Statistical Inference

Introduction

Statistical inference is the part of statistics that lets you use data from a sample to reason about a wider population or about the process that generated the data. Instead of only describing what you observed, you ask questions such as: What population value is plausible? How uncertain is an estimate? Are the data surprising under a proposed model? How strong is the evidence for a difference or relationship?

This aiMOOC is designed for Grades 11–13. You should already be comfortable with averages, percentages, basic probability, graphs, and algebra. The course focuses on frequentist inference because it is widely taught at upper-secondary level, while also introducing bootstrap and Bayesian ideas so that you can compare approaches.

Datei:Sampling distribution.png

A useful way to think about inference is as a chain: question → data collection → model → statistic → uncertainty → conclusion. Each link matters. A sophisticated formula cannot repair biased sampling, weak measurement, confounding, or a question that the study design cannot answer.

Learning Goals

By the end of the course, you should be able to:

distinguish a population, sample, parameter, statistic, estimator, estimate, and sampling distribution;
explain how random sampling creates sampling variability;
use standard error to describe the typical variability of an estimator;
explain why larger random samples usually give more precise estimates;
construct and interpret confidence intervals in context;
formulate null and alternative hypotheses;
interpret a p-value correctly and avoid common misinterpretations;
distinguish Type I error, Type II error, significance level, and statistical power;
connect hypothesis tests with confidence intervals;
separate statistical significance from practical importance;
recognize assumptions behind common inferential procedures;
explain the basic logic of bootstrap resampling and compare frequentist and Bayesian interpretations.

From Populations to Samples

A population is the complete set of units or outcomes you want to learn about. A sample is the subset you actually observe. A parameter is a numerical property of the population, such as a population mean $μ$ or population proportion $p$ . A statistic is a numerical summary calculated from the sample, such as the sample mean $\bar{x}$ or sample proportion $\hat{p}$ .

The statistic changes from sample to sample. The population parameter is treated as fixed but unknown in classical frequentist inference. This gap between what is observed and what is unknown is the reason uncertainty must be quantified.

For example, suppose you want to estimate the average weekly study time of all students in a school. You randomly sample 80 students and calculate a mean of 6.4 hours. The value 6.4 is a sample statistic. The unknown average for all students is the population parameter.

Sampling Design Matters

Inference is only as credible as the process that produced the data. A simple random sample gives each possible sample of a given size an equal chance of selection. Other well-designed probability samples include stratified and cluster samples. Convenience samples can be fast, but they may systematically underrepresent parts of the population.

Random sampling supports generalization from a sample to a population. Random assignment in an experiment supports causal comparison between treatment conditions. They solve different problems. A randomized experiment can provide strong causal evidence for the experimental units even if those units are not a random sample of a broad population. A random sample without random treatment assignment can support population estimates but usually cannot establish causation.

Important threats include:

selection bias: some population members are more or less likely to enter the sample;
nonresponse bias: selected individuals who do not respond differ systematically from responders;
measurement bias: the measurement procedure consistently pushes values away from the truth;
confounding: an outside variable is related to both an explanatory variable and an outcome.

Sampling Variability and Sampling Distributions

If you repeatedly draw random samples of the same size from the same population and compute the same statistic, you will not get exactly the same value every time. This natural sample-to-sample fluctuation is sampling variability.

The sampling distribution of a statistic is the probability distribution of that statistic over all possible samples generated by the sampling process. You usually do not observe this distribution directly, but inferential methods are built around it.

Datei:Effect of sample size on distribution of sample means.svg

For a sample mean based on independent observations from a population with mean $μ$ and standard deviation $σ$ :

the expected value of $\bar{X}$ is $μ$ ;
the standard deviation of the sampling distribution is $σ / \sqrt{n}$ .

That second quantity is the standard error of the sample mean when $σ$ is known. In practice, $σ$ is often unknown, so you estimate the standard error with $s / \sqrt{n}$ .

Datei:Mean of sample vs size of sample.svg

The square-root rule is important: multiplying the sample size by four cuts the standard error roughly in half. Increasing sample size improves precision, but it does not automatically remove bias.

Central Limit Theorem

The central limit theorem explains why normal-based inference works so often. Under suitable conditions, the standardized sampling distribution of the sample mean becomes approximately normal as the sample size grows, even when the population distribution itself is not normal. The speed of that approximation depends on the shape of the original population and on dependence or unusual tail behavior.

