| |

Received: 12 January 2021    Revised: 4 February 2021    Accepted: 8 February 2021

DOI: 10.1002/ece3.7365

ACADEMIC PR ACTICE IN ECOLOGY AND EVOLUTION

When are hypotheses useful in ecology and evolution?

Matthew G. Betts1  | Adam S. Hadley1 | David W. Frey1 | Sarah J. K. Frey1 |

Dusty Gannon1 | Scott H. Harris1  | Hankyu Kim1 | Urs G. Kormann1 |
Kara Leimberger1 | Katie Moriarty2 | Joseph M. Northrup1,3 | Ben Phalan1 |
Josée S. Rousseau1 | Thomas D. Stokely1 | Jonathon J. Valente1 | Chris Wolf1 |
Diego  Zárrate-­Charry1

1
Forest Biodiversity Research Network,
Department of Forest Ecosystems and Abstract
Society, Oregon State University, Corvallis, Research hypotheses have been a cornerstone of science since before Galileo. Many
OR, USA
2 have argued that hypotheses (1) encourage discovery of mechanisms, and (2) reduce
USDA Forest Service, Pacific Northwest
Research Station, Corvallis, OR, USA bias—­both features that should increase transferability and reproducibility. However,
3
Wildlife Research and Monitoring Section, we are entering a new era of big data and highly predictive models where some argue
Ontario Ministry of Natural Resources and
Forestry, Environmental and Life Sciences
the hypothesis is outmoded. We hypothesized that hypothesis use has declined in
Graduate Program, Trent University, ecology and evolution since the 1990s, given the substantial advancement of tools
Peterborough, ON, Canada
further facilitating descriptive, correlative research. Alternatively, hypothesis use
Correspondence may have become more frequent due to the strong recommendation by some journals
Matthew G. Betts, Forest Biodiversity
Research Network, Department of Forest
and funding agencies that submissions have hypothesis statements. Using a detailed
Ecosystems and Society, Forest Biodiversity literature analysis (N = 268 articles), we found prevalence of hypotheses in eco–­evo
Research Network, Oregon State University,
Corvallis, OR 97331, USA.
research is very low (6.7%–­26%) and static from 1990–­2015, a pattern mirrored in an
Email: matt.betts@oregonstate.edu extensive literature search (N = 302,558 articles). Our literature review also indicates

Funding information
that neither grant success nor citation rates were related to the inclusion of hypoth-
National Science Foundation, Grant/Award eses, which may provide disincentive for hypothesis formulation. Here, we review
Number: NSF-­DEB-­1457837
common justifications for avoiding hypotheses and present new arguments based
on benefits to the individual researcher. We argue that stating multiple alternative
hypotheses increases research clarity and precision, and is more likely to address the
mechanisms for observed patterns in nature. Although hypotheses are not always
necessary, we expect their continued and increased use will help our fields move
toward greater understanding, reproducibility, prediction, and effective conservation
of nature.

KEYWORDS

hypothesis, mechanisms, multiple working hypotheses, prediction, scientific method

Matthew G. Betts and Adam S. Hadley contributed equally to this manuscript.

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium,
provided the original work is properly cited.
© 2021 The Authors. Ecology and Evolution published by John Wiley & Sons Ltd.

|
5762     
www.ecolevol.org Ecology and Evolution. 2021;11:5762–5776.
BETTS et al. |
      5763

1 |  I NTRO D U C TI O N hypotheses (sensu Chamberlin,  1890), arguing that without such
hypothesis tests, disciplines would be prone to “stamp collecting”
Why should ecologists have hypotheses? At the beginning of most (Landy,  1986). To constitute “strong inference,” Platt required the
science careers, there comes a time of “hypothesis angst” where scientific method to be a three-­step process including (1) developing
students question the need for the hypothetico-­deductive approach alternative hypotheses, (2) devising a set of “crucial” experiments to
their elders have deemed essential for good science. Why is it not eliminate all but one hypothesis, and (3) performing the experiments
sufficient to just have a research objective or question? Why can't (Elliott & Brook, 2007).
we just collect observations and describe those in our research The commonly touted strengths of hypotheses are two-­fold.
papers? First, by adopting multiple plausible explanations for a phenom-
Research hypotheses are explanations for an observed phe- enon (hereafter “multiple alternative hypotheses”; Box  1), a re-
nomenon (Loehle, 1987; Wolff & Krebs, 2008) (see Box 1) and have searcher reduces the chance that they will become attached to a
been proposed as a central tool of science since Galileo and Francis single possibility, thereby biasing research in favor of this outcome
Bacon in the mid-­1600s (Glass & Hall, 2008). Over the past century, (Chamberlin, 1890); this “confirmation bias” is a well-­known human
there have been repeated calls for rigorous application of hypoth- trait (Loehle, 1987; Rosen, 2016) and likely decreases reproducibil-
eses in science, and arguments that hypothesis use is the corner- ity (Munafò et al., 2017). Second, various authors have argued that
stone of the scientific method (Chamberlin,  1890; Popper,  1959; the a priori hypothesis framework forces one to think in advance
Romesburg,  1981). In a seminal paper in Science, Platt (1964) chal- about—­and then test—­various causes for patterns in nature (Wolff &
lenged all scientific fields to adopt and rigorously test multiple Krebs, 2008), rather than simply examining the patterns themselves
and coming up with explanations after the fact (so called “inductive
research;” Romesburg,  1981). By understanding and testing mech-
anisms, science becomes more reliable and transferable (Ayres &
BOX 1  Definitions of hypotheses and associated
Lombardero,  2017; Houlahan et  al.,  2017; Sutherland et  al.,  2013)
terms
(Figure 1). Importantly, both of these strengths should have strong,
Hypothesis: An explanation for an observed phenomenon. positive impacts on reproducibility of ecological and evolutionary
Research Hypothesis: A statement about a phenomenon studies (see Discussion).
that also includes the potential mechanism or cause of that However, we are entering a new era of ecological and evolutionary
phenomenon. Though a research hypothesis doesn't need science that is characterized by massive datasets on genomes, species
to adhere to this strict framework it is often best described distributions, climate, land cover, and other remotely sensed informa-
as the “if” in an “if-­then” statement. In other words, “if X is tion (e.g., bioacoustics, camera traps; Pettorelli et al., 2017). Exceptional
true” (where X is the mechanism or cause for an observed computing power and new statistical and machine-­learning algorithms
phenomenon) “then Y” (where Y is the outcome of a cru- now enable thousands of statistical models to be run in minutes. Such
cial test that supports the hypothesis). These can also be datasets and methods allow for pattern recognition at unprecedented
thought of as “mechanistic hypotheses” since they link spatial scales and for huge numbers of taxa and processes. Indeed,
with a causal mechanism. For example, trees grow slowly there have been recent arguments in both the scientific literature and
at high elevation because of nutrient limitation (hypothe- popular press to do away with the traditional scientific method and a
sis); if this is the case, fertilizing trees should result in more priori hypotheses (Glass & Hall, 2008; Golub, 2010). These arguments
rapid growth (prediction). go something along the lines of “if we can get predictions right most of
Prediction: The potential outcome of a test that would the time, why do we need to know the cause?”
support a hypothesis. Most researchers call the second In this paper, we sought to understand if hypothesis use in ecol-
part of the if-­then statement a “prediction”. ogy and evolution has shifted in response to these pressures on the
Multiple alternative hypotheses: Multiple plausible expla- discipline. We, therefore, hypothesized that hypothesis use has de-
nations for the same phenomenon. clined in ecology and evolution since the 1990s, given the substan-
Descriptive Hypothesis: Descriptive statements or predic- tial advancement of tools further facilitating descriptive, correlative
tions with the word “hypothesis” in front of them. Typically research (e.g., Cutler et  al.,  2007; Elith et  al.,  2008). We predicted
researchers state their guess about the results they expect that this decline should be particularly evident in the applied conser-
and call this the “hypothesis” (e.g., “I hypothesize trees at vation literature—­where the emergence of machine-­learning models
higher elevation will grow slowly”). has resulted in an explosion of conservation-­oriented species dis-
Statistical Hypothesis: A predicted pattern in data that tribution models (Elith et al., 2006). Our alternative hypothesis was
should occur if a research hypothesis is true. that hypothesis use has become more frequent. The mechanism for
Null Hypothesis: A concise statement expressing the con- such increases is that higher-­profile journals (e.g., Functional Ecology,
cept of “no difference” between a sample and the popula- Proceedings of the Royal Society of London Ser. B) and competitive
tion mean. granting agencies (e.g., the U.S. National Science Foundation) now
require or strongly encourage hypothesis statements.
 |
5764       BETTS et al.

