A regression's overall F-test is highly significant but no coefficient's t-statistic clears 2 — why?
answer
- The two tests have different null hypotheses
- Joint evidence versus incremental contribution
- Each predictor is individually dispensable
- Inflated standard errors shrink every t-statistic
- The sum of the coefficients is estimated precisely
basics
~20 sThat pattern is the classic signature of severe multicollinearity. The predictors jointly explain the outcome well, which the F-test detects, but they overlap so heavily that no single coefficient can be pinned down, so every individual t-statistic stays small.
solid answer
~50 sThe two tests ask different questions. The overall F-test asks whether the predictors *jointly* explain variation in the outcome — the null is that all slopes are zero at once. Each t-test asks whether one predictor adds explanatory power *given all the others already in the model*. When predictors are highly correlated, the joint answer is a confident yes while every individual answer is "can't tell", because dropping any one predictor barely hurts the fit — its twin covers for it. Mechanically, the shared variation inflates each standard error, and inflated standard errors shrink t-statistics even when the estimates themselves are large. To confirm it, compute variance inflation factors for the suspect predictors and refit with one of them removed: if the remaining coefficient suddenly becomes large and precise, collinearity was the cause. Also test the suspects jointly with an F-test on that subset, which usually comes out strongly significant.
go deeper
Know that the overall F-test asks whether the predictors matter as a group while each t-test asks about one predictor given all the others, so the two can legitimately disagree.
Be able to explain the mechanism: overlapping predictors make each one individually removable without hurting the fit, and inflated standard errors pull every t-statistic toward zero even when the estimates are large.
Show a confirmation sequence rather than a guess — variance inflation factors, a refit with one suspect dropped, a joint test on the subset — and name the rival explanation of many weak predictors before settling on collinearity.
Decide what the organisation reports when an effect is real but unattributable. Defend publishing a joint estimate with an honest interval over a per-predictor number that will reverse on the next quarter of data.
## Two tests, two different nulls The apparent contradiction dissolves once you state what each test actually tests. - **The overall F-test** has the null hypothesis that *every* slope in the model is zero simultaneously — that the predictors, taken together, explain nothing. Rejecting it says the predictor block carries real information about the outcome. - **Each t-test** has the null hypothesis that *this one* slope is zero *while every other predictor stays in the model*. It is a question about marginal, incremental contribution: does this predictor add anything the others have not already supplied? These can disagree without any inconsistency. "This set of predictors explains the outcome" and "no individual member of the set is indispensable" are compatible statements — precisely the situation when the members duplicate each other. ## Why collinearity produces exactly this pattern With two heavily overlapping predictors, the fit can explain the outcome using either one, or any weighted blend of them. Remove the first and the second absorbs its role with almost no loss of fit — so the incremental contribution of the first, which is what its t-test measures, is near zero. The same is true of the second. Every predictor looks dispensable because every predictor *is* individually dispensable, while the group is not. On the arithmetic side, the standard error of each affected coefficient is multiplied by the square root of its variance inflation factor. A coefficient whose t-statistic would have been 7 under uncorrelated predictors falls to about 2 at a VIF of 12, and below 1 at a VIF of 50. The estimate has not shrunk; only its measured precision has collapsed. ## Confirming the diagnosis Do not stop at the plausible story. A short checklist that settles it: 1. **Compute the variance inflation factors.** If the suspects come back with large values and the untroubled predictors sit near 1, the explanation is confirmed. 2. **Refit without one suspect.** If the survivor's coefficient jumps to a large value with a small standard error and the model's R-squared barely moves, the two were carrying one shared signal. 3. **Run a joint F-test on the suspect subset.** Testing "both of these coefficients are zero" typically rejects decisively, even though neither individual test does. This is the cleanest formal demonstration that the information is real but unattributable. 4. **Look at the estimated correlation between the two coefficients.** Under collinearity it is strongly negative: the fit trades one against the other, which is why refits slide along that line. 5. **Check the confidence interval for the *sum* of the two coefficients.** It is usually far tighter than either individual interval, because the errors offset. The combined effect is well identified; only its division is not. ## Alternative explanations to rule out Severe collinearity is by far the most common cause, but be able to name the rivals: - **Many weak predictors.** A model with a large number of predictors each contributing a genuinely small amount can produce a significant joint test with no individual test clearing the bar. Here VIFs are near 1 and the pattern is about diffuse signal, not redundancy — the fix is a smaller, better-motivated predictor set, not deduplication. - **A borderline overall test.** If the F-test is only marginally significant and the largest t-statistic is around 1.8, this may be ordinary noise near a threshold rather than a structural finding. The described situation — F *highly* significant, *no* t-statistic near 2 — is much stronger evidence of collinearity. - **Small sample with correlated design.** Limited data amplifies both effects at once, and the honest conclusion may simply be that the design cannot support per-predictor claims. ## What to do about it The response depends entirely on the model's purpose. - **If the model is for prediction**, this pattern is not a problem at all. Joint explanatory power is what drives forecasts, and it is intact. Report the fit and move on. - **If you need one specific effect**, you have a genuine identification limit. Options: drop one of the duplicates on domain grounds and interpret the survivor as the *combined* effect (labelling it honestly); replace the pair with a single deliberately constructed summary measure; report the joint effect and its tight interval rather than the two unstable pieces; or gather data in which the predictors move independently, which is the only fix that truly adds information. - **What not to do**: report the model as if nothing happened, or quietly keep the version whose signs match the story you wanted. Both convert an unidentifiable quantity into a claim. ## The one-line takeaway A significant F with no significant t means *the predictors matter but the data cannot say which one*. It is a statement about attribution, not about whether the model works.
- How would you formally demonstrate that the two suspect predictors matter together?Run an F-test on that subset — the null that both of their coefficients are zero simultaneously. Under collinearity this rejects decisively even though neither individual t-test does, because the joint test asks whether the pair as a whole adds explanatory power rather than whether either is incrementally necessary given the other.
- What causes this pattern other than multicollinearity?A model with many predictors that each contribute a genuinely small amount can show joint significance without any individual coefficient reaching the threshold. The distinguishing check is the variance inflation factors: near 1 across the board means diffuse weak signal rather than redundancy, and the response is a smaller predictor set rather than deduplication.
- Why is the confidence interval for the sum of the two coefficients so much narrower than either one alone?Because the two estimates are strongly negatively correlated: whenever the fit pushes one up it pulls the other down by nearly the same amount. Adding them cancels most of that shared error, so the combined effect is well identified even though the split between the two predictors is not.
- Does this pattern mean the model should not be shipped?Not if its job is prediction. Joint explanatory power drives fitted values and is fully intact here, so forecasting performance is unaffected for cases resembling the training data. It blocks only per-predictor interpretation — attributing an effect to one predictor or setting policy on one lever.
A choir is unmistakably loud, but with everyone singing the same note you cannot say whose voice is carrying it. Mute any one singer and the sound barely changes.
saying these in an interview costs you the question
- Calls the result a contradiction or a computational bug
- Concludes the model has no predictive value
- Deletes every insignificant predictor one at a time
- Claims the F-test and t-tests test the same hypothesis
- Reports the coefficient signs as if they were reliable