Why a Coefficient Test Is Not a Deletion Rule
A fitted ARIMA model may contain a coefficient whose estimate is small relative to its standard error. That is evidence of uncertainty about the coefficient, not automatic proof that the corresponding term should be removed. Conversely, a coefficient that passes a significance threshold does not by itself establish that the full model is adequate or useful for forecasting.
The distinction matters because a test answers a limited question about a coefficient in a specified model. Repeatedly testing terms, deleting those that fail a threshold, and refitting changes the model in response to the same data. The resulting estimates and nominal p-values no longer have the simple interpretation they would have had if the model had been fixed in advance.
Wald Tests and Confidence Intervals
Suppose a candidate ARIMA model has been specified and estimated by a suitable likelihood method. For a coefficient \(\beta_j\), let \(\hat{\beta}_j\) be its estimate and let \(\operatorname{SE}(\hat{\beta}_j)\) be its estimated standard error. Under the null hypothesis \(H_0:\beta_j=0\), the approximate Wald statistic is
When the estimator is approximately normally distributed and the null is true, this statistic is approximately standard normal. A two-sided approximate p-value is \(2[1-\Phi(|z|)]\), where \(\Phi\) is the standard normal cumulative distribution function. A corresponding approximate 95% confidence interval is
These are large-sample approximations, not universal exact results. They rely on a correctly specified model, suitable estimation, and regularity conditions that make the estimator approximately normal. They can be unreliable in a short series, under substantial model misspecification, or when a parameter is near a stationarity or invertibility boundary. A standard error is conditional on the fitted model and estimation procedure; it does not account automatically for searching across many models.
Identify the fitted ARIMA model, the coefficient being tested, the estimation method, and whether the model’s assumptions make a normal approximation reasonable.
Calculate the estimate divided by its standard error, and inspect a confidence interval as well as the p-value.
Interpret the test for this coefficient in this model. Do not treat it as a test of every possible model or of forecast quality.
Why the Wald Approximation Works
The justification starts with the approximate sampling distribution of an estimator. Under the regularity conditions above, an estimator is approximately normal around its true value, with standard deviation estimated by its standard error. If the null value is zero, subtracting zero and dividing by that standard deviation gives an approximately standard normal quantity. That is the reason the ratio is compared with standard normal probabilities.
The same approximation explains the confidence interval: about 95% of standard normal values lie between \(-1.96\) and \(1.96\), so the interval formed by adding and subtracting 1.96 estimated standard errors is an approximate range of plausible values. “Approximate” matters: the interval’s coverage can be poor when the model or large-sample conditions are inadequate.
Worked Example: An Uncertain AR Coefficient
A hypothetical insurer examines a stationary, centered monthly series of changes in a claim-frequency index using the AR(1) model \(Y_t=\phi Y_{t-1}+\epsilon_t\). Assume the innovations are independent \(N(0,\sigma^2)\), with \(\sigma^2=0.64\), and the fitted model satisfies the conditions for approximate Wald inference. The estimate is \(\hat{\phi}=0.42\), with standard error \(0.18\). Test \(H_0:\phi=0\) at the 5% two-sided level and form a 95% interval.
Method. Use the Wald statistic, the standard normal two-sided p-value, and the approximate confidence interval. The model is fixed for this calculation; the test does not assess all possible ARIMA orders.
Test statistic. \(z=0.42/0.18=2.3333\). The two-sided standard normal p-value is approximately \(0.0196\). Since \(0.0196<0.05\), reject \(H_0\) at the 5% level under the stated approximation.
Confidence interval. The margin is \(1.96(0.18)=0.3528\). Thus the interval is \(0.42\mathbin{\pm}0.3528\), or \([0.0672,0.7728]\).
Check and interpretation. The interval excludes zero, consistent with the two-sided test. The result is evidence of a positive AR coefficient in this fitted model, but the interval is broad: it does not pin down the coefficient precisely. A common mistake is to interpret significance as proof that the chosen AR(1) model is adequate. The coefficient test does not replace residual checks or comparison of forecasting performance.
Near-Zero Estimates and Repeated Testing
A small estimate can be statistically uncertain, and that uncertainty should be reported rather than converted into a claim that the true coefficient is zero. The standard error matters: an estimate of \(0.10\) with standard error \(0.02\) is not equivalent evidence to an estimate of \(0.10\) with standard error \(0.20\). Nor does a p-value just above 0.05 mark a sharp scientific boundary. It reflects a chosen threshold applied to a continuous measure of evidence.
Repeated testing creates a second issue. If many null hypotheses are tested, the chance of at least one false rejection can exceed the level used for each individual test. For \(m\) independent tests, each with false-positive probability \(\alpha\), the probability of at least one false positive when all null hypotheses are true is \(1-(1-\alpha)^m\). ARIMA coefficient tests are generally not independent, so this expression is an illustration of the risk, not an exact calculation for every fitted model.
Worked Example: The Risk of Testing Many Terms
Suppose a candidate search for a hypothetical monthly operational index produces six coefficient tests. For this illustration only, assume all six null hypotheses are true, each test has level \(0.05\), and the test results are independent. Find the probability of at least one false rejection.
Method. Find the probability that none of the six tests falsely rejects, then subtract it from one.
For one test, the probability of no false rejection is \(1-0.05=0.95\). Under the stated independence assumption, the probability of no false rejections in all six tests is \(0.95^6=0.7350919\). Therefore, the probability of at least one false rejection is \(1-0.7350919=0.2649081\), or \(0.2649\) to four decimals.
Check and interpretation. The result is greater than \(0.05\), even though each individual test uses a 5% level. Actual ARIMA coefficient tests need not be independent, and this calculation is not an adjustment procedure. It demonstrates why repeatedly searching for individually significant terms makes a nominal single-test p-value an incomplete description of the overall selection process. The mistake to avoid is reporting each selected term’s p-value as if that term had been chosen before examining the data.
