If you're seeing this message, it means we're having trouble loading external resources on our website.

If you're behind a web filter, please make sure that the domains *.kastatic.org and *.kasandbox.org are unblocked.

Main content

Sampling distribution of the sample mean

Take a sample from a population, calculate the mean of that sample, put everything back, and do it over and over. No matter what the population looks like, those sample means will be roughly normally distributed given a reasonably large sample size (at least 30). This is the main idea of the Central Limit Theorem — the sampling distribution of the sample mean is approximately normal for "large" samples. Created by Sal Khan.

Want to join the conversation?

  • blobby green style avatar for user jasen port
    If we know the mean and the standard deviation of the population, then why are we taking samples, if we already have the data?

    Thanks in advance.
    (27 votes)
    Default Khan Academy avatar avatar for user
    • hopper cool style avatar for user Mr. Jones
      Learning statistics can be a little strange. It almost seems like you're trying to lift yourself up by your own bootstraps. Basically, you learn about populations working under the assumption that you know the mean/stdev, which is silly, as you say, but later you begin to drop these assumptions and learn to make inferences about populations based on your samples.

      Once you have some version of the Central Limit Theorem, you can start answering some interesting questions, but it takes a lot of study just to get there!
      (42 votes)
  • blobby green style avatar for user jacob.930321
    Is there any difference if I take 1 "sample" with 100 "instances", or I take 100 "samples" with 1 "instance"?
    (By sample I mean the S_1 and S_2 and so on. With instances I mean the numbers, [1,1,3,6] and [3,4,3,1] and so on.)
    (11 votes)
    Default Khan Academy avatar avatar for user
    • aqualine ultimate style avatar for user Wes Wade
      There is a difference. Your "samples" (random selections of values "x") that are made up of "instances" (referred to as the variable "n") provide what will essentially be the building blocks of your Sampling Distribution of the Sample Mean. Because your "instances" determine the value of the mean of "x", your size of "n" determines the value of "x"'s mean, and the Sampling Distribution of the Sample Mean's standard deviation (Defined as The original dataset's standard deviation divided by the square root of "n").
      For example: If you were to take 1 "sample" with 100 "instances", you would get only one piece of data regarding the mean of 100 items [1,1,3,6,3,6,3,1,1,1,1,1...] from your original data. Your sampling distribution of the Sample mean's standard deviation would have a value of ((The original sample's S.D.)/(The square root of 100)), but that wouldn't really matter, because your data will likely be very close to your original data's mean, and you'd only have one sample.
      Now if you take 100 samples with 1 instance [3], you'll get many pieces of data, but no change in standard deviation from your first sample: ((The original sample's S.D.)/(The square root of 1)). Functionally, with enough samples taken like this, you'll re-create your original dataset! You won't be creating a useful sampling distribution of the sample mean because "x" will equal the mean of "x". With 100 "samples" of 1 "instance", you're randomly picking 100 values of "x" and re-plotting them.
      I hope that helps.
      (9 votes)
  • mr pink red style avatar for user Gigglethorpe
    So if every distribution approaches normal when do I employ say a Poisson or uniform or a Bernoulli distribution? I suppose it's a concept I haven't breached yet but how do I know when or which distribution to employ so I appropriately analyze the data? End goal = solve real world problems!
    (1 vote)
    Default Khan Academy avatar avatar for user
    • leaf blue style avatar for user Dr C
      Not every distribution goes to the Normal. the distribution of the sample mean does, but that's as the sample size increases. If you have smaller sample sizes, assuming normality either on the data or the sample mean may be wholly inappropriate.

      In terms of identifying the distribution, sometimes it's a matter of considering the nature of the data (e.g. we might think "Poisson" if the data collected are a rate, number of events per some unit/interval), sometimes it's a matter of doing some exploratory data analysis (histograms, boxplots, some numerical summaries, and the like).

