked field mouse Apodemus flavicollis or fat dormouse Glis glis) (Lindenmayer et al. 1999; Mortelliti & Boitani 2008; Molinari et al. 2009). Nevertheless, to date no study has investigated the potential bias introduced by visits of non-target species such as mice.

The aim of this work was to evaluate the effect of non-target species (Apodemus sp.) on the detection and occupancy estimates of a target species (red squirrel). The hypothesis we tested was that detection probability of the red squirrel is affected by detection of field mice. Secondly, we investigated the level of bias that occurred in the estimation of key parameters, such as probability of presence and detection probability.

Materials and methods

We carried out a hair-tube survey as part of a project investigating the effects of habitat fragmentation on rodents in a sample of 24 habitat patches in the Province of Siena, central Italy. Patches varied in size, isolation, and age (Table I). Dominant tree species were the downy oak Quercus pubescens and the Turkeyoak Quercus cerris; the surrounding matrix was composed of wheat (Triticum spp.) or alfalfa Medicago sativa fields. Sampled patches were widely distributed across three valleys (Val d'Orcia, Val d'Arbia, Crete Senesi; average nearest patch distance = 1624 m; range 350–6039 m) and were included in an area of 45,000 ha (area of the convex polygon surrounding patches).

Table I. Patch size in hectares, number of hair-tubes used and summary of the distribution of red squirrel (Sciurus vulgari s) and field mice (Apodemus sp.) in 24 patches in the Province of Siena, Central Italy. 0 = none detected, 1 = detected

Patches	Hectares	Number of hair-tubes	Detection/non-detection of Sciurus vulgaris	Detection/non-detection of Apodemus spp.
1	0.48	1	0	1
2	0.41	1	0	1
3	1.38	2	1	1
4	1.07	2	0	1
5	1.32	2	0	1
6	1.15	2	0	0
7	1.80	2	0	1
8	2.20	2	0	1
9	2.24	2	0	1
10	2.22	2	0	1
11	2.15	2	1	1
12	3.00	2	0	1
13	4.51	2	1	1
14	3.86	2	0	0
15	5.09	3	0	1
16	6.37	3	0	1
17	8.59	4	0	1
18	8.17	4	1	1
19	15.65	4	1	1
20	14.79	7	1	1
21	20.85	10	1	1
22	27.24	10	1	1
23	65.56	12	1	1
24	79.86	12	1	1

Hair-tubes were made from commercially available plastic tubes (length 30 cm, diameter 6 cm); a strip of Velcro^© with adhesive material (Alt© rodent glue; Valbrenta Chemicals) and placed on the inner upper side to capture squirrel hairs. A mixture of chocolate, maize and sunflower seeds was used as bait. Hazelnuts were glued in the inner part of the tube in case mice or birds immediately consumed the bait. The number of hair-tubes increased with increasing patch size (Table I), to a total of 95 tubes in the 24 patches. Tubes were randomly distributed in the patch with a minimum distance of 100 m between tubes. The tubes were inspected at 10-day intervals until squirrel hairs were found in the patch (removal design, see MacKenzie et al. 2006 for details). If no hairs were found, the survey was ended after 2 months. The survey was carried out during spring–summer 2007 (March–August).

Hairs were identified to species using procedures described by Teerink (1991). Grey squirrel (Sciurus carolinensis) is not present in the area (pers. obs.). Mice hairs were determined to the genus level as microscope analyses do not allow species determination between Apodemus spp. (Teerink 1991); both Apodemus sylvaticus and Apodemus flavicollis were present in the study area (Mortelliti 2008).

We ran analyses using the program PRESENCE 2.1 (available for download at www.mbr-pwrc.usgov/software.html), utilizing a two-species interaction model, including covariate effects – patch size and tree size – (MacKenzie et al. 2004).

Detection history data for the red squirrel and field mice refer to each hair-tube, therefore the sequence of 0’s and 1’s accounts for presence and absence of the red squirrel and Apodemus spp. hairs each time hair-tubes were checked.

We followed an information-theoretic approach to data analysis (Burnham & Anderson 2002): we defined a set of a-priori models with varying covariates that could explain the patterns of patch occupancy; each model tested corresponds to a specific hypothesis. Models were ranked according to AIC (Akaike Information Criteria) values. We then calculated nAIC and Akaike weights (Burnham & Anderson 2002) using model averaging to account for model selection uncertainty (Burnham & Anderson 2002).

We assumed occupancy was independent. Therefore the probability that the area surrounding the hair-tube was occupied did not depend on the presence or absence of other species, since the species are not known to compete with, or avoid, each other. Both models assuming dependence and independence in detection probability were fitted to data, therefore it was possible to evaluate the relative support of the two contrasting hypotheses: (1) the probability of detecting one species was affected by the detection of the other species, and (2) detection of the two species was independent.

Model description

Red squirrel presence probability (ψ) was modelled as a function of patch size as analyses carried out on data from the previous year showed how the probability of presence increases in larger patches (Mortelliti & Boitani 2008).

