Latest / This Is Why with Dr. Busti / Parametric vs Non-Parametric Statistical Tests Explained

Transcript

Parametric vs Non-Parametric Statistical Tests Explained · This Is Why with Dr. Busti

• 0:00
  SPEAKER_00: Alright, well, welcome to the biostatistics series where we're going to be engaged in this particular discussion about non-parametric and parametric statistical analysis at a high level.
• 0:11
  SPEAKER_00: This is part of a three-part series that's involving biostatistical tests, where part one is about what we do before the study starts, or as you're reading the paper, what should the researcher have done.
• 0:23
  SPEAKER_00: Part two is what happens when you finish the study and now you're analyzing the data, basically those summary statistics.
• 0:30
  SPEAKER_00: And then part three are the diagnostic studies, in particular, looking at precision, accuracy, those kind of points.
• 0:37
  SPEAKER_00: Within part one, we're under the statistical analysis plan just to keep us in the same framework where we covered the how to get oriented, the type of data that you're dealing with, nominal, ordinal, or continuous.
• 0:50
  SPEAKER_00: That helps to drive actually understanding really this picking the right statistical analysis, whether it's parametric or non-parametric.
• 0:59
  SPEAKER_00: So make sure you know those and you know how to pick the right study group.
• 1:03
  SPEAKER_00: Are they independent related?
• 1:05
  SPEAKER_00: Do you know how many there are?
• 1:06
  SPEAKER_00: I know that seems kind of straightforward, maybe you know, elementary, but people mess it up because they're reading the paper, and if they misinterpret the way the methods describe the study, then they get the question or they misinterpret the analysis potentially.
• 1:24
  SPEAKER_00: But we are right here in considering parametric versus non-parametric statistical analysis, and we're going to be using our decision table to guide us with that, to fill in those gaps.
• 1:34
  SPEAKER_00: The other components really are at the completion of the study, where most people are thinking about biostatistics.
• 1:40
  SPEAKER_00: That's our summary statistics, where we do descriptive statistics and our inferential statistics.
• 1:46
  SPEAKER_00: But our measures of central tendency and variability, those become very important to understanding whether our data from the samples that we are assessing are following a normal distribution or is skewed.
• 2:01
  SPEAKER_00: So if it's skewed in either direction right or left, then that typically falls inside of non-parametric statistical analysis.
• 2:10
  SPEAKER_00: If the data and the samples that we're looking at are normally distributed, mean, median, and mode are roughly the same, okay, then it becomes more parametric.
• 2:21
  SPEAKER_00: So that's why it's under this part isn't going to be important when you do finish the analysis to discern are you in that bucket?
• 2:29
  SPEAKER_00: And then within diagnostic tests, then you we will cover those in more details when we talk about precision and accuracy.
• 2:37
  SPEAKER_00: So I have two important tables, and those of you that are joining us on our podcast channel, you can get these at thisiswhy.health or free for download.
• 2:48
  SPEAKER_00: For those of you watching or listening, that's in the description of the episode.
• 2:53
  SPEAKER_00: The link to that direct page at this is why.health is there for you.
• 2:57
  SPEAKER_00: These are great study tools.
• 2:59
  SPEAKER_00: They're also good tools for helping you if you're doing a journal club and you're and you're critically appraising a paper and you're trying to figure out what should they have done and did that paper do that, right?
• 3:10
  SPEAKER_00: Or if you're doing the protocol development stage and you're codifying your own clinical research and what should you be doing.
• 3:18
  SPEAKER_00: That doesn't mean you might still not consult with the biostatistician, but you're at least getting into the proper framework and you're codifying for the IRB and your colleagues, and then when you go to present the results, you're already prepared to do that.
• 3:33
  SPEAKER_00: Makes the study more complete and more robust.
• 3:36
  SPEAKER_00: So we are in the area of thinking about parametric versus or non-parametric versus parametric, and really that's about the way that data is distributed from the sample that we got.
• 3:48
  SPEAKER_00: Remember, we're not studying the whole population.
• 3:49
  SPEAKER_00: So you take a sample out of the population, how is that that that sample's data interacting with other data points inside the sample and as it relates to what we think is the population mean.
• 4:03
  SPEAKER_00: Okay?
