Using RESET Test for Detecting Functional Form Errors

Imagine building a complex model of a house. You meticulously plan the number of rooms, the materials, and the foundation. But what if your fundamental blueprint-the very architectural drawings-is wrong? What if you assumed all walls must be straight, when the design actually requires curves? The house might still stand, but it would be flawed, unstable, and not what you intended. In economics, this is what we call a model specification error. We build a model assuming a simple, linear relationship (a straight wall), when the true, underlying economic reality is complex and non-linear (a curved wall). How do we detect such a fundamental error in our blueprint? This is where a clever diagnostic tool comes in: the Ramsey RESET test.

The RESET test, which stands for Regression Specification Error Test, is one of the most popular and useful tools in an econometrician’s toolkit. It’s like a building inspector who comes to your site, not to check every single nail, but to perform a general stress test. This test doesn’t tell you *exactly* which wall is wrong or how to fix it, but it gives you a crucial warning: “Stop. Something about your fundamental design is incorrect.” It’s a general test designed to sniff out problems, particularly when our model’s “functional form” is wrong.

What is a model specification error? 🔗

In econometrics, we often start with the workhorse of our field: the Ordinary Least Squares (OLS) linear regression. We assume a model looks something like this:

Y = β₀ + β₁X₁ + β₂X₂ + ε

This equation states that our dependent variable (Y) is a simple, straight-line function of our independent variables (X₁ and X₂). A specification error occurs when this assumed model does not match the true, underlying process that actually generates Y. These errors are serious because they can lead to biased and inconsistent coefficient estimates, meaning our conclusions about the world are not just imprecise-they are fundamentally wrong.

The three classic culprits 🔗

Specification errors generally fall into three categories:

1. Omitting a relevant variable: You try to predict a student’s test score using only “hours studied” but forget to include “prior knowledge.” Since prior knowledge affects the test score and is likely correlated with study time, your model is flawed, and the effect you measure for “hours studied” will be biased.
2. Including an irrelevant variable: You try to predict crop yield using “rainfall” and “fertilizer,” but you also include “the farmer’s shoe size.” This is less serious-it won’t bias your other coefficients, but it will increase their standard errors, making your model less precise.
3. Choosing the wrong functional form: This is the big one, and the primary target of the RESET test. You assume the relationship between fertilizer and crop yield is a straight line. But in reality, the law of diminishing returns applies: the first bag of fertilizer helps a lot, the 10th bag helps a little, and the 100th bag might even poison the soil. The true relationship is a curve, not a line. By forcing a line onto a curve, your model is fundamentally misspecified.

The RESET test, developed by James B. Ramsey in 1969, is a general test that is particularly good at catching that third problem. It cleverly checks if your model is missing some non-linear features that it should have included.

How the RESET test cleverly detects non-linearity 🔗

The genius of the RESET test is that it doesn’t require you to *know* the non-linearity in advance. You don’t have to guess whether you’re missing an X² term, an X³ term, or a complex interaction. Instead, it uses the model’s own predictions as a proxy for all these “hidden” non-linearities.

The core idea is this: If your linear model is truly correct, then adding non-linear versions of its *own predictions* should not provide any new explanatory power.

Think about it. Your model’s prediction, called the fitted value (or $\hat{Y}$), is just the linear combination of your variables ($\beta_0 + \beta_1X_1 + \beta_2X_2$). If this linear combination is the *whole story*, then why would the *square* of this prediction ($\hat{Y}^2$) or the *cube* of this prediction ($\hat{Y}^3$) be able to explain *anything* more about Y?

They shouldn’t. They would just be redundant.

But… if the *true* model was, say, $Y = \beta_0 + \beta_1X_1 + \beta_2X_1^2 + \epsilon$, your simple linear model missed the $X_1^2$ term. In this case, the $\hat{Y}^2$ and $\hat{Y}^3$ terms will act as proxies for the missing $X_1^2$ (and other complex terms) and *will* show up as statistically significant. When they do, the test waves a big red flag.

The step-by-step application of RESET 🔗

Let’s walk through exactly how you would perform this test. Imagine you are a real estate analyst with a simple model for house prices.

Step 1: Estimate your original (restricted) model 🔗

First, you run your initial regression-the one you *think* is correct. This is called the restricted model because it’s “restricted” to being purely linear.

Your model: $Price = \beta_0 + \beta_1SquareFeet + \beta_2Bedrooms + \epsilon$

You run this OLS regression. After running it, you ask your statistical software (like R, Python, or Stata) to save the fitted values (the predicted prices, $\hat{Y}$) for every house in your dataset.

Step 2: Create the new (unrestricted) model 🔗

Next, you create new “test” variables. The most common choice is to add the square and the cube of the fitted values you just saved. Why these? Because they are flexible enough to approximate many different kinds of smooth non-linear functions.

Your new variables are:

• $P_{\hat{Y}2} = \hat{Y}^2$ (the squared fitted price)
• $P_{\hat{Y}3} = \hat{Y}^3$ (the cubed fitted price)

Now, you create a new, unrestricted model by adding these new variables to your original one: $Price = \beta_0 + \beta_1SquareFeet + \beta_2Bedrooms + \gamma_1P_{\hat{Y}2} + \gamma_2P_{\hat{Y}3} + \nu$

Note that we use different coefficient names (gamma, $\gamma$) for our new test variables. This is the model we will use to test our original blueprint.

Step 3: Run the F-test for joint significance 🔗

This is the moment of truth. We need to find out if $\gamma_1$ and $\gamma_2$ are, as a group, statistically different from zero. If they are, it means our new variables have real explanatory power, which is bad news for our original model.

