Food insecurity increases energetic efficiency, not food consumption: an exploratory study in European starlings

Subjects and basic husbandry

Subjects were 12 adult European starlings (Sturnus vulgaris) originally sourced from nest boxes in Northumberland, UK as chicks and hand-reared in the lab. The numbers, sexes and ages of birds used in each experiment are given in Table S1. The sample size of 6 birds used in each experiment was chosen based on a previous comparable experiment with starlings that reported an effect of unpredictable food on body mass (Cuthill et al., 2000). All of the birds had previous experience in a range of behavioural experiments. During the current experiments, the birds were group-housed in indoor aviaries. In experiments 1–3 the aviaries were rooms (215 × 340 × 220 cm) and in experiment 4, large walk-in cages (90 × 183 × 183 cm) with one cage per room. In all experiments, room temperature and humidity were maintained at ∼18 °C and ∼40% respectively. Each aviary was furnished with rope perches, wood chippings on the floor, a water bath, a drinker providing ad libitum clean water supplemented with vitamins and either one or two feeding stations (see below). The light cycle, number of birds per aviary and SFS stations per aviary for each experiment are given in Table S1. In all experiments, birds were maintained on a nutritionally complete diet of commercial poultry starter crumb (Special Diets Services Poultry Starter; henceforth ‘food’). Each bird was supplemented with 4 live mealworms daily. Daily husbandry took place at 1700 in experiments 1–3 and at 1100 in experiment 4.

Social foraging system

In all experiments, food was delivered from custom-built feeding stations (the social foraging system, SFS; Fig. 2). An SFS station comprised a retractable food hopper and an illuminable pecking key, both of which could only be accessed via a perch designed to accommodate a single bird (the pecking key was only used in experiment 4). The perch was mounted on a load cell that functioned as an electronic balance that measured to a resolution of 0.01 g. Beneath the perch there was an aerial tuned to read microchips (radio-frequency identification devices, RFIDs), that were mounted on leg rings of the starlings. Each aviary contained either one or two SFS stations depending on the experiment (Table S1). The SFS stations were connected to a computer in an adjacent room running Whisker experimental control software and additional custom-written software. This computer collected continuous data on the identity and masses of birds visiting each station and, in experiment 4, also controlled the operant schedule in place on each station.

Manipulations of food insecurity

In experiments 1–3 food security was created by maintaining the SFS hopper in the raised position, and thus available for foraging from dawn until dusk. Food insecurity was induced by withdrawing the hopper for 12 out of a possible 20 randomly chosen 20-min periods starting two hours after lights-on each day, and ending 1 h and 20 min prior to lights-off.

In experiment 4 the default position of the food hopper was lowered such that food was unavailable. Food security was created via a ratio schedule, whereby a single peck at the illuminated key on the SFS caused the key light to extinguish and the hopper to raise allowing 10-s access to food. At the end of the reinforcement, the hopper lowered and there was a 2-s inter-trial interval before the key re-illuminated signalling the start of the next available trial. The operant sessions began at 15 min after lights-on each day (0615) and ended 15 min before lights-off (2045). This schedule was designed to mimic a starling foraging by probing for hidden soil invertebrates. Food insecurity was induced by reducing the probability that a key peck was reinforced on any given trial first to 0.4 under FI_low, and second to 0.2 under FI_high. Reducing the probability of reinforcement below one increased the variance in the inter-reinforcement interval. On unreinforced trials, a peck caused the key light to extinguish and there was a 10-s time out delay during which the hopper remained lowered followed by a 2-s inter-trial interval. Under food security, the maximum number of reinforcements available per day from the SFS was 4350 (assuming that birds pecked as soon as they key illuminated). Under food insecurity, the expected maximum number of reinforcements available per day dropped to 1740 and then 870 under FI_low and FI_high respectively.