• 4:04
  SPEAKER_00: All that helps us to understand that.
• 4:06
  SPEAKER_00: And so in this table, we break down the data type of data, nominal, ordinal, continuous that no longer meets the criteria or assumptions of being parametrically distributed.
• 4:17
  SPEAKER_00: And then, of course, the bottom one at the down here you see is our parametric statistical analysis and the type of endpoint.
• 4:24
  SPEAKER_00: But we need to know that because it drives the row that we fall into.
• 4:29
  SPEAKER_00: So when we talk about parametric versus non-parametric statistical tests, we're over here on the far left of this standard statistical test decision table that I put together for you.
• 4:42
  SPEAKER_00: Okay, um, that tells you what row you're in.
• 4:46
  SPEAKER_00: Okay, and again, if continuous is a you know parametric typically considered, but remember if it fails to follow through the assumptions, you have to then treat it in a non-parametric way.
• 5:01
  SPEAKER_00: Okay, and so then you start using non-parametric statistical analysis.
• 5:05
  SPEAKER_00: That's why it's so important to make sure that you have these concepts wrapped around in your mind.
• 5:11
  SPEAKER_00: Um, and we're gonna dive a little bit more into that.
• 5:13
  SPEAKER_00: Now, nominal is the nicest, easiest, right?
• 5:16
  SPEAKER_00: It's usually dichotomous, binary, right?
• 5:20
  SPEAKER_00: There's no sense of order, there's no sense of ranking to the data, and it creates no bell curve because you're in one or the other.
• 5:30
  SPEAKER_00: Okay?
• 5:31
  SPEAKER_00: And so you only get really a mode or a frequency of that.
• 5:36
  SPEAKER_00: How many yeses, how many no's.
• 5:38
  SPEAKER_00: You can't be in between, okay?
• 5:41
  SPEAKER_00: Um, and so you get that bimodal kind of distribution of the data when you are doing the descriptive summary statistic statistics there, okay?
• 5:51
  SPEAKER_00: Um, and so that's this row right up here, but it follows in the non-parametric bucket, it's categorical, binary, dichotomous, treated.
• 6:02
  SPEAKER_00: There's no numerically measurable thing, and there's no sense of real order, which is different from ordinal data that does follow a sense of order or rank, which is why it's called ordinal.
• 6:15
  SPEAKER_00: Okay.
• 6:16
  SPEAKER_00: But that becomes important.
• 6:17
  SPEAKER_00: Now, when we're in non-parametric analysis, and let's say we're going over here and we have two independent groups, we would land like in a chi-squared versus a fissure exact.
• 6:27
  SPEAKER_00: So then you might say, well, what you know, Dr.
• 6:29
  SPEAKER_00: Busy, which one am I going to use?
• 6:31
  SPEAKER_00: I've got two listed there.
• 6:33
  SPEAKER_00: And when you see on the table two, typically, those are two separate tests.
• 6:37
  SPEAKER_00: Some of them are spillover from the you know the next line, like uh, for example, a chi-square for k-independent samples, that's all one thing, okay?
• 6:46
  SPEAKER_00: Um, but most cases uh recognize that chi-square and fissures exact are two different non-parametric statistical analysis when that data is not following a normal distribution.
• 6:58
  SPEAKER_00: But then you go, okay, well, well, how do I know uh which one to pick?
• 7:02
  SPEAKER_00: So we have tables like this that helped us to further characterize, you know, the when do we use chi squared over fissures exact?
• 7:09
  SPEAKER_00: And there's some small rules here.
• 7:11
  SPEAKER_00: One of it is based on the uh sample size, okay, typically large, like what number is that exactly?
• 7:18
  SPEAKER_00: I mean, for sure you want to be over 20.
• 7:20
  SPEAKER_00: Um, you know, it really comes down into individual cells and the frequency where the numbers get too low and then it gets a little wonky, right?
• 7:30
  SPEAKER_00: But chi squared is an approximate accura for accuracy.
• 7:33
  SPEAKER_00: Whereas Fisher's exact is an exact number.
• 7:38
  SPEAKER_00: Okay, and so it's best used for smaller sample sizes because it's more difficult, time consuming to calculate, especially back in the old days, by hand than when you now people can use a computer and they can do it very quickly.
