Reading the Sign of the Interval for a Difference

Read the Endpoints to Find the Direction

In “Interpreting a Confidence Interval for \(\mu_1-\mu_2\),” you learned to describe the population quantity and keep the subtraction order explicit. Now use the interval’s sign to say what directions of difference are plausible. The key question is whether every value in the interval is above zero, every value is below zero, or zero is included.

Let \(\mu_1-\mu_2\) represent the true mean for population 1 minus the true mean for population 2. A positive value means the population 1 mean is higher. A negative value means the population 1 mean is lower. A value of zero means the population means are equal. These statements concern population means, not individual observations.

Definition: The sign of a confidence interval for \(\mu_1-\mu_2\) describes the direction of the plausible population mean differences. If both endpoints are positive, the interval is entirely positive; if both are negative, it is entirely negative. If zero is between the endpoints or is an endpoint, zero is included in the interval.

For an interval entirely above zero, every plausible value of \(\mu_1-\mu_2\) is positive, so the interval supports the conclusion that the population 1 mean exceeds the population 2 mean. For an interval entirely below zero, every plausible value is negative, so it supports the conclusion that the population 1 mean is less than the population 2 mean. Keep the order in the parameter: the direction is about population 1 compared with population 2.

\begin{aligned} \text{lower endpoint} > 0 &\quad\Longrightarrow\quad \mu_1-\mu_2 \text{ is positive throughout the interval}\\ \text{upper endpoint} < 0 &\quad\Longrightarrow\quad \mu_1-\mu_2 \text{ is negative throughout the interval}\\ \text{zero is inside or at an endpoint} &\quad\Longrightarrow\quad \text{zero is a value included in the interval} \end{aligned}

The endpoints also describe the plausible size of the difference, in the original variable’s units. For example, an interval from 2 to 5 minutes says more than just that the difference is positive: it gives a range of plausible mean differences from 2 to 5 minutes. It does not say that each person in population 1 has a value 2 to 5 minutes higher than each person in population 2.

When Zero Is Included

If zero lies strictly between the endpoints, the interval includes negative values, zero, and positive values. Thus it contains plausible differences in both directions as well as the no-difference value. The interval does not establish which population mean is larger, and it does not prove that the population means are equal. Instead, the interval leaves zero as one plausible value for the true difference.

There is a useful connection to the two-sided test of \(H_0:\mu_1-\mu_2=0\). As in “Using a t Interval to Decide a Two-Sided Test of a Mean,” the interval and test give corresponding decisions when they use the same data, compatible assumptions, and matching confidence and significance levels. For example, a 95% interval corresponds to a two-sided test at \(\alpha=0.05\). If zero is strictly inside the interval, the matching test fails to reject the null hypothesis. That means the data do not provide convincing evidence of a difference at that test level; it does not mean the means have been shown to be equal.

Pay special attention when zero is exactly an endpoint. With the usual closed-interval convention, an endpoint belongs to the interval. But the test decision is at the significance boundary: the matching two-sided test has a p-value equal to \(\alpha\). Following the AP decision rule, reject \(H_0\) when the p-value is at most \(\alpha\), so a zero endpoint corresponds to rejecting at that boundary—not to failing to reject. If a displayed endpoint has been rounded to zero, its unrounded value might be slightly positive or negative; do not assume an exact boundary from a rounded display.

Boundary reminder: For a matching two-sided test, zero strictly inside the confidence interval corresponds to failing to reject \(H_0:\mu_1-\mu_2=0\). If zero is exactly an endpoint, the test is at the boundary, with \(p=\alpha\); under the \(p\leq\alpha\) rule, reject \(H_0\). Always account for rounding before treating a displayed zero as exact.

This connection is about statistical evidence under the procedure’s assumptions. It does not by itself tell you whether a difference matters in practice. A small difference can be statistically distinguishable from zero yet unimportant for a real decision. And whether a study supports a cause-and-effect claim depends on its design, not simply on the sign of the interval.