Procedure

All experiments began with birds being caught from their home aviary, manually weighed and equipped with two plastic leg rings each of which had a unique RFID microchip attached (birds wore two chips to guard against data loss in the event that one chip fell off, broke or was not read due to poor alignment with the SFS aerial). Birds were then released into experimental aviaries. Food was initially provided ad libitum by raising the hoppers of the SFS stations. Birds were initially encouraged to visit the SFS by placing mealworms around the stations and in the food hopper, but as soon as they were visiting the SFS readily these extra mealworms were withdrawn. The first experimental treatment of experiments 1–3 began once the birds were maintaining stable body masses on food obtained from the SFS.

Experiment 4 required additional training for the birds to learn that food could only be accessed by pecking the illuminated key on the SFS. We used an auto-shaping procedure whereby illumination of the pecking key signalled subsequent unconditional raising of the food hopper. Birds first learned a Pavlovian association between the lit key and food and spontaneously started performing appetitive pecks to the lit key. Pecks to the lit key were reinforced by immediate hopper raising, thus creating an instrumental association between key pecking and food. As soon as key pecking was established, hopper raising was made conditional on key pecking and the birds were moved to continuous foraging, whereby they earned all of their daily food by pecking at the lit key on the SFS. The first experimental treatment of experiment 4 began once the birds were maintaining stable body masses on food earned from the SFS.

The sequence and duration of the treatments experienced in each experiment is shown in Fig. 1. In all experiments, birds were maintained in closed economy, obtaining the majority of their daily food intake from the SFS. The diet was supplemented with four mealworms per bird given during daily husbandry and supplied in two spatially separated bowls to prevent one bird from monopolising the worms. The experiments ran seven days a week with no gap between treatments. The birds’ welfare was monitored daily via visual checks and using the body mass data obtained from the SFS balances; criteria for consulting a vet and considering euthanasia were either presence of puffed feathers and lethargy and/or a mass of <62 g. No bird became unwell or dangerously thin at any point during the study and no birds died or were euthanized. At the end of the final treatment of each experiment the birds were re-caught, manually weighed and returned to their home aviaries where they were retained for further studies. Experiments 1–3 were conducted consecutively on the same group of 6 male birds with a 24-day break in the home aviary between experiments 1 and 2 and an 18-day break between experiments 2 and 3.

Body mass

In all experiments, body masses were recorded by the SFS each day between lights-on and lights-off; each mass was recorded with a bird identity corresponding to the microchip of the bird on the perch. The balances measured masses at a frequency of 6 Hz. A stable mass was recorded for a bird if the balance measured five consecutive masses of >50 g that were within a range of 5 g. These criteria were chosen to eliminate masses from birds that were perching incorrectly (e.g., by placing one foot on the food hopper), but to maximise the number of stable masses recorded from moving birds. Once a stable mass had been recorded another stable mass could not be recorded until the balance had measured a mass <10 g indicating that the current bird had left the perch. Balances were checked with a 100-g test mass a minimum of twice daily and calibrated if necessary. In order to control for build-up of guano on the perch over the day, balances were automatically tared regularly throughout the day when no bird was present on the perch.

The SFSs recorded a mean of 4.64 stable masses per bird per daylight hour over the four experiments. The raw masses showed a clear trend of mass increase over the day as expected (Fig. S1), but there was a lot of random error due to the imprecision of the balances and movement of the birds whilst on the perch, and masses were not always available at all times of every day. To estimate comparable dawn and dusk masses for each bird on each day we used the following procedure to model the available data. As long as a minimum of 10 masses were available, the mass data from each bird on each day were fitted with a cubic polynomial (Fig. S1A). To remove biologically impossible outliers, any masses >3 g from the fitted line were removed and a new cubic polynomial was refitted to the remaining data. This latter fit was used to estimate dawn and dusk masses for that day. To avoid extrapolation beyond the data, a dawn or dusk mass was only estimated if there was a data point within 1 h of the estimate. The above procedure was devised based on a detailed exploration of the data from experiment 4 and then applied unaltered in experiments 1–3. The one exception being that due to differences in daylight hours between experiments, dawn mass was defined as the fitted mass at 0900 for experiments 1–3 and 0615 for experiment 4; dusk mass was defined as the fitted mass at 1800 for experiments 1–3 and 1815 for experiment 4 (Fig. S1B). Thus, the dusk masses for experiments 1–3 and experiment 4 were estimated 9 and 12 h after lights-on respectively. A later time was not chosen for the dusk masses in experiment 4 due to the fact that birds often stopped foraging considerably before lights-off.

