Null & Alternative Hypotheses

A statistical hypothesis is a statement about a parameter of a population — a true, fixed but unknown quantity such as a population mean $\mu$, a proportion $p$, a correlation $\rho$, or a regression slope $\beta$. It is never a statement about the sample itself, because the sample statistic (the sample mean $\bar{x}$, for instance) is already known once the data are collected; there is nothing left to test about it. The null hypothesis, denoted $H_0$, is conventionally the hypothesis of "no effect," "no difference," or "no association" — the default, skeptical position that any observed pattern in the data arose purely from sampling variation. The alternative hypothesis, $H_A$ (sometimes $H_1$), is the hypothesis that some real biological effect exists, and it is what the researcher typically hopes to find support for. Two structural requirements make a hypothesis pair valid. First, $H_0$ and $H_A$ must be mutually exclusive — they cannot both be true. Second, together they must be exhaustive — they must cover every possible true value of the parameter, so that if $H_0$ is false, $H_A$ is necessarily true. A pair like $H_0: \mu = 0$ versus $H_A: \mu = 5$ fails this requirement, because it ignores every other possible value of $\mu$ (negative values, or values between 0 and 5, or values above 5). The valid complement is $H_A: \mu \neq 0$, which sweeps up everything $H_0$ does not claim. The choice of $H_A$'s shape is where biological judgment enters. A two-sided (non-directional) alternative, $H_A: \mu \neq \mu_0$, allows the true parameter to differ from the null value in either direction and is appropriate whenever an effect in either direction is biologically plausible and would be scientifically interesting — most biological questions fall here. A one-sided (directional) alternative, $H_A: \mu > \mu_0$ or $H_A: \mu < \mu_0$, restricts attention to a single direction and is justified only when the researcher has a strong, pre-specified biological or physical reason to care exclusively about one direction — for instance, testing whether a toxin increases mortality (a decrease would be biologically implausible and, even if observed, would not answer the applied question being asked). Critically, the decision between one- and two-sided framing must be made from biological reasoning laid out before the data are examined, never chosen afterward to make a result look more significant.

Step 1 — Translate the biological question into a parameter. Identify exactly which population quantity answers the question. "Does fertilizer increase crop yield?" becomes a statement about $\mu_{\text{fertilized}} - \mu_{\text{control}}$, the difference in population mean yields. "Is beak depth associated with survival?" becomes a statement about a correlation $\rho$ or regression slope $\beta$. Vague biological language must be pinned to one specific, measurable parameter — this step alone eliminates a large fraction of downstream statistical errors, because an ambiguous parameter makes every later step ambiguous too.

Step 2 — Write $H_0$ as the "nothing happening" statement about that parameter. For a two-sample comparison, $H_0: \mu_1 = \mu_2$ (equivalently $\mu_1 - \mu_2 = 0$). For an association, $H_0: \rho = 0$. For a proportion compared to a known baseline, $H_0: p = p_0$. The null is deliberately the boring, skeptical claim — it is the position that must be disproven by the weight of evidence, mirroring the logic of a legal trial where innocence is presumed until evidence forces rejection.

Step 3 — Decide, from biology alone, whether the alternative should be one- or two-sided. Ask: is a change in the "wrong" direction (a) biologically impossible, (b) scientifically uninteresting even if it occurred, or (c) something the study design cannot detect anyway? If yes to any of these and this reasoning existed before data collection, a one-sided $H_A$ is defensible. Otherwise, default to the two-sided $H_A: \mu_1 \neq \mu_2$. Most published biological hypotheses are two-sided, because organisms routinely surprise us — a drug expected to lower blood pressure occasionally raises it (a "paradoxical" effect), and a symmetric test does not blind the researcher to that possibility.

Step 4 — Confirm mutual exclusivity and exhaustiveness. Check that every logically possible parameter value falls under either $H_0$ or $H_A$, with no gap and no overlap. For a two-sided test: $H_0: \mu = \mu_0$ and $H_A: \mu \neq \mu_0$ partition the real line correctly. For a one-sided test: $H_0: \mu \le \mu_0$ and $H_A: \mu > \mu_0$ partition it correctly (note that a one-sided null is often written as an inequality, not a strict equality, precisely to preserve exhaustiveness).

