The Standard Error
The text below is a transcript of the video above, edited with AI assistance for clarity and flow. Where the two differ, the video is the authoritative version.
You’ll learn to
This video is about a statistic that is very important for understanding confidence intervals, p-values and significance: the standard error. You will learn to:
- Explain what the standard error of a slope measures.
- Connect it to the idea of repeated sampling.
- Distinguish the standard error from the standard deviation.
Starting point: a regression from a sample
Let’s start with a regression that we can run with data from the British Election Study. The dependent variable is ideology: left-right preferences on a 0 to 10 scale, where 0 means left and 10 means right. The independent variable is age.
When we run that regression, we get a slope of 0.017 per year of age. Being one year older is associated with moving to the right by this tiny margin. So older people lean slightly more to the right than younger people.
This estimate comes from one particular sample: the British Election Study sample of 2,649 respondents.
This estimate depends on our sample
The estimate depends on our sample. If the British Election Study had surveyed different people altogether, the slope would be slightly different. The estimate of 0.017 is one of many possible outcomes. The question, then, is how much the slope would vary across those different samples.
The thought experiment: many samples
Let’s run a thought experiment. Imagine that the people who run the British Election Study were to draw many samples, and that we ran the regression of ideology on age in each of them.
Imagine drawing 10 new surveys of the same size. Each time you get a different slope: 0.019, 0.015, 0.021, 0.014, 0.018, and so on. The estimate would necessarily vary from sample to sample. If the British Election Study interviewed people again and again, and we ran the regression of these two variables time after time, each time we would get a different slope.
The sampling distribution of the slope
Just as we can imagine the distribution of a variable, we can now imagine the distribution of the slopes. If we did this not 10 times but thousands of times, all of those slopes would form a distribution of regression slopes. This is the sampling distribution of the slope, and it is centred on the true population slope.
The standard error is the standard deviation of this distribution
The standard error is the standard deviation of this distribution. From all of these imagined, hypothetical steps you have a distribution of slopes, and if you apply to those slopes the same formula you learned for the standard deviation, you get the standard error.
The standard deviation measures spread in observed data. The standard error, by contrast, measures spread in estimated slopes across samples. It is the same concept, but the target is different: in the standard deviation case we are looking at data, and in the standard error case we are looking at estimates.
Why this matters for inference
This matters a lot for inference: for confidence intervals, for p-values, and for whether or not our estimates are statistically significant.
A small standard error means that the slope is precisely estimated: the slope is similar across samples. In the British Election Study example, we would be getting similar slopes each time we drew a sample and ran the regression of ideology on age.
A large standard error, by contrast, means that the slope is imprecisely estimated: it varies a lot. It could be 0.017, but it could be 0.92, or it could be −0.8. That is what a large standard error would mean: each time we take a sample and run the regression, we get a completely different slope, so the precision of the slopes we are getting is low.
The standard error is what makes p-values and confidence intervals possible. It is the key statistic that allows us to calculate the p-values and the confidence intervals of these estimates.
Reading the standard error in regression output
| Coefficient | Std. err. | |
|---|---|---|
| Age | 0.017 | 0.002 |
| Intercept | 3.84 | 0.114 |
Here we have a coefficient of 0.017 and a standard error of 0.002. It is a small standard error, so we can be confident about the precision of the coefficient of age, the slope we get when we regress ideology, as the dependent variable, on age, as the independent variable.
Each coefficient has its own standard error. When we divide the coefficient by its standard error, we get the t-statistic.
The t-statistic: how many standard errors from zero?
The t-statistic is the slope divided by its standard error:
\[t = b / SE\]
It tells us how many standard errors away from zero we are: how far the estimate is from zero, measured in standard errors. The t-statistic is what determines significance. If we get a large absolute value of \(t\), roughly above 2, we are in the terrain where p-values are smaller than 0.05 and 95% confidence intervals do not include zero. This is directly related to the ratio of the estimate to its standard error.
Recap
- The standard error measures how much a slope would vary across repeated samples.
- A smaller standard error means a more precise estimate.
Why this matters
Every regression estimate comes from a single sample, and the standard error quantifies the uncertainty in that estimate. Understanding the standard error is very important for understanding where significance and confidence intervals come from: it all depends on how much certainty we have about the slope, or the estimate, that we are calculating.