Use a Consistent Interpretation Process

Before writing a conclusion, identify the population 1 mean minus the population 2 mean, check the endpoint signs, and then translate the result into the context. The confidence interval interpretation still needs the confidence level, both populations, the variable, the subtraction order, the endpoints, and the units, as explained in “Interpreting a Confidence Interval for \(\mu_1-\mu_2\).”

Confirm the order.
Read \(\mu_1-\mu_2\) as the true mean for population 1 minus the true mean for population 2.

Check the signs.
If the lower endpoint is above zero, the interval is entirely positive. If the upper endpoint is below zero, it is entirely negative. Otherwise, check whether zero is inside or at an endpoint.

Translate the direction.
A positive difference means population 1’s mean is higher; a negative difference means it is lower. If zero is strictly inside, the interval includes plausible differences in either direction.

Write the contextual interpretation.
State the confidence level, the true population mean difference in the correct order, the endpoints, and the units. Avoid claiming equality or causation from the interval alone.

Worked Examples

Worked Example: Positive Difference in Practice-Test Time

An invented school project compares the time students spend completing a practice test under two formats. Population 1 is students using the guided format; population 2 is students using the standard format. The variable is completion time in minutes. Suppose the reported 95% confidence interval for \(\mu_1-\mu_2\) is \((2.4,\ 6.8)\) minutes.

Both endpoints are greater than zero, so the interval is entirely positive. Every value in the interval describes a longer mean completion time for students using the guided format than for students using the standard format. In context, the interval says the guided-format population mean exceeds the standard-format population mean by a plausible amount from 2.4 to 6.8 minutes.

A complete interpretation is: We are 95% confident that the true mean practice-test completion time for students using the guided format minus the true mean completion time for students using the standard format is between 2.4 and 6.8 minutes.

Because zero is not in this interval, the matching two-sided test of no difference at \(\alpha=0.05\) would reject \(H_0\), assuming the interval and test use the same data and compatible conditions. This supports evidence of a difference in mean completion time in the stated direction. It does not imply that every guided-format student takes longer, nor does the interval alone show that the format caused a difference.

Worked Example: Negative Difference in Weekly Energy Use

An invented community project compares weekly household energy use for homes with two insulation designs. Population 1 is homes with design Pine; population 2 is homes with design Birch. The quantitative variable is weekly energy use, measured in kilowatt-hours. Suppose a 90% confidence interval for \(\mu_1-\mu_2\) is \((-18,\ -5)\) kilowatt-hours.

Both endpoints are negative, so the interval is entirely below zero. A negative value of \(\mu_1-\mu_2\) means the mean for Pine homes is less than the mean for Birch homes. The plausible differences range from 18 kilowatt-hours lower to 5 kilowatt-hours lower for Pine homes, per week.

A suitable interpretation is: We are 90% confident that the true mean weekly energy use for homes with design Pine minus the true mean weekly energy use for homes with design Birch is between \(-18\) and \(-5\) kilowatt-hours.

The negative signs matter: they show that the first mean is below the second. It would reverse the meaning to say Pine homes use 5 to 18 kilowatt-hours more per week. The interval excludes zero, so the corresponding two-sided test at \(\alpha=0.10\) would reject the no-difference null hypothesis, given matching procedures and assumptions. Whether the difference is practically important requires context beyond the sign alone.

Worked Example: Zero Inside a Difference in Reaction Time

An invented sports-science project compares reaction time under two warm-up routines. Population 1 is athletes using routine North; population 2 is athletes using routine South. The variable is reaction time in seconds. Suppose the reported 95% confidence interval for \(\mu_1-\mu_2\) is \((-0.08,\ 0.03)\) seconds.

The lower endpoint is negative and the upper endpoint is positive, so zero lies strictly inside the interval. The interval includes plausible values in which North’s mean reaction time is lower, values in which the means are equal, and values in which North’s mean is higher. It does not establish a direction for the difference.

A complete interpretation is: We are 95% confident that the true mean reaction time for athletes using routine North minus the true mean reaction time for athletes using routine South is between \(-0.08\) and \(0.03\) seconds.

