Penelitian Korelatif & Regresi

01Core Concepts and Definitions

Correlation refers to the degree to which two continuous variables move together in a linear fashion. It is measured by the correlation coefficient ($rr$), which ranges from $-1-1$ to $+1+1$. A value of $r=+1r\; =\; +1$ indicates a perfect positive linear relationship (as one variable increases, the other increases proportionally), while $r=-1r\; =\; -1$ indicates a perfect negative (inverse) linear relationship. A value of $r=0r\; =\; 0$ means there is no linear association between the two variables. This is a critical distinction: $r=0r\; =\; 0$ does not mean there is no relationship at all. A perfectly curved, parabolic relationship between two variables can yield $r=0r\; =\; 0$ because the correlation coefficient only captures linear trends.
The coefficient of determination ($r^{2}r^2$) is simply the square of the correlation coefficient. It tells you the proportion of variance in the dependent variable that is explained by the independent variable. For example, if $r=0.7r\; =\; 0.7$, then $r^2=0.49r^2\; =\; 0.49$, meaning 49% of the variability in the outcome can be explained by the predictor. This is one of the most commonly tested calculations in biostatistics vignettes.
Regression goes a step further than correlation. While correlation tells you that two variables are associated, regression allows you to model and predict the value of one variable (the dependent or outcome variable) based on the value of another (the independent or predictor variable). Regression also quantifies the magnitude of effect: for every one-unit increase in X, how much does Y change? This is the regression coefficient ($\beta \backslash beta$). Regression is the foundation for adjusting for confounders in clinical research, which is why it appears so frequently on the exam.

02Types of Correlation and Regression

Pearson vs. Spearman is a classic test question. The Pearson correlation requires that both variables be continuous and approximately normally distributed. If the data is ordinal (e.g., pain rated on a scale of 1 to 10) or if the distribution is heavily skewed, you must use the Spearman rank correlation instead. The Spearman method works by converting raw data into ranks and then computing the correlation on those ranks, making it resistant to outliers and non-normal distributions.
Simple linear regression fits a straight line ($Y=a+\beta XY\; =\; a\; +\; \backslash beta\; X$) through a scatterplot. The slope ($\beta \backslash beta$) tells you the predicted change in Y for every one-unit increase in X. The intercept ($aa$) is the predicted value of Y when X equals zero. On the exam, you may be given a regression equation and asked to predict an outcome or interpret the slope.
Multiple linear regression extends this by including more than one predictor variable. The formula becomes $Y=a+{\beta }_1{X}_1+{\beta }_2{X}_2+\mathrm{.}\mathrm{.}\mathrm{.}+{\beta }_n{X}_{n}Y\; =\; a\; +\; \backslash beta\_1\; X\_1\; +\; \backslash beta\_2\; X\_2\; +\; ...\; +\; \backslash beta\_n\; X\_n$. The critical concept here is that each $\beta \backslash beta$ coefficient represents the effect of that variable while holding all other variables constant. This is how researchers adjust for confounders. If a vignette describes a study that "controlled for age, sex, and smoking status," the study almost certainly used some form of multiple regression.
Logistic regression is used when the outcome is binary (e.g., dead vs. alive, diseased vs. healthy). Instead of predicting a continuous number, logistic regression predicts the probability of an event occurring. Its output is expressed as an odds ratio (OR). This is the most commonly used regression method in case-control studies and clinical trials with dichotomous endpoints.
Cox proportional hazards regression is the method of choice for survival analysis. When the outcome is "time until an event" (e.g., time to death, time to relapse), Cox regression produces a hazard ratio (HR). An HR of 2.0 means the exposure group has twice the instantaneous rate of the event compared to the reference group at any given time point.

03Interpreting Outputs and Coefficients

Interpreting confidence intervals around OR and HR is extremely high-yield. For ratios (OR, HR, relative risk), the null value is 1.0. If the 95% confidence interval includes 1.0, the result is not statistically significant at . For differences (mean difference, $\beta \backslash beta$ coefficient), the null value is 0. If the confidence interval includes 0, the result is not statistically significant.
When a vignette presents a multiple regression output table, pay close attention to whether the question asks about the crude (unadjusted) association or the adjusted association. A variable that is significant in crude analysis but becomes non-significant after adjustment suggests that the initial association was due to confounding. Conversely, a variable that becomes significant only after adjustment may have been masked by a confounder (a phenomenon sometimes called negative confounding or suppression).

04Assumptions, Violations, and Important Pitfalls

Regression to the mean deserves special attention because it appears in clinical vignettes disguised as a treatment effect. Imagine a study that enrolls only patients with extremely high blood pressure readings. On repeat measurement, many of those readings will naturally drift downward toward the population average, regardless of any intervention. A question might describe a dramatic drop in blood pressure after starting a new drug in such a selected group and ask you to identify the most likely explanation. The answer is regression to the mean, not a true drug effect. The giveaway is that participants were selected based on extreme initial values without a control group for comparison.
Ecological fallacy is tested when a question presents data at the country or population level (e.g., "countries with higher fat intake have higher rates of heart disease") and then asks whether you can conclude that an individual who eats more fat will develop heart disease. You cannot. Aggregate associations do not translate to individual risk. This is a correlation-level error in reasoning.
Correlation does not imply causation is the single most tested principle in this topic. Even a strong $rr$ value does not establish a cause-and-effect relationship. Establishing causation requires experimental design (randomized controlled trials), temporality, dose-response, and other Bradford Hill criteria. Observational studies, no matter how large, can only demonstrate association.

05Choosing the Right Test

The most common exam trap in this section is confusing when to use correlation vs. regression vs. a $tt$-test vs. chi-square. The decision always comes down to two questions: (1) What type of variables do you have? (continuous, ordinal, categorical/binary) and (2) What are you trying to do? (measure association, predict an outcome, compare groups, or adjust for confounders). If the vignette says the researchers "controlled for" or "adjusted for" other variables, the answer is almost always some form of multiple regression.

06Traps and High-Yield Pearls

The number-one way students lose points on correlation and regression questions is by conflating correlation with causation. When a question describes an observational study showing a strong correlation between two variables and asks what conclusion can be drawn, the correct answer is always limited to "there is an association." Never select an answer choice that uses causal language ("causes," "leads to," "results in") unless the study design is a randomized controlled trial.
The second major trap is regression to the mean. Any time a vignette selects patients based on an extreme initial value (highest blood pressure, worst pain scores, lowest test results) and then shows improvement on follow-up, consider this phenomenon before attributing the change to a treatment. The classic tell is the absence of a control or placebo group.
A third common mistake involves $r^{2}r^2$ interpretation. Students often confuse $rr$ with $r^{2}r^2$. If $r=0.5r\; =\; 0.5$, the proportion of variance explained is $r^2=0.25r^2\; =\; 0.25$, or only 25%, not 50%. The exam frequently tests whether you can perform this conversion and state what it means in plain terms.
Finally, watch for questions that present a confidence interval crossing the null value. For odds ratios and hazard ratios, a 95% CI that includes 1.0 means the finding is not statistically significant. For $\beta \backslash beta$ coefficients or mean differences, a CI that includes 0 is not significant. This is functionally equivalent to $p>0.05p\; >\; 0.05$, and the exam loves to test this concept by giving you a table of regression outputs and asking which variables are independently associated with the outcome. Only those whose confidence intervals do not cross the null value qualify.
The core competency being tested across all correlation and regression questions is your ability to match the correct statistical method to a given study scenario, interpret the output without overclaiming causation, and recognize common analytical pitfalls like regression to the mean, ecological fallacy, and confounding.