A null hypothesis is a statistical or scientific proposition that represents a specified baseline condition against which observed data are evaluated. It commonly states that there is no difference, no association, no effect or no departure from a defined model, although the exact form depends on the research question. For example, a study might define a null hypothesis that two treatments have the same mean outcome, or that two measured variables have no association under a specified statistical model.
The null hypothesis is not necessarily a claim that nothing in the world is happening. It is a formal reference proposition used to determine how compatible the observed evidence is with a particular baseline. Its precise mathematical definition matters because statistical conclusions depend on what the null model actually specifies.
The concept of the null hypothesis is closely associated with frequentist hypothesis testing. Researchers begin by defining a null hypothesis and, where appropriate, an alternative hypothesis representing a different effect or relationship. They then collect data and calculate a test statistic whose behaviour under the null hypothesis is known or approximated.
The resulting p value measures how unusual the observed result, or a result more extreme under the specified test, would be if the null hypothesis and relevant assumptions were correct. A small p value indicates that the data are relatively incompatible with the null model, but it does not establish that the alternative hypothesis is true with a corresponding probability. Nor does it measure the probability that the null hypothesis itself is true.
A common misconception is that failing to reject the null hypothesis means that the null hypothesis has been proved true. In conventional hypothesis testing, failure to reject means that the available evidence does not provide sufficiently strong grounds under the chosen procedure to reject the specified null. This may occur because the null is approximately correct, because the true effect is very small, because the data are noisy or because the study lacks statistical power.
A study with a small sample may fail to detect an important effect that a larger study could identify. Researchers therefore benefit from reporting effect estimates and uncertainty rather than treating a nonsignificant result as evidence of exact equality. Equivalence tests and noninferiority tests can be used when the research question specifically concerns whether differences are sufficiently small to meet a predefined criterion.
The null hypothesis can take forms more complex than a simple statement of equality. In regression analysis, it might specify that a particular coefficient is zero. In a contingency table, it might specify independence between categorical variables. In genetics, it could represent a particular inheritance model. In time series analysis, it could specify the absence of a particular temporal relationship. In model comparison, a null model may represent a simpler mechanism against which a more complex model is evaluated. The important point is that the null hypothesis must be defined precisely enough that the probability model used for testing can be specified. Different null hypotheses can produce different statistical tests and different conclusions from the same dataset.
A null hypothesis is not always the same thing as a scientific claim that an effect does not exist. Researchers sometimes formulate a null hypothesis because it provides a mathematically convenient reference model. Suppose a drug is expected to reduce a symptom score. A null hypothesis might state that the average difference between treatment and control is zero. If the observed difference is sufficiently incompatible with this model, the null may be rejected.
The scientific interpretation then depends on effect magnitude, uncertainty, experimental design, possible biases and biological mechanisms. A statistically detectable difference can be too small to matter in practice, while a biologically important effect can be difficult to detect with limited data. The null hypothesis is therefore one component of evidence evaluation rather than a complete statement of scientific meaning.
The distinction between the null hypothesis and the alternative hypothesis is important. The alternative hypothesis represents the set of effects or relationships that differ from the null, but its exact form depends on the test. A two sided alternative may allow effects in either direction, whereas a one sided alternative specifies a direction. The choice should normally be justified by the scientific question and made before examining the data because selecting the direction after seeing the results can distort statistical interpretation. Researchers must also account for multiple testing when many null hypotheses are examined because the probability of obtaining apparently unusual results increases as the number of tests increases.
Statistical tests depend on assumptions in addition to the null hypothesis. These may include independence, distributional properties, equal variances, correct model specification or appropriate sampling procedures. If assumptions are substantially violated, the nominal p value may not have its intended interpretation. Researchers can use diagnostic procedures, robust methods, permutation tests, nonparametric approaches or alternative models depending on the problem. The null hypothesis itself can also be scientifically unrealistic if it ignores known structure in the data. A useful statistical test therefore requires a null model that corresponds meaningfully to the research question and an analytical procedure appropriate to the data generating process.
The historical development of null hypothesis testing is associated particularly with twentieth century statistical methodology, including work by Ronald Fisher and the separate development of significance testing and hypothesis testing by other statisticians. Modern statistical practice recognises that p values, confidence intervals, effect sizes, likelihood methods and Bayesian approaches answer different questions. Some scientific fields use null hypothesis significance testing extensively, while others emphasise estimation, model comparison or Bayesian inference.
The general concept remains useful because scientists frequently need to compare observations with a defined baseline model. Its correct interpretation requires remembering that rejecting a null hypothesis does not automatically prove a preferred explanation, and failing to reject it does not establish that the null is exactly true.