Because zero is strictly inside the interval, the corresponding two-sided test at \(\alpha=0.05\) would fail to reject \(H_0:\mu_1-\mu_2=0\). In context, the data do not provide convincing evidence of a difference in the population mean reaction times at that level. Do not say that the routines have been proved to produce equal mean reaction times; values on both sides of zero remain plausible.

Worked Example: Zero as an Exact Endpoint

Suppose an invented comparison of two garden watering schedules reports an exact 95% confidence interval of \([0,\ 1.6]\) liters for mean water used per plant per day, schedule Cedar minus schedule Elm. Treat the lower endpoint as exactly zero, not as a rounded display.

Zero is an endpoint and is therefore included in the interval. The interval does not contain negative differences, but it includes the no-difference value and positive differences. Its endpoints alone do not say that the difference is strictly positive throughout the interval: zero is one of the included values.

The confidence interpretation is: We are 95% confident that the true mean daily water use per plant for schedule Cedar minus the true mean daily water use per plant for schedule Elm is between 0 and 1.6 liters.

For the matching two-sided test at \(\alpha=0.05\), this exact endpoint is the boundary case: the p-value is \(\alpha\). Since the decision rule is to reject when \(p\leq\alpha\), the test rejects \(H_0\) at the boundary. This differs from the earlier example where zero was strictly inside the interval and the test failed to reject. If a calculator or report rounds an endpoint to 0, check the unrounded endpoint before making this boundary claim.

Common Mistakes and AP Exam Tips

Reversing the direction: For \(\mu_1-\mu_2\), a positive interval means the population 1 mean is higher, and a negative interval means it is lower. Keep population 1 first in your interpretation.
Calling an interval with zero inside evidence of equality: If zero is strictly inside, say the data do not provide convincing evidence of a difference at the matching test level. Do not say the means are equal or that the null hypothesis has been proved.
Ignoring an endpoint equal to zero: An exact zero endpoint is a boundary case, not the same as zero strictly inside. For the matching test, \(p=\alpha\), and the \(p\leq\alpha\) rule says to reject. If the endpoint is rounded, avoid treating it as exact without more information.
Confusing a mean difference with individual values: The interval estimates a difference between population means. It does not describe how every observation in one group compares with every observation in the other.
Leaving out the context or units: A sign can show direction, but a full-credit interval interpretation still identifies the populations, variable, order, confidence level, endpoints, and units.
Claiming practical importance or causation from the sign: The interval describes plausible population mean differences. Practical importance depends on the setting, and cause-and-effect conclusions depend on study design.

For full credit, state what the interval estimates in context, interpret the direction using the signs of both endpoints, and be precise about zero. If you connect the interval to a test, name the matching two-sided test and significance level. Distinguish zero strictly inside the interval from zero exactly at an endpoint, and do not overstate what either result proves.

Key takeaway: An interval entirely above zero supports a positive \(\mu_1-\mu_2\); an interval entirely below zero supports a negative \(\mu_1-\mu_2\). If zero is strictly inside, the matching test fails to reject no difference. If zero is exactly an endpoint, the test is at the boundary and rejects under the \(p\leq\alpha\) rule. Neither outcome proves that the population means are equal.

Check Your Understanding

For each question, interpret the sign in the stated population order and distinguish an interior zero from an endpoint zero where relevant.

A 95% interval for mean delivery time in region East minus mean delivery time in region West is \((1.2,\ 4.7)\) hours. Which region has the higher mean according to the interval?
A 90% interval for mean monthly library visits, group A minus group B, is \((-2.3,\ -0.4)\) visits. What direction of difference does the interval support?
A matching 95% interval contains zero strictly between its endpoints. What is the decision for the corresponding two-sided test at \(\alpha=0.05\), and what should you say in context about evidence?
A matching 95% interval has an exact endpoint equal to zero. Under the \(p\leq\alpha\) rule, what is the decision for the corresponding test at \(\alpha=0.05\)?
Why does an interval that contains zero not prove that the two population means are equal?