A quarterly review slide reads: “Average order value is ₹850, 95% confidence interval ₹820 to ₹880.” Most people in the room take away “about ₹850” and move on. The interval carries more information than the headline number, and reading it properly changes what you can and cannot claim.
Where the interval comes from
The ₹850 is a sample statistic. Perhaps it comes from 1,537 orders sampled from the quarter, or from a survey of customers, or from one month treated as representative of the year. A different sample would have given a slightly different average, maybe ₹838 or ₹861. The confidence interval puts a range around the estimate that reflects this sampling wobble.
For an average, the building block is the standard error: the standard deviation of the individual values divided by the square root of the sample size. Suppose individual order values have a standard deviation of ₹600. With 1,537 orders, the standard error is 600 divided by about 39.2, roughly ₹15.3. A 95% interval for a mean, with a reasonably large sample, is the estimate plus or minus about 1.96 standard errors. That is ₹850 plus or minus ₹30, which gives ₹820 to ₹880.
The 1.96 comes from the normal distribution: 95% of a normal distribution lies within 1.96 standard deviations of its centre. The Central Limit Theorem is why this works for sample averages even when individual order values are skewed, as they almost always are, provided the sample is reasonably large.
What “95% confidence” actually means
The careful definition is about the method, not about this one interval. If you repeated the whole process many times, drawing a fresh sample each time and building an interval the same way, about 95% of those intervals would contain the true average order value. Your interval is one draw from that process. You do not know whether it is one of the 95% that caught the true value or one of the 5% that missed.
In everyday business language, saying “we are 95% confident the true average is between ₹820 and ₹880” is an acceptable shorthand, and most statisticians will let it pass in a meeting. What is not acceptable are the following readings.
- “95% of orders are between ₹820 and ₹880.” This is the most common misreading. The interval is about the average, not individual orders. With a standard deviation of ₹600, individual orders range from under ₹100 to several thousand rupees. A range that covers most individual orders is a different concept, sometimes called a prediction or reference interval, and it is far wider.
- “Every value inside the interval is equally likely.” Values near the centre are better supported by the data than values near the edges.
- “The true value cannot be outside the range.” It can; that is what the other 5% is.
Width is the useful part
The width of the interval tells you how precise the estimate is, and it is governed by three things.
Sample size. Because the standard error divides by the square root of the sample size, precision improves slowly. To halve the width from plus or minus ₹30 to plus or minus ₹15, you need four times as many orders, about 6,150 instead of 1,537. This is worth knowing before anyone asks for “a tighter number by next week”.
Variability. If order values were more consistent, say a standard deviation of ₹300, the same sample would give an interval half as wide. Segmenting can help here: an interval for “average order value of grocery orders” may be much narrower than one for all orders, because a single category is less variable.
Confidence level. A 99% interval is wider than a 95% interval from the same data, because it must catch the true value more often. A 90% interval is narrower. Higher confidence costs precision.
Using intervals to make decisions
Intervals are most useful when they are compared with a number that matters to the business. Suppose free delivery is offered above ₹800, and the team wants to know if the typical customer is comfortably above that line. An interval of ₹820 to ₹880 says yes with reasonable confidence. If the interval had been ₹780 to ₹920, the honest answer would be “we cannot tell yet”, even though the point estimate is the same ₹850.
For comparisons, such as average order value in an A/B test, build an interval for the difference between the two groups rather than eyeballing two separate intervals. If the interval for the difference excludes zero, the difference is statistically significant at the matching level, which links confidence intervals directly to hypothesis tests. Two separate 95% intervals can overlap somewhat even when the difference between the groups is significant, so overlap is not a reliable test on its own.
The interval for a difference also shows practical importance in a way a p-value cannot. A difference of ₹4 with an interval of ₹1 to ₹7 may be statistically significant and commercially irrelevant. A difference of ₹60 with an interval of minus ₹5 to ₹125 is not significant, yet it hints that a larger test could be worthwhile.
What an interval cannot protect you from
A confidence interval only accounts for random sampling error. It assumes the sample was drawn fairly from the population you care about. If the 1,537 orders all came from the app and none from the website, or all from Diwali week, the interval can be tight and still wrong, because the bias shifts the whole interval rather than widening it. Missing data, duplicate records and measurement errors are likewise invisible to the formula.
So read an interval as “how much this estimate could move from sampling luck alone”. Then ask, separately, whether the sample was the right one.
Written by the RP Strativa Team. Figures in the examples are illustrative.
Leave a Reply