When Should You Add a Quadratic Term in Empirical Research? How to Interpret It? How to Justify Its Inclusion?

1. Introduction: A Straight Line vs. an Inverted U—Your Theoretical Choice Determines the Mathematical Form of Your Regression

You study "the effect of income on happiness." Your intuition is: when poor, an extra dollar greatly boosts happiness; but beyond a certain level of wealth, additional money has little effect on happiness—perhaps even slightly reducing it due to comparison and stress. This implies that the relationship between income and happiness is not a straight line—but a curve that first rises then falls (or first rises then flattens).

You run a linear regression:

happiness = β₀ + β₁ × income + controls + ε

β̂₁ is significantly positive. The reviewer asks: "You assume the effect of income on happiness is linear—that is, each additional 10,000 yuan increases happiness by the same amount, whether the person's income is 10,000 or 1,000,000. Is this assumption reasonable? Have you considered the possibility of diminishing marginal utility?"

You realize—you need a quadratic term.

Core message: Adding a quadratic term means you assume the effect of X on Y is not constant, but varies with the level of X itself. This is the simplest step from "a straight line" to "a curve." After adding the quadratic term, the coefficient is no longer a single number—it becomes a function of X: the marginal effect of X on Y = β₁ + 2β₂·X. When β₂ < 0, the marginal effect is diminishing (inverted U-shape); when β₂ > 0, the marginal effect is increasing (U-shape). Whether you should add a quadratic term depends on whether theory supports "the effect of X should vary with the level of X itself," on whether your scatter plot shows discernible curvature, and on whether you can keep your conclusions interpretable while accepting a "more complex model."

2. When Should You Consider Adding a Quadratic Term?

2.1 Theory Predicts "Diminishing/Increasing Marginal Effects"

This is the strongest reason for adding a quadratic term. If your theoretical hypothesis explicitly states that "the effect of X on Y decreases/increases as X itself increases"—then the theory itself demands a nonlinear model. Classic examples:

If the X-Y relationship you study has an "optimal level" or "turning point"—beyond which the effect of X changes from positive to negative (or from negative to positive)—then the quadratic term is not "icing on the cake" but the natural mathematical expression of your theory.

2.2 Scatter Plots (or Local Smoothing Lines) Show Obvious Curvature

Before running any regression, plot the data first. In Stata:

twoway (scatter y x) (lowess y x, bw(0.5)), ytitle("Y") xtitle("X")

lowess is a locally weighted smoothing line—it does not presuppose any functional form, but fits based on the local density of the data itself. If this line exhibits any of the following patterns:

• Rise then flatten (diminishing marginal effect) → consider adding x² (β₂ < 0)
• Rise then fall (inverted U) → consider adding x² (β₂ < 0)
• Fall then rise (U-shape) → consider adding x² (β₂ > 0)
• Continuously rising but with changing slope (accelerating or decelerating growth) → consider adding x²

If the lowess line and the linear fit line (lfit) essentially coincide → there is no evidence supporting the need for nonlinearity.

But an important caveat: lowess displays nonparametric correlation, not partial correlation (it does not control for other variables). The partial effect after controlling for covariates may differ from the shape shown by lowess. So the scatter plot is "reconnaissance" rather than "conviction"—it tells you "you should check for nonlinearity," but it does not replace formal testing.

2.3 Existing Literature Commonly Includes Quadratic Terms

If your research topic is widely recognized in the literature as having nonlinear relationships—for example, the age-squared term in the Mincer wage equation is almost a "standard component"—then even if you do not find strong evidence of nonlinearity in your particular sample, following the literature's convention by including a quadratic term is generally acceptable. Reviewers would instead ask: "Why didn't you include age²?"

2.4 Your Research Question Itself Is "Whether an Inverted U Relationship Exists"

In some papers, the core research hypothesis is precisely that "X has an inverted U (or U) relationship with Y." In such papers, the quadratic term is not a robustness check—it is your core test. These papers typically need to state this directly in the abstract:

"This paper tests the inverted U-shaped relationship between environmental pollution and GDP per capita (the Environmental Kuznets Curve hypothesis). The regression results support the inverted U hypothesis—the coefficient on the squared term of GDP per capita is significantly negative, with the turning point at approximately USD 8,000 GDP per capita."

