Aversive conditioning in honey bees (Apis mellifera anatolica): a comparison of drones and workers

The present study uses a newly developed protocol to examine aversive learning in honey bee drones and workers. Honey bees provide an excellent model system to elucidate the relationship between brain architecture and social complex behaviors within the same species: workers (females) are highly social and males are more solitary. Patrilineal effects have been correlated with learning ability, aggression and other variables, making the comparative examination of these behaviors significant for further evolutionary and molecular analysis (Bhagavan et al., 1994; Ferguson et al., 2001; Guzman-Novoa et al., 2005).

Previous evidence offers a positive correlation between socio-behavioral roles and changes in brain architecture in worker bees (Fahrbach et al., 1995; Grozinger et al., 2007; Fahrbach et al., 1998; Withers et al., 1993; Farris et al., 2001). Neural architecture has been shown to be influenced by age, task and experience. Principally socially mediated plasticity in the mushroom body (MB) has been observed in the age-induced transition from inside tasks, such as nursing the offspring, to more complex external foraging tasks (Fahrbach et al., 1998; Smith et al., 2010; Farris et al., 2001; Fahrbach et al., 2003). Foraging demands new types of learning, such as identifying and memorizing the colony's location, spatial orientation using environmental cues for navigation and social communication such as the dance language to recruit foragers (Seeley, 1995).

In males, similar changes in the MB have been documented as they begin orientation and mating flights (Fahrbach et al., 1997). The described expansion also coincides with greater expression of dopamine receptors in the mushroom body and an increase in juvenile hormone titers that mirrors those described in workers (Giray and Robinson, 1996; Humphries et al., 2003). Dopamine levels in honey bees have been shown to influence associative (punishment) learning (Vergoz et al., 2007; Agarwal et al., 2011). Juvenile hormone in both males and females has been shown to work as a regulator of behavioral development (Robinson et al., 1989; Fahrbach et al., 1995; Giray and Robinson, 1996; Giray et al., 1999; Giray et al., 2005). These results suggest that the transition to more complex tasks that require learning coincides with broad expansions of neural tissue in the MB.

Similarities and differences in neuroanatomy prompt further study, for an analysis of learning in honey bees would be incomplete if restricted to only the worker honey bees [queens (Aquino et al., 2004); for drones, see below]. Few reported studies examine learning in drones; those available used classical conditioning of the proboscis extension reflex (PER) [first described by Frings (Frings, 1944) and refined over the years (Kuwabara, 1957; Takeda, 1961; Vareschi, 1971; Bitterman et al., 1983; Abramson and Boyd, 2001) to investigate the heritability of conditioning in workers (Bhagavan et al., 1994; Benatar et al., 1995; Chandra et al., 2001; Ferguson et al., 2001)]. These studies used harnessed individuals motivated for food rewards to measure response. Here, we used a newly developed electric shock avoidance assay (Agarwal et al., 2011) to study escape and punishment; in addition, both contingencies have been previously demonstrated using formic acid in a shuttlebox assay (Abramson, 1986). In escape conditioning, an aversive event is presented and is terminated by a response. In addition, the results of escape conditioning are similar to those found with food rewards (see Mackintosh, 1974). Here, an aversive event is presented when the response occurs (Kaczer and Maldonado, 2009), and is ideal for researchers looking for a protocol that produces the opposite effect of escape training.

The study of aversive conditioning is practical and ecologically relevant. Under field conditions, honey bees face numerous challenges related to the escape and avoidance of aversive stimuli, including those related to predators, repellents and pesticides. Free-flying honey bees stopped flying to a target after being punished with a single taste of an essential-oil-based pesticide (Abramson et al., 2006). Negative feedback signaling during forager dances has been observed to reduce visitations to feeders on which peril or competition has been experienced, thus demonstrating not only individual perception of dangers but also a medium for social dissemination (Nieh, 2010). However, the use of aversive stimuli in studies of honey bee learning and memory is extremely rare, and no investigations describe learning in drones. Of 62 honey bee learning citations, none describe experiments using aversive stimuli (Wells, 1973); and in a review of the advantages that honey bees offer students of cognition, none of the 105 citations discuss aversive conditioning (Srinivasan, 2010).

The study of aversive conditioning can provide new paradigms to further advance our understanding of neurobiological and genomic learning mechanisms. Much of what is known about these mechanisms comes from work related to the Pavlovian conditioning of the PER. Issues with PER conditioning, and recent work on the role of biogenic amines in reward and punishment pathways, has stimulated the search for new conditioning paradigms in the area of aversive conditioning (Kaczer and Maldonado, 2009; Abramson et al., 2011; Agarwal et al., 2011; Vergoz et al., 2007; Giray et al., 2007).