Operant foraging behaviour

In experiment 4, key pecks were recorded by the SFS between 0615 and 2045 each day. The time of each key peck was recorded with a bird identity and whether or not the peck was reinforced with food (probability of 1.0 under food security and either 0.4 or 0.2 under food insecurity for FI_low and FI_high respectively).

Food consumption

Total food consumption in each aviary was estimated daily in all experiments by calculating the difference in the mass of the SFS food hoppers at the beginning and end of the day and subtracting any food collected in a spill tray located beneath each hopper. Hence, consumption data were only available at the aviary level.

Energy density of guano

Energy density of guano was measured by bomb calorimetry in experiments 1 and 3. Guano samples were collected daily from plastic trays positioned beneath perches in each aviary avoiding any feathers or wood chippings. Hence, guano data were only available at the aviary level. Samples were immediately frozen at −80 °C for storage. On completion of the experiment, samples were dried in a flow oven at 55 °C for 48 h until stable masses were obtained and finely ground to homogenise. In experiment 1, samples from each aviary were pooled over three consecutive days (yielding pooled samples centred on days 2 and 6 of FS1 and FS2 and days 2, 6, 9 and 13 of FI), whereas in experiment 3 all days were measured separately. For both experiments, each pooled sample was divided into two technical replicates prior to analysis. Bomb calorimetry was outsourced to Pemberton Analytical Services, Shropshire, UK and was conducted blind to the treatment group and technical replicates using a PAR 1261 adiabatic bomb calorimeter with a stated precision of ±0.1% on two determinations.

Physical inactivity

Data on physical inactivity were collected in experiments 1–3. As a metric of daytime physical inactivity we measured roosting behaviour—defined as perching motionless on a high rope perch. Video recordings of the aviaries were made for 15 min starting at 0900 and 1000 every day, before any food insecurity started, using a remotely operated wide-angle surveillance camera mounted in the ceiling of each aviary. The videos were scored manually using BORIS video coding software (Friard & Gamba, 2016). Scoring of roosting was blind to the experimental treatment in place on the previous day (the relevant predictor variable). A scan sampling method was used, whereby the instantaneous behaviour of each bird in the aviary was scored every 30 s for each 15-minute video. Birds were not individually identifiable on the videos, therefore the behavioural metric was the proportion of birds in the aviary observed roosting on each scan.

Statistical analysis

Data were analysed using R version 3.5.1 (R Core Team, 2018). Below we provide an overview of our statistical approach. Further details of statistical models are given in the Tables S2–S9.

For analysis of the effects of food insecurity in individual experiments we used linear mixed models (GLMMs) fitted using the R package ‘lme4’ (Bates et al., 2015). The models contained random effects (intercepts) to account for sources of non-independence in the datasets. In all models we analysed the effect of treatment: FS1 vs FI vs FS2 in experiments 1 and 3; FS1 vs FI in experiment 2; and FS1 vs FI_low vs FI_high in experiment 4. In cases where a treatment continued for two weeks, the data were pooled from the first and second weeks of a treatment for statistical analysis, but the graphs in Fig. 3 show the data broken down by week. Since roosting data were proportions, they were arcsine square root transformed prior to analysis. Following any necessary transformation, all models gave satisfactory distribution of residuals, hence a Gaussian error structure was assumed throughout. We used restricted maximum likelihood estimation (REML) and conducted overall significance tests of treatment in the GLMMs using Satterthwaite’s method.