Detection probability is subdivided in several parameters: p ^a = probability of detecting the red squirrel during the jth survey given only the red squirrel is present; p ^b = probability of detecting Apodemus spp. during the particular survey given only Apodemus spp. are present; r ^a = probability of detecting the red squirrel given both squirrel and mice are present; r ^b = probability of detecting mice given the red squirrel is present and r ^ab = probability of detecting both species given both are present. All these were modelled as a function of dbh (diameter at breast height, a surrogate indicator of tree size) and patch size. Further models that were fitted to data were: (1) models with survey-specific detection probability, (2) a model assuming independent detection probability (with no covariates), and (3) one with red squirrel presence probability as a function of patch size.

Through the analyses, it is possible to estimate a species interaction factor Λ = r ^ab/r ^a*r ^b (MacKenzie et al. 2004). Values of Λ < 1 suggest it is less likely to detect both species in a survey than if the species were detected independently, values Λ > 1 suggest it is more likely to detect both species, while values of Λ = 1 suggest the species are detected independently (MacKenzie et al. 2006).

As common practice in studies following the information-theoretic approach (Burnham & Anderson 2002), we here show results of the top ranked models, including all models with Akaike weight > 0.

Results

Hair-tubes were active a total of 6440 trap-nights. Red squirrels were detected in eight patches (48 detections), whilst Apodemus spp. species were detected in 22 patches (84 detections). The only other non-target species detected was the fat dormouse (Glis glis), found only in one patch.

Results for the first ranked models are shown in Table II. In the first ranked model, ψ (probability of presence) is modeled as a function of patch size, and probability of presence increases in hair-tubes placed in larger patches. The first model assumes detection of the two species is not independent and the value of Λ is quite low (= .30), suggesting that it is less likely to detect both species during a survey than if species were detected independently. The probability of detecting the red squirrel during the jth survey, given only the red squirrel is present (p ^a = 0.67), is higher than the probability of detecting the red squirrel given both species are present (r ^a = 0.55). Conversely, the probability of detecting a field mouse during the jth survey, given only Apodemus spp. are present, is lower than the probability of detecting mice given both mice and red squirrels are present.

Table II. Summary of the top 10 ranked occupancy models obtained through analyses with program PRESENCE, two-species interaction. Models are ranked according to ΔAIC; W = model weight; −2LL = −2 Log-Likelihood. In the table we follow the notation of MacKenzie et al. (2004), in brackets next to the parameter we list the covariates that were modelled as function of the parameter. Ψ = probability of presence; p ^a = probability of detecting the red squirrel given only the red squirrel is present; p ^b = probability of detecting Apodemus spp. given only Apodemus spp. are present; r ^a = probability of detecting the red squirrel given both squirrel and mice are present; r ^b = probability of detecting mice given the red squirrel is present; r ^ab = probability of detecting both species given both are present; log-ha = logarithm of patch size; dbh = diameter breast height

Model	Δ AIC	W	−2LL
Ψ^a(log_ha), Ψ^b(), p^a(), p^b(), r^a(), r^b(), r^ab()	0	0.6611	451.28
Ψ^a(log_ha), Ψ^b(), p^a(), p^b(dbh), r^a(), r^b(), r^ab()	1.75	0.2756	451.03
Ψ^a(log_ha), Ψ^b(log_ha), p^a(), p^b(), r^a(), r^b(), r^ab()	5.95	0.0337	457.24
Ψ^a(log_ha), Ψ^b(), p^a(day), p^b(), r^a(), r^b(), r^ab()	6.26	0.0289	449.55
Ψ^a(), Ψ^b(), p^a(), p^b(day), r^a(), r^b(), r^ab()	14.11	0.0006	459.40
Ψ^a(log_ha), Ψ^b(), p^a(), p^b(), r^a(), r^b(), detection independent	21.45	0	474.74
Ψ^a(log_ha), Ψ^b(), p^a(), p^b(log_ha), r^a(), r^b(),r^ab()	24.78	0	474.07
Ψ^a(log_ha), Ψ^b(), p^a(), p^b(), r^a(dbh), r^b(), r^ab()	25.34	0	492.63
Ψ^a(), Ψ^b(), p^a(), p^b(), r^a(), r^b(), r^ab()	26.36	0	493.65
Ψ^a(), Ψ^b(), p^a(day), p^b(), r^a(), r^b(), r^ab()	33.41	0	478.70

The second ranked model has considerable less weight than the first, ψ (probability of presence) is modelled as a function of patch size, whilst p ^b (the probability of detecting an Apodemus during the jth survey, given only Apodemus are present), is modelled as a function of dbh (diameter breast height), with p ^b increasing slightly in larger trees (p ^b range = 0.37–0.46).

Model-averaged parameter estimates are shown in Table III and account for the first six models (model weight > 0).