Subjects were honey bee foragers (female) and drones (males) (Apis mellifera anatolica Maa 1953) located in the apiary of the Beekeeping Development and Research Center of the Uludağ University in Bursa, Turkey. Foragers were randomly collected from a feeder containing clove-scented 1 mol 1⁻¹ sucrose solution (5 μl 1⁻¹ clove oil). Subjects were experimentally naive and allowed to feed to satiation before capture. Drones were collected from the hive in the mornings and stored in drone crates fitted with queen-/drone-excluding mesh to allow worker entrance and escape but exclude drones from leaving the cell. These drones were stored in their home hive, where they were fed by workers until data collection was initiated. Drones were tested for maturity upon completion of each session by inducing eversion and ejaculation, and those with mucus and sperm on the adeagus were considered ‘reproductively mature’ (Giray and Robinson, 1996). Only mature drones were used for data analysis. Each subject was used for one session and then terminated.

An automated apparatus was used for the aversive conditioning assessments (Fig. 1A). The shuttlebox (14×1.5×0.8 cm) was modeled from the apparatus employed by Agarwal et al. (Agarwal et al., 2011) and was constructed from black plastic and outfitted with a clear plastic roof and two infrared beam ports positioned 1 cm from the center line on both sides of the shuttlebox (Fig. 1B). A stainless steel 29-pin shock grid (14×20 cm useable area) served as the floor of the shuttlebox. This grid was wired bilaterally (14 pins on each side) with a neutral center pin to allow for discrete shock application to one or both sides of the shock grid as needed for individual experimental and trial constraints. Shock was administered via a 1.2 A variable voltage Universal AC Adapter (model number: DX-AC1200, Dynex, Lincoln, UK). When set at 9 V DC, actual measurements on the shock grid were 8.71 V at 1.0 A. A clear plastic sheath was placed below the grid to allow easy cleaning and to prevent bees from attempting to escape through the pins of the shock grid. The shuttlebox lid was coated with a thin layer of petroleum jelly to prevent bees from walking upside-down to ensure the bees remained in contact with the shock grid for the length of the trial.

The shuttlebox was designed to house one subject and contained two side-looking infrared photodiode–phototransistor pairs (512-QEE113, 512-QSE113, Fairchild Semiconductor, San Jose, CA, USA) mounted in the infrared beam ports positioned parallel to the centerline of the shuttlebox. The orientation of one photodiode–phototransistor pair was reversed with respect to the other pair in each shuttlebox so that light from the photodiode in one pair would not inadvertently buffer triggering of the other pair's phototransistor. A mutually exclusive activation circuit was constructed such that responses would only be detected if the subject interrupted the infrared beam farthest from them to ensure the subject must cross the center line to trigger a response. A PIC18F2580-I/SO microcontroller and relay were used to construct this circuit. A plastic top-justified hurdle at the center line ensured that subjects would break the infrared beam. Two shuttleboxes were run in tandem on the same shock grid and placed upon a 17 inch Dell 1704FPTt flat panel monitor (Round Rock, TX, USA) set at default color settings. This monitor was used as the source of visual stimuli when required. To synchronize other aspects of the apparatus with the visual stimuli presented on the computer monitor, a photoresistor-based light-activated relay (Fk401, Backatronics, Meriden, CT, USA) was positioned on the computer monitor. When stimuli were presented on the monitor, the light-activated relay would activate and send a signal to the other equipment.

An experiment controller, developed by Palya and Walter (Palya and Walter, 1993), was interfaced with the previously described equipment. The experiment controller administrated shock in accordance with the experimental design, and was interfaced with the infrared beam circuit to detect responses and with the light-activated relay to synchronize the experiment with the video display. The experiment controller ran custom programs written in ECBASIC (Jacksonville State University, Jacksonville, AL, USA), to define input properties, and organized data and administered stimuli to the specifications of each experiment. After each experimental session, the experiment controller saved labeled time-stamped data files to a computer.

Experimentally naive honey bees were collected for each trial. Prior to each trail, the plastic sheath and shock grid were cleaned and placed upon the powered-off computer monitor. Each shuttlebox was counterbalanced for each session (i.e. a master shuttlebox for one trial would be the yoked shuttlebox for next session). To avoid the deleterious effects of anesthetizing foragers and to allow for between-sex comparisons of non-anesthetized drones, an original translocation device was constructed for transporting and depositing forager bees into the apparatus.

The translocation device consisted of a hollow clear plastic cylinder, 9 cm in length, with a 1.2 cm diameter internal tube and a clear plastic strip 12 cm long and 1.5 cm wide to serve as a shielding extension (Fig. 1C). This device takes advantage of the phototropic nature of honey bees by coaxing them into the tube from their holding cell using an LED flashlight. The tube was briefly plugged and positioned beneath the shuttlebox whereby a puff of air was administered opposite the depositing end to push the bee out of the tube and into the shuttlebox. The depositing end of the tube was affixed with a perpendicular plastic strip to prevent the bee from escaping before the shuttlebox could be lowered into position (Fig. 1C).

