Uji Hipotesis

01Core Concepts and Definitions

Hypothesis testing is the backbone of inferential statistics in clinical research. Every time a clinical trial or epidemiological study reports a result, it is framed within this logic. On the exam, you will not be asked to perform hand calculations. Instead, you will be asked to interpret the output of a study, identify the correct test used, or recognize errors in reasoning.

The null hypothesis ($H_{0}H\_0$) is the default position. It states that there is no difference, no effect, or no association between groups or variables. Think of it as the "nothing is happening" baseline. The alternative hypothesis ($H_{1}H\_1$ or $H_{a}H\_a$) is the research claim that there is a difference, an effect, or an association. The entire framework of hypothesis testing is designed to assess whether the observed data are extreme enough to reject $H_{0}H\_0$ in favor of $H_{1}H\_1$.

A critical concept is the significance level ($\alpha \backslash alpha$), which is the pre-set threshold for how much risk of a false-positive result the investigator is willing to tolerate. By convention, $\alpha =0.05\backslash alpha\; =\; 0.05$ (5%). This means the investigator accepts a 5% probability of incorrectly rejecting a true null hypothesis. The $\alpha \backslash alpha$ is always set before data collection, never adjusted after seeing results.

A one-tailed test evaluates whether a parameter is greater than or less than a stated value (directional hypothesis). A two-tailed test evaluates whether a parameter is simply different from a stated value in either direction (non-directional hypothesis). On the exam, unless stated otherwise, assume a two-tailed test. A one-tailed test has more power to detect an effect in one direction but cannot detect an effect in the opposite direction.

02Types of Errors and Statistical Power

This is one of the most commonly tested areas in biostatistics. The exam will present scenarios and ask you to identify which error is being described or which factor affects power.

Type I Error ($\alpha \backslash alpha$) occurs when you conclude there is an effect when in reality there is none. This is a "false alarm." In clinical terms, this is like approving a drug that does not actually work. The probability of a Type I error is equal to $\alpha \backslash alpha$, which is set by the investigator (typically 0.05).

Type II Error ($\beta \backslash beta$) occurs when you fail to detect a real effect. This is a "missed finding." In clinical terms, this is like concluding a drug does not work when it actually does. The probability of a Type II error is $\beta \backslash beta$, conventionally kept at $\le 0.20\backslash leq\; 0.20$.

Statistical Power is defined as $1-\beta 1\; -\; \backslash beta$, meaning the probability of correctly rejecting a false null hypothesis. A study with 80% power has a 20% chance of missing a real effect. Power is arguably the most testable concept because the exam frequently asks what factors increase or decrease it.

A power analysis is performed before a study begins. Its purpose is to determine the minimum sample size needed to detect a clinically meaningful effect at the desired $\alpha \backslash alpha$ and power level. If a study reports a non-significant result, the first question you should ask is whether the study was adequately powered. A non-significant result in an underpowered study does not mean "no effect." It means the study may not have been large enough to find one.

03Choosing the Right Statistical Test

On the exam, a vignette will describe a study design, and you must identify which statistical test is appropriate. The key decision points are: (1) what type of data is the outcome variable, and (2) how many groups are being compared.

Parametric vs. Nonparametric

Parametric tests (t-test, ANOVA) assume the data are continuous and approximately normally distributed. Nonparametric tests (Mann-Whitney U, Kruskal-Wallis, Wilcoxon) make no such assumption and are used for ordinal data or continuous data that are skewed and cannot be normalized.

Test Selection Table

How to Approach the Question Stem

When you encounter a vignette that asks "which is the most appropriate statistical test," follow this decision sequence:

1. Identify the outcome variable (dependent variable). Is it continuous, categorical, or time-to-event?

2. Determine the number of groups being compared. Is it one group vs. a known value, two groups, or three or more groups?

3. Assess the relationship between groups. Are they independent (different subjects) or paired/matched (same subjects measured twice)?

4. Check distribution assumptions. If the data are stated to be normally distributed, use parametric tests. If the data are skewed, ordinal, or the sample size is very small, use nonparametric tests.