Step 5 — Lock in the rejection region before touching the data. The sidedness of $H_A$ dictates the shape of the rejection region on the sampling distribution of the test statistic: a two-sided test splits the significance level $\alpha$ across both tails (commonly $\alpha/2$ in each tail), while a one-sided test places the entire $\alpha$ in a single tail. This single decision changes the critical value against which the test statistic is compared, so it must be fixed by design, not by convenience after seeing which tail the data happened to fall in.

The modern null/alternative hypothesis framework is not a single "discovery" like Meselson–Stahl; it is a logical structure built by The Logic of Hypothesis Testing literature, chiefly through the fusion of Ronald Fisher's significance testing with Jerzy Neyman and Egon Pearson's decision-theoretic framework in the 1920s–1930s. But its practical necessity — and the real cost of getting $H_0$/$H_A$ wrong — is best shown through a landmark applied case: the controlled agricultural field experiments that Fisher himself ran at Rothamsted Experimental Station, and a much later cautionary tale from clinical and behavioral science about hypotheses chosen after the data were seen.

Fisher's Rothamsted crop trials (1920s). Fisher was hired to analyze decades of uncontrolled fertilizer records at Rothamsted, and found that without a precisely stated null hypothesis and a randomized design, no amount of data could distinguish a genuine fertilizer effect from field-to-field soil variation. He redesigned the experiments so that plots were randomly assigned to fertilizer treatments, and formulated the test as a sharp null — "the fertilizer has no effect on mean yield" — against a two-sided alternative, since both yield increases and unexpected yield decreases (e.g., from soil chemistry interactions) were biologically plausible and worth detecting. This randomized-plot design, married to an explicit $H_0$ versus $H_A$, became the template for essentially all controlled biological experimentation that followed, and it directly demonstrated that vague hypotheses ("fertilizer probably helps") are statistically untestable, while sharp, falsifiable ones are.

The Neyman–Pearson clarification (1928–1933). Neyman and Pearson formalized why the alternative hypothesis must be specified before data collection: without a pre-specified $H_A$, there is no well-defined rejection region and no way to compute the power of a test — the probability of correctly rejecting a false $H_0$. Their lemma (the Neyman–Pearson lemma) proves that, for a given significance level $\alpha$, the most powerful test against a specific alternative uses a likelihood-ratio rejection region tailored to that alternative's direction — mathematically formalizing why one-sided tests are more powerful for a truly one-directional effect, but only when that directionality is genuine and known in advance.

A cautionary empirical result: hypothesis-switching inflates false positives. Simulation and replication studies since the 2010s (e.g., work on "researcher degrees of freedom" by Simmons, Nelson, and Simonsohn, 2011) have quantified what happens when researchers choose the direction of $H_A$, or switch between one- and two-sided tests, after inspecting the data. Simulated analyses show that this kind of post hoc flexibility in hypothesis specification can push the true false-positive rate from a nominal 5% to over 60% across a battery of common "researcher degrees of freedom." This is direct, quantitative evidence that the sequence matters: $H_0$ and $H_A$ must be fixed by biological reasoning before the data are examined, or the entire error-rate machinery of hypothesis testing collapses. Together, these lines of evidence establish the same principle from different angles: a valid test requires a sharply defined, pre-registered $H_0$/$H_A$ pair, and any flexibility introduced after seeing the data destroys the guarantees the whole framework is built to provide.