(a) (b)

(c) (d)

F I G U R E 1   Understanding mechanisms often increases model transferability. Panels (a and b) show snowshoe hares in winter and
summer coloration, respectively. If a correlative (i.e., nonmechanistic) model for hare survival as a function of color was trained only on hares
during the winter and then extrapolated into the summer months, it would perform poorly (white hares would die disproportionately under
no-­snow conditions). On the other hand, a researcher testing mechanisms for hare survival would (ideally via experimentation) arrive at the
conclusion that it is not the whiteness of hares, but rather blending with the background that confers survival (the “camouflage” hypothesis).
Understanding mechanism results in model predictions being robust to novel conditions. Panel (c) Shows x and y geographic locations of
training (blue filled circles) and testing (blue open circles) locations for a hypothetical correlative model. Even if the model performs well on
these independent test data (predicting open to closed circles), there is no guarantee that it will predict well outside of the spatial bounds of
the existing data (red circles). Nonstationarity (in this case caused by a nonlinear relationship between predictor and response variable; panel
d) could result in correlative relationships shifting substantially if extrapolated to new times or places. However, mechanistic hypotheses
aimed at understanding the underlying factors driving the distribution of this species would be more likely to elucidate this nonlinear
relationship. In both of these examples, understanding drivers behind ecological patterns—­via testing mechanistic hypotheses—­is likely to
enhance model transferability

As noted above, many have argued that hypotheses are useful practical recommendations for improving hypothesis use in ecology
and important for overall progress in science, because they facilitate and evolution—­particularly for new practitioners in the field (Box 2).
the discovery of mechanisms, reduce bias, and increase reproduc-
ibility (Platt,  1964). However, for hypothesis use to be propagated
among scientists, one would also expect hypotheses to confer ben- 2 | M E TH O DS
efits to the individual. We, therefore, tested whether hypothesis use
was associated with individual-­level incentives relevant to academic 2.1 | Literature analysis
success: publications, citations, and grants (Weinberg, 2010). If hy-
pothesis use confers individual-­level advantages, then hypothesis-­ To examine hypothesis use over time and test whether hypothesis
based research should be (1) published in more highly ranked presence was associated with research type (basic vs. applied),
journals, (2) have higher citation rates, and (3) be supported by highly journal impact factor, citation rates, and grants, we sampled the
competitive funding sources. ecology and evolution literature using a stratified random sample
Finally, we also present some common justifications for absence of of ecology and evolution journals in existence before 1991. First,
hypotheses and suggest potential counterpoints researchers should we randomly selected 19 journals across impact factor (IF) strata
consider prior to dismissing hypothesis use, including potential bene- ranging from 0.5–­10.0 in two bins (<3 IF and ≥3 IF; see Figure 3
fits to the individual researcher. We hope this communication provides for full journal list). We then added three multidisciplinary journals
BETTS et al. |
      5765

not found, or reviewers were not able to complete their review.

BOX 2  Recommendations for improving Ultimately, our final sample comprised 268 articles.

hypotheses use in ecology and evolution Reviewers were given a maximum of 10  min to find research
hypothesis statements within the abstract or introduction of ar-
Authors: Know that you are human and prone to confirma-
ticles. We chose 10  min to simulate the amount of time that a
tion bias and highly effective at false pattern recognition.
journal editor pressed for time might spend evaluating the intro-
Thus, inductive research and single working hypotheses
ductory material in an article. After this initial 10  min period, we
should be rare in your research. Remember that if your
determined: (1) whether or not an article contained at least one
work is to have a real “impact”, it needs to withstand multi-
hypothesis, (2) whether hypotheses were mechanistic or not (i.e.,
ple tests from other labs over the coming decades.
the authors claimed to examine the mechanism for an observed
Editors and Reviewers: Reward research that is conducted
phenomenon), (3) whether multiple alternative hypotheses were
using principles of sound scientific method. Be skeptical of
considered (sensu Chamberlin, 1890), and (4) whether hypotheses
research that smacks of data dredging, post hoc hypothesis
were “descriptive” (that is, they did not explore a mechanism but
development, and single hypotheses. If no hypotheses are
simply stated the expected direction of an effect; we define this as
stated in a paper and/or the paper is purely descriptive,
a “prediction” [Box 1]). It is important to note that to be identified as
ask whether the novelty of the system and question war-
having hypotheses, articles did not need to contain the actual term
rant this, or if the field would have been better served by a
“hypothesis” under our protocol; we also included articles using
study with mechanistic hypotheses. If only single hypoth-
phrases such as “If X is true, we expected…” or “we anticipated,” both
eses are stated, ask whether appropriate precautions were
of which reflect a priori expectations from the data. We catego-
taken for the researcher to avoid finding support for a pet
rized each article as either basic (fundamental research without ap-
idea (e.g., blinded experiments, randomized attribution of
plications as a focus) or applied (clear management or conservation
treatments, etc.). To paraphrase Platt (1964): beware of the
focus to article). Finally, we also examined all articles for funding
person with only one method or one instrument, either ex-
sources and noted the presence of a national or international-­level
perimental or theoretical.
competitive grant (e.g., National Science Foundation, European
Mentors: Encourage your advisees to think carefully about
Union, Natural Sciences and Engineering Research Council). We
hypothesis use and teach them how to construct sound
assumed that published articles would have fidelity to the hypoth-
multiple, mechanistic hypotheses. Importantly, explain
eses stated in original grant proposals that funded the research,
why hypotheses are important to the scientific method,
therefore, the acknowledgment of a successful grant is an indi-
the individual and group consequences of excluding them,
cator of financial reward for including hypotheses in initial pro-
and the rare instances where they may not be necessary.
posals. Journal impact factors and individual article citation rates
Policymakers/media/educators/students/readers: Read
were gleaned directly from Web of Science. We reasoned that
scientific articles with skepticism; have a scrutinous eye out
many researchers seek out journals with higher impact factors
for single hypothesis studies and p-­hacking. Reward multi-­
for the first submission of their manuscripts (Paine & Fox,  2018).
hypothesis, mechanistic, predictive science by giving it
Our assumption was that studies with more careful experimental
greater weight in policy decisions (Sutherland et al., 2013),
design—­including hypotheses—­should be published where initially
more coverage in the media, greater leverage in education,
submitted, whereas those without may be eventually published, on
and more citations in reports.
average, in lower impact journals (Opthof et al., 2000). Ideally, we
could have included articles that were rejected and never published
in our analysis, but such articles are notoriously difficult to track
that regularly publish ecology and evolution articles (Proceedings (Thornton & Lee, 2000).
of the National Academy of Sciences, Science, and Nature). From this To support our detailed literature analysis, we also tested for tem-
sample of 22 journals, we randomly selected ecology and evolu- poral trends in hypothesis use within a broader sample of the ecology
tion articles within 5-­year strata beginning in 1991 (3 articles/ and evolution literature. For the same set of 22 journals in our detailed
journal per 5-­year bin) to ensure the full date range was evenly sample, we conducted a Web of Science search for articles containing
sampled. We removed articles in the following categories: edi- “hypoth*” in the title or abstract. To calculate the proportion of articles
torials, corrections, reviews, opinions, and methods papers. In with hypotheses (from 1990–­2018), we divided the number of articles
multidisciplinary journals, we examined only ecology, evolution, with hypotheses by the total number of articles (N = 302,558). Because
and conservation biology articles, as indicated by section head- our search method does not include the main text of articles and ex-
ers in each journal. Once selected, articles were randomly distrib- cludes more subtle ways of stating hypotheses (e.g., “We expected…,”
uted to the authors of the current paper (hereafter “reviewers:” “We predicted…”), we acknowledge that the proportion of papers
MGB, ASH, DF, SF, DG, SH, HK, UK, KL, KM, JN, BP, JSR, TSS, JV, identified is likely to be an underestimate of the true proportions.
DZC) for detailed examination. On rare occasions, an article was Nevertheless, we do not expect that the degree of underestimation
 5766       | BETTS et al.