3. Should You Only Add a Quadratic Term for the Core Explanatory Variable? Do Control Variables Need One Too?

3.1 In General, Add Quadratic Terms Only for the Core Explanatory Variable

In the vast majority of empirical papers, quadratic terms appear only on the core explanatory variable. Control variables could in principle also have nonlinear effects, but the paper's focus is on identifying the effect of the core X on Y—if you tentatively add quadratic terms for every control variable, you will quickly fall into "overfitting" and "cherry-picking" (data-mining).

Standard practice:

• Core explanatory variable → If theory supports it, add the quadratic term. Explain clearly.
• Control variables → Usually keep linear form. Unless the nonlinearity of a particular control variable is widely confirmed in the literature (e.g., age almost always enters quadratically in wage equations), or the nonlinearity of a control variable directly affects the estimation of the core coefficient.

3.2 But There Is an Important Exception—If a Control Variable's Quadratic Term Significantly Affects the Estimation of the Core Coefficient

Suppose you study the effect of "education (X) on wage (Y)." Your core variable is education, and control variables include age. The effect of age on wage is known to be inverted U-shaped (wage rises then falls with age).

If you only use age (linear), the effect of age in the model is incorrectly constrained to a straight line. Could this affect the coefficient estimate on education? Possibly—if education and age are correlated (younger people have more education), then misspecifying the functional form of age can indirectly bias the coefficient on education.

In this case, the quadratic term on age, although a "control variable's quadratic term," is included to correctly specify the model's functional form to protect the estimation of the core coefficient. You do not need to over-interpret age² in the paper—it is an auxiliary specification serving the main objective.

3.3 When Is a Quadratic Term Absolutely Unnecessary for Control Variables?

• Dummy variables (0/1 variables)—squaring them yields the identical variable, which is meaningless.
• Categorical variables taking only a few discrete values—the quadratic term only adds redundant degrees of freedom.
• Variables with no theoretical basis for a curvilinear relationship with your core variable.
• When your sample size is small (n < 100) and the model already has many control variables—avoid overfitting and data mining.

4. Correct Interpretation of the Quadratic Term Coefficient—Three Things You Must Clarify

4.1 After Adding the Quadratic Term, the "Coefficient" on X Is No Longer a Constant

The model becomes:

$Y = \beta_0 + \beta_1 X + \beta_2 X^2 + \gamma' \mathbf{W} + \varepsilon$

The marginal effect of X on Y is a linear function of X itself:

$\frac{\partial Y}{\partial X} = \beta_1 + 2\beta_2 X$

This is not a constant—it takes different values at different levels of X.

Therefore, after adding a quadratic term, you can no longer simply say "the coefficient on X is β̂₁." You must report "the marginal effect of X on Y at the mean of X is XX; when X increases from the 25th percentile to the 75th percentile, the change in the marginal effect is XX." Many papers fail precisely here—they add a quadratic term but still interpret the coefficients using linear model logic—this is the most fatal interpretation error.

4.2 The Correct Way to Report—The Four-Part Package

When reporting quadratic term regression results in a paper, you cannot simply present a coefficient table and be done. You need the four-part package:

(1) Coefficient table: Report the point estimates, standard errors, and significance of β̂₁ and β̂₂ as usual.

(2) Turning point: If β̂₂ < 0 (inverted U), calculate the turning point:

$X^* = -\frac{\hat{\beta}_1}{2\hat{\beta}_2}$

The turning point must fall within the sample range of X to have economic meaning. If $X^*$ lies beyond the upper tail of the sample (e.g., X ranges from 1 to 10, but the turning point is 15), then what you observe in the sample is only the "left half of the rising segment"—you have not actually observed an inverted U.

