A hypothesis is a tentative proposition, stated in advance of the evidence, which asserts a relationship between two or more variables in a form that observation can support or refute. It is neither a guess nor a conclusion: it is a provisional answer framed so precisely that it directs what shall be observed and specifies what result would count against it. A variable is a concept that varies — a property of cases which can take more than one value across the units studied. Caste is a variable because cases differ in it; a property identical in every case is a constant and can explain nothing.
The pair is the elementary machinery of the hypothetico-deductive design. Theory supplies concepts; operationalisation converts each concept into a measurable variable with a stated indicator; the hypothesis states the expected relation between those variables; and the data determine whether the relation holds. Without variables a hypothesis cannot be tested, and without a hypothesis data collection lacks direction — which is why Goode and Hatt described the hypothesis as the researcher's guide, telling one what to look for and what to ignore.
Sources and criteria
Hypotheses come from four principal sources. Theory is the most valued, since a hypothesis deduced from a general proposition returns a result that bears on the theory itself. Prior research yields hypotheses through replication, extension to a new population, or the resolution of contradictory findings. Observation and personal or field experience suggest relationships that have not been formulated. And analogy transfers a relationship established in one domain to another. General culture, folk wisdom and administrative concern supply many as well, though these require restatement in sociological terms.
Not every statement qualifies. A serviceable hypothesis must have conceptual clarity — its terms defined unambiguously and used consistently. It must have an empirical referent, so that each concept can be tied to something observable; statements about divine will or about the innate spirit of a people are not hypotheses. It must be specific, naming the variables, the direction of the expected relation and the population to which it applies, since a vague claim can never be falsified. It must be testable with available techniques and resources. And it should be related to a body of theory, so that whatever the result, the finding contributes to cumulative knowledge rather than remaining an isolated fact.
In statistical testing the substantive claim is expressed as two complementary statements. The null hypothesis asserts no relationship — no difference between groups, no association between variables — and the alternative hypothesis asserts that a relationship exists, either directionally or not. Testing proceeds by attempting to reject the null; failure to reject is not proof of no relationship, only absence of sufficient evidence against it. This asymmetry mirrors the logic of falsification: evidence may refute decisively but confirmation is always provisional.
Types of variable
The independent variable is the presumed cause, the condition whose variation is held to produce variation elsewhere. The dependent variable is the presumed effect, the thing to be explained. An intervening variable stands between the two and transmits the effect, specifying the mechanism: education may raise income by way of occupational access, which is the intervening term. An extraneous or confounding variable lies outside the stated hypothesis but influences the dependent variable and may account for the observed association. A control variable is an extraneous variable that the researcher deliberately holds constant — physically in an experiment, or statistically by examining the relationship separately within each of its categories.
The distinction matters because correlation is not causation. Two variables may covary because one causes the other, because the second causes the first — the problem of the direction of causation — because both are effects of a common third variable, or by chance. A spurious relationship is one that disappears when the confounding variable is controlled. Paul Lazarsfeld's technique of elaboration formalised the response: introduce a third variable and re-examine the original association within its categories. If the relationship vanishes, it was spurious; if it persists at the same strength in every category, the third variable is irrelevant; if it varies across categories, the relationship is specified — conditional on that variable. The minimal conditions for a causal claim remain association, temporal order with the cause preceding the effect, and the elimination of plausible alternative explanations.
Levels of measurement
Following S. S. Stevens, variables are measured at four levels, and the level determines what statistics are permissible. A nominal variable classifies without ordering — religion, mother tongue, marital status — and admits only frequencies, the mode and the chi-square test. An ordinal variable ranks but with unequal or unknown intervals — Likert agreement, class of degree, socio-economic rank — and admits medians and rank-order correlation. An interval variable has equal intervals but an arbitrary zero, so differences are meaningful but ratios are not. A ratio variable has equal intervals and an absolute zero — age, income, years of schooling, number of children — and permits the full range of arithmetic operations. A common error is to treat ordinal codes as though they were ratio numbers and compute means of them.
An Indian illustration
Consider the hypothesis that higher female literacy is associated with lower fertility. Here female literacy, operationalised as the district percentage of literate women aged seven and above from Census data, is the independent variable, and fertility, operationalised as the total fertility rate, is the dependent variable. The relationship is strongly negative across Indian districts, and Kerala with high female literacy and below-replacement fertility contrasts sharply with several districts of the northern plains.
The example also shows why controls are indispensable. Districts with high female literacy also tend to have higher household income, greater urbanisation, better health infrastructure and lower child mortality, and each of these independently depresses fertility, so a bivariate correlation cannot establish that literacy is doing the work. The intervening variables must also be specified: schooling plausibly operates through later age at marriage, greater autonomy in reproductive decisions and knowledge of and access to contraception. The direction of causation requires attention too, since low fertility itself frees girls for schooling. Multivariate analysis of Indian district data, notably by Dreze and Murthi, found that female literacy and child mortality retained substantial independent effects on fertility once income, urbanisation and other district characteristics were controlled — while male literacy did not, a specification that no uncontrolled correlation could have produced.
For the UPSC answer
Define the hypothesis as a testable statement of relationship between variables and the variable as a concept that varies, and make the link to operationalisation explicit, since that is where concepts become measurable. List the sources and then the criteria of a good hypothesis — clarity, empirical referent, specificity, testability, theoretical relation — and add the null and alternative pair to show command of the testing logic. Set out the five types of variable and use the independent–dependent–control triad to explain spuriousness, naming Lazarsfeld's elaboration if the question allows. Close with the female literacy and fertility illustration, where controls for income, urbanisation and child mortality demonstrate concretely why correlation is not causation.
References & further reading
- Goode, W. J. and Hatt, P. K. (1952). Methods in Social Research. McGraw-Hill.
- Young, P. V. (1966). Scientific Social Surveys and Research. Prentice-Hall.
- Lazarsfeld, P. F. (1955). Interpretation of Statistical Relations as a Research Operation. In P. F. Lazarsfeld and M. Rosenberg (eds.), The Language of Social Research. Free Press.
- Stevens, S. S. (1946). On the Theory of Scales of Measurement. Science, 103(2684), 677–680.
- Dreze, J. and Murthi, M. (2001). Fertility, Education and Development: Evidence from India. Population and Development Review, 27(1), 33–63.
- Bryman, A. (2012). Social Research Methods. Oxford University Press.