(a) (b) F I G U R E 2   Trends in hypothesis use

1.00 1.00 from 1991–­2015 from a sample of the
ecological and evolutionary literature

Mechanistic hypotheses
(N = 268, (a) multiple alternative
Multiple alternative

0.75 0.75 hypotheses, (b) mechanistic hypotheses,

(c) descriptive hypotheses [predictions],
0.50 and (d) no hypotheses present). We
0.50
detected no temporal trend in any of
these variables. Lines reflect LOESS
0.25 smoothing with 95% confidence intervals.
0.25
Dots show raw data with darker colors
0.00 indicating overlapping data points. The
0.00
total number of publications in ecology
1990 1995 2000 2005 2010 2015 1990 1995 2000 2005 2010 2015 and evolution in selected journals
has increased (e), but use of the term
(c) (d) “hypoth*” in the title or abstracts of these
1.00
302,558 articles has remained flat, and at
1.00
very low prevalence (f)
Descriptive hypotheses

0.75
0.75
No hypothesis

0.50
0.50

0.25
0.25
0.00

0.00
1990 1995 2000 2005 2010 2015 1990 1995 2000 2005 2010 2015

(e) (f)
0.5
60000
Proportion with hypotheses
Number of publications

0.4

40000 0.3

0.2
20000
0.1

0 0.0
1992 1994

1995 1999

2000 2004

2005 2009

2010 2014

2015 2018
1992 1994

1995 1999

2000 2004

2005 2009

2010 2014

2015 2018

Year

would change over time, so temporal trends in the proportion of pa- The predictor variable (i.e., year) was scaled to enable convergence.
pers containing hypotheses should be unbiased. Similarly, we tested for differences in hypothesis prevalence between
basic and applied articles using GLMMs with “journal” as a random
effect. Finally, we tested the hypothesis that hypothesis use might
2.2 | Statistical analysis decline over time due to the emergence of machine-­learning in the
applied conservation literature; specifically, we modeled “hypothesis
We used generalized linear mixed models (GLMMs) to test for change presence” as a function of the statistical interaction between “year”
in the prevalance of various hypothesis types over time (descriptive, and “basic versus applied” articles. We conducted this test for all hy-
mechanistic, multiple, any hypothesis). Presence of a hypothesis was pothesis types. GLMMs were implemented in R (version 3.60) using
modeled as dichotomous (0,1) with binomial error structure, and the lme4 package (Bates et al., 2018). In three of our models, the “jour-
“journal” was included as a random effect to account for potential nal” random effect standard deviation was estimated to be zero or
lack of independence among articles published in the same outlet. nearly zero (i.e., 10–­8). In such cases, the model with the random effect
BETTS et al. |
      5767

is exceptionally difficult to estimate, and the random effect standard 3.2 | Do hypotheses “pay?”
deviation being estimated as approximately zero indicates the random
effect was likely not needed. We found little evidence that presence of hypotheses increased
We tested whether the presence of hypotheses influenced the paper citation rates. Papers with mechanistic (LMM: ̂
𝛽  =  −0.109
likelihood of publication in a high-­impact journal using generalized [95% CI: −0.329, 0.115], t  =  0.042, p  =  0.97, Figure  4a, middle
linear models with a Gaussian error structure. We used the log of panel) or multiple alternative hypotheses (LMM: ̂ 𝛽  =  −0.008 [95%
journal impact factor (+0.5) as the response variable to improve nor- CI: −0.369, 0.391], t = 0.042, p = 0.96, Figure 4a, bottom panel) did
mality of model residuals. We tested the association between major not have higher average annual citation rates, nor did papers with at
competitive grants and the presence of a hypotheses using gener- least one hypothesis type (LMM: ̂
𝛽 = −0.024 [95% CI: −0.239, 0.194],
alized linear models (logistic regression) with “hypothesis presence” t = 0.218, p = 0.83, Figure 4a, top panel).
(0,1) as a predictor and presence of a grant (0,1) as a response. On the other hand, journal articles containing mechanistic hy-
Finally, we tested whether hypotheses increase citation rates potheses tended to be published in higher impact journals (GLM:
using linear mixed effects models (LMMs); presence of various hy- 𝛽  =  0.290 [95% CI: 0.083, 0.497], t  =  2.74, p  =  0.006) but only
̂
potheses (0,1) were predictors in univariate models and average slightly so (Figure  4b, middle panel). Including multiple alternative
citations per year (log-­transformed) was the response. “Journal” hypotheses in papers did not have a statistically significant effect
was treated as a random effect, which assumes that articles within (GLM: = 0.339 [95% CI: −0.029, 0.707], t = 1.80, p = 0.072, Figure 4b,
a particular journal are unlikely to be independent in their citation bottom panel).
rates. LMMs were implemented in R using the lme4 package (Bates Finally, we found no association between obtaining a competi-
et al., 2015). tive national or international grant and the presence of a hypothesis
(logistic regression: mechanistic: ̂
𝛽 = −0.090 [95% CI: −0.637, 0.453],
z = −0.36, p =0 .745; multiple alternative: ̂
𝛽 = 0.080 [95% CI: −0.891,
3 |  R E S U LT S 1.052], z  =  0.49, p  =  0.870; any hypothesis: ̂
𝛽  =  −0.005 [95% CI:
−0.536, 0.525], z = −0.02, p = 0.986, Figure 4c).
3.1 | Trends in hypothesis use in ecology and
evolution
4 | D I S CU S S I O N
In the ecology and evolution articles we examined in detail, the preva-
lence of multiple alternative hypotheses (6.7%) and mechanistic hy- Overall, the prevalence of hypothesis use in the ecological and evo-
potheses (26%) was very low and showed no temporal trend (GLMM: lutionary literature is strikingly low and has been so for the past
multiple alternative: ̂
𝛽  =  0.098 [95% CI: −0.383, 0.595], z  =  0.40, 25  years despite repeated calls to reverse this pattern (Elliott &
p  =  0.69, mechanistic: ̂
𝛽  =  0.131 [95% CI: −0.149, 0.418], z  =  0.92, Brook, 2007; Peters, 1991; Rosen, 2016; Sells et al., 2018). Why is
p  =  0.36, Figure  2a,b). Descriptive hypothesis use was also low this the case?
(8.5%), and although we observed a slight tendency to increase over Clearly, hypotheses are not always necessary and a portion
time, 95% confidence intervals overlapped zero (GLMM: ̂ 𝛽  = 0.351 of the sampled articles may represent situations where hypothe-
[95% CI: −0.088, 0.819], z = 1.53, p = 0.13, Figure 2c). Although the ses are truly not useful (see Box  3: “When Are Hypotheses Not
proportion of papers containing no hypotheses appears to have de- Useful?”). Some authors (Wolff & Krebs,  2008) overlook knowl-
clined (Figure 2d), this effect was not statistically significant (GLMM: edge gathering and descriptive research as a crucial first step for
𝛽 = −0.201 [95% CI: −0.483, 0.074], z = −1.41, p = 0.15). This overall
̂ making observations about natural phenomena—­from which hy-
pattern is consistent with a Web of Science search (N = 302,558 arti- potheses can be formulated. This descriptive work is an important
cles) for the term “hypoth*” in titles or abstracts that shows essentially part of ecological science (Tewksbury et  al.,  2014), but may not
no trend over the same time period (Figure 2e,f). benefit from strict use of hypotheses. Similarly, some efforts are
Counter to our hypothesis, applied and basic articles did not simply designed to be predictive, such as auto-­recognition of spe-
show a statistically significant difference in the prevalence of either cies via machine learning (Briggs et al., 2012) or for prioritizing con-
mechanistic (GLMM: ̂ 𝛽  =  0.054 [95% CI: −0.620, 0.728], z  =  0.16, servation efforts (Wilson et al., 2006), where the primary concern
p  =  0.875) or multiple alternative hypotheses (GLMM: ̂
𝛽  = 0.517 is correct identification and prediction rather than the biological or
[95% CI: −0.582, 1.80], z = 0.88, p = 0.375). Although both basic and computational reasons for correct predictions (Box 3). However, it
applied ecology and evolution articles containing hypotheses were would be surprising if 75% of ecology since 1990 has been purely
similarly rare overall, there was a tendency for applied ecology ar- descriptive work from little-­k nown systems or purely predictive in
ticles to show increasing prevalence of mechanistic hypothesis use nature. Indeed, the majority of the articles we observed did not fall
over time, whereas basic ecology articles have remained relatively into these categories.
unchanged (Table  S1, Figure  S1). However, there was substantial Alternatively, researchers may not include hypotheses because
variation across both basic and applied journals in the prevalence of they see little individual-­level incentive for their inclusion. Our re-
hypotheses (Figure 3). sults suggest that currently there are relatively few measurable
 |
5768       BETTS et al.