If β̂₂ > 0 (U-shape), the turning point formula is the same, and its meaning is the minimum point.

(3) Marginal effects plot: This is the most important visualization after adding a quadratic term. Plot the marginal effect curve of X (vertical axis = $\hat{\beta}_1 + 2\hat{\beta}_2 X$, horizontal axis = X), with a 95% confidence interval. This graph allows readers to see intuitively—over which range of X is the effect of X on Y significantly positive? Over which range is it significantly negative? Where is the turning point?

Stata commands (margins + marginsplot):

reg y c.x##c.x controls    // c.x##c.x is equivalent to x + x²
margins, dydx(x) at(x=(min(1)max))
marginsplot

(4) Confidence interval for the turning point: You do not just report a point estimate $\hat{X}^*$—you also report the 95% confidence interval for this turning point. The turning point is a nonlinear function of two coefficients, β̂₁ and β̂₂—its standard error cannot be computed with a simple analytical formula; you need the Delta method or Bootstrap.

If the 95% confidence interval for the turning point is too wide (e.g., extending from X = 2 all the way to X = 50), this indicates that the data do not precisely pin down the shape of the curve—the very existence of the "turning point" may be unreliable.

4.3 Interpretation Examples—Correct vs. Incorrect

❌ Incorrect interpretation (reporting only coefficients):

"The coefficient on X is 3.20, and the coefficient on X² is −0.15. Both are significant. This indicates that X has an inverted U-shaped effect on Y."

This statement provides no substantive information—the reader does not know the magnitude of the marginal effect, where it is positive, where it is negative, or where the turning point is.

✅ Correct interpretation (complete four-part package):

"Column (2) of Table X reports the regression results with the quadratic term. The coefficient on the linear term of X is 3.20 (p < 0.01), and the coefficient on the quadratic term is −0.15 (p < 0.01), supporting an inverted U-shaped relationship. The turning point is located at X* = 3.20 / (2 × 0.15) ≈ 10.7. In the sample, the 25th percentile of X is 4 and the 75th percentile is 18, so the turning point of 10.7 lies within the interquartile range, indicating that the inverted U-shaped relationship exists within a reasonable range covered by the sample. The marginal effects analysis (Figure X) shows that when X < 10.7, the marginal effect is significantly positive; when X > 10.7, the marginal effect is negative, but it only becomes significantly negative at the 95% confidence level once X exceeds 16. The 95% Bootstrap confidence interval for the turning point of 10.7 is [8.3, 14.1], further supporting the existence of the inverted U-shaped relationship in the sample."

5. How to Justify Adding a Quadratic Term?—Three Things You Need to Do in Your Paper

5.1 Theoretical Justification—Why Should It Be Curved? (Place in the "Theoretical Analysis" or "Research Hypotheses" Section)

Your paper needs to explain in the theory or literature review section—why should the effect of X on Y not be a straight line? This is typically based on one of the following logics:

• Diminishing marginal utility: Each additional unit of X has a weaker incremental effect. This is the most common argument in economics.
• Optimal level theory: Below some "optimal" level of X, increases in X are beneficial; beyond this level, increases in X begin to have adverse effects. You need to draw on existing theory or literature to argue for the existence of this optimal level.
• Two opposing forces: There exist two forces operating in opposite directions—one strengthens as X increases, and the other also strengthens as X increases, but their relative strength differs at different levels of X. For example, "firm size → economies of scale (positive) vs. growing managerial complexity (negative)."

The key point of theoretical justification is: you cannot simply say "I guess it's not linear"—you need to provide a mechanism. This mechanism need not be complicated—"economies of scale imply increasing marginal returns, but beyond a certain point, coordination costs exceed scale benefits" suffices. One or two sentences, clearly articulating the mechanism.

5.2 Empirical Justification—Data and Tests (Place in the "Empirical Results" or "Robustness Checks" Section)

You cannot justify solely by "the quadratic coefficient β̂₂ is significant"—you also need:

(a) Plot a lowess smoothed scatter plot. This is the most intuitive evidence—without any functional form assumption, does the data itself exhibit curvature?