Datei:Standard Normal Distribution en svg.svg

You should not turn the central limit theorem into the slogan “large samples are always normal.” The theorem concerns the distribution of a properly standardized statistic, not necessarily the raw data. Strong dependence, severe skewness, extreme outliers, or infinite variance can require special methods or much larger samples.

Estimation

An estimator is a rule for using sample data to estimate a parameter. An estimate is the numerical result obtained from one sample. Good estimators are often judged by several properties:

bias: whether the estimator systematically misses the target;
variance: how much the estimator changes from sample to sample;
standard error: the standard deviation of the estimator's sampling distribution;
consistency: whether the estimator tends to get closer to the true parameter as sample size increases.

A point estimate alone hides uncertainty. Reporting “the mean is 42 minutes” can sound more certain than the data justify. An interval estimate adds a range of plausible parameter values produced by a calibrated procedure.

Confidence Intervals

A two-sided confidence interval often has the form:

$estimate \pm critical value \times standard error$

For a population mean when the population standard deviation is unknown and a one-sample t procedure is appropriate:

$\bar{x} \pm t^{*} \frac{s}{\sqrt{n}}$

The critical value $t^{*}$ depends on the confidence level and the degrees of freedom. A 95% confidence interval comes from a procedure designed so that, under its assumptions, about 95% of intervals built from repeated random samples would contain the true parameter.

Datei:Confidenceinterval.svg

Datei:95% confidence interval.svg

A frequentist 95% confidence interval is not literally a statement that there is a 95% probability that the already fixed population parameter lies inside this one computed interval. The 95% describes the long-run coverage performance of the method. In everyday school-level language, it is reasonable to say that the interval gives a range of parameter values compatible with the data and model, while remembering what the confidence level formally means.

Margin of Error and Precision

The margin of error is the critical value multiplied by the standard error. It becomes:

smaller when the sample size increases;
larger when the data are more variable;
larger when you demand a higher confidence level, all else equal.

Precision is not the same as accuracy. A very narrow interval around a biased estimate can be precisely wrong. Good inference needs both a defensible design and a suitable statistical model.

Hypothesis Testing

A statistical hypothesis test asks whether the observed data are sufficiently incompatible with a specified null model to reject that model at a chosen significance level.

The null hypothesis $H_{0}$ usually states a reference value or no-effect condition. The alternative hypothesis $H_{A}$ states the competing possibility. A test statistic measures how far the observed data are from what the null hypothesis predicts, relative to expected sampling variability.

For a one-sample t test of a population mean: $t = \frac{\bar{x} - μ_{0}}{s / \sqrt{n}}$

A large absolute value of $t$ means the sample mean is many estimated standard errors away from the null value.

The p-Value

A p-value is the probability, assuming the null hypothesis and the test model are correct, of obtaining a test statistic at least as incompatible with the null hypothesis as the one observed.

Fehler beim Erstellen des Vorschaubildes:

A small p-value means the observed result would be unusual under the null model. It does not tell you:

the probability that the null hypothesis is true;
the probability that the result happened “by chance”;
the size or importance of an effect;
whether the study design was unbiased;
whether the finding will replicate.

The American Statistical Association has emphasized that scientific conclusions should not be based only on whether a p-value crosses a fixed threshold. Context, study quality, effect size, uncertainty, prior evidence, and the consequences of decisions all matter.

Significance Level, Errors, and Power

Before examining the data, a researcher may choose a significance level $α$ , often 0.05. In a simple decision framework:

a Type I error occurs when a true null hypothesis is rejected;
a Type II error occurs when a false null hypothesis is not rejected;
power is the probability of rejecting the null hypothesis when a particular alternative is true.

Power usually increases when the true effect is larger, the sample size is larger, the measurements are less noisy, or a larger significance level is used. Raising $α$ may increase power, but it also increases the Type I error rate. This is a trade-off, not a free improvement.

Connecting Confidence Intervals and Tests

For many standard two-sided procedures, a 95% confidence interval and a two-sided hypothesis test at significance level 0.05 give matching decisions. If the null value lies outside the 95% confidence interval, the corresponding two-sided test rejects the null hypothesis at the 0.05 level. If the null value lies inside the interval, the test does not reject at that level.

This connection is useful because a confidence interval shows more than a reject-or-not-reject decision. It displays an estimate of effect size and the uncertainty around it.