Ambient light was minimized to reduce interference with the apparatus detection circuitry and to reduce unintended phototaxic responses. Both foragers and drones were run throughout the day from 09:00 to 16:00 h to limit the effects of circadian-influenced behavioral variations. Following random selection in groups of three bees from the feeder, two of the three bees were chosen based on their similarity in activity levels. Bees that did not move for more than 50% of the trial time were considered non-responsive and were discarded. Each experiment was run to completion before beginning the next.

A sample size of 40 honey bees (20 drones and 20 foragers) was used for the escape experiment. Half of each group was designated as master and half as yoked. Each subject was placed in a shuttlebox atop a stainless steel shock grid positioned on a powered-off computer monitor. Each subject was exposed to a 10 min trial period in which an unavoidable shock was administered by a decrementing 30 s timer. Upon reaching zero, the timer would trigger shock (DC 8.71 V, 1.0 A) on both sides of the shuttlebox, for both master and yoked subject, then restart the timer. Master subjects could terminate this indefinite shock for both the master and yoked subject by crossing to the other side of the shuttlebox. At the onset of shock, the master subject was required to break the infrared beam farthest from them to deactivate the shock. The yoked subject had no control over the shock.

A sample size of 40 honey bees (20 drones and 20 foragers) was chosen for this place preference experiment. Each subject was placed in a shuttlebox atop a stainless steel shock grid positioned on a powered-off computer monitor to maintain environmental consistency with other experiments.

Following a 2 min habituation period, each subject was exposed to a 10 min trial period where shock (DC 8.71 V, 1.0 A) was continuously administered to one half of the shock grid; shock areas were counterbalanced. Spatial positioning of the shock was static and its spatial orientation was the only discriminative cue. When placed in the shuttlebox, a bee would repeatedly shuttle from end to end. A decrease in this baseline shuttling behavior would allow the bee to avoid punishment. The subject's shuttling behavior was recorded via interruption of the infrared beams. Each 10 min trial was partitioned into 60 s bins for data analysis.

A sample size of 80 honey bees [40 drones (half yellow, half blue) and 40 foragers (half yellow, half blue)] was chosen for this place preference experiment. Each subject was placed in a shuttlebox atop a stainless steel shock grid positioned on a powered-off computer monitor. Each subject was exposed to a 10 min training period, followed by a 2 min testing period. Following a 2 min habituation period, the computer monitor was turned on and displayed two colors. One color was paired with the shock side (DC 8.71 V, 1.0 A), the other color with the no-shock side of the shuttlebox. To remain consistent with Agarwal et al. (Agarwal et al., 2011), the selected colors were Microsoft Paint default swatches: blue (R: 0, G: 0, B: 255, Hue: 160, Sat: 240, Lum: 120) and yellow (R: 255, G: 255, B: 0, Hue: 40, Sat: 240, Lum: 120). These were counterbalanced and displayed on a 17 inch Dell 1704FPTt flat panel monitor set at factory default color settings. Each 10 min trial was partitioned into 60 s bins for data analysis. To account for spatial place preference, these colors and paired shock positioning would be transposed once per 60 s bin.

Following the 10 min training period, the computer monitor and shock were turned off for a 20 s loading period. During this time, no responses were recorded. At the end of the loading period, the computer monitor was turned back on to display the same colors as during the training period; however, no shock was administered. Response recording was continued at this point in an extinction phase. During these two bins of data collection, one spatial color switch occurred after 65 s following the initial presentation.

We used observation oriented modelling (OOM) (Grice, 2011; Grice et al., 2012) to analyze our data. OOM is a data analysis technique that allows comparisons of observed results with expected patterns of outcomes for each bee and group, and the evaluation of these differences with an accuracy index and a randomization test. OOM assesses the individual subject observations and does not rely on traditional summaries of data such as measures of central tendency. By using these methods, we were able to eschew the assumptions of null hypothesis significance testing (e.g. homogeneity, normality) as well as avoid construing learning as an abstract population parameter to be estimated from our data. We have successfully used this approach in recent experiments (e.g. Craig et al., 2012).

Within OOM, we used an ordinal analysis that produces a percent correct classification (PCC) value and a chance value (a probability statistic). The PCC value is computed by comparing an a priori ordinal prediction with the observed data, and is the ratio of the observed data that matches the expected pattern compared with the number of comparisons that were made. A chance value (c-value) ranging from zero to one displays how many randomized versions of the observed data yielded higher PCCs compared with the observed data. A c-value of 0.01 indicates a 99% chance that the PCC value is not due to chance based on a range of values obtained from randomized versions of the data. In a two-order comparison, a c-value could be considered as conceptually similar to a binomial probability.