(b) Compare the fit of linear and quadratic models. Report the adjusted R² or AIC/BIC of both models, showing whether adding the quadratic term leads to a substantial improvement in overall fit. If adjusted R² goes from 0.35 to 0.351, this does not support the necessity of the quadratic term—your improvement is negligible, and you are merely fitting noise.

(c) Ramsey RESET test (model specification test). The RESET test is a general test for detecting omitted nonlinear relationships or interaction effects in the model. Under the null hypothesis that "the model is correctly specified" (i.e., the linear form suffices), the RESET test statistic follows an F distribution. If the RESET test rejects the null, it means your linear model omits important nonlinear features—this provides formal test support for adding a quadratic term.

reg y x controls
estat ovtest    // Ramsey RESET test

(d) Joint significance test. Test H₀: β₁ = 0 and β₂ = 0 (i.e., X has no effect on Y at all—whether linear or nonlinear). This is not a direct test of "whether a quadratic term is needed," but it is the correct significance reporting that replaces the "single t-test" after adding a quadratic term—the joint F-test is more rigorous than looking at the t-test on β̂₂ alone.

reg y c.x##c.x controls
test x c.x#c.x    // joint significance

5.3 Robustness Justification—How Sensitive Is Your Conclusion to the Quadratic Specification?

In the robustness checks section, you can do the following to show that adding the quadratic term is not merely a statistical trick:

• Report the baseline regression without the quadratic term (purely linear) → show whether the core conclusions are consistent.
• If your variable has a zero point or natural reference point, re-run the quadratic model with centered X → show whether the coefficient magnitudes are robust.
• If theory predicts a more complex nonlinearity (e.g., a cubic curve), try adding a cubic term and show that its incremental contribution is insignificant → indirectly arguing that the quadratic form is sufficient.

6. Practical Implementation and Common Issues with Adding Quadratic Terms

6.1 Centering—To Do or Not to Do?

Before adding a quadratic term, many textbooks recommend centering X first ($X_i - \bar{X}$), then squaring. The reason is simple: X and X² are naturally highly correlated (a person with 16 years of education has a square of 256)—after centering, $(X - \bar{X})$ and $(X - \bar{X})^2$ are approximately orthogonal, which can substantially reduce collinearity.

But note:

• Centering only changes the meaning of β̂₁—from "the marginal effect when X = 0" to "the marginal effect when X = $\bar{X}$." β̂₂ and the turning point are unaffected.
• Centering does not affect the overall model fit (R² is unchanged).
• If your variable has a natural zero point (e.g., "years since policy implementation"—X = 0 in the year of implementation), centering would make this zero point lose its intuitive meaning.

Recommendation: If your variable's zero point has a natural interpretive meaning (e.g., distance, age, income), you may not need to center; if the zero point has no natural meaning (e.g., scale scores, standardized indices), centering is recommended to reduce collinearity.

6.2 What If the Turning Point Falls Outside the Sample Range?

If $\hat{X}^*$ exceeds the sample range of your X, you need to honestly tell the reader: what you observe in the sample is only one arc segment of the curve, not the entire U or inverted U. You cannot claim to have "found a U-shape"—you can only say "within the sample range, the marginal effect of X on Y is diminishing (or increasing)."

Example: The sample range of X is [2, 15], and the turning point is $\hat{X}^* = 22$. In your sample, you only observe the rising segment—the marginal effect gradually declines from 3.0 at X = 2 to 1.5 at X = 15, but remains positive throughout. What you observe is not an "inverted U" but "diminishing but still positive marginal effects"—these two have completely different theoretical implications.

6.3 After Adding the Quadratic Term, Does the Original β̂₁ Coefficient Still Have Independent Meaning?

Yes—but the interpretation must be tied to the value of X. The meaning of β̂₁ is: the marginal effect of X on Y when X = 0. (In a centered model, β̂₁ is the marginal effect when X = $\bar{X}$.)