Suppose a difference in average scores is estimated as 0.8 points with a 95% confidence interval from 0.1 to 1.5 points. The interval excludes zero, so the corresponding two-sided test would be statistically significant at the 0.05 level. Yet a difference of less than two points might still be too small to matter educationally. Statistical significance and practical significance are different questions.

Worked Example: Mean Commute Time

A school wants to estimate the mean one-way commute time of its students. A simple random sample of 36 students gives:

sample mean $\bar{x} = 42.0$ minutes;
sample standard deviation $s = 12.0$ minutes;
sample size $n = 36$ .

The estimated standard error is $12 / \sqrt{36} = 2.0$ minutes. Using a 95% t critical value of about 2.03 gives a margin of error of about $2.03 \times 2.0 = 4.06$ minutes. The confidence interval is therefore approximately 37.94 to 46.06 minutes.

Now test $H_{0} : μ = 45$ against $H_{A} : μ \neq 45$ . The test statistic is:

$t = \frac{42 - 45}{2} = - 1.5$

With 35 degrees of freedom, the two-sided p-value is about 0.14. At the 0.05 significance level, you would not reject the null hypothesis. This agrees with the confidence interval because 45 minutes lies inside the interval.

Notice the wording: do not reject is not the same as prove the null hypothesis true. The sample is compatible with a range of mean commute times, and the interval communicates that range more clearly than a binary decision alone.

Assumptions and Diagnostic Thinking

Every inferential method depends on assumptions. You should identify them before choosing a formula.

For common one-sample mean procedures, important considerations include:

observations should come from a defensible random or approximately independent sampling process;
the sample should be small relative to the population when sampling without replacement, unless a finite-population correction is used;
for small samples, severe skewness or extreme outliers can make t procedures unreliable;
for larger samples, the sampling distribution of the mean is often approximately normal under broad conditions.

For proportions, expected counts must be large enough for normal approximations unless you use exact or simulation-based methods. For regression, assumptions concern the form of the relationship, independence, variability of residuals, and the behavior of errors. Diagnostics should be tied to the method and the scientific question, not treated as a ritual checklist.

Resampling and Bootstrap Inference

The bootstrap estimates sampling uncertainty by resampling from the observed data with replacement. Each bootstrap sample has the same size as the original sample. You compute the statistic for each resample and use the resulting bootstrap distribution to estimate standard errors, bias, or confidence intervals.

Datei:Illustration bootstrap.svg

Bootstrap methods are especially useful when a statistic has a complicated theoretical sampling distribution. However, the bootstrap cannot magically repair a badly biased original sample. It also requires care with very small samples, dependent data, extreme tails, and parameters that behave irregularly.

A simple bootstrap workflow is:

Start with the observed sample.
Draw a new sample of the same size with replacement.
Compute the statistic of interest.
Repeat the resampling many times.
Examine the distribution of bootstrap statistics.
Use that distribution to quantify uncertainty.

A Brief Bayesian Comparison

Frequentist and Bayesian inference answer uncertainty questions differently.

In frequentist inference, parameters are treated as fixed unknown quantities and probability describes the behavior of data or procedures under repeated sampling. A 95% confidence interval is calibrated by long-run coverage.

In Bayesian inference, unknown parameters are represented with probability distributions. A prior distribution is combined with a likelihood from the observed data to produce a posterior distribution. A 95% credible interval can then be interpreted as containing 95% of the posterior probability for the parameter, conditional on the model and prior.

These approaches can sometimes give similar numerical answers, but the interpretations are not interchangeable. At Grades 11–13, the main goal is to recognize that the word “probability” refers to different objects in the two frameworks.

Responsible Statistical Inference

Inference is not only about calculation. Responsible practice includes transparency about data collection, preprocessing, model choice, missing data, multiple comparisons, and uncertainty.

Important habits include:

report effect sizes and uncertainty, not only p-values;
distinguish exploratory analysis from confirmatory testing;
avoid choosing hypotheses after seeing the results and then presenting them as preplanned;
consider whether multiple tests inflate the chance of false positive findings;
examine whether missing data or excluded observations could change the conclusion;
make graphs that reveal variation rather than hide it;
separate statistical association from causal claims;
ask whether the result is practically meaningful for the people or system being studied.

Statistical inference is strongest when mathematics, study design, subject knowledge, and honest communication support one another.

