1 Foundations of inferential statistics
Inferential statistics is the part of statistics concerned with drawing conclusions about a larger population from information observed in a sample. Because a sample represents only part of the whole, any conclusion carries uncertainty. The field uses probability to describe that uncertainty and to judge how confidently a result can be extended beyond the data that were directly collected.
At its core, inferential work links observed measurements to broader claims. It includes estimation of unknown quantities, tests of proposed explanations, and quantitative modeling of relationships among variables. These methods are used across the sciences, social research, engineering, medicine, business, and other areas where complete population data are impractical or impossible to obtain.
1.1 Population and sample
A population is the full set of individuals, events, measurements, or objects of interest. A sample is a smaller subset taken from that population. In practice, researchers study samples because measuring every member of a population may be too costly, time-consuming, or impossible.
The usefulness of inference depends on how well the sample reflects the population. If the sample is representative, results from it are more likely to generalize. If it is distorted by selection problems or other sources of imbalance, conclusions may be unreliable.
1.2 Parameters and statistics
A parameter is a numerical characteristic of a population, such as a mean, proportion, or variance. Because parameters are usually unknown, they are estimated from sample data. A statistic is any numerical summary computed from the sample, such as the sample mean or sample proportion.
Statistics serve as observable evidence, while parameters are the underlying quantities being inferred. Inferential methods use statistics to estimate parameters and to evaluate how plausible particular population values are.
1.3 Sampling variability
Different samples drawn from the same population rarely produce identical results. This natural fluctuation is called sampling variability. It arises because each sample contains a different mix of observations, even when drawn under the same procedure.
Sampling variability is central to inferential statistics. It explains why two samples from the same population can produce different means, percentages, or regression slopes. Quantifying this variation allows statisticians to assess the reliability of estimates and tests.
1.4 Probability and uncertainty
Probability provides the mathematical language for uncertainty. In inferential statistics, probability models are used to describe how likely sample outcomes are under assumed conditions. This makes it possible to evaluate whether observed data are ordinary fluctuations or unusually strong evidence for a conclusion.
Uncertainty in inference is not treated as mere ignorance; it is measured and incorporated into the method itself. Confidence intervals, p-values, and posterior probabilities are examples of tools that express uncertainty in different ways.
2 Data collection and sampling
Inferential conclusions are only as sound as the data on which they are based. Careful data collection aims to reduce distortions and ensure that the sample provides a useful basis for generalization. Sampling design, representativeness, and sample size all shape the quality of the inference.
2.1 Sampling methods
Sampling methods are the procedures used to select units from a population. Good sampling methods seek to avoid systematic distortion while making efficient use of available resources. The choice of method often depends on the population structure and the practical constraints of the study.
2.1.1 Simple random sampling
In simple random sampling, every member of the population has an equal chance of being selected. This design is often treated as a basic reference standard because it minimizes selection bias when implemented correctly.
Its main advantage is conceptual simplicity. However, it may be difficult to carry out when populations are large, dispersed, or not fully listed in advance.
2.1.2 Stratified sampling
Stratified sampling divides the population into subgroups, or strata, based on an important characteristic, then samples from each stratum. This approach helps ensure that key segments of the population are represented.
It can increase precision when the strata are internally similar but differ from one another. It is especially useful when some subgroups are small yet important for analysis.
2.1.3 Cluster sampling
Cluster sampling selects groups rather than individual units at first. Entire clusters may then be studied, or a sample may be taken from within selected clusters. This method is often used when a population is widely spread or difficult to enumerate individually.
It can reduce cost and logistical burden, but it may also increase variability if the members of each cluster are very similar to one another.
2.2 Sampling bias
Sampling bias occurs when the selection process systematically favors some outcomes over others. It can arise from incomplete population lists, nonresponse, convenience sampling, or procedures that exclude certain groups.
Bias is especially problematic because it does not disappear simply by increasing the sample size. A large but distorted sample can still lead to misleading conclusions.
2.3 Sample size
Sample size affects the precision and stability of estimates. Larger samples usually produce smaller random error and narrower confidence intervals, although the improvement eventually becomes less dramatic. The required sample size depends on the desired precision, the variability of the data, and the strength of the effect being studied.
Researchers also consider practical limits. Collecting more data may improve accuracy, but beyond a point the additional benefit may not justify the extra time or expense.
