When doing an investigation, scientists follow steps similar to those in the flowchart below. These steps are called the scientific method: a structured cycle of observation, question, hypothesis, prediction, experiment, analysis, and communication. The method is not a rigid checklist but a loop that repeats as evidence accumulates. It exists to make claims testable and to reduce the role of bias in drawing conclusions. This explainer covers the mechanism behind each step, the numbers that govern good design, and where the method breaks down.
The mechanics of hypothesis testing
A hypothesis is a proposed explanation that generates predictions you can observe. The key property is falsifiability: a useful hypothesis must specify what result would prove it wrong. "Plants grow better with light" is weak because almost any outcome fits it. "Bean seedlings given 12 hours of light grow taller than those given 6 hours" is strong because it names a measurable comparison and a direction.
The mechanism that makes this work is the controlled comparison. Scientists change one variable at a time while holding others constant. The control group establishes what happens without the treatment. The experimental group receives the treatment. The difference between them, adjusted for random variation, is the effect being tested.
This is why the flowchart shows an arrow looping backward. If results contradict the prediction, the hypothesis is revised or rejected, and the cycle restarts. In practice, working scientists rarely march through the steps in a straight line. They often run exploratory observations first, then formalize a hypothesis once a pattern emerges. The flowchart is a model of rigor, not a literal transcript of daily lab work.
How the steps map to real research
The canonical sequence has seven stages. Each has a distinct failure mode worth understanding.
Observation: noticing a pattern or anomaly in prior data or the natural world. Failure mode: noticing only patterns that confirm existing belief.
Question: framing a specific, answerable inquiry. Failure mode: questions so broad that no experiment can address them.
Hypothesis: a testable, falsifiable explanation. Failure mode: hypotheses that predict nothing specific.
Prediction: the observable outcome the hypothesis implies. Failure mode: predictions stated after the data are seen.
Experiment: a controlled test that isolates variables. Failure mode: confounds, where a second variable changes alongside the one studied.
Analysis: statistical evaluation of results. Failure mode: testing many outcomes and reporting only the favorable ones.
Communication: peer review and publication. Failure mode: selective reporting that hides null results.
Timeline: a realistic research cycle
| Stage | Typical duration | Output | Common bottleneck |
|---|---|---|---|
| Literature review | 2–8 weeks | Research gap identified | Access to paywalled studies |
| Hypothesis design | 1–4 weeks | Falsifiable statement | Confusing correlation with cause |
| Protocol and ethics review | 2 weeks–6 months | Approved protocol | Institutional review board queue |
| Data collection | 1 month–3 years | Raw dataset | Recruitment or equipment limits |
| Statistical analysis | 2–8 weeks | Effect estimates | Multiple-comparison correction |
| Peer review | 3–12 months | Reviewer reports | Finding qualified reviewers |
| Publication | 1–6 months | Indexed article | Journal backlog |
The totals matter. A clinical trial from idea to publication commonly spans three to seven years. A short ecology field study may complete in under a year. The method's steps are fixed, but their duration is highly variable.
Quantified context: what good design looks like
Statistical power is the probability of detecting a real effect if one exists. Conventional targets are 80 percent power or higher, with a false-positive threshold of 0.05. Underpowered studies waste resources and inflate false positives. A study testing many hypotheses at once must correct for multiple comparisons, because at a 0.05 threshold, roughly 1 in 20 tests will appear significant by chance alone.
Sample size is governed by effect size, variability, and desired power. Detecting a large effect needs far fewer participants than detecting a subtle one. This is why fields studying small effects, such as genetics, require enormous cohorts. Peer review typically involves two to four reviewers, and acceptance rates at selective journals often fall below 15 percent. Pre-registration, where the protocol is filed publicly before data collection, reduces the freedom to cherry-pick results after the fact.
Edge cases and counterintuitive findings
A counterintuitive professional caveat: replication is not optional polish. Landmark analyses have shown that a substantial fraction of published psychology and biomedical findings fail to replicate at their original effect sizes. Many fields now treat a single study as provisional until independent labs reproduce it.
Edge cases complicate the clean flowchart. Observational studies cannot randomize exposure, so they rely on statistical adjustment for confounders, and causation claims remain weaker. Field studies sacrifice control for realism. Exploratory "fishing" analyses can generate hypotheses but cannot confirm them. Null results, though often unpublished, are essential because they correct the record.
Misconceptions that undermine investigations
A hypothesis is not a guess. It is a specific, testable prediction that can fail. A p-value of 0.05 does not mean a 5 percent chance the hypothesis is wrong. It describes how often results this extreme would appear if no real effect existed. Correlation is not causation. A co-occurring trend may reflect a hidden third factor. A failed experiment is not a failed method. Disproven predictions refine theory. Peer review is not a guarantee of truth. It is a filter, not a proof.
Practical actions for running a sound investigation
State the question as a single measurable comparison before collecting any data. Write down the prediction and the disconfirming outcome in advance, then pre-register the protocol. Calculate required sample size using your expected effect size and an 80 percent power target. Keep a detailed lab notebook so another person can repeat the analysis from raw data. Report all results, including nulls, and correct for multiple comparisons explicitly. Seek independent replication before treating a finding as established.
Limitations and scope
The scientific method describes ideal practice, not universal reality. It cannot answer questions that are intrinsically untestable, such as purely moral or aesthetic claims. Effect sizes depend on context, so a result from one population may not transfer to another. Measurement error, publication bias, and funding pressures all distort the published record. The method's strength is that it builds in correction: every claim carries the conditions under which it would be revised. That is why the loop keeps looping rather than ending at a single confirmed answer.