However, as c-values are calculated from randomizations of the observed data points, each PCC value's likelihood of being due to chance is assessed on an adaptable distribution that is based on observed data rather than a hypothetical distribution (e.g. the standard normal curve). Two procedures of ordinal pattern analysis were conducted to compare between and within groups to thoroughly analyze each dependent variable in each experiment.

To assess learning within subjects and trials, each respective variable was binned, and these bins were compared with every other bin for an individual bee (e.g. bin 1 versus bin 2, bin 1 versus bin 3…bin 1 versus bin 20, etc.); consequently, the number of observations that fit the ordinal pattern can range from zero to combination _nC₂ (n choose 2). For example, chunking responses into 20 bins would result in 190 bin comparisons. A PCC value of the comparisons matching the expected patterns is computed for each bee, and a randomization c-value is obtained by comparing the observed data with 100 randomizations of the observed data.

To assess learning between groups, each respective variable was binned for all subjects, and combinations of each group's individual's bins were compared against another group's individual's bins. We avoided simple mean comparisons via this method, for every response bin of every individual in a first group was compared against every response bin of every individual in a second group. Instead of relying on t-tests, which simply indicate whether differences in means were observed, we were able to predict the ordinal direction of the pairwise comparison.

Two dependent variables were assessed to investigate how quickly subjects would escape a shock occurring every 30 s. Because of dimorphism in base response rates, a direct comparison between workers and drones could not be made; thus, a relative comparison was performed. Two possible strategies of shock termination were analyzed: continuous responding (response frequency) and reactive responding (latency to respond). Response frequency and first response latencies from 20 trials were compared between each individual master subject and its yoked counterpart using the between-group pairwise method to assess responding differences between the pair. This assessment was performed under the prediction that master animals would have both higher rates of responding and lower latency to respond following the onset of shock. Tables 1 and 2 display the individual and group master–yoke comparisons for response frequency and master first response latency, respectively. Response rates and latencies to respond varied greatly across individuals. Visual representations of response rates and latency comparisons of master–yoked pairs of aggregated worker and drone groups are displayed in Figs 2 and 3, respectively.

In the ordinal analysis, workers were predicted to have shorter latencies and higher response rates than drones. Workers' response frequency ranged from PCC values of 1% (c-value: 1.00) to 98% (c-value: 0.01). Workers' latency to respond ranged from PCC values of 14% (c-value: 1.00) to 97%, (c-value: 0.01). Sixty percent of the worker master subjects had higher response rates than yoked subjects; 70% of the worker master subjects had shorter latencies to respond than yoked subjects. Individual drone response frequency was varied: PCC values ranged from 0% (c-value: 1.00) to 83% (c-value: 0.01); master response latency PCC ranged from 13% (c-value: 1.00) to 80% (c-value: 0.01). Sixty percent of the drone master subjects had higher response rates than yoked subjects, and 60% of the drone master subjects had shorter latencies to respond than yoked subjects. When assessing group response frequency and latency, worker master subjects performed better than yoked worker subjects (PCC response: 52%, c-value: 0.01; PCC latency: 63%, c-value: 0.01), as did drone master subjects (PCC response: 43%, c-value: 0.01; PCC latency: 49%, c-value: 0.01).

We compared worker and drone master subjects to assess sex differences under the prediction that workers would have higher response rates and lower latencies to respond. Consistent with our prediction, workers had both higher response rates (PCC: 76%, c-value: 0.01) and lower latencies to respond (PCC: 73%, c-value: 0.01) compared with drones. We then compared worker and drone yoked subjects to assess sex differences under the prediction that workers would have higher response rates and lower latencies to respond. Consistent with our prediction, workers had both higher response rates (PCC: 62%, c-value: 0.01) and lower latencies to respond (PCC: 60%, c-value: 0.01) compared with drones.