Example 1 — A genetics cross with an asymmetric expectation. A researcher crosses two heterozygous pea plants for a gene showing classic Mendelian dominance, expecting a 3:1 ratio of dominant (tall) to recessive (short) phenotypes in the $F_2$ generation. Out of 200 offspring, 138 are tall and 62 are short. Step 1 — parameter: let $p$ = true population proportion of tall offspring. Step 2 — $H_0$: $p = 0.75$ (the Mendelian 3:1 expectation holds). Step 3 — sidedness: deviations could plausibly occur in either direction (e.g., linked lethal alleles could reduce the tall proportion, or incomplete penetrance/scoring bias could increase it) — no biological reason restricts this to one side, so: $H_A: p \neq 0.75$ (two-sided). Observed proportion: $\hat p = 138/200 = 0.69$. Expected count of tall = $200 \times 0.75 = 150$; expected short = $50$. A chi-square goodness-of-fit statistic:

With 1 degree of freedom, the critical value at $\alpha = 0.05$ (two-sided by construction of the chi-square test) is 3.84 — so this result sits almost exactly at the boundary, illustrating why the sidedness built into $H_A$ (here, "any deviation from 3:1 counts") changes what counts as significant, and why a borderline result like this demands replication rather than a confident verdict either way.

Example 2 — A one-sided clinical/physiological scenario. A new insecticide is tested on a crop pest with a known baseline mortality rate under the current standard insecticide of $p_0 = 0.40$ within 48 hours. The manufacturer's biological claim is specific: the new formulation contains a more potent neurotoxin and is predicted, on mechanistic grounds established in prior enzyme-inhibition assays, to increase mortality — a decrease is not biologically plausible given the compound's known mode of action, and even if a decrease were somehow observed, it would carry no different practical implication than "no improvement." Step 1 — parameter: $p$ = true population mortality proportion under the new insecticide. Step 2 — $H_0$: $p \le 0.40$ (no improvement over standard). Step 3 — sidedness: because only "does it increase mortality" is scientifically and economically actionable, and the mechanism rules out a plausible decrease, a one-sided alternative is justified a priori: $H_A: p > 0.40$. In a trial of 150 insects, 78 die within 48 hours: $\hat p = 78/150 = 0.52$. Test statistic (large-sample proportion z-test):

For a one-sided test at $\alpha = 0.05$, the critical z is 1.645 (all 5% in the upper tail); $z = 3.00$ exceeds this comfortably, so $H_0$ is rejected in favor of the pre-specified directional claim. Note that if the researcher had instead used a two-sided critical value (±1.96) chosen only because it "felt more rigorous" after seeing a strong positive result, that would be a misuse — the one-sided framing here was justified by mechanism before the trial, not chosen afterward to inflate significance.

Pre-specifying $H_0$ and a correctly sided $H_A$ is now a formal requirement in clinical trial registration (e.g., ClinicalTrials.gov) precisely because regulators learned — from decades of trials that quietly switched primary endpoints or test directions — that undisclosed flexibility in hypothesis framing manufactures false positives. In drug development, a Phase III trial testing whether a new antihypertensive lowers blood pressure uses a one-sided $H_A$ only when regulatory guidance and prior mechanistic data justify ignoring the paradoxical-increase possibility; more often two-sided tests are mandated specifically to catch unexpected harms. In conservation biology, tests of whether a management intervention (e.g., a wildlife corridor) increases genetic diversity are typically two-sided, because a poorly designed corridor could plausibly decrease diversity by funneling gene flow from an unintended source population — an outcome regulators and biologists need the test to be able to detect. In agricultural biotechnology, GMO crop trials testing yield or pest-resistance claims against conventional varieties routinely specify two-sided hypotheses in registration dossiers, precisely to guard against unanticipated yield penalties. And in public health surveillance, tests of whether a new diagnostic assay's sensitivity differs from an existing gold standard are two-sided by default, because both false-positive and false-negative drift are safety-relevant and neither can be assumed away. In every one of these domains, the discipline of writing $H_0$ and $H_A$ correctly — before data collection — is what separates a trustworthy inference from a statistically laundered coincidence, and it is the direct foundation for interpreting The p-Value & Its Interpretation and managing Type I & Type II Errors in any downstream analysis.

Ask the tutor: How do I decide whether my biological question calls for a one-sided or two-sided alternative hypothesis?