For meta-analysis we used multi-level random effects meta-analysis models fitted using the R package ‘metafor’ (Viechtbauer, 2010). The models contained random effects of aviary nested in experiment. The effect sizes used in the meta-analyses were obtained from GLMMs fitted to the data from each aviary, using the same models for the analysis of individual experiments described above (with the exception that random effects of aviary were no longer required). Effects of food-insecurity were the parameter estimates (β values) and associated standard errors corresponding to: the difference between FS1 and FI for experiments 1–3 and the difference between FS1 and FI_high for experiment 4. Tests of modifiers were conducted using meta-regression, whereby both baseline mass and competition (number of birds per SFS station) were added to the meta-analytic models. Estimation was by REML in all meta-analytic models. We used aviary-level data as the unit of analysis for the meta-analyses, because in the experiments involving multiple aviaries (1 and 4) there was considerable heterogeneity in results obtained from the different aviaries. Furthermore, increasing the number of replicates available for the meta-analysis increased the power available for the meta-regressions, making it possible to test for effects of baseline mass and competition.

Ethical statement

The study adhered to ASAB/ABS guidelines for the use of animals in research. Birds were taken from the wild under Natural England permit 20121066 and the research was completed under UK Home Office licence PPL 70/8089 with approval of the Animal Welfare and Ethical Review Body at Newcastle University. Prior to each individual experiment, a detailed protocol was approved by Newcastle University’s Comparative Biology Centre, but these protocols were not publicly pre-registered.

Description of the dataset

Data were collected on 12 starlings comprising 6 males used in experiments 1–3 and 6 females used in experiment 4. Data were analysed at the individual bird level for body mass and key pecking, and at the aviary level for food consumption, energetic efficiency, guano energy density and roosting behaviour. In total, the experimental treatments ran for a total of 91 days in 8 separate aviaries across the four experiments. Data were collected on 210 aviary days and 546 bird days. On a few days, data were lost for one or more outcome variables due to experimenter error, equipment failure or insufficient mass data for the fitting algorithm (see Methods). Details of the unit of analysis and final sample size for each analysis are provided in Table S2–S8.

Body mass

Two measures of body mass were estimated for each bird on each day of each experiment: dawn mass, which is the mass before any foraging commenced and reflects changes in body composition including fat reserves, and dusk mass, which is the mass after a day of foraging and thus additionally includes gut content. To investigate the effects of food insecurity in each experiment, mixed-effects general linear models (GLMMs) were used to compare body mass on all days during the initial period of food security (FS1) with body mass on all days during the subsequent period of food insecurity (FI in exp. 1–3 and FI_high in exp. 4; Figs. 3A–3H and Fig. S2; Tables S2 and S3). To summarise the effects from the four experiments combined, we used random effects (RE) meta-analysis. Overall, dawn mass increased by 0.52 g under FI (95% CI [−0.12 to 1.17]), but this change was not significant (RE meta-analysis: z = 1.59, p = 0.113; Fig. 4A; Table S4). Dusk mass increased by 1.51 g under FI (95% CI [0.78–2.25]; RE meta-analysis: z = 4.05, p < 0.001; Fig. 4B; Table S4). The increase in dusk mass corresponds to ∼3% of the mean dusk mass during FS1.