To facilitate a comparison in data analysis strategies, we performed a repeated-measures ANOVA with an a priori alpha level of 0.05 to assess differences in latency scores. Mauchly's test of sphericity was significant (W=0.007, χ²₁₈₉=349.759, P<0.000). A within-subject trial Greenhouse–Geisser correction (ε=0.678) was not significant (F_{12.890,979.644}=1.60, P<0.080, η²=0.021), and neither was a Huynh–Feldt correction (ε=0.858, F_{16.308,1239.416}=1.60, P<0.060). Additionally, the subject sex and bin interaction was not significant with a Greenhouse–Geisser correction (F=1.191, P<0.281, η²=0.015), nor was the interaction significant with a Huynh–Feldt correction (P<0.267). Moreover, the subject's master or yoked status and bin interaction was not significant with a Greenhouse–Geisser correction (F=1.372, P<0.167, η²=0.018), nor was the interaction significant with a Huynh–Feldt correction, P<0.146. However, a significant between-sex difference was observed (F_1,76=25.724, P=0.000, η²=0.253); drones (M=17.005, 95% CI=14.971, 19.040) had higher latencies than workers (M=9.678, 95% CI=7.643, 11.712). Moreover, a significant difference was observed between a subject's master or yoked status, (F_1,76=5.378, P=0.023, η²=0.066), even though master subjects' confidence intervals (M=11.666, 95% CI=9.632, 13.701) overlapped with those of yoked subjects (M=15.017, 95% CI=12.982, 17.051). This inconsistency may be related to a subject's master or yoked status existing as a truly dependent between-subjects comparison, for a repeated-measures ANOVA can only assume independent between-subject comparisons; our master and yoke subjects are dependent. A significant between-subject interaction between sex and master or yoked status was not observed (F=1.476, P<0.228, η²=0.019).

Additionally, we performed a repeated-measures ANOVA with an a priori alpha level of 0.05 to assess differences in the number of escape responses. Mauchly's test of sphericity was significant (W=0.004, χ²₁₈₉=390.116, P<0.000). A within-subject trial Greenhouse–Geisser correction (ε=0.551) was significant (F_{10.477,796.257}=1.839, P<0.047, η²=0.024), as was a Huynh–Feldt correction (ε=0.672, F_{12.761,969.868}=1.839, P<0.035). However, the subject sex and bin interaction was not significant with a Greenhouse–Geisser correction (F=0.822, P<0.613, η²=0.011), nor was the interaction significant with a Huynh–Feldt correction (P<0.635). Moreover, the subject's master or yoke status and bin interaction was not significant with a Greenhouse–Geisser correction (F=1.437, P<0.155, η²=0.019), nor was the interaction significant with a Huynh–Feldt correction (P<0.138). However, a significant between-sex difference was observed (F_1,76=49.157, P=0.000, η²=0.393): workers (M=3.104, 95% CI=2.734, 3.474) made more responses than drones (M=1.263, 95% CI=0.893, 1.632). A significant difference was not observed between a subject's master or yoked status (F_1,76=3.827, P=0.054, η²=0.048): master subjects (M=2.440, 95% CI=2.070, 2.810) did not make more responses than yoke subjects (M=1.926, 95% CI=1.556, 2.296). A significant between-subject interaction between sex and master or yoke status was not observed (F=1.615, P<0.208, η²=0.021).

To demonstrate the complications of continuity assumptions in null hypothesis significance testing, we also performed a series of Friedman's tests to assess within-subject changes in escape behavior across trials. Friedman's test assumes the sample is a single group; hence, we performed four assessments. Master worker subjects (χ²₁₉=30.057, P<0.051) did not significantly change response performance across trials, nor did yoked worker subjects (χ²₁₉=22.578, P<0.256), master drone subjects (χ²₁₉=19.423, P<0.430) or yoked drone subjects (χ²₁₉=26.627, P<0.114).

We measured avoidance learning by creating ten 60 s bins to develop a percentage of the subject's correct responding (i.e. time spent not being shocked) across each 60 s bin. To assess within-subject improvement across the 10 min trial, we predicted a monotonic increase in correct response length percentages across the 10 bins while separating for each subject. Group PCC values were calculated by pooling each subject within a group to allow assessment of group differences of within-subject improvement across the trials without relying on measures of central tendency. Individual and group ordinal PCC values and corresponding c-values are displayed in Table 3; subjects did not tend to monotonically improve avoidance performance across bins and no systematic differences in monotonic improvement between workers and drones were observed. A visual representation of an aggregated worker–drone comparison of correct responding is portrayed in Fig. 4.

To assess between-sex differences, we predicted that workers would have higher PCC responding values compared with drones, and observed that workers successfully avoided shock for longer periods of time than drones (PCC: 64%; c-value: 0.01). Additionally, we assessed subject performance by comparing each individual's percentage of correct responding to a chance performance of 50% correct responding, or avoiding shock for 30 s of the 60 s bin. We predicted that workers and drones would both perform better than 50% correct responding. Consistent with our between-sex analysis, drones (PCC: 52%, c-value: 0.13) performed at a lower responding PCC value compared with workers (PCC: 64%, c-value: 0.01).

