Skip to main content
Cross Validated

Return to Answer

edited typos in body
Source Link
jginestet
jginestet
  • 5.4k
  • 1
  • 6
  • 19

Like for any other data analysis, the very first thing to do is to look at the data (so called IOTT - IntraInter Occular Trauma Test- see if the data tells you something which hits you between the eyes).
In this case, it would be a simple scatter plot, ploting $R$'s against $S$'s (or vice versa). That plot alone will tell you a lot; can you see any kind of correlation between the 2 counts of events? If they look independent, then they are (yes, you can run statistical tests to formalize this, but you already know the answer). If there is some kind of trend, then you also know the answer; they are not independent.
Now, you can quantify this in/dependence. The simplest is a linear regression, $S=b_0+b_1.R$. You can add higher order terms ($R^2, R^3$) to your regression, if needed. You will get coefficients of (linear) correlation R, you will get confidence intervals (CI) and p-values on the coefficients $b_0$ and $b_1$, etc. and from that you can determine if the correlation is statistically significant.
But I am afraid that what I described above may be a bit naive. This treats all the users as a single homogenuous population. But it would seem that a frequent user of the platform may behave differently than a very occasional user, that a long time user may behave differently than a newbie, that a gen X user may behave differently than a boomer, etc... If you find no correlation, it may not be because there is none, but because, when "averaged" over all user categories, the dependencies cancel out. So you may want to categorize your users, and either analyze each category separately, or add this predictor (user group) to the regression (it will give you different equations, and results, for the different categories). It is also not clear that your data corresponds to the exact same time period for all users (there may be 10 years of data for 1 user, and only 10 days for another). So the duration of the observation should be made equal, or the users should be inversely weighted by their duration of observations. And users at the launch of the platform may not behave as they do today... The whole point is that I doubt you have homogenuous data (aka identically distributed), so any formal statistical test is basically pointless; you will need to slice and dice it until you can be reasonably sure it is homogenuous.

Like for any other data analysis, the very first thing to do is to look at the data (so called IOTT - Intra Occular Trauma Test- see if the data tells you something which hits you between the eyes).
In this case, it would be a simple scatter plot, ploting $R$'s against $S$'s (or vice versa). That plot alone will tell you a lot; can you see any kind of correlation between the 2 counts of events? If they look independent, then they are (yes, you can run statistical tests to formalize this, but you already know the answer). If there is some kind of trend, then you also know the answer; they are not independent.
Now, you can quantify this in/dependence. The simplest is a linear regression, $S=b_0+b_1.R$. You can add higher order terms ($R^2, R^3$) to your regression, if needed. You will get coefficients of (linear) correlation R, you will get confidence intervals (CI) and p-values on the coefficients $b_0$ and $b_1$, etc. and from that you can determine if the correlation is statistically significant.
But I am afraid that what I described above may be a bit naive. This treats all the users as a single homogenuous population. But it would seem that a frequent user of the platform may behave differently than a very occasional user, that a long time user may behave differently than a newbie, that a gen X user may behave differently than a boomer, etc... If you find no correlation, it may not be because there is none, but because, when "averaged" over all user categories, the dependencies cancel out. So you may want to categorize your users, and either analyze each category separately, or add this predictor (user group) to the regression (it will give you different equations, and results, for the different categories). It is also not clear that your data corresponds to the exact same time period for all users (there may be 10 years of data for 1 user, and only 10 days for another). So the duration of the observation should be made equal, or the users should be inversely weighted by their duration of observations. And users at the launch of the platform may not behave as they do today... The whole point is that I doubt you have homogenuous data (aka identically distributed), so any formal statistical test is basically pointless; you will need to slice and dice it until you can be reasonably sure it is homogenuous.

Like for any other data analysis, the very first thing to do is to look at the data (so called IOTT - Inter Occular Trauma Test- see if the data tells you something which hits you between the eyes).
In this case, it would be a simple scatter plot, ploting $R$'s against $S$'s (or vice versa). That plot alone will tell you a lot; can you see any kind of correlation between the 2 counts of events? If they look independent, then they are (yes, you can run statistical tests to formalize this, but you already know the answer). If there is some kind of trend, then you also know the answer; they are not independent.
Now, you can quantify this in/dependence. The simplest is a linear regression, $S=b_0+b_1.R$. You can add higher order terms ($R^2, R^3$) to your regression, if needed. You will get coefficients of (linear) correlation R, you will get confidence intervals (CI) and p-values on the coefficients $b_0$ and $b_1$, etc. and from that you can determine if the correlation is statistically significant.
But I am afraid that what I described above may be a bit naive. This treats all the users as a single homogenuous population. But it would seem that a frequent user of the platform may behave differently than a very occasional user, that a long time user may behave differently than a newbie, that a gen X user may behave differently than a boomer, etc... If you find no correlation, it may not be because there is none, but because, when "averaged" over all user categories, the dependencies cancel out. So you may want to categorize your users, and either analyze each category separately, or add this predictor (user group) to the regression (it will give you different equations, and results, for the different categories). It is also not clear that your data corresponds to the exact same time period for all users (there may be 10 years of data for 1 user, and only 10 days for another). So the duration of the observation should be made equal, or the users should be inversely weighted by their duration of observations. And users at the launch of the platform may not behave as they do today... The whole point is that I doubt you have homogenuous data (aka identically distributed), so any formal statistical test is basically pointless; you will need to slice and dice it until you can be reasonably sure it is homogenuous.

Source Link
jginestet
jginestet
  • 5.4k
  • 1
  • 6
  • 19

Like for any other data analysis, the very first thing to do is to look at the data (so called IOTT - Intra Occular Trauma Test- see if the data tells you something which hits you between the eyes).
In this case, it would be a simple scatter plot, ploting $R$'s against $S$'s (or vice versa). That plot alone will tell you a lot; can you see any kind of correlation between the 2 counts of events? If they look independent, then they are (yes, you can run statistical tests to formalize this, but you already know the answer). If there is some kind of trend, then you also know the answer; they are not independent.
Now, you can quantify this in/dependence. The simplest is a linear regression, $S=b_0+b_1.R$. You can add higher order terms ($R^2, R^3$) to your regression, if needed. You will get coefficients of (linear) correlation R, you will get confidence intervals (CI) and p-values on the coefficients $b_0$ and $b_1$, etc. and from that you can determine if the correlation is statistically significant.
But I am afraid that what I described above may be a bit naive. This treats all the users as a single homogenuous population. But it would seem that a frequent user of the platform may behave differently than a very occasional user, that a long time user may behave differently than a newbie, that a gen X user may behave differently than a boomer, etc... If you find no correlation, it may not be because there is none, but because, when "averaged" over all user categories, the dependencies cancel out. So you may want to categorize your users, and either analyze each category separately, or add this predictor (user group) to the regression (it will give you different equations, and results, for the different categories). It is also not clear that your data corresponds to the exact same time period for all users (there may be 10 years of data for 1 user, and only 10 days for another). So the duration of the observation should be made equal, or the users should be inversely weighted by their duration of observations. And users at the launch of the platform may not behave as they do today... The whole point is that I doubt you have homogenuous data (aka identically distributed), so any formal statistical test is basically pointless; you will need to slice and dice it until you can be reasonably sure it is homogenuous.