Correlation, Regression and Econometric Reasoning

Introduction

International business decisions often depend on relationships among variables: whether advertising is associated with export sales, how exchange rates affect import costs, or whether market income predicts product demand. Correlation and regression provide systematic ways to quantify such relationships. Econometric reasoning extends these tools by asking whether estimated relationships are credible, stable and potentially causal.

A regression coefficient is not automatically a causal effect. Confounding variables, reverse causality, measurement error and selection bias may produce statistically convincing but managerially misleading results. Sound analysis therefore combines calculation with theory, diagnostics, validation and transparent interpretation.

Learning Objectives

After completing this chapter, students should be able to:

• calculate and interpret covariance and correlation;
• estimate and interpret simple and multiple linear regression models;
• evaluate statistical significance, model fit and prediction error;
• diagnose violations of major regression assumptions;
• model categorical variables, interactions and nonlinear relationships;
• distinguish association, prediction and causal inference;
• recognize confounding, bias and endogeneity; and
• apply responsible econometric reasoning to international business decisions.

1. Association Between Quantitative Variables

1.1 Covariance

Covariance measures whether two variables tend to move together. For a sample of size n, the sample covariance is

Here, x_i and y_i are observations, while \bar{x} and \bar{y} are sample means. A positive covariance indicates that larger values of x tend to accompany larger values of y. A negative covariance indicates opposite movement.

Covariance has compound units. If advertising is measured in thousands of euros and sales in thousands of units, covariance is measured in thousands of euros multiplied by thousands of units. Its magnitude is consequently difficult to compare across datasets.

1.2 Pearson correlation

Pearson’s sample correlation standardizes covariance:

The coefficient satisfies -1 \le r \le 1 and has no units.

• r>0 indicates positive linear association.
• r<0 indicates negative linear association.
• r=0 indicates no linear association, although a nonlinear relationship may still exist.
• Values near 1 or -1 indicate strong linear association.

Correlation is symmetric: the correlation between exports and GDP equals the correlation between GDP and exports. It is also sensitive to outliers and range restriction. A correlation calculated only among high-income countries may differ substantially from one calculated across all countries.

Spearman’s rank correlation is often preferable when the relationship is monotonic but nonlinear, observations are ordinal, or extreme values make Pearson correlation unreliable.

2. Simple Linear Regression

2.1 Population and estimated models

A simple population regression model is

where:

• Y_i is the dependent or outcome variable;
• X_i is the explanatory variable;
• β₀ is the population intercept;
• β₁ is the population slope; and
• u_i contains unobserved influences on Y_i.

If Y is sales in millions of dollars and X is advertising expenditure in millions of dollars, β₁ is measured in millions of dollars of sales per million dollars of advertising.

The estimated sample regression is

The residual is e_i=Y_i-\hat{Y}_i. A residual is an observed sample estimate; the disturbance u_i is an unobserved population quantity.

2.2 Ordinary least squares

Ordinary least squares (OLS) chooses the coefficients that minimize the sum of squared residuals:

Differentiating with respect to b₀ and b₁, and setting the derivatives equal to zero, gives the normal equations. Solving them produces

These formulas require variation in X; if every observation has the same X, the denominator is zero and the slope cannot be estimated.

2.3 Worked regression calculation

Suppose a market analyst has computed the following verified summaries from a sample relating promotional expenditure X to monthly sales Y:

The cross-product sum is 62.6. Therefore,

The intercept is

Thus,

At X=6,

If both variables are measured in thousands, the model predicts approximately 30.60 thousand units of sales at promotional expenditure of 6 thousand currency units. The slope means that an additional thousand currency units of promotion is associated with approximately 2.75 thousand additional units of monthly sales. This is an association unless the research design justifies causal interpretation.

The intercept predicts sales when promotion is zero. It should not be interpreted substantively if zero lies outside the observed range or is economically implausible.

