Regression output hands you a column of t-statistics and one overall F-statistic, and most problem sets ask a question none of them answers directly. If you are taking econometrics at Georgetown, George Washington University, American University, Johns Hopkins or the University of Maryland, this t and F tests lesson from a Washington DC econometrics tutor starts with exactly that kind of question. It builds the tests you need to answer it: the t test on one coefficient, the t test on a difference, the restricted-regression F test, and the joint test of several restrictions.
Two advertising budgets, one awkward question
A drinks brand records 30 weeks of sales, in thousands of units, along with what it spent on TV advertising and on social-media advertising, both in thousands of USD. Its analyst estimates
sales = β0 + βTV TV + βS social + u
and gets the following.
| Coefficient | Standard error | t | |
|---|---|---|---|
| Intercept | 18.79 | 1.36 | 13.82 |
| TV spend | 2.214 | 0.232 | 9.54 |
| Social spend | 1.224 | 0.271 | 4.52 |
The regression has n = 30 observations, R² = 0.907 and a sum of squared residuals of 33.68. The software also reports the estimated covariance between the two slope estimates: −0.0386.
Both coefficients are clearly significant. The marketing director asks the obvious next question: does a thousand USD on TV really sell more than a thousand USD on social media? The point estimates say 2.214 against 1.224. The analyst draws both 95% confidence intervals, sees that they overlap, and says the data cannot tell the two apart. That answer is wrong, and seeing why is the cleanest route into how t and F tests actually work.
The t test, coefficient by coefficient
Every t test in regression has the same shape. Take the estimate, subtract the value the null hypothesis claims, and divide by the standard error.
t = (b − β0) / se(b)
Under the classical assumptions, including normal errors with constant variance, and when the null is true, this statistic has a t distribution with n − k − 1 degrees of freedom, where k is the number of slope coefficients. Here that is 30 − 2 − 1 = 27. The two-sided 5% critical value is 2.052.
For TV, the null βTV = 0 gives t = 2.214/0.232 = 9.54, with a p-value well below 0.001. For social media, t = 1.224/0.271 = 4.52, again with p < 0.001. A 95% confidence interval is the set of null values a two-sided 5% t test would not reject: the estimate plus or minus 2.052 standard errors. That gives 1.74 to 2.69 for TV and 0.67 to 1.78 for social media. They overlap between 1.74 and 1.78.
If the logic of nulls, p-values and error types needs a refresher, hypothesis testing from first principles covers it. The standard errors themselves come from the OLS variance formula; the OLS derivation and the Gauss–Markov assumptions show where it comes from.
Overlapping intervals do not settle it
The director’s question is about one number, βTV − βS. So test that number directly. Its estimate is 2.214 − 1.224 = 0.990. Its standard error needs the variance of a difference, and that includes the covariance:
Var(bTV − bS) = Var(bTV) + Var(bS) − 2 Cov(bTV, bS)
Plugging in gives 0.232² + 0.271² − 2 × (−0.0386), so the standard error is 0.452. Then t = 0.990/0.452 = 2.19, above the critical value of 2.052, with a two-sided p-value of 0.037. The 95% interval for the difference runs from 0.06 to 1.92, which excludes zero. At the 5% level, the TV coefficient is significantly larger than the social-media coefficient.
Why do the intervals mislead? Two intervals fail to overlap only when the estimates are more than 2.052 × (seTV + seS) apart. But the standard error of the difference is smaller than the sum of the two standard errors for any correlation between the estimates short of a perfect negative one. So the overlap check is always conservative, and overlap can happen even when the difference is clearly significant. The covariance moves the standard error of the difference up or down relative to the uncorrelated case, but never above the sum. Here TV and social spend move together, with a correlation of about 0.6, which (with two regressors and constant error variance) makes their estimates negatively correlated. That raises the standard error of the difference, from 0.356 if you wrongly ignore the covariance to 0.452. Even so, the difference is significant.
The same question as a restricted regression
There is a second route to the same answer, and it generalises. Impose the null hypothesis on the model and see how much worse it fits. If βTV = βS, the model becomes
sales = β0 + β (TV + social) + u
so you regress sales on total ad spend. The restriction costs fit: the sum of squared residuals rises from 33.68 to 39.67. The F statistic scales that rise by the number of restrictions q and by the unrestricted model’s residual variance.