So whether β̂₁ has meaning in isolation depends on whether "X = 0" has meaning. If X never takes the value 0 in your data (e.g., "years of firm establishment" actually ranges from 1 to 50), then β̂₁ has no independent economic meaning—it is an extrapolated value outside the sample range. In this case, you should report "at the mean of X, the marginal effect is XX" rather than interpreting β̂₁ in isolation.

6.4 What If the Quadratic Term Is Not Significant?

If you add a quadratic term and β̂₂ is not significant—this could mean:

• The relationship between X and Y is indeed approximately linear. You should state in the paper: the quadratic coefficient is not significant, and the RESET test cannot reject the linear specification; therefore, subsequent analysis uses the linear model as the baseline.
• Your sample size is insufficient to detect curvature (curvature requires more data to estimate two parameters rather than just one).
• Curvature exists but is not quadratic—it could be logarithmic, exponential, or piecewise linear. The quadratic form is only one type of nonlinearity, not the only one.

Do not insist on keeping an insignificant quadratic term because "I added it, so my model must be better." An insignificant quadratic term merely consumes a degree of freedom, weakens your core inference—and reviewers will question your motives.

7. How Is This Tested in Exams?

7.1 Common Exam Formats

Format 1: Given regression output, interpret it.

"ln(wage) = 0.05 + 0.12 × age − 0.0015 × age² + 0.08 × educ. Please explain the meaning of the age coefficient, and calculate at what age the marginal effect of age on wage changes from positive to negative."

Answer:

• The marginal effect of age = 0.12 − 0.003 × age
• At age = 30, the marginal effect = 0.12 − 0.09 = 0.03 (each additional year increases wage by approximately 3%)
• The turning point is age* = 0.12 / (2 × 0.0015) = 40 years
• Before age 40, the marginal effect is positive; after age 40, it is negative

Format 2: Theoretical analysis.

"When studying the impact of environmental regulation on firm innovation, the Porter hypothesis argues that moderate regulation incentivizes innovation, but overly stringent regulation may stifle it. Please specify an appropriate econometric model to test this hypothesis."

Should include: linear term + quadratic term; expected linear coefficient > 0, quadratic coefficient < 0 (inverted U); report the turning point; discuss whether the turning point falls within the sample range.

Format 3: Given a scenario, judge whether a quadratic term is reasonable.

"A researcher studies the effect of urban population density on commuting time. She finds that the coefficient on the linear term of population density is positive, the coefficient on the quadratic term is also positive, and both are significant. What does this mean? Is this result reasonable?"

Answer: Both coefficients being positive means the marginal effect is continuously increasing—the higher the population density, the larger the additional effect of one more unit of density on commuting time (rather than diminishing). This could be theoretically reasonable—when density exceeds a certain critical threshold, congestion deteriorates nonlinearly (each additional car causes more congestion than the previous one).

8. Summary

Five core takeaways about quadratic terms:

1. A quadratic term = you assume the effect of X on Y is not a constant but a linear function of X. The marginal effect = β₁ + 2β₂·X. At different values of X, the marginal effect differs. This is the simplest and most commonly used relaxation of the "linearity assumption."

2. The justification for adding a quadratic term comes from three levels: theory (diminishing marginal utility/optimal level/two opposing forces), data (curvature in scatter plots), and literature (whether your field has this tradition). The weakest justification is "I tried it and it was significant"; the strongest is "my theory requires the effect to reverse direction beyond some point."

3. Generally, add quadratic terms only for the core explanatory variable. Control variables remain linear, unless the nonlinear form of a specific control variable is widely confirmed in the literature (e.g., age), or misspecifying its functional form would affect the estimation of the core coefficient.

4. After adding a quadratic term, you cannot say "the coefficient on X is XX." You must report the four-part package: coefficient table + turning point + marginal effects plot + confidence interval for the turning point. Reporting β̂₁ alone is insufficient—the marginal effect is a function, not a number.

