Brief Summary
This video provides a practical overview of Analysis of Variance (ANOVA), focusing on its purpose, how to use it, and how to interpret its results, while avoiding the complex mathematics often associated with it. Key points include:
- ANOVA tests if a quantitative variable is independent from a categorical variable, looking at group means.
- It compares variability between and within groups to assess differences.
- ANOVA requires three main assumptions and leverages technology for execution, showcasing results in an ANOVA table.
The video introduces Analysis of Variance (ANOVA) as a statistical method used to determine if there are significant differences between the means of three or more independent groups. It explains that ANOVA helps test the null hypothesis, which states that the means of a quantitative variable are equal across different levels of a categorical variable. The presenter emphasizes using practical examples rather than the underlying mathematics to make the concept more accessible.
ANOVA evaluates the variability of data both within groups and between groups. The presenter illustrates this by discussing two scenarios with sample data—one showing distinct group means with little overlap, and the other showing similar group means with significant overlap in data. The first scenario would provide strong evidence that group means differ, while the second offers weaker support due to higher variability within the groups.
The video outlines three essential assumptions of ANOVA: the independence of observations, normal distribution of response variables within each group, and equal variances across groups. It notes that ANOVA is robust against violations of normality and equal variances as long as sample sizes are reasonable. However, the independence assumption is critical; lack of independence renders ANOVA inappropriate for analysis.
An example is presented where researchers measure the lengths of flower petals across three groups. The data from the groups reveals variability, with overlapping distributions. The visual representation helps to illustrate how the differences in sample means should prompt the researchers to determine if the average petal lengths differ significantly in the broader population.
Today, ANOVA is typically conducted using statistical software, with R being a common choice. The video introduces the command "aov" used in R for performing ANOVA. The resulting output is presented in an ANOVA table, highlighting the importance of the p-value, which indicates the strength of evidence against the null hypothesis.
The presenter explains how to interpret p-values. A p-value of 0.335 suggests no strong evidence against the null hypothesis, indicating that the group means may be consistent across the sample. If the null hypothesis is rejected, it is essential to note that the result only indicates that at least one group mean is different; additional analyses are needed to identify which means differ.
Finally, the video discusses the necessity of post hoc testing following ANOVA to explore specific differences between group means. The Tukey's Honest Significant Difference (HSD) test is highlighted as a common method, which compares all pairs of groups while controlling for Type I error risk. The presenter mentions plans for a future video dedicated to explaining Tukey's HSD in more detail.
Summarize any YouTube video in seconds
Get summaries like this one for any video, in your language, with the free Summ app.
Transcript
Full transcript of “Understand ANOVA in 8 minutes”. Click a timestamp to open the video at that moment. Need another video? Use the YouTube transcript generator.
0:01 hey everybody today we're approaching analysis of variance or an NOA from a very practical point of view leaving the math aside almost entirely to focus on what an NOA is when to use it and how to interpret the results an Nova has a reputation of being a very difficult topic because the math underneath the hood really is difficult but when we leave that aside and just focus on how to use it it's really not that bad so a NOA is used to test whether a quantita variable is independent from a categorical variable a bit more mathematically we're testing the null hypothesis that the mean of the quantitative variable is the same in for every level of the categorical variable that is its mean is the same across groups for instance doctors might wonder
0:48 whether people of different age groups respond equally to a new blood pressure medicine or whether several different medications are all equally effective the basic idea of analysis of variance is compare the variability of the data within the groups to the variability between the groups and this all sounds very abstract until we get some pictures so here I'm imagining two different studies we have sample data in each case and in each case the sample means are the same within each group so in the first case on the left we have three groups with these different means there's very little overlap of the data and so this is going to be providing pretty decent evidence that in the larger population the group means are going to be different in the second case
1:35 on the other hand in the right plot we have three different groups where there is a lot of overlap between the data even though the group means are exactly the same as in the first case because there's so much more spread in the data there's much more overlap in these distributions the evidence that the group means are different in the larger population is actually much weaker the difference between the group means is more significant in the first case because of the lower variability within the groups more specifically Anova looks at the ratio of the between groups variance and the within groups variance both appropriately defined and weighted I'm going to flash the math here but not really talk about it I have a longer vid that goes into the math in more detail
2:22 I'll throw a link up top if the null hypothesis is true and the quantitative variable is the same on average between the different groups in the larger population then this thing that we're calculating this ratio that we're calculating has an F distribution and we can use that to compute A P value and make a decision analysis of variance has three basic assumptions first of all all the observations have to be independent of one another the response variable must be normally distributed within each group in the larger population and the variances within the groups all have to be equal in the larger population now of these three an NOA is pretty robust
3:07 against the second two assumptions as long as your sample size is reasonable particularly normality and this is all because the central limit theorem ends up kicking in so um you don't have to worry about those two unless there's extreme um uh departure from equal variances and from normal distributions within each group The first Assumption of Ascend of Independence is more important this is usually having more to do with the design of the experiment and has and um if you don't have independent observations an NOA is probably just not the appropriate method of statistical inference to use um a quick example to have in mind before we actually get the results of an anova and talk about how to interpret those and use those researchers are
3:52 measuring the lengths in millimeters of a certain species of flower of Petals of those flowers so here's some samples data we've got uh five six and seven observations so very small sample sizes for each of these three um different groups if we plot that for the three different groups just doing a very simple Dot Plot for each of these groups um we can see both the group means here I've labeled them y1 bar Y2 bar and Y3 bar we can see that there are differences between them apparently and so we might think this would give us some evidence that the pedal L the average pedal length for these three different groups are going to be be different in the larger population we can also see that there's a lot of variability Within These groups and in particular these three distributions all overlap a fair amount the variability
4:40 within the groups is fairly large compared to the variability between the groups so when we run in an Nova we're going to try and compare those variances and take um a ratio of the overall variability between groups to the overall variability within groups now an NOA is always performed using technology these days you almost certainly will not have to deal with the math unless you are in a specifically mathematically oriented statistics class I use R that's the typical um one very typical way of doing it in R the command Is AO and uh whether you're using r or any other sort of Technology the results will be given to you in an anova table like this and in that longer more mathy video I talk about what these things
5:26 mean a bit more specifically in practice the only one you're going to care about is going to be the P value at the end and this is representing the probability of data like what you got in your sample occurring by chance when the null hypothesis is true that is when in the larger population the means are the same in each of the different groups in this case we have a P value of 335 P values are always going to be between 0er and one lower P values indicate stronger evidence against the null hypothesis that is against the claim that the quantitative variable is independent from the categorical one and that the group means are the same for the quantitative variable across all the different groups P equals 335 is saying that there is not strong evidence for
6:13 that we're going to um say that the data is consistent with that null hypothesis and draw no further conclusions now even when the null hypothesis is rejected the conclusion of an analysis of variance is pretty weak the alternative hypothesis is just that at least one of the group means is different in one group from a group mean in at least one other group um so it doesn't say which groups are potentially different and it doesn't say how different they potentially are to answer those kinds of questions you need an after Thea or post Hawk test the most common and simplest one is the Tuki honest significant differences test which roughly speaking performs um a lot of different pairwise um um T tests so
7:01 for every pair of categories you get um a paired T Test between those two um because you have all of those repeated tests the probability of a type 1 error ends up being pretty high and the Tuki on a significant differences test accounts for that by making the cut off value for statistical significance much lower at some point I'll record record a whole video on the Tuki honest significance differences test when that happens I'll make sure that there's a link up top for that