      For actually analyzing data: I would suggest hiring someone with more extensive training in Statistics to actually do such. Taking one course in Stats, which is basically what KhanAcademy goes through, isn't really enough to prepare someone to be a data analyst. I see the primary goal of taking one or two stats courses as giving you enough information to allow you to understand the results of statistical analyses. You can better tell the statistician what you want in his/her own terms, and you can better understand what s/he gives back to you.
      (11 votes)
  • aqualine tree style avatar for user Emily Kusel
    Do your sample sizes have to be the same size? E.G, at (ish) there are a bunch of samples with a sample size of four. Would it mess up any calculations if you took a sample of four and then, say, a sample of ten?
    (2 votes)
    Default Khan Academy avatar avatar for user
    • leaf blue style avatar for user Dr C
      Yes, the sample sizes should be the same. The sample size is not considered to be a variable, it's considered to be a constant. The sampling distribution of the sample mean can be thought of as "For a sample of size n, the sample mean will behave according to this distribution." Any random draw from that sampling distribution would be interpreted as the mean of a sample of n observations from the original population.
      (7 votes)
  • piceratops ultimate style avatar for user Parthik Patel
    What is the difference between "sample distribution" and "sampling distribution"?
    (2 votes)
    Default Khan Academy avatar avatar for user
    • aqualine ultimate style avatar for user JaniceHolz
      The sample distribution is what you get directly from taking a sample. You plot the value of each item in the sample to get the distribution of values across the single sample. When Sal took a sample in the previous video at and got S1 = {1, 1, 3, 6}, and graphed the values that were sampled, that was a sample distribution. The 2nd graph in the video above is a sample distribution because it shows the values that were sampled from the population in the top graph.

      The sampling distribution is what you get when you compare the results from several samples. You plot the mean of each sample (rather than the value of each thing sampled). In the previous video, Sal did that starting at , when he plotted the mean of each sample. The 3rd and 4th graphs above are sampling distributions because each shows a distribution of means from the many samples of a particular size.

      http://www.psychstat.missouristate.edu/introbook/SBK19.htm also has an explanation.
      (4 votes)
  • aqualine ultimate style avatar for user George Lodewijk Bertram
    Is it possible to determine the sample variance without the population variance? I have an assignment that requires me to show the sampling distribution of the mean with only a population proportion and sample size.
    (3 votes)
    Default Khan Academy avatar avatar for user
    • blobby green style avatar for user robshowsides
      If a question talks about a "population proportion" then you are dealing with a binomial distribution, except that you divide by the sample size to get sample proportion rather than the sample count. If the population proportion is p, then the mean value of sample proportions will be also be p (as usual, the mean of the sampling distribution is just the same as for the whole population), and the variance will be p(1 - p)/n, where n is the size of the sample. You can read about this distribution here (note they use the letter pi for population proportion. It does NOT mean 3.14159...):
      (2 votes)
  • blobby green style avatar for user REPUBLICAMAGINIST
    why can we say that the sampling distribution of mean follows a normal distribution for a large enough sample size even though the population is may not be normally distributed?
    (2 votes)
    Default Khan Academy avatar avatar for user
    • aqualine tree style avatar for user Ted Fischer
      Properly, the sampling distribution APPROXIMATES a normal distribution for a sufficiently large sample (sometimes cited as n > 30). A coin flip is not normally distributed, it is either heads or tails. But 30 coin flips will give you a binomial distribution that looks reasonably normal (at least in the middle).
      (2 votes)
  • blobby green style avatar for user Smit Shah
    at , it has been said that even for single samples the central limit theorem is true. It is not so, central limit theorem is applicable only for sample MEANS. For example, out of a population of 5000 if I have taken the sample of n=50, central limit theorem does NOT apply to that. It applies only when I have taken (e.g.)40 samples of n=50. However, this is as per my understanding. Please correct me if I am wrong.
    (2 votes)
    Default Khan Academy avatar avatar for user
  • piceratops ultimate style avatar for user Jet Simon
    What is the difference between X-bar and mu? Like when do you know which to use what?
    (1 vote)
    Default Khan Academy avatar avatar for user
  • duskpin tree style avatar for user xinyuan lin
    How is the sampling distribution of a sample mean related to the sampling distribution of a sample proportion?
    (2 votes)
    Default Khan Academy avatar avatar for user