5. Whether the turning point falls within the sample range determines whether you can claim to have "found a U/inverted U shape." If the turning point is outside the sample range, you have only observed one arc segment of the curve—what you see is "diminishing marginal effects" rather than "a reversal in direction." These two have vastly different implications.

One-sentence conclusion:

"Adding a quadratic term means you are no longer satisfied with the answer 'when X increases by one unit, Y changes by a fixed β'—you are asking a more honest but also more difficult question: 'Does the effect of a one-unit increase in X on Y depend on how large X already is?' Once this question is raised, the linear model is no longer sufficient. The quadratic term is the first step into the nonlinear world—the simplest step, but also the one most easily misinterpreted."

9. Presentation Suggestions for Bilibili/WeChat Official Account

• Bilibili video: Suggest using the visual metaphor of "from a straight ruler to a tape measure." Opening: a researcher uses a straight ruler (linear model) to measure the "income → happiness" relationship—but the data points are clearly curved, and the straight ruler cannot fit. Voiceover: "If your theory says 'diminishing marginal utility,' but you use a regression that assumes a linear effect by default—you are using a straight ruler to measure a curve." Then introduce the quadratic term—the straight ruler becomes a bendable tape measure (quadratic curve), and the same ruler now fits the data. Then develop in three acts: Act 1 "When to add"—switch through five classic inverted U/U examples (Environmental Kuznets Curve, age-wage, inflation-growth, competition-innovation, income-happiness), in each scenario data points scatter away from the straight line → after adding the quadratic term, the tape measure fits. Act 2 "How to interpret"—marginal effect animation: horizontal axis is X, vertical axis is ∂Y/∂X, showing how the marginal effect changes from positive to negative (inverted U) or from negative to positive (U-shape) as X increases, with the turning point highlighted and flashing. Act 3 "How to justify"—visual flow of three justifications (theory → scatter plot curvature → RESET test → joint F-test → turning point confidence interval), with "✅ Pass" or "❌ Fail" markers appearing on each.
• WeChat Official Account: The marginal effect formula and turning point formula should be displayed in separate formula boxes. The correct vs. incorrect interpretation should be made into a two-column comparison card (incorrect: only reporting coefficients and saying "inverted U exists"; correct: reporting the four-part package). Three curve shapes (diminishing marginal effects, inverted U, U-shape) should each be paired with a marginal effects plot plus a schematic diagram with the turning point annotated. The three-step Stata commands (regression → margins → marginsplot) should be presented in code blocks. The two scenarios of whether the turning point falls within the sample range (inside vs. outside) should be made into a comparison figure, using different colors to mark the observed arc segment and the extrapolated portion.
• Recommended titles:
  • Main title: 《When Should You Add a Quadratic Term in Empirical Research? How to Interpret the Coefficients After Adding It?》
  • Alternative title: 《Marginal Effect = β₁ + 2β₂·X — Four Things You Must Know After Adding a Quadratic Term》
  • New media title: 《Have You Added a Quadratic Term to Your Regression?—A Complete Guide to Nonlinear Relationships, from Theory to Justification》
• Key quotes:

  "After adding a quadratic term, the coefficient is no longer a constant—it becomes a function of X. You no longer have a 'magnitude of effect'—you have an 'effect curve.' Your reporting must also be upgraded from 'one sentence' to a 'four-part package.'"

  "Only when the turning point falls within the sample range can you say you have 'found a U-shape.' If the turning point is outside—what you see is only an arc segment, where the marginal effect is diminishing but does not change sign. The theoretical implications of these two are completely different—don't conflate them."

  "A quadratic term is not a seasoning you add whenever you like. You need a theoretical mechanism—why should the effect differ at different levels of X? If your answer is 'because I saw the scatter plot curve'—that's not wrong, but it's not enough. Find a story, not just a curve."

  "The best nonlinearity is not statistical significance but theoretical necessity. The Kuznets curve does not need the t-statistic of the quadratic term to prove itself—it is already curved in economic intuition."