Any hypothesis
Functional Ecology
Ecol Monographs
American Naturalist
PNAS
Science
J. Applied Ecology
Theoretical Pop. Biol.
J. Ecology
Ecol Applications
Nature
Northwest Science
Biotropica
Ecology
J. Wildlife Management
J. Animal Ecology
Biological Conservation
Proc. Royal Soc. B
Acta Oecologica
Polar Biology
J. Soil and Water Cons.
J. Freshwater Biology
Conservation Biology 0.0

0.2

0.4

0.6
Mechanistic hypothesis
Functional Ecology
Ecol Monographs
American Naturalist
Science
PNAS
Theoretical Pop. Biol.
J. Ecology
J. Applied Ecology
Ecol Applications
Nature
Biotropica
Ecology
J. Wildlife Management
Northwest Science
Biological Conservation
Proc. Royal Soc. B
Acta Oecologica
Polar Biology
J. Animal Ecology
J. Soil and Water Cons.
J. Freshwater Biology
Conservation Biology
0.0

0.2

0.4

0.6

Multiple alternative hypotheses

Science
Northwest Science
Ecol Monographs
J. Applied Ecology
J. Animal Ecology
Functional Ecology
American Naturalist
Nature
Ecology
PNAS
J. Ecology
Theoretical Pop. Biol.
Proc. Royal Soc. B
Polar Biology
J. Wildlife Management
J. Soil and Water Cons.
J. Freshwater Biology
Ecol Applications
Conservation Biology
Biotropica
Biological Conservation
Acta Oecologica
0.0

0.2

0.4

0.6

Proportion with hypotheses

F I G U R E 3   Frequency distributions showing proportion of various hypotheses types across ecology and evolution journals included
in our detailed literature search. Hypothesis use varied greatly across publication outlets. We considered J. Applied Ecology, J. Wildlife
Management, J. Soil, and Water Cons., Ecological Applications, Conservation Biology, and Biological Conservation to be applied journals;
both applied and basic journals varied greatly in the prevalence of hypotheses
BETTS et al. |
      5769

Average times cited Impact factor Grant probability

6 0.7
log (Average cites per year)

Any hypothesis
3
0.6

Grant probability
4

Impact factor
2
0.5
2
1

0.4
0
0
0.3
No hypothesis Hypothesis No hypothesis Hypothesis No Hypothesis Hypothesis

6 0.7

Mechanistic
log (Average cites per year)

3
0.6

Grant probability
4
Impact factor

2
0.5
2
1

0.4
0
0
0.3
No hypothesis Hypothesis No hypothesis Hypothesis No Hypothesis Hypothesis

6 0.7

Multiple alternative
log (Average cites per year)

3
0.6
Grant probability

4
Impact factor

2
0.5
2
1

0.4
0
0
0.3
No hypothesis Hypothesis No hypothesis Hypothesis No Hypothesis Hypothesis

F I G U R E 4   Results of our detailed literature search showing the relationship between having a hypothesis (or not) and three commonly
sought after scientific rewards (Average times a paper is cited/year, Journal impact factor, and the likelihood of having a major national
competitive grant). We found no statistically significant relationships between having a hypothesis and citation rates or grants, but articles
with hypotheses tended to be published in higher impact journals

benefits to individuals. Articles with mechanistic hypotheses do Here we address some common justifications for hypotheses
tend to be published in higher impact factor journals, which, for bet- being unnecessary and show how one's first instinct to avoid hy-
ter or worse, is one of the key predictors in obtaining an academic potheses may be mistaken. We also present four reasons that use of
job (van Dijk et  al.,  2014). However, few of the other typical aca- hypotheses may be of individual self-­interest.
demic metrics (i.e., citations or grant funding) appear to reward this
behavior. Although hypotheses might be “useful” for overall progress
in science (Platt, 1964), for their use to be propagated in the popu- 5 | R E S P O N S E S TO CO M M O N
lation of scientists, one would also expect them to provide benefits J U S TI FI C ATI O N S FO R TH E A B S E N C E O F
to the individuals conducting the science. Interestingly, the few ex- H Y P OTH E S E S
isting papers on hypotheses (Loehle, 1987; Romesburg, 1981; Sells
et al., 2018) tended to explain the advantages in terms of benefits to During our collective mentoring at graduate and undergraduate lev-
the group by offering arguments such as “because hypotheses help els, as well as examination of the literature, we have heard a number
the field move forward more rapidly”. of common justifications for why hypotheses are not included. We
 |
5770       BETTS et al.

BOX 3 (Continued)
BOX 3  When are hypotheses not useful?
predict efficient methods or contexts for conserving spe-
Of course, there are a number of instances where hypothe-
cies (Myers et al., 2000; Wilson et al., 2006). Perhaps this
ses might not be useful or needed. It is important to recog-
is the reason for such low prevalence of hypotheses in con-
nize these instances to prevent the pendulum from swinging
servation journals (e.g., Conservation Biology).
in a direction where without hypotheses, research ceases
to be considered science (Wolff & Krebs, 2008). Below are
several important types of ecological research where for-
mulating hypotheses may not always be beneficial.
must admit that many of us have, on occasion, rationalized absence
When the goal is prediction rather than understand-
of hypotheses in our own work using the same logic! We understand
ing. Examples of this exception include species distribu-
that clearly formulating and testing hypotheses can often be chal-
tion models (Elith et  al.,  2008) where the question is not
lenging, but propose that the justifications for avoiding hypotheses
why species are distributed as they are, but simply where
should be carefully considered.
species are predicted to be. Such results can be useful in
conservation planning (Guisan et  al.,  2013; see below).
1. “But I do have hypotheses”. Simply using the word “hypothesis”
Another example lies in auto-­recognition of species (Briggs
does not a hypothesis make. A common pattern in the literature
et  al.,  2012) where the primary concern is getting identi-
we reviewed was for researchers to state their guess about
fication right rather than the biological or computational
the results they expect and call this the “hypothesis” (e.g., “I
reasons for correct predictions. In such instances, complex
hypothesize trees at higher elevation will grow slowly”). But
algorithms can be very effective at uncovering patterns
these are usually predictions derived from an implicit theoretical
(e.g., deep learning). A caveat and critical component of
model (Symes et  al.,  2015) or are simply descriptive statements
such efforts is to ensure that such models are tested on
with the word “hypothesis” in front of them (see Box  1). A
independent data. Further, if model predictions are made
research hypothesis must contain explanations for an observed
beyond the spatial or temporal bounds of training or test
phenomenon (Loehle,  1987; Wolff & Krebs,  2008). Such ex-
data, extreme caution should be applied (see Figure 4).
planations are derived from existing or new theory (Symes
When the goal is description rather than understanding.
et al., 2015). Making the link between the expected mechanism
In many applications, the objective is to simply quantify
(hypothesis) and logical outcome if that mechanism were true
a pattern in nature; for example, where on Earth is forest
(the prediction), is a key element of strong inference. Similarly,
loss most rapid (Hansen et al., 2013)? Further, sometimes
using “statistical hypotheses” and “null hypothesis testing” is
so little is known about a system or species that formulat-
not the same as developing mechanistic research hypotheses
ing hypotheses is impossible and more description is nec-
(Romesburg,  1981; Sells et  al.,  2018).
essary. In rare instances, an ecological system may be so
2. “Not enough is known about my system to formulate hypotheses”.
poorly known and different to other systems that generat-
This is perhaps the most common defense against needing hy-
ing testable hypotheses would be extremely challenging.
potheses (Golub, 2010). The argument goes that due to lack of
Darwin's observations while traveling on the Beagle are
previous research no mature theory has developed, so formal
some of the best examples of such “hypothesis generating”
tests are impossible. Such arguments may have basis in some
science; these initial observations resulted in the formula-
truly novel contexts (e.g., exploratory research on genomes)
tion of one of the most extensively tested hypotheses in
(Golub, 2010). But on close inspection, similar work has often
biology. However, such novelty should be uncommon in
been conducted in other geographic regions, systems, or with
ecological and evolutionary research where theoretical
different taxa. If the response by a researcher is “but we re-
and empirical precedent abounds (Sells et al., 2018). In the
ally need to know if X pattern also applies in this region” (e.g.,
field of biogeography, there is the commonly held view
does succession influence bird diversity in forests of Western
that researchers should first observe and analyze patterns,
North America the same way as it does in Eastern forests), this
and only then might explanations emerge (“pattern before
is fine and it is certainly useful to accumulate descriptive stud-
process”); however, it has frequently been demonstrated
ies globally for future synthetic work. However, continued ef-
that mechanistic hypotheses are useful even in disciplines
forts at description alone constitute missed opportunities for
where manipulative experiments are impossible (Crisp
understanding the mechanisms behind a pattern (e.g., why
et al., 2011).
does bird diversity decline when the forest canopy closes?).
When the objective is a practical planning outcome such
Often with a little planning, both the initial descriptive local
as reserve design. In many conservation planning efforts,
interest question (e.g., “is it?”) and the broader interest ques-
the goal is not to uncover mechanisms, but rather simply to
tion (i.e., “why?”) can both be tackled with minimal additional
effort.
BETTS et al. |
      5771