3. Multiple Linear Regression

Business outcomes normally have several determinants. The multiple regression model is

Here, k is the number of explanatory variables excluding the intercept. The coefficient β_j measures the expected change in Y associated with a one-unit increase in X_j, holding the other included explanatory variables constant.

For example,

If β₁=-4.2, a one-unit increase in price is associated with 4.2 fewer units of sales, conditional on income and advertising. “Holding constant” is a model-based comparison, not proof that actual firms can change price while everything else remains unchanged.

3.1 Omitted-variable bias

Suppose the true model contains Z, but the analyst regresses Y only on X. In a simple setting, the expected omitted-variable bias in the estimated slope is

This expression applies to a linear model in which Z is the omitted determinant. Bias requires both:

1. Z affects Y, so β_Z\ne0; and
2. Z is correlated with X.

For example, advertising may appear to increase international sales partly because larger firms both advertise more and possess stronger distribution networks.

4. Statistical Inference

4.1 Standard errors and t-tests

An estimated coefficient varies across possible samples. Its standard error measures this sampling uncertainty. To test H₀:β_j=β_j,0, use

Under the classical assumptions and normally distributed disturbances, the statistic follows a t distribution with n-k-1 degrees of freedom. In large samples, asymptotic approximations often support inference without exact normality.

A two-sided 100(1-α)\% confidence interval is

A p-value measures compatibility between the data and the null hypothesis under the assumed model. It is not the probability that the null hypothesis is true and does not measure commercial importance.

4.2 Overall significance

An F-test can evaluate the joint null hypothesis that several slope coefficients are zero. This matters when variables are meaningful as a group—for example, a set of regional indicators or seasonal controls—even if individual t-tests are inconclusive.

Statistical significance should be reported alongside coefficient magnitude, units, confidence intervals and managerial relevance.

5. Model Fit and Predictive Performance

5.1 Variance decomposition and R²

Define

For an OLS regression that includes an intercept, the residuals are orthogonal to fitted values and

This exact decomposition is not generally guaranteed for a regression estimated without an intercept.

The coefficient of determination is

For an intercept model, R² is the proportion of sample variation in Y explained by the fitted regression. A high R² does not prove causality, eliminate bias or guarantee accurate future predictions.

Adding regressors cannot reduce in-sample R². Adjusted R² penalizes model size:

5.2 Residual standard error

The estimated disturbance standard deviation is

It has the same units as Y. The denominator reflects degrees of freedom used to estimate the intercept and k slopes.

5.3 Out-of-sample prediction error

Prediction should normally be evaluated on validation or test observations not used to estimate the model. For test-set errors e_i=Y_i-\hat{Y}_i, common measures include

where m is the number of test observations. Both measures have the same units as Y. RMSE gives relatively greater weight to large errors; MAE is less sensitive to extreme errors.

Suppose test-set forecast errors are 2, −1, 1 and −2 million dollars. Then

The model misses sales by about $1.50 million in absolute terms on average. RMSE is slightly higher because it penalizes the two-unit errors more strongly.

Mean absolute percentage error can be intuitive but is undefined when an actual value is zero and unstable near zero. Cross-validation is preferable to judging many candidate models on one favorable split. Time-dependent international business data should be divided chronologically rather than randomly to avoid using future information to predict the past.

A confidence interval estimates the mean response at a specified X. A prediction interval concerns a new individual outcome and is wider because it incorporates both uncertainty in the estimated mean and new disturbance variation.

6. Assumptions and Diagnostics

The main linear regression conditions serve different purposes.

6.1 Nonlinearity

A curved pattern in a residual plot suggests functional-form error. Responses include logarithms, polynomial terms and theoretically justified interactions. Adding complexity solely to improve in-sample fit risks overfitting.

6.2 Heteroskedasticity

Heteroskedasticity occurs when Var(u_i\mid X_i) changes across observations. Cross-country sales errors, for example, may be larger in large markets. If the zero conditional mean condition holds, OLS coefficients remain unbiased, but conventional standard errors may be wrong. Heteroskedasticity-robust standard errors provide more reliable large-sample inference.