We state our hypotheses clearly:

• Null Hypothesis (H₀): $\gamma_1 = 0$ and $\gamma_2 = 0$.
  This means our new terms are useless. The original model is correctly specified.
• Alternative Hypothesis (H_A): At least one $\gamma$ is not equal to 0.
  This means our new terms *are* significant. The original model is misspecified.

We use an F-test to compare the fit of the restricted model (from Step 1) against the unrestricted model (from Step 2). The F-statistic measures how much the $R^2$ (the model’s explanatory power) improved by adding the new $\hat{Y}$ terms, adjusting for the number of terms we added. Statistical software will do this calculation for you and give you the most important number: the p-value.

Step 4: Interpret the results 🔗

The p-value tells you the probability of seeing an F-statistic this large *if the null hypothesis were true*.

• If p-value > 0.05 (or your chosen significance level): You fail to reject the null hypothesis. This is good news! It means there is no statistical evidence to suggest your model’s functional form is wrong. Your simple linear “blueprint” appears to be adequate.
• If p-value < 0.05: You reject the null hypothesis. This is your “Check Engine” light. The test provides strong evidence that your original model is misspecified. The added non-linear terms *are* significant, which implies your original model was improperly capturing the non-linear nature of the relationships.

The crucial limitations of the RESET test 🔗

A significant RESET test is a call to action, but it’s famously unhelpful in telling you *what* action to take. This is its single greatest weakness.

Limitation 1: It’s a test of ‘something’s wrong,’ not ‘what’s wrong’ 🔗

The RESET test is a *diagnostic*, not a *cure*. It’s a “test of general misspecification.” It only signals a problem; it does not provide any guidance on the correct functional form. As a diagnostic tool, its job is to alert you to a fever, not to tell you if the cause is the flu or an infection.

If you reject the null hypothesis, you are left with several questions:

• Should I add a quadratic term ($X_1^2$) for one of my variables?
• Should I use a logarithmic transformation (e.g., $\log(Y)$ or $\log(X_1)$)?
• Should I add an interaction term ($X_1 \times X_2$)?

The test is completely silent on this. You, the econometrician, must use economic theory, common sense, and other diagnostic tools (like residual plots) to figure out *how* to fix the model.

Limitation 2: It can’t reliably distinguish functional form from omitted variables 🔗

This is a subtle but critical point. The test is *designed* to detect functional form misspecification. However, it can also be triggered by a “bad” case of omitted variable bias. If you leave out an important variable (X₃) that is *also* non-linearly related to your included variables (X₁ and X₂), the $\hat{Y}$ terms might just be acting as a proxy for that missing X₃.

Therefore, a “failed” RESET test (a low p-value) could mean:

1. Your functional form is wrong (e.g., you need $X_1^2$).
2. You are missing an important variable (e.g., you need X₃).
3. Both.

It’s generally considered more powerful at detecting functional form issues, but it’s not a clean-cut diagnosis. This ambiguity means you can’t just blindly “fix” the functional form; you must also reconsider your entire theoretical model and variable selection.

Limitation 3: It has low power in some situations 🔗

The test isn’t perfect. It relies on the assumption that $\hat{Y}^2$ and $\hat{Y}^3$ are good proxies for the *true* hidden non-linearity. If the true relationship is very complex or unusual (e.g., involves trigonometric functions or sharp “kinks”), these simple polynomial terms might not be very good at capturing it. In that case, the test might (incorrectly) fail to reject the null, giving you a false sense of security.

What to do after a failed RESET test 🔗

So, your p-value is 0.01. The “Check Engine” light is on. What now? Don’t panic. This is where the “art” of econometrics begins.

1. Go back to theory: First, stop and think. Does economic theory suggest a non-linear relationship? (e.g., diminishing returns to education, U-shaped average cost curves). Theory should always be your primary guide.
2. Use graphical analysis: This is your most powerful tool. Create a residual plot, plotting your model’s errors (residuals) against its fitted values ($\hat{Y}$). If your model is well-specified, this plot should look like random, patternless “static.” If you see a U-shape, a funnel shape, or any clear pattern, that’s a massive clue that you’re missing a non-linear term.
3. Try alternative specifications: Based on theory and your plots, start experimenting.
   • Add a quadratic term ($X^2$) for a variable that you suspect has a non-linear (e.g., U-shaped) effect.
   • Try logarithmic transformations. A log-log model ($\log(Y)$ on $\log(X)$) measures elasticity, while a log-lin model ($\log(Y)$ on $X$) measures semi-elasticity. These are very common ways to handle non-linearity and skewness.
   • Add interaction terms ($X_1 \times X_2$) if you believe the effect of one variable depends on the level of another.
4. Test again: After you create a new, theoretically-sounder model, run the RESET test on *it*. If your new model is a better specification, the RESET test should now “pass” (with a high p-value).

The Ramsey RESET test is an essential, if imperfect, part of the econometric process. It doesn’t give you the answer, but it forces you to ask the right questions. It stops you from blindly trusting a simple linear model in a world that is almost never that simple.

What do you think? Have you ever trusted a simple model, only to suspect that the real-world relationship was far more complex? When a diagnostic test like RESET gives you a “red flag,” what’s your first step: do you trust it immediately, or do you look for other evidence (like plots or theory) first?

References

1. https://online.stat.psu.edu/stat462/node/179/
2. https://www.jstor.org/stable/1914102
3. https://www.economics.mcmaster.ca/courses/econ4g03/lecture10.pdf
4. https://www.imf.org/external/pubs/ft/wp/2004/wp0498.pdf
5. https://www.statsmodels.org/stable/generated/statsmodels.stats.diagnostic.linear_reset.html