3. “What about Darwin? Many important discoveries have been research. Perhaps because the bar of falsification and testable mu-
made without hypotheses.” Several authors (and many students) tually exclusive hypotheses is so high, many have opted to ignore the
have argued that many important and reliable patterns in na- need for hypotheses altogether. If this is the case, our response is
ture have emerged outside of the hypothetico-­deductive (H-­D) that in ecology and evolution we must not let Popperian perfection
method (Brush, 1974). For instance, Darwin's discovery of natu- be the enemy of strong inference. With sufficient knowledge of a
ral selection as a key force for evolution has been put forward system, formal a priori hypotheses can be formulated that directly
as an example of how reliable ideas can emerge without the address the possibility of nonlinear relationships and interactions
H-­D method (May, 1981; Milner, 2018). Examination of Darwin's among variables. An example from conservation biology is the well-­
notebooks has suggested that he did not propose explicit hy- explored hypothesis that the effects of habitat fragmentation should
potheses and test them (Brush, 1974). However, Darwin himself be greatest when habitat amount is low due to dispersal limitation
wrote “all observation must be for or against some view if it is (i.e., there should be a statistical interaction between fragmentation
to be of any service!” (Ayala, 2009). In fact, Darwin actually put and habitat loss (Andrén, 1994)).
forward and empirically tested hypotheses in multiple fields,
including geology, plant morphology and physiology, psychol- 5. “But I am not a physiologist.” A common misconception has
ogy, and evolution (Ayala, 2009). This debate suggests that, like to do with the hierarchical aspect of mechanisms (Figure  5).
Darwin, we should continue to value systematic observation Many think that they are not testing the mechanism for a
and descriptive science (Tewksbury et al., 2014), but whenever pattern because they have not managed to get to the bottom
possible, it should be with a view toward developing theory and of a causal hierarchy (which reflects a sort of physics envy
testing hypotheses that commonly occurs in ecology and evolution (Egler,  1986)).
However, hierarchy theory (O'Neill et  al.,  1989), states that
The statement that “many important discoveries have been the cause of a given phenomenon usually occurs at the level
made without hypotheses” stems from a common misconcep- of organization just below the observed phenomenon. So, for
tion that somehow hypotheses spring fully formed into the mind, example, species distributions might be best understood by
and that speculation, chance and induction play no role in the H-­D examining hypotheses about the spatial composition and con-
method. As noted by Loehle (1987; p. 402) “The H-­D method and figuration of landscapes (Fahrig,  2003), explanations for pop-
strong inference, however, are valid no matter how theories are ob- ulation regulation might be best explored through observing
tained. Dreams, crystal balls, or scribbled notebooks are all allowed. the reproductive success and survival of individual organisms
In fact, induction may be used to create empirical relations which (Lack, 1954), and to understand individual variation in fecundity,
then become candidates for hypothesis testing even though induc- one might test hypotheses relating to individual behavior or
tion cannot be used to prove anything”. So, although induction has physiology. Hypothesis generation is possible at all levels of
frequently been used to develop theory, it is an unreliable means organization (Figure  5). Support for a hypothesis at one level
to test theory (Popper, 1959). As is well-­known, Darwin's theory of often generates a subsequent question and hypotheses at the
natural selection was heavily debated in scientific circles at the time, next (e.g., Observation: variation in animal densities can best
and it is only through countless hypothesis tests that it remains the be explained by forest patch size; Question: why are densities
best explanation for evolution even today (Mayr,  2002). lower in small patches? H1: small patches have more edge, and
predation rates are higher at the edge). However, in a single
4. “Ecology is too complex for hypotheses”. In one of the most research project it is not necessary to develop hypotheses
forcefully presented arguments for the H-­D method, Karl Popper that address mechanisms at all scales.
(1959) argued that science should be done through a process 6. “But my model predicts patterns well”. An increasingly common
of falsification; that is, multiple hypotheses should be con- justification for not presenting and testing research hypotheses
structed and the researcher's role is to successively eliminate seems to be the notion that if large datasets and complex mod-
these one at a time via experimentation until a single plausible eling methods can predict outcomes effectively, what is the need
hypothesis remains. This approach has caused some consterna- for hypothesizing a mechanism (Glass & Hall, 2008; Golub, 2010)?
tion among ecologists because the idea of single causes to Indeed, some have argued that prediction is a gold standard in
phenomena doesn't match most of our experiences (Quinn & ecology and evolution (Houlahan et  al.,  2017). However, under-
Dunham,  1983); rather, multiple interacting processes often lying such arguments is the critical assumption that the relation-
overlap to drive observed patterns. For example, Robert Paine ship between predictors (i.e., independent variables, 'x's) and
found that the distribution of a common seaweed was best responses ('y's) exhibit stationarity in time and space. Although
explained by competition, physical disturbance, and dispersal this appears to be the case in cosmology (e.g., relativity is thought
ability (Paine,  1966). to apply wherever you are in the universe (Einstein, 1920)), the as-
sumption of stationarity has repeatedly been shown to be violated
It would be interesting if Popperian logic has inoculated ecol- in ecological and evolutionary studies (Betts et al., 2006; Osborne
ogy and evolution against the frequent application of hypotheses in et  al.,  2007; Thompson,  2005). Hence the well-­known maxim
 |
5772       BETTS et al.

F I G U R E 5   Hypothesis generation is possible at all levels of organization, and does not need to get to the bottom of a causal hierarchy to
be useful. As illustrated in this case study (after Betts et al., 2015), using published work by the authors, support for a hypothesis at one level
often generates a subsequent question and hypotheses at the next. After each new finding we had to return to the white board and draw
out new alternative hypotheses as we progressed further down the hierarchy. Supported hypotheses are shown in black and the alternative
hypotheses that were eliminated are in grey. A single study is not expected to tackle an entire mechanistic hierarchy. In fact, we still have yet
to uncover the physiological mechanisms involved in this phenomenon
BETTS et al. |
      5773

