Technical Comments

Comment on “Number-space mapping in the newborn chick resembles humans’ mental number line”

See allHide authors and affiliations

Science  26 Jun 2015:
Vol. 348, Issue 6242, pp. 1438
DOI: 10.1126/science.aaa9565


Rugani et al. (Reports, 30 January 2015, p. 534) presented evidence that domestic chicks employ a “mental number line.” I argue that the hypothesis testing used to support this claim unjustifiably assumes that domestic chicks are unbiased when choosing between identical stimuli presented to their left and right.

Rugani et al. (1) presented evidence in support of the claim that domestic chicks (Gallus gallus domesticus) employ a “mental number line” similar to that used by most, but not all, humans. This Comment will focus on a critical flaw in the study’s design and statistical analysis. I argue that Rugani et al. failed to measure, model, or control for the significant side biases displayed by precocial birds, including the domestic chicks used in their study (29).

Precocial chicks possess highly lateralized brains and, consequently, show strong motor and perceptual biases (10)—biases often more extreme than human analogs (e.g., pseudoneglect). For example, chicks show significant bias for turning left at the intersection of a T maze (4, 6, 9, 11), in addition to task-dependent preferences for foot and eye use (2, 3, 5). Such biases are sculpted by normally occurring pre- and perinatal experience, including both differential exposure of the hemispheres to audiovisual stimulation prenatally (4, 7, 11, 12) and the repetitive turning movements necessary for chicks to successfully hatch from the egg (6, 9). As a consequence, as many as 67 to 90% of domestic chicks display significant side biases in tasks requiring a locomotor response (4, 6, 9).

Side biases are less consequential in experiments involving choice between nonidentical stimuli, because counterbalancing can be used to distribute error stemming from such bias between stimuli and competing hypotheses. In contrast, in designs involving choice between identical stimuli presented to the left and right of the subject, such as employed by Rugani et al., side biases require far closer attention, as the problem of dissociating side bias from experimental effect becomes central to the ability to draw meaningful conclusions. There are several acceptable ways of accomplishing this. One is to provide independent trials in which bias is measured (e.g., a series of trials similar, although not identical, to testing, covering the range of stimuli in the study) and correcting for bias on a chick-by-chick basis. Another approach is to test a more sophisticated model that includes a bias parameter, as is common in many operant paradigms involving choice [e.g., 13)]. Finally, the null hypotheses tested could be adjusted to reflect the average bias in a particular population. In the latter case, a result would be declared significant only if it exceeded this average expected bias.

None of these approaches were taken by Rugani et al., who instead relied on single-sample t tests evaluated against a null hypothesis of 0.5 (50% left, 50% right) for all of their key comparisons. As noted previously, there is nevertheless a strong a priori reason to expect motor and perceptual biases under the conditions used in this study [e.g., (4, 6, 9)]. Evidence of side bias is also detectable in Rugani et al. For example, the average absolute deviation from 50/50 responding across all 10 trials for each chick was 12.7% ± 1.1%, indicating that most chicks exhibited some degree of bias (t = 11.5, P < 0.00001). If the conservative approach of adjusting the null hypotheses tested to reflect the average absolute bias displayed by chicks in each experiment (i.e., testing a chance-plus-bias model) were taken, a significant difference would have been found only for the small numerosity trial of experiment 3b (t = 2.12, P < 0.03) and large numerosity trials of experiments 2, 3a, and 3c (t = –2.61, P < 0.02; t = –3.66, P < 0.002; and t = –1.79, P < 0.05, respectively; one-tailed testing).

If a stricter bias criterion were applied—for example, ≥70% choices on a single side, then 60% of chicks (9 out of 15) in experiment 1 and 34.4% of chicks across all experiments exhibited bias. Theoretically, this would not present a serious challenge to the conclusions of the study if such biases were symmetrical around a mean of 0.50 and affected both major hypotheses equally. With respect to the first consideration, the majority of chicks (63.6%) that met the stricter bias criterion (i.e., ≥70%) exhibited rightward bias. Considering all subjects, the average proportion of choices made to the right across all experiments was 0.522 ± 0.019, indicating a similar, albeit weaker bias (t = –1.147, P = 0.26). This bias was largest in experiment 3a, in which the proportion of choices to the right was 0.571 ± 0.037—a borderline significant difference from 0.5 (t = –1.92, P = 0.081). Did bias affect both hypotheses equally? Figure 1 displays the percentage of chick choices to the left and right on small and large trials, respectively, plotted both against the absolute value of chick bias (A and B) and side bias relative to predicted outcomes (C and D). As can be seen, the absolute value of chick bias appears to have significantly influenced the outcome on small (r = –0.277, P < 0.03) but not large (r = –0.042, P = 0.742) trials (upper panels), whereas the direction of bias correlated significantly with predicted outcomes on both trial types (lower panels).

Fig. 1 An illustration of the effects of side bias on the two predicted outcomes of the study.

The choice of “left” on small numerosity trials (A and C) and the choice of “right” on large numerosity trials (B and D). In (A) and (B), the x axis depicts the absolute value of chick bias (i.e., the absolute value of 0.5 minus the proportion of choices made to the left). In (C) and (D), the x axis depicts the strength of side bias relative to the predicted direction. An asterisk indicates a significant Pearson’s correlation. As can be seen, the absolute value of chick bias appears to have had a significant effect on the outcome of small but not large numerosity trials [(A) and (B)], whereas the direction of chick bias correlates strongly with predicted outcomes on both trial types [(C) and (D)].

The foregoing analyses demonstrate that there was unaccounted bias in Rugani et al. that differentially affected the study’s two major predictions. That is, rightward bias likely inflated chick performance on large numerosity trials and interfered with chick performance on small numerosity trials. Given that there is no choice test between identical stimuli presented to the left and right of a domestic chick in which significant bias would not be expected a priori, the hypotheses tested by Rugani et al. are closer to statistical “straw men” than empirically meaningful. Both the training that chicks underwent before testing and the logic of the series of experiments are irrelevant to this consideration. Nevertheless, it should be noted that the analyses presented here are by no means ideal, given that the measures of bias employed are necessarily confounded by any experimental (i.e., training-induced) effect present.

References and Notes

  1. Acknowledgments: I thank M. Casey, J. Spencer, D. Landy, M. Blumberg, B. Lickliter, C. Allen, and J. de Leeuw for helpful feedback on a previous draft.

Stay Connected to Science


Navigate This Article