AP Statistics: Conditions for Inference

Statistical inference allows you to use sample data to draw conclusions about a wider population or to test a claim. However, these powerful conclusions are only trustworthy if the procedures are built on a solid foundation. Jumping straight to a t-test or confidence interval without verifying the underlying assumptions is like building a house on sand—the results may look impressive but could collapse under scrutiny. Checking the conditions for randomness, normality, and independence isn't just a procedural step; it's the essential act of validating that your mathematical tools are appropriate for your data.

The Pillars of Valid Inference: Randomness, Normality, and Independence

Every formal inference procedure—whether for proportions or means—rests on three interconnected pillars. Failing to satisfy even one compromises the validity of your p-values and confidence intervals. The first and most critical pillar is randomness. This comes in two distinct forms: random sampling and random assignment.

Random sampling involves selecting individuals from a population in such a way that every possible sample has a known, non-zero chance of being chosen. This is the cornerstone for generalizing your findings from the sample back to the population. For example, if you want to estimate the proportion of all high school students who prefer digital textbooks, you need a random sample of students, not just the students in your AP Statistics class. Without random sampling, you cannot claim your inference applies to a broader group.

Random assignment, on the other hand, is used in experiments. It involves using a chance process, like a random number generator, to assign subjects to treatment or control groups. The goal here is not generalization, but causation. Random assignment creates roughly equivalent groups at the outset, so that any significant difference in outcomes can be attributed to the treatment itself. It’s crucial to identify which goal you have: to infer about a population (requiring random sampling) or to infer cause-and-effect (requiring random assignment).

Verifying the Normality Condition

The second pillar is the approximate normality of the sampling distribution. The Central Limit Theorem (CLT) is your primary tool here, but its application differs for proportions and means. For both, you must check this condition using your sample data to justify that the underlying sampling distribution is approximately normal.

For inference about a population proportion $p$, you check two requirements related to sample size. You need to verify that the expected number of successes $np$ and the expected number of failures $n(1-p)$ are both at least 10. Since you don't know the true population proportion $p$, you use your sample proportion $\hat{p}$ as an estimate. So, in practice, you check: $n\hat{p}\ge 10$ and $n(1-\hat{p})\ge 10$. If you are testing a hypothesized proportion $p_0$, you use that value instead.

For inference about a population mean $\mu$, the path to normality depends on your sample size and the shape of the population distribution.

1. If the population distribution is known to be normal, the sampling distribution of $\bar{x}$ is normal for any sample size $n$.
2. If the population distribution is unknown or not normal, you rely on the CLT. The rule of thumb is that if your sample size $n\ge 30$, the sampling distribution of $\bar{x}$ will be approximately normal. For $n<30$, you must examine your sample data graphically (using a dotplot, boxplot, or histogram) to check for strong skew or extreme outliers that suggest the underlying population is not normal. Always report how you are checking normality—via the CLT ($n\ge 30$) or via a graph showing no strong skew/outliers.

Ensuring Independence of Observations

The third pillar is the independence of observations. This condition has two layers. The first layer is that individual observations within a sample must not influence each other. Practically, this is often verified using the 10% rule: the sample size $n$ must be no more than 10% of the population size $N$ (when sampling without replacement). If you are sampling 50 students from a school of 400, you are sampling 12.5% of the population, violating the 10% rule. In such cases, the observations are not independent because the probability of selecting each subsequent student changes meaningfully.

The second layer relates to study design. Data collected over time (time series) or from related individuals (e.g., siblings) may violate independence. For experiments, independence means the outcome for one subject does not affect the outcome for another. Independence is fundamentally about the data collection method, not something you can fix with a formula after the data is collected.

A Step-by-Step Application: One-Sample t-Interval

Let's walk through verifying conditions for a concrete scenario. Suppose you want to estimate the mean commute time for employees at a large company of 5,000 people. You take a random sample of 35 employees and record their times (in minutes): 22, 25, 30, 28, ... (and so on). The sample mean $\bar{x}=27.1$ minutes and the sample standard deviation $s_x=5.8$ minutes. You plan to construct a one-sample t-interval for the population mean.

1. Randomness: The data comes from a random sample of employees. This is stated and is required to generalize to all 5,000 employees.
2. Independence: Because sampling was random and the sample size (35) is less than 10% of the population (5,000), the 10% rule is satisfied ($35<0.10\times 5000=500$). We can assume individual commute times are independent.
3. Normality: The sample size ($n=35$) is $\ge 30$. By the Central Limit Theorem, the sampling distribution of the sample mean will be approximately normal, even if we don't know the shape of the population distribution of commute times.

Since all three conditions are met, proceeding with a one-sample t-interval for the mean is justified.

Common Pitfalls

Confusing Random Sampling with Random Assignment. This is a fundamental conceptual error. Remember: random sampling supports generalization to a population; random assignment supports causal conclusions in an experiment. You cannot use the results from a non-random sample to make inferences about a larger population, no matter how fancy the statistical test.

Misapplying the Normality Check for Means. A common mistake is checking the normality of the sample data itself for a mean. The condition concerns the shape of the sampling distribution. For $n\ge 30$, the CLT assures us this distribution is normal, even if the sample data is skewed. You only need to graph the sample data and worry about its shape when $n<30$. Conversely, for a proportion, you always check the success/failure condition using counts ($n\hat{p}\ge 10$).

Overlooking the Independence Condition. Students often focus solely on randomness and normality, forgetting to check the 10% rule. If your sample is more than 10% of the population, the independence condition fails because the probability of selecting each individual changes too much. This invalidates the standard deviation formula used in your inference procedure. Always ask: "Is $n\le 0.1N$?"

Using Sample Size Rules Blindly. The $n\ge 30$ rule for means and the $np\ge 10$ rule for proportions are guidelines, not absolute laws. With extreme population skew (e.g., income data), a sample size of 30 might not be sufficient for the CLT to kick in. Similarly, if $\hat{p}$ is very close to 0 or 1, the success/failure condition might be met mathematically, but the approximation may still be poor. A brief comment acknowledging this limitation shows deeper statistical understanding.

Summary

• Valid inference is built on three non-negotiable conditions: randomness in data collection, approximate normality of the sampling distribution, and independence of observations.
• Always verify randomness first: use random sampling to generalize or random assignment to establish cause.
• Check normality correctly: for proportions, use the success/failure condition ($n\hat{p}\ge 10$ and $n(1-\hat{p})\ge 10$); for means, use the Central Limit Theorem ($n\ge 30$) or examine a graph of the sample data if $n<30$.
• Enforce independence using the 10% rule: when sampling without replacement, ensure your sample size $n$ is no more than 10% of the population size $N$.
• State your condition checks clearly and in context as part of your answer; it demonstrates the statistical reasoning that underpins your conclusion.