3 Estimation
Estimation is the process of using sample information to infer unknown population values. It is one of the central aims of inferential statistics and appears in both descriptive summaries and formal modeling. Estimation can be expressed as a single best guess or as a range of plausible values.
3.1 Point estimation
A point estimate is a single numerical value used to represent an unknown parameter. For example, the sample mean may be used to estimate a population mean, and a sample proportion may estimate a population proportion.
Point estimates are easy to interpret, but they do not by themselves show how uncertain the estimate is. For that reason, they are often paired with interval estimates.
3.2 Interval estimation
Interval estimation provides a range of values believed to contain the parameter with a specified level of confidence. Instead of claiming one exact value, the method acknowledges uncertainty by giving a plausible interval.
This approach is especially useful when sample data are limited or variable. A well-constructed interval conveys both the estimate and the degree of imprecision.
3.2.1 Confidence intervals
A confidence interval is a range computed from sample data using a procedure that would capture the true parameter a specified proportion of the time in repeated sampling. A 95% confidence interval is one of the most common forms.
The interval should not be read as a direct probability statement about the parameter in a single completed study. Rather, it reflects the long-run performance of the method used to generate it.
3.2.2 Margin of error
The margin of error describes the amount by which an estimate may differ from the true population value within a confidence interval framework. It is often presented as the distance from the point estimate to the upper or lower bound of the interval.
Smaller margins of error indicate greater precision. They are influenced by sample size, variability, and the confidence level chosen for the analysis.
3.3 Estimator properties
Estimators are judged by mathematical properties that describe how they behave across repeated samples. These properties help analysts compare methods and choose procedures with desirable long-run performance.
3.3.1 Bias
Bias is the difference between an estimator’s expected value and the true parameter. An unbiased estimator centers on the correct value on average, though individual samples may still differ substantially.
Bias can result from flawed measurement, poor sampling design, or a method that systematically overshoots or undershoots the target.
3.3.2 Consistency
Consistency means that an estimator tends to approach the true parameter as sample size increases. A consistent estimator becomes more accurate in large samples, even if it is somewhat imperfect in small ones.
This property is important because it reflects whether the method improves with more information.
3.3.3 Efficiency
Efficiency refers to the relative precision of an estimator. Among estimators with similar bias, a more efficient one has less variability and therefore yields tighter estimates.
Efficiency matters because two methods may be correct in principle but differ substantially in how stable they are from sample to sample.
4 Hypothesis testing
Hypothesis testing is a formal framework for evaluating claims about a population using sample data. It begins with a proposed statement and assesses whether the observed evidence is compatible with that statement. The method does not prove a claim; it measures how strongly the data support or challenge it.
4.1 Null and alternative hypotheses
The null hypothesis is the default claim, often representing no effect, no difference, or no association. The alternative hypothesis expresses the competing possibility that some effect or relationship exists.
These hypotheses are set up so that the data can be used to judge whether the null appears implausible. The framework is especially common in experimental and observational research.
4.2 Test statistics
A test statistic is a computed value that summarizes the evidence in the sample relative to the null hypothesis. It measures how far the observed result is from what the null would predict, taking into account variability.
Different tests use different statistics, such as z, t, chi-squared, or F values. Each has a known reference distribution under the null model.
4.3 Significance level
The significance level, often denoted by alpha, is the threshold used to decide whether evidence against the null is strong enough to reject it. Common choices are 0.05 or 0.01.
This threshold is selected before seeing the data in order to control the rate of false positives under repeated use of the procedure.
4.4 P-values
A p-value is the probability of obtaining a result at least as extreme as the observed one, assuming the null hypothesis is true. Small p-values indicate that the observed data would be unusual under the null model.
A p-value does not measure the probability that the null hypothesis is true. It is a conditional probability tied to the assumptions of the test.
4.5 Type I and Type II errors
A Type I error occurs when a true null hypothesis is wrongly rejected. This is a false positive. A Type II error occurs when a false null hypothesis is not rejected. This is a false negative.
These two kinds of error reflect different risks in statistical decision-making. Reducing one often increases the other, so researchers must balance them according to the context.
4.6 Statistical power
Statistical power is the probability that a test will correctly reject a false null hypothesis. Higher power means a greater chance of detecting an actual effect when one exists.
Power depends on effect size, sample size, variability, and the significance level. Studies with low power may miss important findings even when a real relationship is present.