• 7:52
  SPEAKER_00: Um, but if you want a more accurate, exact number, uh nowadays it's a little bit easier with computer programming.
• 7:59
  SPEAKER_00: But it's but you know, is it really exact?
• 8:02
  SPEAKER_00: You know, we get close, uh, but it's an approximation at this level.
• 8:06
  SPEAKER_00: But when you start having cells that have more than 20% of the cells have a frequency of less than five, okay, and cells being your data cells, you know, the two groups and the two, you know, the endpoints that you have, the yes or no's.
• 8:20
  SPEAKER_00: Um and so if you fall in this, okay, then you want to use the fissures exact versus a chi-square.
• 8:29
  SPEAKER_00: And that's most commonly seen when the sample size is smaller, and that's why that's listed up there.
• 8:35
  SPEAKER_00: But those are some of the then nuances that guide you into which which one do you pick over the other.
• 8:43
  SPEAKER_00: Okay.
• 8:44
  SPEAKER_00: Um, so again, reminding people of non-parametric versus parametric, remember the bell curve in the middle, the gray line, is normally distributed.
• 8:53
  SPEAKER_00: That's where the mean, median, and mode are the same.
• 8:55
  SPEAKER_00: But when you have skewed data, whether it's to the right, like in this case, or to the left in this case, it really has to do with where does the mean fit in relation to the median.
• 9:07
  SPEAKER_00: And so data, sorry, data that is skewed to the right over here, to the right, the mean is much greater than the median, because the remember the middle most, which is a median or in a quartile range, is gonna be over here.
• 9:22
  SPEAKER_00: Um and so the mean, something is skewing it.
• 9:24
  SPEAKER_00: There's some sort of outliers in the data.
• 9:27
  SPEAKER_00: Whereas negatively skewed or skewed to the left, the mean is going to be smaller than the median.
• 9:33
  SPEAKER_00: And that's really important as you assess your data because again, it's it's orienting you to this particular bucket.
• 9:42
  SPEAKER_00: Now, again, that would even apply to continuous data that no longer meets those assumptions.
• 9:48
  SPEAKER_00: So you do the you know evaluation of the central tendency of the data, and you get a median, uh median, the mode, median, and mean, and you do a comparison, how far away are they?
• 10:01
  SPEAKER_00: And it, you know, they may not be exact, but if they're really close and that follows that bell curve, you can do visuals like histograms and things to help guide you to that.
• 10:10
  SPEAKER_00: Well, that puts you, you know, which what are you treating that continuous data as parametric or non-parametric when it falls outside of the assumptions?
• 10:18
  SPEAKER_00: All right, and that's why I have that listed on those tables.
• 10:20
  SPEAKER_00: But parametric is going to follow that normal distributed information.
• 10:24
  SPEAKER_00: And that when we think about that in continuous data, remember the continuous, the magnitude of difference is the same.
• 10:32
  SPEAKER_00: And it can go on to infinity.
• 10:34
  SPEAKER_00: There's a sense of order, there's a sense of rank, and the magnitude of the difference is the same.
• 10:40
  SPEAKER_00: Okay, so you can get an average, you can actually generate a mean blood pressure in that kind of scenario.
• 10:48
  SPEAKER_00: And so that's where the magnitude of difference is the same.
• 10:51
  SPEAKER_00: It's one millimeter of mercury pressure at every level that you go versus something like New York Heart functional classification of heart failure, class one, two, and three, and four, there's a sense of order, there's a sense of rank, but the magnitude of difference is not the same.
• 11:08
  SPEAKER_00: And so if you're treating blood pressure as a continuous variable, it's usually in the context of average changes at the end of a study.
• 11:19
  SPEAKER_00: If you take that continuous variable and you you can now turn it into a nominal data endpoint, i.e., did this drug or intervention, right?
• 11:32
  SPEAKER_00: Remember the independent variable, the intervention, did it result in the achievement of the LDL goal at the end of the study?
• 11:41
  SPEAKER_00: Or did it achieve their blood pressure goal?
• 11:44
  SPEAKER_00: And so at the end of the study, it's a yes or no, it's dichotomous, it's binary.