6.3 Multicollinearity

Strong correlation among explanatory variables increases coefficient uncertainty and can make estimates unstable. It does not, by itself, bias OLS. Remedies include collecting more data, combining conceptually overlapping measures or focusing on joint tests. Removing an essential confounder merely to reduce multicollinearity can increase omitted-variable bias.

6.4 Outliers and influential observations

An outlier has an unusual outcome, while a high-leverage observation has unusual explanatory-variable values. An influential observation materially changes fitted results. Analysts should check data quality, studentized residuals, leverage and Cook’s distance. Valid observations should not be deleted merely because they weaken a preferred conclusion.

6.5 Dependence

Repeated observations from the same country, company or region may have correlated errors. Cluster-robust standard errors can permit arbitrary error correlation within defined clusters when there are enough independent clusters. Time-series residuals may exhibit serial correlation and require specialized models or robust standard errors.

7. Categorical Variables, Interactions and Transformations

7.1 Indicator variables

A categorical variable with g categories is usually represented by g-1 indicator variables when an intercept is included. One category becomes the reference group. Including all g indicators plus an intercept creates perfect multicollinearity, known as the dummy-variable trap.

For example,

where EU_i=1 for European Union markets and 0 otherwise. At a given advertising level, β₂ is the estimated mean difference between EU and reference markets.

7.2 Interactions

An interaction permits one variable’s association to depend on another:

For reference markets, the advertising slope is β₁. For EU markets, it is β₁+β₃. The main effects should normally remain in the model when their interaction is included.

7.3 Logarithmic models

Provided logged variables are strictly positive:

• In \ln(Y)=β₀+β₁X+u, a one-unit increase in X is associated approximately with a 100β₁ percent change in Y for small β₁. The exact percentage is 100(e^β₁-1).
• In Y=β₀+β₁\ln(X)+u, a 1% increase in X is associated approximately with a β₁/100-unit change in Y.
• In \ln(Y)=β₀+β₁\ln(X)+u, β₁ is an elasticity: a 1% increase in X is associated with a β₁% change in Y.

8. Correlation, Causation and Endogeneity

A causal claim asks what would happen to the same unit under alternative actions. Because only one outcome is observed, causal inference requires assumptions or research designs that construct a credible comparison.

Figure: Market wealth confounds the relationship between advertising and sales when it affects both variables.

Endogeneity means an explanatory variable is correlated with the disturbance: Cov(X,u)\ne0. Major sources include:

• omitted variables, such as market wealth or managerial quality;
• reverse causality, such as firms advertising more because sales are expected to rise;
• measurement error in explanatory variables;
• simultaneous determination of price and quantity; and
• non-random sample selection.

Possible strategies include randomized experiments, credible natural experiments, fixed-effects panel models, instrumental variables, regression discontinuity and difference-in-differences. Each strategy requires specific assumptions. For example, a valid instrument must predict the endogenous explanatory variable and affect the outcome only through that variable, conditional on controls. Statistical software cannot verify the exclusion restriction by itself.

9. Applied Case: International Digital Promotion

Background

This is an illustrative case. Meridian Foods sells packaged products in 30 national markets. Management wants to estimate whether increased digital advertising raises quarterly sales.

Problem

A simple regression shows a strong positive advertising coefficient and a high R². However, advertising budgets are concentrated in wealthy markets with established distributors. Managers also increase advertising before quarters in which they expect strong demand.

Analysis

The initial model is vulnerable to confounding and reverse causality. Analysts assemble a panel dataset containing sales, advertising, prices, income, distribution coverage, exchange rates, market indicators and quarterly indicators. Monetary values are converted consistently, and nominal measures are distinguished from inflation-adjusted measures.