The meta-analysis showed heterogeneity among the effects of food insecurity on dawn and dusk mass across the four experiments (tests for heterogeneity: τ² ≥ 0.95, Q7 ≥ 29.79, p < 0.001; Table S4), suggesting variation in how birds responded to FI. Two possible moderators of the effect of food insecurity on mass gain were explored: the baseline mass of the birds during FS1 (mean dawn mass) and the degree of competition in the aviary (low, where 2 birds shared one SFS station, or high, where 6 birds shared two SFS stations; see Fig. 1). Addition of these two moderators to the meta-analytic model explained ∼82% of the heterogeneity in the effects of FI on dawn mass (omnibus test of meta-regression: Q_M2 = 18.52, p < 0.001) and the test for residual heterogeneity was not significant (Q_E5 = 5.56, p = 0.352). The FI-induced increase in dawn mass was greater for aviaries with lower baseline body mass (β_{baseline mass} = −0.45 95% CI [−0.67 to −0.23], z = −4.06, p < 0.001; Fig. 5) and for aviaries with higher competition for access to food (β_{high competition} = 1.76 95% CI [0.55 to 2.96], z = 2.86, p = 0.004; Fig. 5). The results for dusk mass were similar. The moderators explained ∼76% of the heterogeneity in the effects on dusk mass (omnibus test of meta-regression: Q_M2 = 16.34, p < 0.001). The effect of FI on dusk mass gain was greater for aviaries with lower baseline body mass (β_{baseline mass} = −0.50 95% CI [−0.67 to −0.23], z = −2.90, p = 0.004) and in aviaries with higher competition (β_{high competition} = 3.60 95% CI [1.76 to 5.44], z = 3.84, p = 0.001). However, the test for residual heterogeneity was still significant in this case (Q_E5 = 12.63, p = 0.027). Note that the moderating effect of baseline mass is not an artefact of measurement error and regression to the mean, because the mass estimates are based on multiple measurements and are therefore highly precise.

To investigate the effect of reinstating food security in experiments 1 and 3, dawn and dusk mass in FS2 were compared with those in FS1 and FI. Reinstating food security had no effect on dawn mass in either experiment (Figs. 3A, 3C; Table S2), but dusk mass declined in both experiments relative to FI back to a level not significantly different from that in FS1 (Figs. 3E, 3G; Table S3).

Foraging behaviour

Key pecking data from the operant schedules in experiment 4 were used to investigate how birds responded to the reduced probabilities of reinforcement in FI_low and FI_high. Birds increased their frequency of key pecking during both periods of FI compared to FS1 (GLMM: F_2,118 = 87.16, p < 0.001; Fig. 6A, Table S5). The birds also exploited each 10-s reinforcement more effectively, consuming more food per second in FI (GLMM: F_2,55 = 23.07, p < 0.001; Fig. 6B, Table S5). These responses were graded, with birds pecking and eating faster under FI_high than under FI_low (Table S5). However, this increased foraging effort was insufficient to compensate for the reduced probabilities of reinforcement in FI. The birds earned fewer reinforcements per day (GLMM: F_2,118 = 32.95, p < 0.001; Table S5) and overall ate less food per day in FI (GLMM: F_2,54 = 8.77, p < 0.001; Fig. 3L; Table S6).

Food consumption

The mean consumption per bird each day was computed (by dividing the measured consumption in an aviary by the number of birds present) and compared on all days during FS1 with all days during FI for each experiment (Figs. 3I–3L; Table S6). Overall, food consumption dropped by 2.50 g bird⁻¹ day⁻¹ (95% CI: [−3.13 to −1.87]) when birds moved from FS1 to FI (RE meta-analysis: z = −3.08, p = 0.002; Fig. 4C; Table S4). This corresponds to a ∼13% drop in daily consumption compared to the initial period of food security. There was heterogeneity among the effects of FI on food consumption (tests for heterogeneity among associations: τ² = 3.20, Q7 = 61.76, p < 0.001; Table S3). Addition of baseline body mass and competition to the meta-analytic model as moderators did not explain a significant percentage of this heterogeneity (omnibus test of meta-regression: Q_M2 = 5.76, p = 0.056), although given the marginal nature of this result there is potentially some evidence for an effect here.

Reinstating food security had no effect on consumption in experiment 1, but caused an immediate increase in consumption in experiment 3 (Fig. 3K; Table S6). However, consumption during FS2 remained lower than during FS1 in both experiments (Figs. 3I, 3K; Table S6).