Repeated deletion can also make the selected coefficient estimates too optimistic in magnitude: terms tend to survive precisely when their estimates look unusually strong in the observed sample. After selection, ordinary standard errors and p-values generally do not account for that selection. In addition, refitting after deletion can change the remaining coefficient estimates and their standard errors because the terms in the model are estimated together. A coefficient’s test in one specification does not simply carry over unchanged to another.
Worked Example: A Near-Zero MA Estimate in Context
A hypothetical, centered monthly service-volume series is fitted with an ARMA(1,1) model, equivalent to an ARIMA model with no differencing. Assume independent Gaussian innovations with mean zero and estimated variance \(\sigma^2=0.49\); the fitted parameters are within the stationary and invertible region, and approximate Wald inference is considered reasonable. The fitted full model gives \(\hat{\phi}=0.48\) with standard error \(0.20\), and \(\hat{\theta}=-0.17\) with standard error \(0.14\), using the convention \(Y_t=\phi Y_{t-1}+\epsilon_t+\theta\epsilon_{t-1}\). A refitted AR(1) model gives \(\hat{\phi}=0.31\), standard error \(0.12\). Residual Ljung–Box diagnostic p-values are \(0.27\) for the full model and \(0.03\) for the reduced model at the chosen diagnostic lag set.
Test the full-model coefficients. For \(\phi\), \(z=0.48/0.20=2.4000\), giving a two-sided approximate p-value of \(0.0164\). For \(\theta\), \(z=-0.17/0.14=-1.2143\), giving a two-sided approximate p-value of \(0.2248\). The 95% interval for \(\theta\) is \(-0.17\mathbin{\pm}1.96(0.14)=-0.17\mathbin{\pm}0.2744\), or \([-0.4444,0.1044]\).
Compare the refit and diagnostics. In the reduced AR(1) model, \(z=0.31/0.12=2.5833\), with approximate two-sided p-value \(0.0098\). The smaller p-value for its AR coefficient does not establish that the reduced model is better: the coefficient estimate changed after refitting, and the reduced model’s diagnostic p-value of \(0.03\) indicates evidence against residual white noise at the stated diagnostic settings. The full model’s diagnostic p-value is \(0.27\), which does not provide that evidence.
Conclusion and trap. The MA estimate is uncertain, but the coefficient test alone does not settle whether to remove it. In this hypothetical comparison, deleting it is not an automatic improvement, because the reduced fit has a concerning residual diagnostic. A common mistake is to delete \(\theta\) solely because \(0.2248>0.05\), then choose the reduced model because its remaining AR coefficient has a smaller p-value. The appropriate decision considers the candidate model’s purpose and diagnostics as well as coefficient uncertainty.
A Disciplined Use of Coefficient Tests
Coefficient tests are most useful as part of a structured assessment, not as a mechanical search algorithm. Start from a defensible candidate model and understand what each term represents. Then interpret estimates and uncertainty, check whether residual behavior is acceptable, and assess the model using selection and forecast-evaluation methods appropriate to the task. A coefficient that is not individually significant may still be retained for a reason, but retaining it without justification can also add unnecessary complexity.
| Observation | What it supports | What it does not establish |
|---|---|---|
| A coefficient’s p-value is below the chosen level | Evidence against the coefficient’s null value in the fitted model, under test assumptions | That the full model is adequate or forecasts well |
| A coefficient’s p-value is above the chosen level | Insufficient evidence against the null at that level | That the coefficient is zero or its term must be deleted |
| Many terms were tested and removed | That the final model was selected using the observed data | That ordinary post-selection p-values retain their original interpretation |
The central principle is to distinguish uncertainty about a coefficient from the decision to select a model. The first is what a coefficient test addresses. The second requires a broader comparison, including whether the model captures the series’ dependence and serves the forecasting or explanatory purpose. This approach is aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus.
Check Your Understanding
For each problem, use the stated normal approximation where appropriate and distinguish coefficient evidence from a model-selection conclusion.
- A fitted AR(1) model for a hypothetical centered monthly index has Gaussian white-noise innovations and gives \(\hat{\phi}=0.30\) with standard error \(0.12\). Find the Wald statistic, approximate two-sided p-value, and 95% confidence interval. (Answer: \(z=2.5000\), \(p=0.0124\), interval \([0.0648,0.5352]\).)
- In a fitted ARMA model with Gaussian innovations and valid Wald conditions, an MA coefficient estimate is \(-0.05\) with standard error \(0.10\). Calculate the approximate two-sided p-value and 95% interval, and state whether zero is excluded. (Answer: \(z=-0.5000\), \(p=0.6171\), interval \([-0.2460,0.1460]\); zero is not excluded.)
- Under the simplifying assumption of five independent tests, each at level \(0.05\), calculate the probability of at least one false rejection when all five null hypotheses are true. (Answer: \(1-0.95^5=0.2262\).)
- A fitted ARIMA term has estimate \(0.14\) and standard error \(0.11\), while removing it leads to residual autocorrelation evidence under the chosen diagnostic. Why is deletion based only on the coefficient p-value inadequate? (Answer: the approximate test gives \(z=1.2727\) and \(p\approx0.2031\), but that is not proof the coefficient is zero; model adequacy and the consequences of deletion must also be assessed.)
- After trying several candidate ARIMA specifications, a coefficient has a nominal p-value of \(0.04\) in the selected model. What qualification belongs with that result? (Answer: it is a nominal test conditional on the fitted specification; the repeated search is not accounted for, so it should not be interpreted as though the model had been fixed in advance.)