Interactive Tasks

Quiz: Test Your Knowledge

Which statement best describes a parameter? (A numerical property of a population) (!A graph made from sample data) (!A list of all observed values) (!A rule for drawing a histogram)

What usually happens to the standard error of a sample mean when sample size increases? (It becomes smaller) (!It becomes larger) (!It always becomes zero) (!It becomes the population mean)

What does the central limit theorem mainly describe in this course? (The sampling distribution of a standardized mean) (!The shape of every raw data set) (!The exact value of a population mean) (!The removal of all sampling bias)

What does a frequentist confidence level describe? (Long run coverage of an interval procedure) (!Probability that one fixed parameter changes) (!Percentage of data inside every interval) (!Probability that a sample has no bias)

What is a p value calculated under? (The null hypothesis model) (!The alternative hypothesis only) (!A guarantee of replication) (!A proof that no bias exists)

Which statement about a small p value is correct? (The observed result is unusual under the null model) (!The null hypothesis has a small probability of being true) (!The effect must be practically important) (!The study design must be unbiased)

What is a Type I error? (Rejecting a true null hypothesis) (!Failing to reject a false null hypothesis) (!Estimating a parameter without a sample) (!Using a confidence interval instead of a test)

What is statistical power? (The chance to reject a false null hypothesis) (!The chance that every estimate is exact) (!The size of the sample mean) (!The width of every confidence interval)

What does bootstrap resampling use? (Samples drawn with replacement from observed data) (!Only new observations from the population) (!A fixed list of population parameters) (!Only normally distributed raw data)

What is the best response to a statistically significant result? (Interpret effect size uncertainty and study quality) (!Assume the effect is important) (!Assume the null hypothesis is impossible) (!Ignore the sampling design)

Memory Game

Parameter	Numerical characteristic of a population
Statistic	Numerical summary calculated from a sample
Standard error	Typical sampling variability of an estimator
Confidence interval	Range produced by a procedure designed to cover an unknown quantity at a stated rate
Null hypothesis	Reference claim tested against the data
Power	Probability of rejecting a false reference claim
Bootstrap	Resampling method that draws with replacement from observed data

Drag and Drop

Match the correct terms.	Topic
Population quantity	Parameter
Sample summary	Statistic
Typical estimator variability	Standard error
Evidence measure under a null model	p value
Repeated resampling with replacement	Bootstrap

...

Crossword Puzzle

Parameter	What is an unknown numerical feature of a population called?
Statistic	What is a numerical summary calculated from a sample called?
Estimator	What is a rule for estimating a population quantity called?
Sampling	What word completes the phrase distribution of a statistic over repeated samples?
Power	What is the probability of rejecting a false null hypothesis called?
Bootstrap	What resampling method repeatedly samples with replacement from observed data?

LearningApps

Cloze Text

Statistical inference uses a

to learn about a wider population. A numerical feature of the population is called a

. A numerical summary calculated from observed data is a

. Repeated random samples produce a

for a statistic. The standard deviation of that distribution is the

. Larger samples usually reduce sampling uncertainty because the standard error often contains the square root of

in the denominator. A confidence interval combines an estimate with a measure of

. In frequentist inference the confidence level describes the long run

of the interval procedure. A hypothesis test begins with a

that specifies a reference model. The p value measures how surprising the observed or more extreme result would be under the

. Rejecting a true null hypothesis is a

. The probability of rejecting a false null hypothesis is called

. Bootstrap inference creates many resamples with

from the observed data. Statistical significance should be interpreted together with effect size, uncertainty, design quality, and

.

Open-Ended Tasks

Easy

Population and sample: Choose a question about your school or local community, define the target population, propose a feasible sample, and explain one way the sample could become biased.
Sampling variability: Roll two dice thirty times, record the mean of every group of five rolls, create a graph of those means, and explain why the means vary even though the same dice are used.
Confidence interval infographic: Create a one-page image that explains estimate, standard error, margin of error, and confidence level to a student who has not yet studied inference.
Statistical communication video: Record a two-minute video explaining why “not statistically significant” does not mean “no effect exists,” using your own example.