Video transcript

In the last video, we learned about what is quite possibly the most profound idea in statistics, and that's the central limit theorem. And the reason why it's so neat is, we could start with any distribution that has a well defined mean and variance-- actually, I wrote the standard deviation here in the last video, that should be the mean, and let's say it has some variance. I could write it like that, or I could write the standard deviation there. But as long as it has a well defined mean and standard deviation, I don't care what the distribution looks like. What I can do is take samples-- in the last video of say, size four-- that means I take literally four instances of this random variable, this is one example. I take their mean, and I consider this the sample mean from my first trial, or you could almost say for my first sample. I know it's very confusing, because you can consider that a sample, the set to be a sample, or you could consider each member of the set is a sample. So that can be a little bit confusing there. But I have this first sample mean, and then I keep doing that over and over. In my second sample, my sample size is four. I got four instances of this random variable, I average them, I have another sample mean. And the cool thing about the central limit theorem, is as I keep plotting the frequency distribution of my sample means, it starts to approach something that approximates the normal distribution. And it's going to do a better job of approximating that normal distribution as n gets larger. And just so we have a little terminology on our belt, this frequency distribution right here that I've plotted out, or here, or up here that I started plotting out, that is called-- and it's kind of confusing, because we use the word sample so much-- that is called the sampling distribution of the sample mean. And let's dissect this a little bit, just so that this long description of this distribution starts to make a little bit of sense. When we say it's the sampling distribution, that's telling us that it's being derived from-- it's a distribution of some statistic, which in this case happens to be the sample mean-- and we're deriving it from samples of an original distribution. So each of these. So this is my first sample, my sample size is four. I'm using the statistic, the mean. I actually could have done it with other things, I could have done the mode or the range or other statistics. But sampling distribution of the sample mean is the most common one. It's probably, in my mind, the best place to start learning about the central limit theorem, and even frankly, sampling distribution. So that's what it's called. And just as a little bit of background-- and I'll prove this to you experimentally, not mathematically, but I think the experimental is on some levels more satisfying with statistics-- that this will have the same mean as your original distribution. As your original distribution right here. So it has the same mean, but we'll see in the next video that this is actually going to start approximating a normal distribution, even though my original distribution that this is kind of generated from, is completely non-normal. So let's do that with this app right here. And just to give proper credit where credit is due, this is-- I think was developed at Rice University-- this is from onlinestatbook.com. This is their app, which I think is a really neat app, because it really helps you to visualize what a sampling distribution of the sample mean is. So I can literally create my own custom distribution here. So let me make something kind of crazy. So you could do this, in theory, with a discrete or a continuous probability density function. But what they have here, we could take on one of 32 values, and I'm just going to set the different probabilities of getting any of those 32 values. So clearly, this right here is not a normal distribution. It looks a little bit bimodal, but it doesn't have long tails. But what I want to do is, first just use a simulation to understand, or to better understand, what the sampling distribution is all about. So what I'm going to do is, I'm going to take-- we'll start with-- five at a time. So my sample size is going to be five. And so when I click animated, what it's going to do, is it's going to take five samples from this probability distribution function. It's going to take five samples, and you're going to see them when I click animated, it's going to average them and plot the average down here. And then I'm going to click it again, and it's going to do it again. So there you go, it got five samples from there, it averaged them, and it hit there. So what I just do? I clicked-- oh, I wanted to clear that. Let me make this bottom one none. So let me do that over again. So I'm going to take five at time. So I took five samples from up here, and then it took its mean and plotted the mean there. Let me do it again. Five samples from this probability distribution function, plotted it right there. I could keep doing it. It'll take some time. But you can see I plotted it right there. Now I could do this 1,000 times, it's going to take forever. Let's say I just wanted to do it 1,000 times. So this program, just to be clear, it's actually generating the random numbers. This isn't like a rigged program. It's actually going to generate the random numbers according to this probability distribution function. It's going to take five at a time, find their means, and plot the means. So if I click 10,000, it's going to do that 10,000 times. So it's going to take five numbers from here 10,000 times and find their means 10,000 times and then plot the 10,000 means here. So