• 11:48
  SPEAKER_00: So you're taking blood pressure, which is a commonly treated continuous data variable, and you're now treating it as a nominal endpoint.
• 12:00
  SPEAKER_00: Again, it changes the statistical test that you use, and you go outside of parametric analysis at that point when you do that.
• 12:08
  SPEAKER_00: And so that's why I spend so much time really trying to die, you know, codify this process.
• 12:16
  SPEAKER_00: So as you go through this biostatistical series, depending on what your purpose is, right?
• 12:23
  SPEAKER_00: Don't make any assumptions.
• 12:25
  SPEAKER_00: Keep diving into it.
• 12:26
  SPEAKER_00: Some of the repetition allows it to stick.
• 12:28
  SPEAKER_00: Sometimes one lecture adds another little element that fills in a little bit of a gap.
• 12:34
  SPEAKER_00: All of this is important to establishing that right foundation for not only literature evaluation and critically appraising a paper correctly so that you can apply it to your patients and follow the process of the five A's of EBM, which is, you know, ask, you know, acquire, assess, apply to your patient appropriately in the context of the evidence, but then you're supposed to assess your own practice.
• 13:00
  SPEAKER_00: Right?
• 13:01
  SPEAKER_00: To do that, you have to be able to interpret the literature that you're reading appropriately.
• 13:06
  SPEAKER_00: Is it internally valid in your opinion?
• 13:08
  SPEAKER_00: Because if it's not internally valid, then it's not externally valid, and trying to extrapolate it to the general population that you're serving would be inappropriate if it's not valid.
• 13:18
  SPEAKER_00: You have got to ask that question.
• 13:19
  SPEAKER_00: It's you that are going to defend your own license and your own medical decisions.
• 13:23
  SPEAKER_00: Okay?
• 13:24
  SPEAKER_00: If you're the researcher, you obviously want to get published.
• 13:27
  SPEAKER_00: You want to produce quality information that helps people.
• 13:30
  SPEAKER_00: The clinicians, the front line, yourself.
• 13:33
  SPEAKER_00: If you're doing your own research to help your own patients, right?
• 13:36
  SPEAKER_00: So you want valid information, and that's why codifying it and establishing a framework of what kind of data are you dealing with.
• 13:45
  SPEAKER_00: Now you might make assumptions at the beginning of the study that you presume, because of the sample size based on power analysis, that you're going to get all those patients.
• 13:55
  SPEAKER_00: And you presume that the randomization or the methods of how you try to control for confounders worked.
• 14:01
  SPEAKER_00: But that's why you do descriptive statistical analysis.
• 14:06
  SPEAKER_00: Those central tenancy.
• 14:08
  SPEAKER_00: What is the mean, media, mode?
• 14:09
  SPEAKER_00: What does your data look like around that sample?
• 14:12
  SPEAKER_00: How does it interact with other data points?
• 14:14
  SPEAKER_00: All of that drives this bigger concept of understanding parametric versus non-parametric data so that we then apply their appropriate statistical analysis and interpretation so that we can translate those results to something meaningful clinically.
• 14:32
  SPEAKER_00: Alright, guys, I hope that was helpful.
• 14:34
  SPEAKER_00: Alright.
• 14:35
  SPEAKER_00: Thank you so much for joining me and on this series.
• 14:37
  SPEAKER_00: Again, if you find it helpful, please share this with your friends, colleagues.
• 14:40
  SPEAKER_00: Give us a like, share any comments, perspectives.
• 14:44
  SPEAKER_00: Obviously, this is not the you know everything that can be talked about.
• 14:49
  SPEAKER_00: You know, so we all learn from each other, but I will respond to comments.
• 14:53
  SPEAKER_00: If you need something you know appropriate, then we'll we'll let's let's learn from each other.
• 14:57
  SPEAKER_00: It's great.
• 14:59
  SPEAKER_00: Also, other people on the channel can also learn.
• 15:02
  SPEAKER_00: So thanks so much for joining us.
• 15:03
  SPEAKER_00: If you need the tables, go to thisiswhy.elf.
• 15:06
  SPEAKER_00: It's available there for you for free.
• 15:07
  SPEAKER_00: Take care.
• 15:08
  unknown: Bye.