Standard

Random sampling survey: Design and conduct a small probability-based survey in your class or year group, calculate a proportion, estimate its uncertainty, and discuss limitations of generalizing beyond the sampled group.
Interview a data user: Interview a teacher, technician, health worker, business employee, researcher, or sports coach about how they use sample data to make decisions, then identify the population, sample, and uncertainty in one real example.
Simulation experiment: Use a spreadsheet or programming tool to simulate repeated random samples from a skewed population and show how the distribution of sample means changes as sample size increases.
p-value critique: Find a news article, school report, or public claim that mentions statistical significance, summarize the claim, and write a critique that distinguishes evidence, effect size, uncertainty, and study design.

Advanced

Bootstrap project: Collect a numerical data set of at least thirty observations, create a bootstrap distribution for the median or mean, construct an interval estimate, and compare it with a theory-based interval when appropriate.
Power investigation: Simulate hypothesis tests under several sample sizes and effect sizes, estimate statistical power in each scenario, and explain the trade-off between false positive risk and missed effects.
Replication study: Reproduce a simple published or classroom analysis with a new or openly available data set, compare estimates and uncertainty, and discuss reasons why the results may differ.
Inference field study: Visit a library, science center, university open day, local business, laboratory, or public institution and document one situation where sample data are used to support a wider conclusion; present your findings as a report, poster, podcast, or video.

Learning Assessment

Design and inference: Given three different study designs, decide which conclusions can be generalized to a population and which can support causal claims, justifying each decision from the sampling and assignment procedures.
Confidence interval reasoning: Compare two studies that estimate the same parameter but have different sample sizes and variability, predict which interval should be wider, and explain your reasoning before calculating.
Hypothesis test interpretation: Given a null hypothesis, test statistic, p value, effect estimate, and confidence interval, write a conclusion that separates statistical evidence from practical importance.
Error trade-offs: For a realistic decision such as a medical screening rule, manufacturing alarm, or spam filter, explain the consequences of Type I and Type II errors and argue for an appropriate balance.
Model assumptions: Inspect a data description and graphs, identify which assumptions of a proposed inferential method are questionable, and recommend either a defensible modification or a different method.
Bootstrap transfer: Explain when bootstrap inference may be useful for a statistic whose theoretical sampling distribution is difficult to derive, and identify at least two situations in which naive bootstrapping could be misleading.
Responsible inference: Review a short statistical claim and identify any missing information about design, effect size, uncertainty, multiplicity, missing data, or generalizability that would be needed before making a decision.

Evidence of Learning

Knowledge: You can accurately define population, sample, parameter, statistic, estimator, sampling distribution, standard error, confidence interval, p-value, significance level, Type I error, Type II error, power, and bootstrap.

Skills: You can select a suitable basic inferential method, calculate and interpret standard errors and interval estimates, formulate hypotheses, interpret p-values, connect tests and confidence intervals, and check whether study design and assumptions support the intended conclusion.

Products: Strong evidence may include a reproducible spreadsheet or program, simulation graphs, a survey report, an infographic, a short explanatory video, a bootstrap analysis, or a critique of a real statistical claim.

Reasoning: You can explain why uncertainty arises, how sample size affects precision, why a low p-value is not the probability that the null hypothesis is true, and why statistical significance does not automatically imply practical importance.

Transfer: You can apply inferential thinking to unfamiliar contexts in science, social research, business, engineering, medicine, environmental studies, sports, and everyday decision-making while communicating limits and uncertainty honestly.

OERs on the Topic

Reliable open or freely accessible references for deeper study include:

Media Review and Concept Connections

The following media were selected because each one supports a distinct inferential idea. Use them actively: pause a video before an explanation, predict what should happen, or describe a graph in your own words.

The sampling-distribution video reinforces the idea that a statistic itself behaves like a random variable across repeated samples.

The family of normal curves helps you separate the roles of center and spread and prepares you to reason about standardized statistics.

This diagram connects standard deviation with areas under a normal curve. In inference, such standardized areas become probability statements about statistics under a model.

Linked Learning Areas

Statistical Inference

Statistical inference connects mathematics with data analysis, probability, scientific reasoning, experimental design, computer science, economics, biology, psychology, social science, and engineering. It is especially important when you must make decisions from incomplete data and communicate uncertainty responsibly.

aiMOOC Projects

Diesen aiMOOC als Battle spielen: Mathematik, Klasse 11-13, G-Niveau

Inhalte werden geladen ...

Mediathek wird aus dem Wiki geladen ...

Frag mich zum aktuellen Lerninhalt oder wähle eine kurze Hilfe.

Textmodus bereit

Sprache

English:Statistical Inference