“correlation does not equal causation;” correlates of a phenom- hypotheses” are top reasons a paper ends up in the editor's reject pile
enon often shift, even if the underlying cause remains the same. (Eassom, 2018; Elsevier, 2015). If editors have to struggle as often as
we did to determine the purpose of a paper, this does not bode well
The advantage of understanding mechanism is that the relation- for future publication. Clearly, communication through succinctly
ship between cause and effect is less likely to shift in space and time stated hypotheses is likely to enhance publication success.
than between the correlates of a phenomenon (Sells et  al.,  2018) Hypotheses also provide crucial direction during study design.
(Figure 1). For instance, climate-­envelope models are still commonly Nothing is more frustrating than realizing that your hard-­earned data
used to predict future species distributions (Beale et al., 2008) despite cannot actually address the key study objectives or rule out alter-
the fact that links between correlates often fail (Gutiérrez et al., 2014) native explanations. Developing clear hypotheses and, in particular,
and climate per se may not be the direct driver of distributions. In multiple alternative hypotheses ensures that you actually design
an example from our own group, predictions that fit observed data your study in a way that can answer the key questions of interest.
well in the region where the model was built completely failed when
predicted to a new region only 250  km away (Betts et  al.,  2006). 2. Personal Fulfillment
Although it is true that mechanisms can also exhibit nonstationarity,
at least in these instances logic can inform decisions about whether Second, science is more likely to be fulfilling and fun when
or not causal factors are likely to hold in a new place or time. the direction of research is clear, but perhaps more importantly,
when questions are addressed with more than one plausible an-
swer. Results are often disappointing or unfulfilling when the
6 |  W H Y S H O U LD YO U H AV E study starts out with a single biological hypothesis in mind (Symes
H Y P OTH E S E S ? ( A S E LF- ­I NTE R E S TE D et al., 2015)—­particularly if there is no support for this hypothesis. If
PE R S PEC TI V E ) multiple alternative hypotheses are well crafted, something interest-
ing and rewarding will result regardless of the outcome. This results
We have already described two arguments for hypothesis use, both in a situation where researchers are much more likely to enjoy the
of which should have positive influences on reproducibility and process of science because the stress of wanting a particular end is
therefore progress in science: (1) multiple alternative hypotheses de- removed. Subsequently, as Chamberlin (1890) proposed, “the dan-
veloped a priori prevent attachment to a single idea, and (2) hypoth- gers of parental affection for a favorite theory can be circumvented”
eses encourage exploration of mechanisms, which should increase which should reduce the risk of creeping bias. In our experience
the transferability of findings to new systems. Both these argu- reviewing competitive grant proposals at the U.S. National Science
ments have been made frequently in the eco-­evolutionary literature Foundation, it is consistently the case that proposals testing sev-
for decades (Elliott & Brook, 2007; Loehle, 1987; Rosen, 2016; Sells eral compelling hypotheses were more likely to be well received—­
et al., 2018), but our results show that such arguments have been lost presumably because reviewers are risk-­averse and understand that
on the majority of researchers. One hypothesis recently proposed ultimately finding support for any of the outcomes will pay-­off. Why
to explain why “poor methods persist [in science] despite perennial bet on just one horse when you can bet on them all?
calls for improvements” is that such arguments have largely failed
because they do not appeal to researcher self-­interest (Smaldino 3. Intrinsic Value to Mechanism
& McElreath,  2016). In periods of intense competition for grants
and top-­tier publications, perhaps arguments that rely on altruism Mechanism seems to have intrinsic value for humans—­regardless
fall short. However, happily, there are at least four self-­interested of the practical application. Humans tend to be interested in acquir-
reasons that students of ecological and evolutionary science should ing understanding rather than just accumulating facts. As a species,
adopt the hypothetico-­deductive method. we crave answers to the question “why.” Indeed, it is partly this de-
sire for mechanism that is driving a recent perceived “crisis” in ma-
1. Clarity and Precision in Research chine learning, with the entire field being referred to as “alchemy”
(Hutson,  2018); algorithms continue to increase in performance,
First, and most apparent during our review of the literature, hy- but the mechanisms for such improvements are often a mystery—­
potheses force clarity and precision in thinking. We often found it even to the researchers themselves. “Because our model predicts
difficult to determine the core purpose of papers that lacked clear well” is the unsatisfying scientific equivalent to a parent answering
hypotheses. One of the key goals of scientific writing is to communi- a child's “why?” with “because that's just the way it is.” This problem
cate ideas efficiently (Schimel, 2011). Increased clarity through use is beginning to spawn a new field in artificial intelligence “AI neu-
of hypotheses could potentially even explain the pattern for man- roscience” which attempts to get into the “black-­box” of machine-­
uscripts using hypotheses getting published in higher impact jour- learning algorithms to understand how and why they are predictive
nals. Editors are increasingly pressed for time and forced to reject (Voosen, 2017).
the majority of papers submitted to higher impact outlets prior to Even in some of our most applied research, we find that manag-
detailed review (AAAS, 2018). “Unclear message” and “lack of clear ers and policymakers when confronted with a result (e.g., thinning
 |
5774       BETTS et al.

trees to 70% of initial densities reduced bird diversity) want to know reward use of hypotheses. Therefore, we propose that in order to
why (e.g., thinning eliminated nesting substrate for 4 species); If the promote hypothesis use we may need to provide additional incen-
answer to this question is not available, policy is much less likely to tives (Edwards & Roy, 2016; Smaldino & McElreath, 2016). We sug-
change (Sells et  al., 2018). So, formulating mechanistic hypotheses gest editors reward research conducted using principles of sound
will not only be more personally satisfying, but we expect it may also scientific method and be skeptical of research that smacks of data
be more likely to result in real-­world changes. dredging, post hoc hypothesis development, and single hypotheses.
If no hypotheses are stated in a paper and/or the paper is purely
4. You Are More Likely To be Right descriptive, editors should ask whether the novelty of the system
and question warrant this, or if the field would have been better
In a highly competitive era, it seems that in the quest for high served by a study with mechanistic hypotheses. Eleven of the top
publication rates and funding, researchers lose sight of the original 20 ecology journals already indicate a desire for hypotheses in their
aim of science: To discover a truth about nature that is transferable to instructions for authors—­with some going as far as indicating “prior-
other systems. In a recent poll conducted by Nature, more than 70% ity will be given” for manuscripts testing clearly stated hypotheses.
of researchers have tried and failed to reproduce another scientist's Although hypotheses are not necessary in all instances, we expect
experiments (Baker, 2016). Ultimately, each researcher has a choice; that their continued and increased use will help our disciplines move
put forward multiple explanations for a phenomenon on their own toward greater understanding, higher reproducibility, better predic-
or risk “attachment” to a single hypothesis and run the risk of bias tion, and more effective management and conservation of nature.
entering their work, rendering it irreproducible, and subsequently We recommend authors, editors, and readers encourage their use
being found wrong by a future researcher. Imagine if Lamarck had (Box 2).
not championed a single hypothesis for the mechanisms of evolu-
tion? Although Lamarck potentially had a vital impact as an early pro- AC K N OW L E D G M E N T S
ponent of the idea that biological evolution occurred and proceeded Funding from the National Science Foundation (NSF-­DEB-­1457837)
in accordance with natural laws (Stafleu, 1971), unfortunately in the to MGB and ASH supported this research. We thank Rob Fletcher,
modern era he is largely remembered for his pet hypothesis. It may Craig Loehle and anonymous reviewers for thoughtful comments
be a stretch to argue that he would have necessarily come up with early versions of this manuscript, as well as Joe Nocera and his
natural selection, but if he had considered natural selection, the idea graduate student group at the University of New Brunswick for con-
would have emerged 50 years earlier, substantially accelerating sci- structive comments on the penultimate version of the paper. The
entific progress and limiting his infamy as an early evolutionary bi- authors are also grateful for A. Dream for providing additional re-
ologist. An interesting contemporary example is provided by Prof. sources to enable the completion of this manuscript.
Amy Cuddy's research focused on “power posing” as a means to suc-
ceed. The work featured in one of the most viewed TED talks of all C O N FL I C T O F I N T E R E S T
time but rather famously turned out to be irreproducible (Ranehill The authors have no conflicts of interests to declare.
et al., 2015). When asked in a TED interview what she would do dif-
ferently now, Prof. Cuddy noted that she would include a greater AU T H O R C O N T R I B U T I O N S
diversity of theory and multiple potential lines of evidence to “shed Matthew G. Betts: Conceptualization (lead); data curation (lead);
light on the psychological mechanisms” (Biello, 2017). formal analysis (lead); funding acquisition (lead); investigation (lead);
methodology (equal); project administration (lead); resources (lead);
supervision (lead); visualization (lead); writing-­original draft (lead);
7 |  CO N C LU S I O N writing-­review & editing (lead). Adam S. Hadley: Conceptualization
(lead); data curation (lead); funding acquisition (equal); investiga-
We acknowledge that formulating effective hypotheses can feel tion (equal); methodology (lead); project administration (equal);
like a daunting hurdle for ecologists. However, we suggest that resources (supporting); software (supporting); supervision (lead);
initial justifications for absence of hypotheses may often be un- validation (lead); visualization (lead); writing-­original draft (equal);
founded. We argue that there are both selfish and altruistic reasons writing-­review & editing (equal). David W. Frey: Conceptualization
to include multiple alternative mechanistic hypotheses in your re- (supporting); data curation (supporting); formal analysis (sup-
search: (1) testing multiple alternative hypotheses simultaneously porting); funding acquisition (supporting); writing-­review & edit-
makes for rapid and powerful progress which is to the benefit of ing (supporting). Sarah J. K. Frey: Conceptualization (supporting);
all (Platt,  1964), (2) you lessen the chance that confirmation bias Investigation (equal); writing-­review & editing (equal). Dusty Gannon:
will result in you publishing an incorrect but provocative idea, (3) Conceptualization (supporting); Investigation (equal); writing-­review
hypotheses provide clarity in design and writing, (4) research using & editing (equal). Scott H. Harris: Conceptualization (supporting);
hypotheses is more likely to be published in a high-­impact journal, Investigation (equal); methodology (equal); writing-­review & editing
and (5) you are able to provide satisfying answers to “why?” phe- (equal). Hankyu Kim: Conceptualization (supporting); Investigation
nomena occur. However, few current academic metrics appear to (equal); Methodology (equal); writing-­review & editing (equal). Kara
BETTS et al. |
      5775