F = [(SSRr − SSRu) / q] / [SSRu / (n − k − 1)]
With q = 1, this gives F = 4.80, against a 5% critical value of 4.21 for F(1, 27). Reject, as before. That is no coincidence: 2.19² = 4.80. With one restriction, F is exactly t², and the two tests give the same p-value. Which route to use is a matter of convenience. The t route needs the covariance; the F route needs a second regression.
Testing two coefficients at once
Now ask whether advertising matters at all: βTV = βS = 0. That is two restrictions, so it needs an F test. The restricted model is just the intercept, whose sum of squared residuals is the total sum of squares, 363.99. Then
F = [(363.99 − 33.68)/2] / [33.68/27] = 132.4
far above the 5% critical value of 3.35 for F(2, 27). When the dependent variable and the sample are the same in both models, the same number, up to rounding of R², comes from the R² form, [(R²u − R²r)/q] / [(1 − R²u)/(n − k − 1)], with R²r = 0 here. This particular test, all slopes equal to zero, is the overall F statistic that standard OLS output reports.
Joint tests answer questions that a string of separate t tests cannot. Rejecting whenever either t statistic exceeds 2.052 gives a test whose size is above 5%: about 9.75% if the two estimates were independent, and how far above depends on how correlated they are. The F test uses their joint distribution, so its size is exactly 5%. It can also reject when neither t does, and the reverse.
Habits that cost marks
A few errors recur every term. The first is forgetting that a significant coefficient is not necessarily a large one. With enough data, a trivially small effect earns a big t. Report the size and the interval, not just the stars.
Another is testing against zero when the question is about some other value. A hypothesis such as “the elasticity equals 1” needs (b − 1)/se(b), not b/se(b).
Then there is using the R² form when the restricted model has a different dependent variable, which happens whenever a restriction is imposed by moving a term to the left-hand side. Use the SSR form then.
Finally, every critical value on this page assumes the classical error structure. With heteroskedastic errors the usual standard errors are wrong, and the SSR-based F test is no longer valid. Switch to heteroskedasticity-robust standard errors and a robust Wald test of the same restrictions, remembering that its critical values are large-sample approximations.
Three problems with answers
Problem 1. A regression on 30 observations with two slope coefficients reports b = 0.84 with standard error 0.35. Test H0: β = 1 against a two-sided alternative at 5%, and then test β = 0.
Answer. Against 1: t = (0.84 − 1)/0.35 = −0.46. Its absolute value is below the 27-degree-of-freedom critical value of 2.052, so do not reject. Against 0: t = 0.84/0.35 = 2.40, so reject. The data are consistent with a coefficient of 1 and inconsistent with 0.
Problem 2. A wage regression with four slope coefficients on 60 workers has R² = 0.62. Dropping three of the regressors lowers it to 0.55, with the same dependent variable. Are the three jointly significant at 5%?
Answer. F = [(0.62 − 0.55)/3] / [(1 − 0.62)/(60 − 4 − 1)] = 3.38. The 5% critical value of F(3, 55) is 2.77, and the p-value is 0.025. Reject: the three regressors matter jointly.
Problem 3 (harder). A log Cobb–Douglas production function estimated on 40 firms gives a capital coefficient of 0.38 (standard error 0.07) and a labour coefficient of 0.71 (standard error 0.09), with an estimated covariance of −0.004. Test constant returns to scale, that the two coefficients sum to 1.
Answer. The sum is 1.09. Its variance is 0.07² + 0.09² + 2 × (−0.004) = 0.005, so its standard error is 0.0707. Then t = (1.09 − 1)/0.0707 = 1.27. With 37 degrees of freedom the critical value is 2.026, so constant returns are not rejected. The equivalent F test, run as a restricted regression, gives F = t² = (0.09/0.0707)² = 1.62.
Tests like these are where econometrics marks are won or lost at Georgetown, GWU, American, Johns Hopkins and Maryland, because the question rarely names the test it wants. Econometrics tutoring for Georgetown and GWU students starts from your own regression output: which hypothesis the question is really asking about, which standard error it needs, and how to write the answer a grader expects. See how sessions work.
Work with a metrics tutor in Washington DC on your next problem set
Send the regression output or the problem set that is holding you up, and you will be matched with a PhD tutor for one-on-one sessions online. Sessions with a metrics tutor in Washington DC run on DC hours, so they fit around your classes and your Stata or R deadlines. Use the booking form, or the WhatsApp button at the top of this page.