Energetic efficiency

We calculated energetic efficiency as the body mass maintained per mass of food consumed, using dawn mass and the amount of food consumed on the previous day. We compared all days during FS1 against all days during FI for all 4 experiments (Figs. 3M–3P, Table S7). Overall, energetic efficiency increased by 0.85 (95% CI [0.65 to 1.06]) when birds moved from FS1 to FI (RE meta-analysis: z = 8.17, p < 0.001; Fig. 4D, Table S4). This corresponds to a ∼18% increase in energetic efficiency compared to FS1. The effect of FI on energetic efficiency (1.15 sd) was larger than that on dawn or dusk mass alone (0.17 and 0.42 sd respectively). Furthermore, the effect of FI on energetic efficiency was less heterogeneous between experiments and aviaries than the effects on body mass, although the heterogeneity was still significant (τ² = 0.20, Q7 = 29.35, p < 0.001). Addition of baseline body mass and competition to the meta-analytic model as moderators did not explain a significant percentage of the heterogeneity (omnibus test of meta-regression: Q_M2 = 5.26, p = 0.072).

Reinstating food security had no effect on efficiency in experiment 1, but caused an immediate decrease in efficiency in experiment 3 (Figs. 3M, 3O; Table S7). However, efficiency during FS2 remained higher than FS1 in both experiments (Figs. 3M, 3O; Table S7).

Energy density of guano

In experiments 1 and 3 bomb calorimetry was used to measure the effect of FI on the energy density of homogenised pooled samples of guano that were collected to be representative of all birds in an aviary. Energy density was compared on all days during FS1 with all days during FI for each experiment (Fig. 7, Table S8). Overall, energy density dropped by 0.12 MJ. kg⁻¹ (95% CI: [−0.22 to −0.01]) when birds moved from food security to food insecurity (RE meta-analysis: z = −2.18, p = 0.030; heterogeneity: τ² = 0, Q₃ = 1.50, p = 0.681; Fig. 7C). This corresponds to a ∼1% drop in energy density compared to FS1. In experiment 1, in which food insecurity lasted for 14 days, the energy density of guano decreased during FI relative to FS1 (βFI = −0.17 95% CI [−0.33–−0.01], t-test, t₁₉ = −2.12, p = 0.047), whereas in experiment 3, in which FI lasted for only 7 days, there was no difference between FI and FS1 (βFI = −0.07 95% CI [−0.22 to 0.09], t-test, t₅₃ = −0.84, p = 0.403).

Reinstating food security had no effect on energy density of guano in experiment 1 (βFS2 = −0.06 95% CI [−0.22 to −0.10], t-test, t₁₉ = −0.77, p = 0.452), whereas it was associated with a decrease in energy density in experiment 3 (βFS2 = −0.17 95% CI [−0.31 to −0.03], t-test, t₅₃ = −2.45, p = 0.018). In both experiments, energy density was lower during FS2 than FS1 (exp. 1: βFS2 = −0.23 95% CI [−0.41 to −0.05], t-test, t₁₉ = −2.51, p = 0.022; exp 3: βFS2 = −0.24 95% CI [−0.37 to −0.10], t-test, t₅₃ = −3.42, p = 0.001).

Roosting behaviour

In experiments 1–3 the proportion birds that were observed roosting in instantaneous scans of the aviary made every 30 s was scored from video as a metric of physical inactivity. Scans were made during the period between 0900 and 1100 when ad libitum food was always available regardless of the treatment in place that day. We modelled the effects of the treatment experienced the previous day on roosting (Figs. 8A–8C; Table S9). Overall, the proportion of birds observed roosting did not change significantly under FI (RE meta-analysis on arcsine square root-transformed proportions: estimate = 0.03 95% CI [−0.14 to 0.08], z = −0.46, p = 0.645; heterogeneity: τ² = 0.04, Q₄ = 33.59, p < 0.001; Fig. 8D). In experiment 3 roosting increased significantly under FI (βFI = 0.17, 95% CI [0.08 to 0.27]; t-test, t₂₄ = 3.47, p = 0.002), but in aviary 1 of experiment 1 the effect was in the opposite direction.