To facilitate a comparison in data analysis strategies, we performed a repeated-measures ANOVA with an a priori alpha level of 0.05 to assess within-trial differences in avoidance learning. Mauchly's test of sphericity was significant (W=0.113, χ²₄₄=76.015, P<0.002). A within-subject trial Greenhouse–Geisser correction (ε=0.692) was not significant (F_{6.224,236.522}=1.997, P<0.064, η²=0.050), but a Huynh–Feldt correction (ε=0.864) was significant (F_{7.772,236.522}=1.997, P<0.048). These differences in corrections complicate interpretations of this within-subject trial assessment; however, overlapping confidence intervals surrounding mean performances of workers' and drones' first and last trials dissuades a rejection of the null hypothesis. Additionally, the subject sex and trial interaction was not significant with a Greenhouse–Geisser correction (F=0.764, P<0.604, η²=0.020), nor was the interaction significant with a Huynh–Feldt correction (P<0.631). However, significant between-sex differences were observed (F_1,38=4.393, P<0.043, η²=0.104): workers (M=37.710, 95% CI=31.765, 43.655) had higher percentages of correct responding than drones (M=29.005, 95% CI=23.060, 34.950), even though overlapping confidence intervals around the means were observed.

Using the same methods in Experiment 2, we assessed the percentage of a subject's correct responding to determine whether honey bees could discriminate between colors. To assess within-subject improvement across the trial, we predicted a monotonic increase in correct response length percentages across the 10 bins while separating for each subject. Additionally, group PCC values were calculated by pooling each subject within a group to allow assessment of group differences of within-subject improvement across the trials without relying on measures of central tendency. Drones' individual and group ordinal PCC values and corresponding c-values are displayed in Table 4 while workers' individual and group ordinal PCC values and corresponding c-values are displayed in Table 5; subjects did not tend to monotonically improve avoidance performance across bins, and no systematic group differences in monotonic improvement between colors, or in workers and drones were observed. A visual representation of an aggregated worker–drone comparison of correct responding for both yellow and blue is portrayed in Fig. 5.

Table 6 displays our between-group assessments for both training and extinction conditions. Subjects that did not respond during the extinction trial were not included in the group extinction analysis. Before assessing sex differences, we investigated if any color discrimination bias was observable in workers and drones. We first separated which color was the punished S+ for each sex before assessing general between-sex differences under the prediction that workers would have higher percentages of correct responding than drones. Based on these between-group ordinal comparisons reported in Table 4, workers with blue as the S+ performed better than each of the three other groups.

Additionally, we assessed subject performance by comparing each individual's responding PCC value to a chance performance of 50% correct responding, or avoiding shock for 30 s of the 60 s bin. We predicted that workers and drones would both perform better than 50% correct responding. Consistent with our between-sex analysis, drones (PCC: 53%, c-value: 0.20) performed at a lower percentage of correct responding compared with workers (PCC: 78%, c-value: 0.00) when shock was paired with blue. However, when shock was paired with yellow there was no observable difference between drones (PCC: 46%, c-value: 0.92) and workers (PCC: 45%, c-value: 0.95).

To facilitate a comparison in data analysis strategies, we performed a repeated-measures ANOVA with an a priori alpha level of 0.05 to assess within-bin differences and between sex and color differences. Mauchly's test of sphericity was significant (W=0.327, χ²₄₄=81.351, P<0.001). A within-subject bin Greenhouse–Geisser correction (ε=0.804) was not significant (F_{7.235,549.896}=0.850, P<0.549, η²=0.011), and a Huynh–Feldt correction (ε=0.932) was not significant (F_{8.389,637.533}=0.850, P<0.563). Additionally, the subject sex and bin interaction was not significant with a Greenhouse–Geisser correction (F=0.975, P<0.450, η²=0.013), nor with a Huynh–Feldt correction (P<0.457). The bin and color interaction was not significant with a Greenhouse–Geisser correction (F=1.318, P<0.238, η²=0.017), nor was the interaction significant with a Huynh–Feldt correction (P<0.229). The bin, subject and color interaction was also not significant with a Greenhouse–Geisser correction (F=1.245, P<0.275, η²=0.016), nor was the interaction significant with a Huynh–Feldt correction (P<0.268). However, significant between-sex differences were observed (F_1,76=27.986, P<0.000, η²=0.269): drones (M=29.145, 95% CI=27.755, 30.534) had lower percentages of correct responding than workers (M=34.363, 95% CI=32.974, 35.752). Moreover, significant between-color differences were observed (F=47.667, P<0.000, η²=0.385): the percentages of correct responding on yellow (M=28.348, 95% CI=26.959, 29.738) were lower than percentages of correct responding on blue (M=35.159, 95% CI=33.770, 36.548). A significant interaction between sex and color was observed (F=14.930, P<0.000, η²=0.164).