Leimberger: Conceptualization (supporting); Investigation (equal); Betts, M. G., Hadley, A. S., & Kress, J. (2015). Pollinator recognition in
a keystone tropical plant. Proceedings of the National Academy of
Methodology (equal); writing-­review & editing (equal). Katie Moriarty:
Sciences, 112, 3433–­3 438.
Conceptualization (supporting); Investigation (equal); methodol- Biello, D. (2017). Inside the debate about power posing: A Q & A with Amy
ogy (equal); writing-­review & editing (equal). Joseph M. Northrup: Cuddy. Ideas.TED.com. Retrieved from https://ideas.ted.com/insid​e-­
Investigation (equal); methodology (equal); writing-­review & edit- the-­debat​e-­about​-­power​-­posin​g-­a-­q-­a-­with-­amy-­cuddy/
ing (equal). Ben Phalan: Investigation (equal); Methodology (equal); Briggs, F., Lakshminarayanan, B., Neal, L., Fern, X. Z., Raich, R., Hadley,
S. J. K., Hadley, A. S., & Betts, M. G. (2012). Acoustic classification
writing-­review & editing (equal). Josée S. Rousseau: Investigation
of multiple simultaneous bird species: A multi-­instance multi-­label
(equal); Methodology (equal); writing-­review & editing (equal). Thomas approach. The Journal of the Acoustical Society of America, 131, 4640–­
D. Stokely: Investigation (equal); methodology (equal); writing-­review 4650. https://doi.org/10.1121/1.4707424
& editing (equal). Jonathon J. Valente: Investigation (equal); meth- Brush, S. G. (1974). Should the history of science be rated X? Science,
183, 1164–­1172.
odology (equal); writing-­review & editing (equal). Urs G. Kormann:
Chamberlin, T. C. (1890). The method of multiple working hypotheses.
Methodology (supporting); resources (equal); writing-­review & editing Science, 15, 92–­96.
(supporting). Chris Wolf: Formal analysis (supporting); writing-­review Crisp, M. D., Trewick, S. A., & Cook, L. G. (2011). Hypothesis testing in
& editing (supporting). Diego Zárrate-­Charry: Investigation (equal); biogeography. Trends in Ecology & Evolution, 26, 66–­72. https://doi.
org/10.1016/j.tree.2010.11.005
Methodology (equal); writing-­review & editing (equal).
Cutler, D. R., Edwards, T. C., Beard, K. H., Cutler, A., Hess, K. T., Gibson,
J., & Lawler, J. J. (2007). Random forests for classification in ecology.
E T H I C A L A P P R OVA L Ecology, 88, 2783–­2792. https://doi.org/10.1890/07-­0539.1
The authors adhered to all standards for the ethical conduct of Eassom, H. (2018). 9 common reasons for rejection. Wiley: Discover the
Future of Research. Retrieved from https://hub.wiley.com/commu​
research.
nity/excha​n ges/disco​ver/blog/2018/2001/2031/2019-­c ommo​n -­
reaso​ns-­for-­rejec​tion
DATA AVA I L A B I L I T Y S TAT E M E N T Edwards, M. A., & Roy, S. (2016). Academic research in the 21st century:
Data for the analysis of hypothesis use in ecology and evolution pub- Maintaining scientific integrity in a climate of perverse incentives
and hypercompetition. Environmental Engineering Science, 34, 51–­61.
lications is available at https://figsh​are.com/artic​les/datas​et/Betts_
https://doi.org/10.1089/ees.2016.0223
et_al_2021_When_are_hypot​h eses_useful_in_ecolo​g y_and_evolu​ Egler, F. E. (1986). “Physics envy” in ecology. Bulletin of the Ecological
tion_Ecolo​g y_and_Evolu​tion/14110289. Society of America, 67, 233–­235.
Einstein, A. (1920). Relativity: The special and general theory (78 pp.).
ORCID Henry Holt and Company.
Elith, J., Graham, C. H., Anderson, R. P., Dudík, M., Ferrier, S., Guisan,
Matthew G. Betts  https://orcid.org/0000-0002-7100-2551
A., Hijmans, R. J., Huettmann, F., Leathwick, J. R., Lehmann, A., Li,
Scott H. Harris  https://orcid.org/0000-0002-4832-8393 J., Lohmann, L. G., Loiselle, B. A., Manion, G., Moritz, C., Nakamura,
M., Nakazawa, Y., Overton, J. M. M., Townsend Peterson, A., …
REFERENCES Zimmermann, N. E. (2006). Novel methods improve prediction of
species' distributions from occurrence data. Ecography, 29, 129–­151.
AAAS (2018). What percentage of submissions does Science accept?. AAAS
Elith, J., Leathwick, J. R., & Hastie, T. (2008). A working guide to boosted
Science Contributors. Retrieved from http://www.scien​cemag.org/
regression trees. Journal of Animal Ecology, 77, 802–­813. https://doi.
site/featu​re/contr​ibinf​o/faq/index.xhtml​-­pct_faq
org/10.1111/j.1365-­2656.2008.01390.x
Andrén, H., & Andren, H. (1994). Effects of habitat fragmentation
Elliott, L. P., & Brook, B. W. (2007). Revisiting Chamberlin: Multiple
on birds and mammals in landscapes with different proportions
working hypotheses for the 21st century. BioScience, 57, 608–­614.
of suitable habitat: A review. Oikos, 71, 355–­366. https://doi.
https://doi.org/10.1641/B570708
org/10.2307/3545823
Elsevier (2015). 5 ways you can ensure your manuscript avoids the desk re-
Ayala, F. J. (2009). Darwin and the scientific method. Proceedings of
ject pile. Elsevier Connect: https://www.elsev​ier.com/autho​rs-­updat​
the National Academy of Sciences, 106, 10033–­10039. https://doi.
e/story/​p ubli​s hing​-­t ips/5-­w ays-­you-­c an-­e nsur​e -­your-­m anus​c ript​
org/10.1073/pnas.09014​0 4106
-­avoid​s-­the-­desk-­rejec​t-­pile
Ayres, M. P., & Lombardero, M. J. (2017). Forest pests and their manage-
Fahrig, L. (2003). Effects of habitat fragmentation on biodiversity. Annual
ment in the Anthropocene. Canadian Journal of Forest Research, 48,
Review of Ecology, Evolution, and Systematics, 34, 487–­515.
292–­3 01. https://doi.org/10.1139/cjfr-­2017-­0 033
Glass, D. J., & Hall, N. (2008). A brief history of the hypothesis. Cell, 134,
Baker, M. (2016). 1,500 scientists lift the lid on reproducibility. Nature,
378–­381. https://doi.org/10.1016/j.cell.2008.07.033
533, 452–­454.
Golub, T. (2010). Counterpoint: Data first. Nature, 464, 679. https://doi.
Bates, D., Mächler, M., Bolker, B., & Walker, S. (2015). Fitting linear mixed-­
org/10.1038/464679a
effects models using lme4. Journal of Statistical Software, 67, 1–­48.
Guisan, A., Tingley, R., Baumgartner, J. B., Naujokaitis-­Lewis, I.,
Bates, D., Maechler, M., & Bolker, B. (2018). ‘lme4’ Linear mixed-­effects
Sutcliffe, P. R., Tulloch, A. I. T., Regan, T. J., Brotons, L., McDonald-­
models using S4 classes. R Core Team. Retrieved from https://cran.r-­
Madden, E., Mantyka-­Pringle, C., Martin, T. G., Rhodes, J. R.,
proje​c t.org/web/packa​ges/lme4/lme4.pdf
Maggini, R., Setterfield, S. A., Elith, J., Schwartz, M. W., Wintle, B.
Beale, C. M., Lennon, J. J., & Gimona, A. (2008). Opening the climate en-
A., Broennimann, O., Austin, M., … Buckley, Y. M. (2013). Predicting
velope reveals no macroscale associations with climate in European
species distributions for conservation decisions. Ecology Letters, 16,
birds. Proceedings of the National Academy of Sciences, 105, 14908–­
1424–­1435. https://doi.org/10.1111/ele.12189
14912. https://doi.org/10.1073/pnas.08035​06105
Hansen, M. C., Potapov, P. V., Moore, R., Hancher, M., Turubanova, S. A.,
Betts, M., Diamond, T., Forbes, G. J., Villard, M.-­A., & Gunn, J. (2006). The
Tyukavina, A., Thau, D., Stehman, S. V., Goetz, S. J., Loveland, T. R.,
importance of spatial autocorrelation, extent and resolution in predict-
Kommareddy, A., Egorov, A., Chini, L., Justice, C. O., & Townshend,
ing forest bird occurrence. Ecological Modeling, 191, 197–­224.
 |
5776       BETTS et al.