They estimate a multiple regression with country and quarter controls, cluster standard errors by country, inspect residual plots and test whether advertising effects differ between mature and emerging markets. A chronological holdout sample is used to calculate RMSE and MAE. The richer model has a slightly lower in-sample adjusted R² than an overfitted alternative but produces lower holdout RMSE.

Approach and outcome

The estimated advertising association falls after distribution coverage and market income are controlled. An interaction suggests a weaker association in mature markets. Management therefore rejects the original claim that the simple coefficient is a causal return on advertising.

The firm launches a randomized regional campaign where feasible. In markets where randomization is not legally or operationally possible, results are reported as conditional associations rather than causal effects.

Learning points

• A strong correlation can reflect budget allocation and market characteristics.
• Better prediction does not necessarily identify causal effects.
• Controls should follow business theory rather than automated variable selection alone.
• Out-of-sample error is essential when the model will support forecasts.
• Transparent uncertainty protects the firm from overconfident investment decisions.

10. Governance, Ethics and Contemporary Analytics

Regression models can influence prices, market entry, lending, supplier selection and workforce decisions. Analysts should document data provenance, missing-data treatment, exclusions, model versions and validation results. Cross-border projects must respect privacy, data-localization and anti-discrimination requirements.

Automated machine-learning systems can identify interactions and nonlinear patterns, but flexible prediction is not equivalent to causal inference. Regularization can improve prediction under many correlated variables, while explainability tools can summarize model behavior. Neither removes endogeneity or replaces institutional knowledge.

Best practice is to predefine the business question, distinguish prediction from causal estimation, preserve a test sample, report uncertainty, investigate subgroup performance and monitor models for structural change. Exchange-rate regimes, regulation, consumer behavior and supply chains can change, causing relationships estimated from historical data to deteriorate.

Chapter Summary

Correlation measures the direction and strength of linear association but does not establish causation. OLS regression estimates the line or hyperplane that minimizes squared residuals. Simple regression describes a bivariate relationship, while multiple regression provides conditional comparisons among variables.

Coefficient interpretation depends on units, functional form, reference categories and interactions. Statistical inference uses standard errors, confidence intervals, t-tests and F-tests, but statistical significance must be distinguished from economic importance. With an intercept, OLS supports the decomposition SST=SSR+SSE and the conventional interpretation of R². Adjusted R², RMSE and MAE provide complementary assessments of fit and prediction.

Credible analysis requires diagnostics for nonlinearity, heteroskedasticity, multicollinearity, dependence and influential observations. Causal claims require more than regression control because omitted variables, reverse causality, measurement error and selection can create endogeneity. Ethical and professionally responsible practice combines theory, defensible research design, out-of-sample validation, transparent reporting and continuing model governance.

Key Terms

Covariance: A measure of joint variation between two variables.

Correlation: A standardized measure of linear association ranging from −1 to 1.

Ordinary least squares: Estimation method that minimizes the sum of squared residuals.

Dependent variable: The outcome the regression seeks to explain or predict.

Explanatory variable: A variable used to explain or predict the outcome.

Residual: The difference between an observed and fitted outcome.

Coefficient: An estimated conditional association between a regressor and the outcome.

Standard error: A measure of sampling uncertainty in an estimator.

Confidence interval: A range constructed by a procedure with a stated long-run coverage rate.

R²: The proportion of sample outcome variation explained by an intercept regression.

Adjusted R²: A fit measure that penalizes the inclusion of additional regressors.

RMSE: The square root of mean squared out-of-sample prediction error.

MAE: The mean absolute out-of-sample prediction error.

Heteroskedasticity: Nonconstant conditional variance of regression disturbances.

Multicollinearity: Strong linear dependence among explanatory variables.

Indicator variable: A binary variable representing category membership.

Interaction: A model term allowing one variable’s association to depend on another.

Confounder: A variable related to both the explanatory variable and outcome.

Endogeneity: Correlation between an explanatory variable and the model disturbance.

Omitted-variable bias: Bias caused by excluding a relevant variable correlated with an included regressor.