Reinstating food security was associated with an increase in roosting in experiment 1 (βFS2 = 0.27 95% CI [0.16–0.38], t-test, t₇₆ = 4.87, p < 0.001), but no change in roosting in experiment 3 (βFS2 = −0.05, 95% CI −0.13 to 0.03; t-test, t₂₃ = −1.19, p = 0.245). In both experiments, roosting was more common during FS2 than FS1 (exp. 1: βFS2 = 0.21 95% CI [0.09–0.34], t₇₆ = 3.39, p = 0.001; exp 3.: βFS2 = 0.12 95% CI [0.04–0.20], t-test, t₂₃ = 2.79, p = 0.011).

Limitations and strengths

Each of the experiments in this study used only six birds. Small sample sizes simultaneously increase the probability of false negative results and the probability that any positive results obtained are false positives (Bateson & Martin, 2021). The probability of obtaining false negative results will have been further exacerbated by our decision to use all days of data from each treatment in the analyses presented. Since birds adjusted to changes in treatment over a period of days (see Fig. S2 for an example from the body mass data), this was a conservative decision, resulting in a likely underestimation of treatment effects that will have been greatest when treatments lasted for only 7 days. While this decision is likely to have contributed to false negative results in some individual experiments, it was made to avoid the complexity and added researcher degrees of freedom that would have been introduced by identifying and implementing stability criteria for each of the output variables within each treatment.

To mitigate against false negatives, we used longitudinal within-subjects designs that increased our power to detect treatment effects. We measured individual body mass in unprecedented detail, collecting a mean of 4.64 stable masses per bird per daylight hour across the four experiments. These data permitted modelling of individual within-day trajectories in body mass, reducing the impact of measurement error and increasing confidence in mass estimates. Finally, our headline conclusions are based on summary estimates of effect sizes obtained from meta-analysis of the results obtained from multiple experiments (some of which were individually inconclusive). Since our conclusions are positive findings, concerns about false negatives are irrelevant.

Could our headline results be false positives? The probability that a positive finding is a true positive is known as the positive predictive value (PPV). All else being equal small studies have lower PPV, but PPV is increased by the probability that a given finding was true before the study was conducted (Button et al., 2013). Given the prior evidence that unpredictable food should and does increase body mass in starlings and other species, the result that food-insecurity increased dusk mass is likely to be true. In contrast, the result that food insecurity reduced food consumption and hence increased energetic efficiency was not predicted beforehand. Therefore, despite our significant meta-analytic findings, these results have lower PPV and should be treated with greater scepticism. However, there are reasons for confidence in these findings, as we explore below.

The experiments differed in a number of methodological details including the age, sex and previous experience of the birds, the number of birds in an aviary, the competition for feeder access, the method used to induce food insecurity and the number of weeks for which each treatment was maintained. This heterogeneity undoubtedly contributed to variability in the results and might be seen as a limitation of the dataset. However, there is a trade-off between standardisation to reduce variability and heterogenisation to increase the generality of findings. Over-standardisation of protocols for animal experiments has been highlighted as a likely cause of replication failure (Voelkl & Würbel, 2016). Triangulation of conclusions from multiple different experiments, as we have done in the current study, is recognised as a powerful strategy for reducing false-positive findings due either to chance, or caused by a specific confounding variable or moderator present in a single experiment (Munafò & Smith, 2018).

Due to constraints of our methodology, food consumption (and hence energetic efficiency), energy density of guano and roosting behaviour were only available at the aviary level. This is undoubtedly a limitation, because aviary-level effects could have been due to effects in only a subset of the birds present. This concern is somewhat mitigated by the fact that the effects of food insecurity on food consumption, energetic efficiency and energy density of guano all emerged from meta-analytic summaries of multiple experiments.