5 Common inferential procedures
A wide range of statistical tests and models are used in inference. Some compare group means or proportions, others assess association between variables, and still others evaluate complex patterns through modeling. The choice of procedure depends on the question being asked and the structure of the data.
5.1 Z-tests and t-tests
Z-tests and t-tests are used to compare sample results with hypothesized values or with other groups. They are most common when working with means or proportions under conditions where the sampling distribution is well understood.
5.1.1 One-sample tests
One-sample tests compare a sample statistic with a known or hypothesized population value. They are used, for example, to ask whether a sample mean differs from a stated benchmark.
The choice between z and t methods often depends on whether the population variability is known and on the size of the sample.
5.1.2 Two-sample tests
Two-sample tests compare the means or proportions of two independent groups. They are widely used in experimental and observational comparisons.
These tests help determine whether any observed difference is likely to be more than random fluctuation.
5.1.3 Paired tests
Paired tests are used when observations are naturally matched, such as measurements taken before and after an intervention on the same subjects. Each pair is treated as a single unit of comparison.
This design can increase sensitivity because it focuses on within-pair differences rather than variation between unrelated cases.
5.2 Chi-squared tests
Chi-squared tests are used with categorical data to compare observed counts with expected counts. They are especially helpful for tables of frequencies and counts.
5.2.1 Goodness-of-fit tests
Goodness-of-fit tests examine whether observed category frequencies match a specified distribution. They ask whether the data follow a proposed pattern closely enough to be considered plausible.
These tests are common when checking whether outcomes conform to a theoretical model.
5.2.2 Tests of independence
Tests of independence assess whether two categorical variables are associated. They are often applied to contingency tables to determine whether the variables vary together more than would be expected by chance.
A significant result suggests association, but not necessarily causation.
5.3 Analysis of variance
Analysis of variance, or ANOVA, compares the means of three or more groups. It evaluates whether group differences are larger than would be expected from random variation alone.
Although the method is named for variance, its practical purpose is often to test differences in average outcomes across multiple conditions or categories.
5.4 Correlation and regression
Correlation and regression examine relationships between variables. Correlation measures the strength and direction of association, while regression models how one variable changes in relation to another.
These methods are used for description, prediction, and explanation, though they do not by themselves establish cause and effect.
5.4.1 Simple linear regression
Simple linear regression describes the relationship between one predictor variable and one outcome variable using a straight-line model. The fitted line summarizes the average change in the outcome associated with changes in the predictor.
This method is useful when the pattern is approximately linear and the main goal is to estimate or predict.
5.4.2 Multiple regression
Multiple regression extends the idea to several predictors at once. It estimates the relationship between each predictor and the outcome while holding the others constant.
This makes it possible to study more complicated data structures, though interpretation requires attention to overlapping relationships among predictors.
6 Distribution-based methods
Many inferential procedures rely on probability distributions that describe how statistics behave across repeated sampling. These distributions provide the reference framework for estimation and testing. Understanding them is essential for interpreting uncertainty and calculating probabilities.
6.1 Sampling distributions
A sampling distribution is the distribution of a statistic over all possible samples of a given size from a population. It shows how a statistic would vary if the sampling process were repeated many times.
Sampling distributions are the basis for standard errors, confidence intervals, and hypothesis tests. They connect individual sample outcomes to long-run behavior.
6.2 Normal distribution
The normal distribution is a symmetric, bell-shaped probability distribution that plays a central role in statistics. Many natural measurements and many sample statistics are approximately normal under suitable conditions.
Its mathematical convenience makes it a foundation for numerous inferential procedures and approximations.
6.3 t-distribution
The t-distribution resembles the normal distribution but has heavier tails. It is used when estimating means with extra uncertainty, especially when the population standard deviation is unknown and the sample is small.
As sample size grows, the t-distribution becomes closer to the normal distribution.
6.4 Chi-squared distribution
The chi-squared distribution arises in methods involving sums of squared standardized values. It is widely used in tests for categorical data and in variance-related inference.
Its shape depends on degrees of freedom, and it becomes less skewed as those degrees increase.
6.5 F-distribution
The F-distribution is based on ratios of variances or mean squares. It appears prominently in analysis of variance and in some regression tests.
Like the chi-squared distribution, it depends on degrees of freedom and is typically right-skewed.
7 Model assessment and assumptions
Inferential procedures are built on assumptions about the data and the model. Checking whether those assumptions are reasonable is an important part of statistical analysis. When assumptions are poorly met, results may be less accurate or less trustworthy.