Review Questions

Short Answer Questions

1. What is the principal difference between covariance and correlation?
2. How is an OLS residual calculated?
3. Why is adjusted R² sometimes preferred to R² when comparing models?
4. What does heteroskedasticity mean?
5. Why does a categorical variable with four categories usually require only three indicators in an intercept model?

Long Answer Questions

1. Derive the OLS slope and intercept formulas for simple linear regression and explain their interpretations.
2. Explain the assumptions supporting unbiased OLS estimation and conventional statistical inference.
3. Compare R², adjusted R², residual standard error, RMSE and MAE.
4. Discuss the sources and consequences of endogeneity in international business research.
5. Explain how categorical variables, interaction terms and logarithmic transformations alter coefficient interpretation.

Analytical / Application Questions

1. A company finds a strong positive correlation between foreign-market advertising and sales. Develop three non-causal explanations and propose a research design for estimating a causal effect.
2. A test sample produces forecast errors of 3, −2, 0, 1 and −4 million dollars. Calculate RMSE and MAE, then interpret both measures.
3. Design a multiple regression for explaining export sales across countries. Specify variables, expected coefficient signs, possible confounders, diagnostics and an out-of-sample validation strategy.

Multiple Choice Questions with Answers

Question 1

In the worked simple regression example, what is the OLS slope when the centered cross-product sum is 62.6 and the centered sum of squares for X is 22.8?

A. 2.25

B. 2.75

C. 3.64

D. 14.12

Correct Answer: B. 2.75

Question 2

Using the worked model \hat{Y}=14.1215+2.7456X, what is the predicted value at X=6?

A. 16.87

B. 28.84

C. 30.60

D. 32.34

Correct Answer: C. 30.60

Question 3

Under what stated condition does the OLS identity SST=SSR+SSE apply in the chapter?

A. The regression includes an intercept

B. All slope coefficients are positive

C. The dependent variable is standardized

D. The residuals are normally distributed

Correct Answer: A. The regression includes an intercept

Question 4

Test-set errors are 2, −1, 1 and −2. What is the RMSE?

A. 1.00

B. 1.50

C. 1.58

D. 2.50

Correct Answer: C. 1.58

Question 5

Which statement best distinguishes MAE from RMSE?

A. MAE has no units

B. MAE gives relatively less emphasis to large errors

C. MAE can be calculated only in sample

D. MAE always exceeds RMSE

Correct Answer: B. MAE gives relatively less emphasis to large errors

Question 6

In a multiple regression, a slope coefficient represents the expected change in the outcome associated with a one-unit increase in its regressor:

A. holding the other included regressors constant

B. after proving a causal relationship

C. only when R² equals one

D. without regard to measurement units

Correct Answer: A. holding the other included regressors constant

Question 7

Which situation is a source of endogeneity?

A. An explanatory variable is correlated with the disturbance

B. The regression includes an intercept

C. The outcome is measured in dollars

D. The sample correlation is negative

Correct Answer: A. An explanatory variable is correlated with the disturbance

Question 8

A categorical predictor has five categories, and the model includes an intercept. How many indicator variables are normally required?

A. Three

B. Four

C. Five

D. Six

Correct Answer: B. Four

Question 9

In the model \ln(Y)=β₀+β₁\ln(X)+u, how is β₁ interpreted?

A. The unit change in Y from a one-unit change in X

B. The percentage change in Y from a one-unit change in X

C. The percentage change in Y associated with a 1% change in X

D. The predicted value of Y when X is zero

Correct Answer: C. The percentage change in Y associated with a 1% change in X

Question 10

Which conclusion follows from a high in-sample R²?

A. The estimated relationship is necessarily causal

B. The model cannot contain omitted-variable bias

C. The model explains a large proportion of sample variation under the fitted specification

D. The model must have low out-of-sample RMSE

Correct Answer: C. The model explains a large proportion of sample variation under the fitted specification