Additionally, we performed a repeated-measures ANOVA with an a priori alpha level of 0.05 to assess within-bin differences and between-sex and -color differences during extinction trials. Nine subjects did not respond during extinction; hence their data were not included in the extinction analysis. A significant within-subject bin effect was not observed (F_1,67=0.318, P<0.575, η²=0.005). Additionally, the subject sex and bin interaction was not significant (F=0.838, P<0.450, η²=0.001). The bin and color interaction was not significant (F=0.245, P<0.622, η²=0.004). The bin, subject and color interaction was also not significant (F=0.158, P<0.692, η²=0.002). However, significant between-sex differences were observed (F_1,67=16.559, P<0.000, η²=0.198): drones (M=23.785, 95% CI=21.238, 26.332) had lower percentages of correct responding than workers (M=31.551, 95% CI=28.718, 34.384). Significant between-color differences were not observed (F=1.453, P<0.232, η²=0.021): correct responding on yellow (M=28.818, 95% CI=26.018, 31.618) did not differ from correct responding on blue (M=26.518, 95% CI=23.935, 29.101). However, a significant interaction between sex and color was observed (F=25.927, P<0.000, η²=0.279).

In addition to utilizing OOM, we also analyzed our data via null hypothesis significance testing (NHST) to demonstrate four main differences in data analysis methods and results between OOM and NHST. First, assumptions of continuity are not made in OOM; discriminating between ordinal data or ratio data and determining whether a parametric or non-parametric assessment should be performed is not paramount, for both forms of data are assessed in the same manner in OOM and the ordinal analysis. To exemplify the difficulties of continuity concerns in NHST, we performed a series of Friedman's tests and a repeated-measures ANOVA and argue that responding can be conceived as ordinal data or ratio data, for a response count per trial can be argued to be ordinal and ratio data. Each trial in Experiment 1 lasted for 30 s; dividing each trial's ordinal response frequency by a constant 30 s to create a trial response rate produces truly continuous data. Even though NHST conclusions are not affected by constant multiplicative scale changes at a trial level, the same data can undergo two radically different assessment procedures and could result in different rejection decisions of the null hypothesis. Indeed, while using the same response data points from Experiment 1, a repeated-measures ANOVA reported significant within-subject differences in response rate while a Friedman's test failed to report significant within-subject differences in response numbers. Second, NHST assumptions about sphericity or homogeneity are eschewed in OOM because means, variances and sums of squares are not compared. As such, concerns about corrections and the observed inconsistent results from different corrections are avoided in OOM (e.g. in Experiment 2, a Greenhouse–Geisser correction was not significant but a Huynh–Feldt correction was significant). Third, OOM tests do not utilize critical alpha levels; hence, there are no concerns about alpha-level adjustments following numerous tests. In focusing on the individual observations of the collected data, generalizations to population parameters are not made; rather, uniqueness of the specific observations of the data are assessed. Fourth, testing between-group dependency or between-group independency does not involve inherently different methods. In Experiment 1, latency data were assessed within subjects (trial) in a repeated-measures design, and between subjects independently (sex) and dependently (master status). The assumptions and requirements to properly run a between-subjects dependent t-test while controlling for within-subject trials in a single assessment are not met in NHST.

Several clear advantages of OOM are identifiable when compared with comparisons of measures of central tendency. Concerns of unrepresentative aggregates because of outlier effects or dichotomous trends in individual performances are irrelevant in OOM. Adjusting critical alpha levels after performing multiple tests is unneeded in OOM. Complications with missing trial data do not result in all of the subject's data remaining unassessed in OOM. Abstract, often impossible population parameters are not compared as if they are concrete individual observations in OOM. Instead of providing a probability value of a dataset's extremity based on pre-determined alpha levels, OOM provides a chance value of the observed dataset's uniqueness compared with a series of randomizations of the dataset. Finally, the PCC value indicates the percentage of data points in a group that are larger or smaller than an alternative group; we believe the information in a PCC value offers an easily comprehensible summary of the dataset compared with the required hodgepodge of tests required to thoroughly assess a between-subject repeated-measures design.

The neural plasticity and complex social interactions within a colony make honey bees an advantageous model. Female workers constantly interact and transition from tasks inside the nest when they are young. In contrast, a drone's life history primarily constitutes finding and memorizing specific congregation sites in which mating will occur with no involvement in other colony-associated tasks. This system provides an ideal opportunity for comparative examination of learning. Further analyses of learning across the more solitary drones and the more social worker bees sets the stage for later phylogenetic studies of learning in solitary and social bees (Fischman et al., 2011; Woodard et al., 2011). Here, we present such an analysis using two forms of learning: escape and punishment.

The escape experiment consisted of two measurements recorded in a master–yoked paradigm: frequency to respond and latency to respond following administration of shock. For both workers and drones and for both measurements, there was large individual response variation. Worker individual master–yoked comparisons of PCC values ranged from 1 to 98% for response frequency and 14 to 97% for latency to respond. Drone individual master–yoked comparisons of PCC values ranged from 0 to 83% for response frequency and 13 to 80% for latency to respond. Despite the range of variation between individuals for workers and drones, master subjects as a whole for both sexes performed better under both measurements. When comparing workers with drones, workers performed better under both measurements.