7.1 Independence
Independence means that one observation does not determine another. Many statistical methods assume that data points are not linked in ways that would distort variability.
Dependence can occur in repeated measures, clustered samples, time series, and other structured settings. Special methods are often needed in such cases.
7.2 Normality
Normality refers to the assumption that data or residuals follow a normal distribution, or are close enough for a method to work well. Many classical procedures are fairly robust to moderate departures from normality, especially with large samples.
Severe nonnormality may affect p-values, interval estimates, and model fit.
7.3 Homoscedasticity
Homoscedasticity means that the spread of errors or residuals is roughly constant across levels of the predictor or fitted values. It is an important assumption in regression and related models.
If the variability changes noticeably, estimates of uncertainty may become unreliable.
7.4 Model diagnostics
Model diagnostics are techniques used to evaluate how well a statistical model fits the data. They may include residual plots, influence measures, and checks for unusual observations.
Diagnostics help identify violations of assumptions, outliers, and areas where the model simplifies the data too much.
7.5 Robust methods
Robust methods are designed to perform reasonably well even when assumptions are imperfect or when data contain outliers. They reduce sensitivity to unusual values or mild departures from ideal conditions.
Such methods are useful when the data are messy or when classical assumptions are difficult to justify.
8 Bayesian inference
Bayesian inference is an approach to statistical reasoning that updates beliefs about unknown quantities using observed data. It combines prior information with evidence from the sample to produce revised conclusions. This framework is widely used in settings where prior knowledge is meaningful and probabilistic interpretation is desired.
8.1 Prior and posterior distributions
A prior distribution expresses beliefs about a parameter before new data are analyzed. After observing data, the prior is updated to a posterior distribution, which reflects the revised state of knowledge.
The posterior combines prior information with the likelihood of the observed data. It serves as the basis for estimation and prediction in Bayesian analysis.
8.2 Bayes’ theorem
Bayes’ theorem is the mathematical rule that links prior beliefs, data, and posterior conclusions. It describes how to reverse conditional information and update probabilities when evidence becomes available.
This theorem is the foundation of Bayesian inference and allows probability to be interpreted as a degree of belief under uncertainty.
8.3 Credible intervals
A credible interval is a range within which a parameter lies with a specified posterior probability. For example, a 95% credible interval indicates that, given the model and data, there is a 95% posterior probability that the parameter falls in that range.
This differs from a frequentist confidence interval in interpretation, though the numerical results may sometimes be similar.
8.4 Comparison with frequentist inference
Bayesian and frequentist approaches share many computational tools but differ in interpretation. Frequentist methods focus on long-run sampling behavior, while Bayesian methods assign probabilities directly to parameters after updating prior beliefs.
The two frameworks are often used for similar practical tasks, but they answer questions in different conceptual languages.
9 Interpretation and reporting
Statistical results must be interpreted carefully to avoid overstating what the data show. Clear reporting helps readers understand both the evidence and its limitations. Sound interpretation considers magnitude, uncertainty, and the context of the study.
9.1 Statistical significance
Statistical significance indicates that an observed result would be unlikely under the null hypothesis at a chosen threshold. It is a statement about evidence, not about importance or real-world impact.
A statistically significant result may still be small in magnitude, and a nonsignificant result may still be meaningful in a practical sense.
9.2 Practical significance
Practical significance concerns whether a result has real-world relevance or consequences. A finding can be statistically detectable yet too small to matter in practice.
This distinction is important in applied research, where decision-making often depends on effect size and context rather than on significance alone.
9.3 Effect size
Effect size measures the magnitude of a relationship, difference, or association. Unlike significance tests, effect sizes emphasize how large an effect is rather than whether it is simply detectable.
They help compare results across studies and provide a more substantive sense of importance.
9.4 Multiple comparisons
Multiple comparisons arise when many tests are performed at once. As the number of tests increases, the chance of at least one false positive also increases.
To address this issue, analysts may adjust significance thresholds or use methods that control error rates across the set of tests.
9.5 Reproducibility and transparency
Reproducibility means that an analysis can be repeated with the same data and methods to obtain the same results. Transparency refers to clear reporting of data collection, processing, assumptions, and analytical choices.
These practices support trust in statistical conclusions and make it easier for others to evaluate, replicate, and build on the work.