Stepwise Approach to Statistical Analysis

So as I mentioned, let’s look at a power example to get a little bit better of idea of what I was talking about before. So you can see what I’ve highlighted here they set a sample size of 205 patients per treatment group was needed to achieve 92 percent power to detecting difference of 0.4% in change in HbA1c from baseline. So you can see I listed the four components of power on the right. So it’s the first one is sample size. Well, from this you can tell because remember I told you it’s a sample size in the final analysis. So you need to go look at the results to see how many patients they actually had.

And so we really would have to look and so you can see I want to look for us and it showed that the results had 233 and 223 patients respectively in each group, so that criteria for sample size was meant needed to be linked to the primary outcome, well you’d have to look somewhere else and change in HbA1c was their primary outcome. So that was correct. And the power of 80-90% but then linked to a clinically appropriate clinical marker. If you look in the guidelines, usually 0.5 is the the least amount of change that you see is clinically relevant. So this one is considered a little bit low.

So it’s easier to find a significant difference the lower they affect sizes. So I know there’s a lot on this slide and I’ll go over each part separately. But I wanted to put it all in place for you. So when we’re looking to determine if the correct statistical test was used, there’re three main things that you need to look at and that’s study design, data types and the number of groups. So the first one is looking at is the study independent or dependent? And you’ll say well why does that matter? Well, when they’re calculating the statistics, an independent test will have a smaller N. Then you will end a dependent test because you have double the data when it’s dependent.

So you just need to make sure you can determine whether it was baseline to endpoint in each group, it was looking at two groups separately, to make sure you’re picking the right statistical tests. And that’ll become obvious when I show you the lists of the drugs that or the statistical test that you’ll use. The next one is determining the data types and this will determine this statistical tests that are used. And again I’ll go over those a little bit more. And you can see that there’re some other little caveats, too. To that, besides determining, the datatype that I’ll also go over as we get to those sections. And the last one is the number of groups.

Now this is important because you’ll have some statistical tests can handle two groups and some can handle two or more groups. So if you can use a study for for two or more for three or more groups, you can also use it for the two. But if it can handle two groups it can’t handle three. So you’ll see that as we go along as we list the options for each statistical test that could be used. So the data types as I mentioned after determining this study design is very important. And even though there’re four types of data, it’s really convenient that you can list them is interval/ratio together. Because the same statistical test will be used for them.

And I know people always want to know what’s the difference between interval and ratio? Well, interval and ratio are both continuous data, so it’s most of the things that you think about that we see in medicine, blood pressure, height, weight. Ratio is just an absolute zero, so it’s using Fahrenheit versus Celsius type of a thing for a zero scale. So there’s no reason to differentiate between them. So you just need to know is it a continuous scale or not and that’s listed as interval/ratio data. And this is considered the most specific.

The middle data type is ordinal which is rankings and ratings, a lot of the subjective data with that we have with Likert scales and psychiatric scales and things like that. And then the least specific type of data is nominal and you’ll see that that’s primarily yes/no, although it could be yes no no no if you’ve got more than two groups. So it’s very important that we be able to differentiate between these and we’ll go over some examples of them to help you to get a better feel for what they are. Now converting data types.

Now you’re not going to be doing this, but it’s something that the statisticians in the article might do and there’re appropriate conversions and there’re not appropriate conversions. Now interval/ratio data, and we’ll get to this in a minute with the previous slide, in order to be able to use a test on interval/ratio data, the data must be normally distributed. And sometimes and you’ve got very small sample sizes or even sometimes in larger sample sizes it may not be. So you can’t use a certain statistical test on that. So it’s appropriate for the statistician to convert that interval/ratio data to ordinal. You’re not doing it, they will be, and use an ordinal test.

So we’ll see that when we go through finding things that if you can use an interval/ratio test, you can also use an ordinal test on that interval/ratio data, and that’s not a violation. Now again that’s the only conversion that’s acceptable. And the other way, you’re losing too much of your data specifics so you can’t convert ordinal to nominal that again loses too much of your data types. Now one conversion that they’ll sometimes make which is okay as long as they present it both ways. If they take their interval/ratio data and change it into response versus nonresponse, and that would be nominal nominal data. That’s all right as long as they provide you with the original specifics of the data.

Because otherwise, you’re not getting as robust of results if they only report it as a response and nonresponse.