A research hypothesis is a concise, testable statement about how a system behaves or why a phenomenon occurs. It should propose an explanation, lead to specific, observable predictions; and be possible to accept or reject with attainable data.
What makes a good hypothesis
A strong hypothesis in the BIRD Lab is:
- Grounded in prior work. Based on literature, theory, or existing data and models, so you can point to why the expectation is reasonable.
- Explicit about how expectation leads to predictions. It states what you expect (a relationship, pattern, or mechanism), and that expectation implies measurable predictions.
- Testable and falsifiable. Some conceivable data would count against it, and you can distinguish it from competing hypotheses.
Common hypothesis forms
Use whichever best matches your question.
Relationship-based. If X changes, Y will change in [direction], because [reason]. Example: “If gust intensity increases, peak roll rate during recovery will increase because larger corrective torques are required.”
Pattern-based. [Behavior or field] will show [specific pattern] under [conditions], consistent with [reason]. Example: “In formation flight, follower positions will cluster near the leader’s upwash region.”
Mechanism-based. [Phenomenon] occurs because [mechanism], leading to [observable consequence]. Example: “Rapid yaw turns are generated primarily by asymmetric wingbeat kinematics, leading to characteristic asymmetries in wingtip paths.”
What is not a hypothesis
Some statements sound research-y but are not hypotheses:
- Goals, not expectations: “We want to understand how birds steer in clutter.”
- Methods, not predictions: “We will use high-speed cameras to record wing motion.”
- Descriptive plans: “We will measure wingbeat frequency at different speeds.” A hypothesis must add an expectation about how the measurements will differ.
- Vague or unfalsifiable claims: “Birds use complex strategies to fly efficiently.” Nothing would clearly show it wrong.
- Bare predictions with no context: “Wingbeat frequency will increase.” A hypothesis needs conditions and reasoning.
- Tautologies: “Higher lift allows the bird to stay aloft.” A definition, not an explanation.
- Design wishes: “Our controller should be robust across all conditions.” A goal, not a testable claim.
From question to hypotheses
- Define the phenomenon. What exactly are you trying to explain or predict?
- Summarize what’s known. A short paragraph or bullet list from the literature and prior data.
- List plausible explanations. Morphology, aerodynamics, control, sensing, environment, task demands.
- Write competing hypotheses. \(H_0\) (null), then \(H_1\), \(H_2\) as alternative mechanisms or relationships.
- Derive discriminating predictions. For each, “if this is true, we should see ___,” focusing on predictions that differ across hypotheses.
- Check feasibility. Can you measure what you need with available tools and time? Plan how you’ll test each prediction; see Data analysis.
Hypothesis checklist
Before you commit a hypothesis to your thesis or proposal:
- Grounded in literature, data, or theory
- States a clear expectation (relationship, pattern, or mechanism)
- Implies specific, measurable predictions
- Is falsifiable (you can say what would contradict it)
- Has at least one competing hypothesis, plus a null
- Your planned experiment or analysis can distinguish between hypotheses
- Wording is precise, neutral, and concise
- You can state the core hypothesis in one or two presentation-ready sentences
Worked example: what predicts flight behavior?
A competing-hypotheses example based on Altshuler et al. (2025).
Question: which traits best predict variation in flight behavior across bird species? Framed as competing hypotheses:
- \(H_0\) (null): flight behavior is not meaningfully associated with the measured traits; classification accuracy is similar to shuffled data.
- \(H_1\) (wing shape): static wing shape is the main predictor, and shape-based models classify behavior better than chance.
- \(H_2\) (range of motion): wing range of motion is the main predictor, and ROM-based models outperform shape- and mass-based models.
- \(H_3\) (body mass): body mass is the main predictor, and mass-based models show the highest accuracy.
Each hypothesis implies different predictions about classification accuracy, how each trait compares against a shuffled null, and relative importance. Writing those predictions as comparisons between classification accuracies \(a\) makes explicit what each hypothesis stakes itself on, and therefore what would falsify it:
\[H_0:\ a_\text{shape} \approx a_\text{ROM} \approx a_\text{mass} \approx a_\text{shuffled} \qquad H_2:\ a_\text{ROM} > \max\left(a_\text{shape},\, a_\text{mass}\right) > a_\text{shuffled}\]If ROM-based models beat shape- and mass-based models and their shuffled controls, that supports \(H_2\) over \(H_1\) and \(H_3\). Note that the null is a claim about the shuffled baseline, not about zero accuracy: a classifier can look accurate simply because some behaviors are more common than others, which is exactly what the shuffled comparison controls for.
Spot something to improve? Anyone in the lab can edit this guide. Edit this page suggest a change how edits work
Last updated