Statistics Is Not a Coin Toss: One Sample, Measured Uncertainty

Every eligible voter already has a response at the time of study. The finite population and its outcomes are real and fixed; our knowledge of them is incomplete.

A census can be too slow, costly, or inaccessible. Statistics exists because complete observation is often impractical, not because the population itself must be random.

From a large fixed population, the survey selects and questions one smaller group. That observed sample is evidence; it is not the whole population and not a sequence of repeated elections.

No. A sampling design must give relevant regions and groups a credible path into the sample. Representation comes from the design, not merely from having observations.

Simple random, stratified, systematic, cluster, and multistage designs create different known selection probabilities. Equal probability is one option, not a requirement.

Once five responses are observed, their values are known. Uncertainty is no longer about those five answers; it is about what they imply for the unobserved population.

For person i, the outcome Y_i is fixed while the inclusion indicator I_i is random under the sampling design. An estimate changes because selected people change, not because the population moves.

Selecting one unit removes one available slot, so distinct inclusion indicators have negative covariance. The finite population correction comes from dependent selection without replacement, not from dependent voting decisions.

For responses 1, 1, 1, 0, 0, the sample total is 3, the sample mean and proportion are 0.60, and the sample standard deviation is about 0.548. These summarize only the observed five.

Mean answers average support, standard deviation describes variation, proportion answers how many support, and sample size records how many were observed. Different questions require different statistics.

The sample proportion p-hat=3/5=0.60 is used as an estimate of the unknown population proportion p. Estimation is a disciplined prediction about unseen units, not a declaration that sample and population are identical.

Yes. With the population fixed, different samples permitted by the design can yield proportions such as 0.20, 0.40, 0.60, or 0.80. These are possible surveys, not repeated elections.

Apply the same statistic to every possible sample allowed by the design. Comparing the resulting estimates creates the sampling distribution used to study accuracy and uncertainty.

Different selected units produce different estimates. Sampling variance measures their spread around the design expectation, while standard error is the typical size of that variation in the estimator's original units.

Across four possible samples, the mean ranges from 0.20 to 0.80, the standard deviation moves only from about 0.45 to 0.55, and sample size stays 5. Variability belongs to a specific statistic under a specific design.

Use a sample mean for a population mean, a sample proportion for a binary proportion, Horvitz-Thompson for totals under unequal selection, and ratio or regression estimators when useful auxiliary information exists.

The sample mean suits equal-probability designs, sample proportion suits binary outcomes, Horvitz-Thompson corrects unequal inclusion probabilities, and regression uses related auxiliary data. The best estimator is the justified one.

No. A real election or survey may be observed only once. Repetition lives in the set of samples the design could have selected, allowing uncertainty to be studied without rerunning the event.

They are tied to a time, population, or treatment context that cannot simply be replayed unchanged. Statisticians therefore quantify uncertainty from one observed sample and a defensible data-generating or sampling design.

Define the fixed population, choose a probability design, observe one sample, compute a matching statistic, estimate the population quantity, compare possible estimates, measure uncertainty, and apply finite-population correction when sampling without replacement is substantial.