let's do that. So there you go. And notice it's already looking a lot like a normal distribution. And like I said, the original mean of my crazy distribution here was 14.45, and after doing 10,000 samples-- or 10,000 trials-- my mean here is 14.42. So I'm already getting pretty close to the mean there. My standard deviation, you might notice, is less than that. We'll talk about that in a future video. And the skew and kurtosis, these are things that help us measure how normal a distribution is. And I've talked a little bit about it in the past, and let me actually just diverge a little bit, it's interesting. And they're fairly straightforward concepts. Skew literally tells-- so if this is-- let me do it in a different color-- if this is a perfect normal distribution-- and clearly my drawing is very far from perfect-- if that's a perfect distribution, this would have a skew of zero. If you have a positive skew, that means you have a larger right tail than you would otherwise expect. So something with a positive skew might look like this. It would have a large tail to the right. So this would be a positive skew, which makes it a little less than ideal for normal distribution. And a negative skew would look like this, it has a long tail to the left. So negative skew might look like that. So that is a negative skew. If you have trouble remembering it, just remember which direction the tail is going. This tail is going towards a negative direction, this tail is going to the positive direction. So if something has no skew, that means that it's nice and symmetrical around its mean. Now kurtosis, which sounds like a very fancy word, is similarly not that fancy of an idea. So once again, if I were to draw a perfect normal distribution. Remember, there is no one normal distribution, you could have different means and different standard deviations. Let's say that's a perfect normal distribution. If I have positive kurtosis, what's going to happen is, I'm going to have fatter tails-- let me draw it a little nicer than that-- I'm going to have fatter tails, but I'm going to have a more pointy peak. I didn't have to draw it that pointy, let me draw it like this. I'm going to have fatter tails, and I'm going to have a more pointy peak than a normal distribution. So this right here is positive kurtosis. So something that has positive kurtosis-- depending on how positive it is-- it tells you it's a little bit more pointy than a real normal distribution. And negative kurtosis has smaller tails, but it's smoother near the middle. So it's like this. So something like this would have negative kurtosis. And maybe in future videos we'll explore that in more detail, but in the context of the simulation, it's just telling us how normal this distribution is. So when our sample size was n equal 5 and we did 10,000 trials, we got pretty close to a normal distribution. Let's do another 10,000 trials, just to see what happens. It looks even more like a normal distribution. Our mean is now the exact same number, but we still have a little bit of skew, and a little bit of kurtosis. Now let's see what happens if we do the same thing with a larger sample size. And we could actually do them simultaneously. So here's n equal 5. Let's do here, n equals 25. Just let me clear them. I'm going to do the sampling distribution of the sample mean. And I'm going to run 10,000 trials-- I'll do one animated trial, just so you remember what's going on. So I'm literally taking first five samples from up here, find their mean. Now I'm taking 25 samples from up here, find its mean, and then plotting it down here. So here the sample size is 25, here it's five. I'll do it one more time. I take five, get the mean, plot it. Take 25, get the mean, and then plot it down there. This is a larger sample size. Now that thing that I just did, I'm going to do 10,000 times. And remember, our first distribution was just this really crazy very non-normal distribution, but once we did it-- whoops, I didn't want to make it that big. Scroll up a little bit. So here, what's interesting? I mean they both look a little normal, but if you look at the skew and the kurtosis, when our sample size is larger, it's more normal. This has a lower skew than when our sample size was only five. And it has a less negative kurtosis than when our sample size was five. So this is a more normal distribution. And one thing that we're going to explore further in a future video, is not only is it more normal in its shape, but it's also tighter fit around the mean. And you can even think about why that kind of makes sense. When your sample size is larger, your odds of getting really far away from the mean is lower. Because it's very low likelihood, if you're taking 25 samples, or 100 samples, that you're just going to get a bunch of stuff way out here, or a bunch of stuff way out here. You're very likely to get a reasonable spread of things. So it makes sense that your mean-- your sample mean-- is less likely to be far away from the mean. We're going to talk a little bit more about in the future. But hopefully this kind of satisfies you that-- at least experimentally, I haven't proven it to you with mathematical rigor, which hopefully we'll do in the future. But hopefully this satisfies you, at least experimentally, that the central limit theorem really does apply to any distribution. I mean, this is a crazy distribution. And I encourage you to use this applet at onlinestatbook.com and experiment with other crazy distributions to believe it for yourself. But the interesting things are that we're approaching a normal distribution, but as my sample size got larger, it's a better fit for a normal distribution.