Obtaining accurate individual data for variables such as food consumption is currently only possible in individually housed birds. Due to space constraints, this usually implies housing birds in smaller cages that restrict natural behaviour (and compromise animal welfare). Our aim in this study was to create a refined and more ecologically valid model of body mass regulation in which birds could interact socially, walk, fly, water bathe and forage more naturally. Many of these behaviours are either impossible or severely limited in individual cages, meaning that caged birds have less latitude to modify their energy expenditure via changes in physical activity. Therefore, there is a trade-off between the level at which some variables can be measured and the ecological validity of experiments that has implications for the types of effects that are likely to be detected. Given the established importance of the social and physical environment for body mass (Miles et al., 2009), we believe that a move towards more naturalistic experimental paradigms in which opportunities for social interaction and physical exercise are present is critical, even if this comes at the cost of limiting individual-level measurement of some variables.

A final limitation is that our analyses of guano energy density were based on representative guano samples collected from the aviaries, rather than on the total amount of guano produced. In order to calculate the proportion of ingested energy that is lost in guano it would be necessary to know the relative amount of guano produced in the different treatments. Due to the size of the aviaries and the use of wood chips on the floor, accurately collecting total guano production was simply not feasible.

Implications for human obesity

Our findings highlight the dangers of extrapolating from short-term laboratory studies showing increased food consumption in food-insecure humans (Nettle et al., 2018; Stinson et al., 2018) to the assumption that energy intake is increased by food insecurity in the home environment. The results from experiment 4 show that it is possible for food-insecure birds to show increased foraging motivation, yet still consume less food overall in a food insecure environment. The lack of evidence for increased energy intake in food-insecure humans from analysis of 24-hour food recall data has been dismissed as under-reporting of food intake by the obese (Stinson et al., 2018). However, an alternative hypothesis suggested by our results is that mass gain in food-insecure humans is explained by reduced energy expenditure or increased assimilation of energy from food, rather than increased intake.

Only women show increased body mass associated with food insecurity (Nettle, Andrews & Bateson, 2017), meaning that much research on human food-insecurity has been restricted to women. However, our results raise the question of whether increased energetic efficiency has been missed in men. A longitudinal study showing that girls responded to becoming food insecure by gaining body mass, despite no change in energy intake, also reported that boys responded by maintaining their body mass, despite reducing their energy intake (Jansen et al., 2017). These results support a common effect of food insecurity on energetic efficiency in human children of both sexes. Further work is needed to establish whether this result is also found in food-insecure adults.

As a final note, we have encountered some scepticism over whether passerine birds are appropriate animal models for understanding human strategic mass regulation (e.g., Speakman, 2018). Critics focus on two related arguments concerning the relevance of the insurance hypothesis (which was originally developed for birds) to humans. The first is that the starvation-predation trade-off assumed to underpin strategic mass regulation in birds does not apply to humans, because predation has not been an important selection pressure in our recent evolutionary history. The second argument is that the costs of excess body mass present in birds do not apply in humans, because humans do not fly. In response to these arguments we emphasise that the insurance hypothesis does not rely on either the existence of a starvation-predation trade-off underpinned by mass-dependent predation, or the costs of flight. The theory simply requires that there is an asymmetrical inverted U-shaped relationship between reserves and Darwinian fitness, such that fitness declines rapidly as reserves decrease below their optimum level and less rapidly as reserves increase above their optimal level (Nettle, Andrews & Bateson, 2017). Humans may not fly, but we are adapted to walk and run long distances (Pontzer, 2017), suggesting that endurance locomotion was important for fitness during our recent evolutionary history. Furthermore, human athletic performance in distance events declines continuously with higher body mass (Sedeaud et al., 2014). Therefore, the costs of being overweight will be incurred in any human activity involving endurance locomotion, whether this is fleeing predators, foraging for food, fighting wars or attracting mates.