A common trap is confusing paired vs. unpaired designs. If the same patients are measured before and after an intervention, the data are paired. Using an unpaired test on paired data inflates the variance estimate and can cause you to miss a real effect.

04Interpreting Results: p-Values, Confidence Intervals, and Significance

The p-Value

The $pp$-value is the probability of observing data as extreme as, or more extreme than, the results actually obtained, assuming that $H_{0}H\_0$ is true. It is not the probability that $H_{0}H\_0$ is true, and it is not the probability that the results are due to chance. These are common misinterpretations that the exam will exploit.

If , we reject $H_{0}H\_0$ and declare the result statistically significant. If $p\ge 0.05p\; \backslash geq\; 0.05$, we fail to reject $H_{0}H\_0$. Note the language: we never "accept" $H_{0}H\_0$. Failing to reject is not the same as proving $H_{0}H\_0$ is true; it simply means we did not find sufficient evidence against it.

Confidence Intervals (CI)

A 95% confidence interval provides a range of values within which the true population parameter is expected to lie, with 95% confidence across repeated sampling. The CI carries more information than the $pp$-value alone because it communicates both the direction and magnitude of the effect, along with the precision of the estimate.

The critical rule linking CI to hypothesis testing: if a 95% CI for a difference does not include zero (or does not include 1.0 for ratios like odds ratio or relative risk), then and the result is statistically significant.

Statistical Significance vs. Clinical Significance

This distinction is heavily tested. A result can be statistically significant but clinically meaningless when the study has a very large sample size that detects a tiny difference with no practical importance. For example, a study of 50,000 patients finds that Drug A lowers blood pressure by 0.5 mmHg more than placebo ($p=0.003p\; =\; 0.003$). The result is statistically significant, but a 0.5 mmHg difference has no meaningful impact on patient outcomes.

Conversely, a result can be clinically meaningful but not statistically significant when the study is underpowered. A small trial of 30 patients finds that a new chemotherapy regimen improves 5-year survival by 15% ($p=0.08p\; =\; 0.08$). The effect size is clinically important, but the study lacked the sample size to achieve statistical significance.

05Common Exam Distractors and Pitfalls

06Traps and High-Yield Pearls

The most common way students get questions on hypothesis testing wrong is by confusing what the $pp$-value actually means. The exam loves to offer answer choices that restate the $pp$-value as "the probability that the null hypothesis is true" or "the probability that the result is due to chance." Both are wrong. The $pp$-value is strictly the probability of obtaining the observed data (or more extreme) assuming $H_{0}H\_0$ is true.

The second major trap involves power and sample size. When a study reports a non-significant result, students often conclude that the treatment does not work. The correct interpretation is that the study failed to demonstrate a difference, and the next step is to evaluate whether the study had adequate power. If power was low (small sample, small effect size, high variance), a Type II error is likely.

Another frequently tested trap is the multiple comparisons problem. When investigators perform many statistical tests on the same dataset (e.g., comparing 10 subgroups), the probability of at least one false-positive finding increases dramatically. With 20 independent tests at $\alpha =0.05\backslash alpha\; =\; 0.05$, the probability of at least one Type I error is approximately $1-(0.95{)}^{20}\approx 0.641\; -\; (0.95)^\{20\}\; \backslash approx\; 0.64$, or 64%. The correction for this is the Bonferroni correction, which divides $\alpha \backslash alpha$ by the number of comparisons (e.g., $0.05\mathrm{/}20=0.00250.05\; /\; 20\; =\; 0.0025$). The trade-off is reduced power.

Students also confuse paired and unpaired designs. If the vignette says "patients were measured at baseline and again after 8 weeks of treatment," the data are paired, and you must use a paired t-test (or Wilcoxon signed-rank if non-normal). Using an unpaired test on paired data introduces between-subject variability as noise, which inflates the variance and can mask a real treatment effect.

Finally, remember the hierarchy of what the exam is truly testing when it asks a biostatistics question about hypothesis testing: (1) can you identify the correct test for a given study design, (2) can you interpret $pp$-values and confidence intervals without falling for common misstatements, and (3) do you understand the relationship between sample size, power, effect size, and error types. Master these three competencies and you will handle the majority of hypothesis-testing questions correctly.