J. R. G. (2013). High-­resolution global maps of 21st-­century forest Ranehill, E., Dreber, A., Johannesson, M., Leiberg, S., Sul, S., & Weber, R. A.
cover change. Science, 342, 850–­853. https://doi.org/10.1126/scien​ (2015). Assessing the robustness of power posing: No effect on hormones
ce.1244693 and risk tolerance in a large sample of men and women. Psychological
Houlahan, J. E., McKinney, S. T., Anderson, T. M., & McGill, B. J. (2017). Science, 26, 653–­656. https://doi.org/10.1177/09567​97614​553946
The priority of prediction in ecological understanding. Oikos, 126, Romesburg, H. C. (1981). Wildlife science: Gaining reliable knowl-
1–­7. edge. The Journal of Wildlife Management, 45, 293–­313. https://doi.
Hutson, M. (2018). AI researchers allege that machine learning is al- org/10.2307/3807913
chemy. Science Posted in: Technology May 3, 2018. https://doi. Rosen, J. (2016). Research protocols: A forest of hypotheses. Nature, 536,
org/10.1126/scien​ce.aau0577 239–­241. https://doi.org/10.1038/nj761​5-­239a
Illán, J. G., Thomas, C. D., Jones, J. A., Wong, W.-­K ., Shirley, S. M., & Betts, Schimel, J. (2011). Writing science: How to write papers that get cited and
M. G. (2014). Precipitation and winter temperature predict long-­term proposals that get funded. Oxford University Press.
range-­scale abundance changes in Western North American birds. Sells, S. N., Bassing, S. B., Barker, K. J., Forshee, S. C., Keever, A. C., Goerz, J.
Global Change Biology, 20, 3351–­3364. https://doi.org/10.1111/ W., & Mitchell, M. S. (2018). Increased scientific rigor will improve reli-
gcb.12642 ability of research and effectiveness of management. Journal of Wildlife
Lack, D. (1954). The natural regulation of animal numbers. Clarendon Press. Management, 82, 485–­494. https://doi.org/10.1002/jwmg.21413
Landy, F. (1986). Stamp collecting versus science -­Validation as hypoth- Smaldino, P. E., & McElreath, R. (2016). The natural selection of bad science.
esis testing. American Psychologist, 41, 1183–­1192. Royal Society Open Science, 3(9), 160384. https://doi.org/10.1098/
Loehle, C. (1987). Hypothesis testing in ecology: Psychological aspects rsos.160384
and the importance of theory maturation. The Quarterly Review of Stafleu, F. (1971). Lamarck: The birth of biology. Taxon, 20, 397–­4 42.
Biology, 62, 397–­4 09. https://doi.org/10.2307/1218244
May, R. M. (1981). The role of theory in ecology. American Zoologist, 21, Sutherland, W. J., Spiegelhalter, D., & Burgman, M. A. (2013). Twenty tips
903–­910. for interpreting scientific claims. Nature, 503, 335–­337.
Mayr, E. (2002). What evolution is. Basic Books. Symes, L. B., Serrell, N., & Ayres, M. P. (2015). A pactical guide for mento-
Milner, S. (2018). Newton didn't frame hypotheses. Why should we? Real ring scientific inquiry. The Bulletin of the Ecological Society of America,
Clear Science: Posted in: Physics Today April 25, 2018. https://doi. 96, 352–­367.
org/10.1063/PT.2016.2013.20180​424a Tewksbury, J. J., Anderson, J. G. T., Bakker, J. D., Billo, T. J., Dunwiddie, P.
Munafò, M., Nosek, B., Bishop, D., Button, K. S., Chambers, C. D., du Sert, W., Groom, M. J., Hampton, S. E., Herman, S. G., Levey, D. J., Machnicki,
N. P., Simonsohn, U., Wagenmakers, E.-­J., Ware, J. J., & Ioannidis, J. N. J., del Rio, C. M., Power, M. E., Rowell, K., Salomon, A. K., Stacey,
P. A. (2017). A manifesto for reproducible science. Nature Human L., Trombulak, S. C., & Wheeler, T. A. (2014). Natural history's place in
Behavior, 1, 0021. science and society. BioScience, 64, 300–­310. https://doi.org/10.1093/
Myers, N., Mittermeier, R. A., Mittermeier, C. G., da Fonseca, G. A. B., biosc​i/biu032
& Kent, J. (2000). Biodiversity hotspots for conservation priorities. Thompson, J. N. (2005). The geographic mosaic of coevolution. University
Nature, 403, 853. of Chicago Press.
O'Neill, R. V., Johnson, A. R., & King, A. W. (1989). A hierarchical frame- Thornton, A., & Lee, P. (2000). Publication bias in meta-­analysis: Its
work for the analysis of scale. Landscape Ecology, 3, 193–­205. https:// causes and consequences. Journal of Clinical Epidemiology, 53, 207–­
doi.org/10.1007/BF001​31538 216. https://doi.org/10.1016/S0895​- ­4356(99)00161​- ­4
Opthof, T., Furstner, F., van Geer, M., & Coronel, R. (2000). Regrets van Dijk, D., Manor, O., & Carey, L. B. (2014). Publication metrics and
or no regrets? No regrets! The fate of rejected manuscripts. success on the academic job market. Current Biology, 24, R516–­R517.
Cardiovascular Research, 45, 255–­258. https://doi.org/10.1016/ https://doi.org/10.1016/j.cub.2014.04.039
S0008​-­6363(99)00339​-­9 Voosen, P. (2017). The AI detectives. Science, 357, 22–­27. https://doi.
Osborne Patrick, E., Foody Giles, M., & Suárez-­Seoane, S. (2007). Non-­ org/10.1126/scien​ce.357.6346.22
stationarity and local approaches to modelling the distributions Weinberg, R. (2010). Point: Hypotheses first. Nature, 464, 678. https://
of wildlife. Diversity and Distributions, 13, 313–­323. https://doi. doi.org/10.1038/464678a
org/10.1111/j.1472-­4642.2007.00344.x Wilson, K. A., McBride, M. F., Bode, M., & Possingham, H. P. (2006).
Paine, C. E. T., & Fox, C. W. (2018). The effectiveness of journals as arbi- Prioritizing global conservation efforts. Nature, 440, 337. https://doi.
ters of scientific impact. Ecology and Evolution, 8, 9666–­9685. org/10.1038/natur​e 04366
Paine, R. T. (1966). Food web complexity and species diversity. American Wolff, J. O., & Krebs, C. J. (2008). Hypothesis testing and the scientific
Naturalist, 100, 65–­75. https://doi.org/10.1086/282400 method revisited. Acta Zoologica Sinica, 54, 383–­386.
Peters, R. H. (1991). A critique for ecology. Nature. Cambridge University
Press.
Pettorelli, N., Nagendra, H., Rocchini, D., Rowcliffe, M., Williams, R., S U P P O R T I N G I N FO R M AT I O N
Ahumada, J., de Angelo, C., Atzberger, C., Boyd, D., Buchanan, G., Additional supporting information may be found online in the
Chauvenet, A., Disney, M., Duncan, C., Fatoyinbo, T., Fernandez, N., Supporting Information section.
Haklay, M., He, K., Horning, N., Kelly, N., … Wegmann, M. (2017).
Remote sensing in ecology and conservation: Three years on. Remote
Sensing in Ecology and Conservation, 3, 53–­56.
How to cite this article: Betts MG, Hadley AS, Frey DW, et al.
Platt, J. R. (1964). Strong inference. Science, 146, 347–­353.
Popper, K. (1959). The logic of scientific discovery. Basic Books. When are hypotheses useful in ecology and evolution?. Ecol
Quinn, J. F., & Dunham, A. E. (1983). On hypothesis testing in ecology Evol. 2021;11:5762–­5776. https://doi.org/10.1002/ece3.7365
and evolution. The American Naturalist, 122, 602–­617. https://doi.
org/10.1086/284161