Table III. Model-averaged parameter estimates with standard errors in brackets. Ψ^A = probability of red squirrel presence; ψ^B = probability of Apodemus spp. Presence; p ^A = probability of detecting the red squirrel given only the red squirrel is present; p ^B = probability of detecting Apodemus spp. given only Apodemus spp. are present; r ^A = probability of detecting the red squirrel given both squirrel and mice are present; r ^B = probability of detecting mice given the red squirrel is present; Λ = species interaction factor

Ψ^A	Ψ^B	p ^A	p ^B	r ^A	r ^B	Λ
0.41 (0.06)	0.74 (0.08)	0.66 (0.17)	0.40 (0.04)	0.55 (0.08)	0.49 (0.08)	0.29 (0.1)

Models assuming non-independent detection probability rank higher than the model assuming independent detection probability (ΔAIC = 21.45, model weight = 0; Table II).

A comparison between the estimate of ψ for the models assuming dependent detection probability (model-averaged parameter estimate) and the estimate for the model assuming independent detection probability show little difference: the estimate obtained with independent detection probability is slightly higher, but the standard errors overlap (0.42 SE = 0.06 vs. 0.45 SE = 0.07).

Discussion

Our results show how detection of the red squirrel and Apodemus spp. using hair-tubes surveys is not independent: models assuming non-independent detection probability have more support than models assuming independent detection probability. Our results demonstrated that the probability of detecting the red squirrel was higher when Apodemus spp. did not visit the hair tubes. Nevertheless, there was no bias between occupancy probability estimates; therefore, if the focus of a study is to establish the probability of presence, relatively accurate estimates can still be obtained.

Application of these models to hair-tube survey data should consider model assumptions carefully. First, models assume that populations are closed during surveys (Mackenzie et al. 2004), therefore the duration of the survey should be limited. Second, models assume detection between hair-tubes is independent (MacKenzie et al. 2006); given the minimum distance of 100 m, commonly used in hair-tube studies (Bertolino et al. 2009), this requirement should have been met for these three species, nevertheless in smaller patches (less than 1 ha) closer spacing was necessary.

We underline that detection probability is the probability that a hair-tube will detect a red squirrel hair during a 10-day period. Therefore, a detection event (a 1 value in a detection history) is independent from the number of actual visits; in other words, different individuals visiting a single hair-tube will count as one individual. Another consequence of not being able to recognize individuals is that ‘tube-addicted’ squirrels or mice (individuals that repeatedly visit tubes between our visits) will lead to overestimation of detection probability. We also acknowledge that a possible underestimation of Apodemus spp. detection probability may have occurred due to the fact that these species are smaller than the red squirrel and may pass through the hair-tubes without being detected. Nevertheless, the large number of detections (84) and the fact that Apodemus spp. were detected in 91% of the patches suggest that this might have happened in a small number of cases with the same rate in tubes where the red squirrel was detected and in tubes where the red squirrel was not detected, therefore it would not affect our conclusions.

The models with Apodemus spp. detection probability as a function of tree size rank higher; the interpretation of the second ranked model is that when only Apodemus spp. are present in the area surrounding the tube, detection probability slightly increases in larger trees. We underline that this is the second ranked model, with considerably less weight. However, these results suggest that placing hair-tubes in relatively smaller trees may help to minimize interference.

In addition, we stress that arboreality of Apodemus species may differ between Apodemus flavicollis and Apodemus sylvaticus, with the former being more arboreal (Hoffmeier 1973).

Consequently, differential detection probabilities and different levels of interaction between the red squirrel and field mice (Apodemus spp.) may arise in different environmental contexts, years and seasons. This could also explain an interesting pattern we have observed: the fact that the probability of detecting a field mouse during the jth survey, given only Apodemus spp. were present, was lower than the probability of detecting mice, given both mice and red squirrels were present. Since Apodemus flavicollis presence probability is higher in larger patches (Mortelliti 2008) where the presence probability of the red squirrel is also higher, the higher detection probability of Apodemus spp. may reflect a higher abundance of the relatively more arboreal Apodemus flavicollis.

In our case, despite non-independent detection, there was no bias in parameter estimates; therefore, if the focus of a study is to obtain accurate probability of presence estimates, these may still be unbiased. Nevertheless, we suggest adopting a cautionary approach since, as discussed above, in different study areas or sampling periods the observed dependence in detection probability could produce biased parameter estimates. This means that: (1) data analysis should account for detection of non-target species, and (2) effort should be undertaken to minimize visits by non-target species, since visits by non-target species (Apodemus spp.) imply bait consumption.

In conclusion, our results show how detection of the red squirrel and field mice using hair-tubes surveys is not independent. This may not bias parameter estimates; nevertheless, we suggest that great effort should be undertaken to: (1) minimize detection of non-target species; and (2) account for non-target species during the analyses.

A pilot study may help in determining the level of interference and determine appropriate strategies to minimize it.

Acknowledgements

This project was financed by the Province of Siena ‘Ufficio Risorse Faunistiche e Riserve Naturali’. Thanks to Joyce Keep for language revision. Hair-tube surveys comply with current Italian legislation.