Escape behavior is a basic and ecologically significant strategy for avoiding predation. As such, our results match expectations that both drones and workers are capable of escape learning. Foraging workers experience increased mortality rates as they age. Dukas (Dukas, 2008) argues both age-dependent and independent mortality rates are due primarily to predation and posits that learning contributes to some decrease in the observed age-specific mortality rate. Additionally, male honey bees experience similar increases in mortality rates once flight is initiated, which is, at least in part, attributed to predation (Rueppell et al., 2005); one should expect little to no difference it their ability to learn under this paradigm, as our results demonstrate. Disparity in the response of workers and drones was observed and does not seem to be constrained to learning. In one set of studies, aggressive response of workers and drones were analyzed, and results showed that worker reaction is more threshold dependent compared with drone reaction when under exposure to the same aversive stimuli (A.A. and T.G., unpublished). Together, these observations might indicate a difference in threshold for reaction to negative stimuli, which prompts further examination.

In the punishment experiment, there was no indication of monotonic increase in correct response patterns across bins for either drones or workers. A between-sex comparison revealed that workers performed better than drones and effectively avoided shock for longer periods of time. In addition to the improvement measure and the between-sex comparison, a flat ‘to chance’ comparison was performed comparing correct responses per bin with the 50% chance response value. Workers performed better than chance and outperformed drones in this measure. Drones did not perform significantly better than chance.

A lack of monotonic increase in response pattern across bins for both drones and workers was also observed in the punishment with discrimination experiment. For this discrimination experiment, the color of the discriminative stimulus played a pivotal role on the avoidance potential of the bee. As with the punishment experiment, correct response values per bin were compared with a chance correct response value of 50%. When blue was paired with shock, subjects performed better than subjects whose shock was paired with yellow. Between-sex comparisons revealed workers performed better than drones when shock was paired with blue; however, there was no observable difference between workers and drones when the shock was paired with yellow. Both workers and drones performed worse than chance when shock was paired with yellow and performed better than chance when shock was paired with blue. Both sexes displayed a preference for spending time on the yellow side; this bias inhibits a clear conclusion on learning ability and demonstrates the importance of counterbalancing discriminative stimuli.

Negative associations have been shown to be an important component with regards to resource patch visitation and signaling by honey bee workers (McNally and Westbrook, 2006; Abramson et al., 2006; Nieh, 2010). Similarities can be drawn between the importance of resource patches for workers and mating congregation sites for drones. As such, we expected associative (punishment) learning ability to be similar, but our results did not support this expectation. An explanation for workers outperforming drones in this type of learning might stem from our methods. The described complications might also account for the differences in response demonstrated in Experiment 1: escape.

A point to consider in interpreting our drone data is the suitability of the shuttlebox. In contrast to the smooth performance of the worker bees, the movement of the drones often appeared ‘clumsy’. To keep the experimental design as similar as possible between the worker and drones, we opted to use the same dimensions. Our pilot work showed both the workers and drones could turn around in the apparatus; hence, we attempted to address inherent locomotion differences between drones and workers. While they could easily turn around in the shuttlebox, they would often run to the end of the compartment before turning in the opposite direction. We believe this behavior may have been caused, in part, by their momentum and is reminiscent of what the early comparative psychologist Schneirla called ‘centrifugal swing’ in his discussion of the errors made by ants in a complex maze (Abramson, 1997).

Our experiment represents the first time that a drone has been tested in a shuttlebox; however, the development of a suitable apparatus to test the learning of drones in non-appetitive situations is still required [for a review of apparatus used for the study of invertebrate learning, see Abramson (Abramson, 1994)]. The popular PER in honey bee research has been shown to have methodological inconsistencies across laboratories (Abramson et al., 2011; Matsumoto et al., 2012; Frost et al., 2012); these inconsistencies have led to the renewed interest in developing aversive conditioning shuttlebox situations for honey bees (Agarwal et al., 2011).

In the punishment with discrimination experiment, color significantly affected response behavior. Foraging workers experience and use color as a significant environmental cue. Past experience by the examined individuals to specific color associations might explain the observed bias. Constancy towards an experienced color has been described (Hill et al., 1997). Furthermore, subspecies of bees have been shown to vary in their experience-dependent preference and constancy towards flower/resource coloration (Cakmak et al., 2010). Previous analysis of punishment learning with discrimination on a gentle Africanized hybrid (gAHB) found in Puerto Rico did not find differences in color preferences (Agarwal et al., 2011). We therefore conclude that in future experiments that use color as a discriminatory cue, care must be taken in the selection process.

We thank Ian Finley and Corey Vyhlidal of Oklahoma State University for their assistance with circuit design and construction.