3 Replies Latest reply on Nov 28, 2011 12:34 PM by Scott Tennican

    Custom trend lines i.e. null model

    Helen Greatrex

      Hello,

       

      I'm busy using Tableau software for the first time to do some rainfall analysis, so sorry if this is a very basic question!  I have also had a look through the forum & help files but I'm sure I might have just missed the answer.  The aim of my current work is to let students do some statistical analysis of rainfall data.  To do this, I want them to be able to play with trendlines and understand the p-values etc

       

      As far as I can tell, in the edit trend line box you can currently add

      A linear trend i.e. y = mx + c  ,    A polynomial trend i.e. y= b1x + b2x^2 .... bnx^n    ,    or a logarithmic trend

       

      For much of my data, there is no trend at all.  Therefore is there any way I could fit a null model e.g. y = C and get a p-value for that?  (so a straight line where the intercept has been fitted but it has zero slope)

       

      Or is there any way I could fit a custom model where I describe the fit I require?  Finally, is there any way to compare the different lines i.e. anova? 

       

      Don't worry if not, but it would be very handy if that's the case!  To be honest I didn't expect Tableau software to offer any statistical analysis, so I'm already pleasantly surprised by the analysis I can do and anything else is a bonus!

       

      Best wishes

      Helen Greatrex

        • 1. Re: Custom trend lines i.e. null model
          Scott Tennican

          Hi Helen,

           

          It sounds like you are looking for a seasonal model.

          For example, you might use a linear regression model with ARIMA as the error term as you can do with the arima function in R.

          Tableau doesn't have any sort of seasonal model yet.

          A linear trend through your data will give you an insignificant p-value if there is no significant slope.

          However, in 7.0, we have added t-tests for the significance of each term in the regression in addition to the f-test which gives you the p-value for the line. With 7.0, you would see that your intercept term has a significant p-value.

          But, I recommend adding reference lines instead. You could add a reference line for the average and another for a confidence band at the 95% level. This gives you the information you would get from a constant model. You could also add the standard deviation band and show your students how the confidence band changes width as either the number of observations or the standard deviation of the observations changes.

          With regards to ANOVA, Tableau currently only has multi-factor ANOVA for regressions. This is used to test the significance of categorical variables in the trend line model when you use them to partition your data.

           

          If you are using a 7.0 beta, I have five workbooks that I use for teaching statistics with Tableau.

          A couple of them use references lines as I described above and another shows how to use trend line ANOVA feature to select the correct categorical variable to use in your model.

          I've attached a couple just in case.

          Remember, you need the 7.0 beta to view the workbooks.

           

          good luck, Scott

          • 2. Re: Custom trend lines i.e. null model
            guest contributor

            Hi Scott,

             

            That's fantastic - thanks for the help and advice

             

            Cheers

            Helen

            • 3. Re: Custom trend lines i.e. null model
              Scott Tennican

              Hi Helen,

               

              The materials from the Tableau Customer conference 2011 (TCC11) Session presentations are now online.

              Here is the link to a zip of the powerpoint and all the demos for my Statistics session:

              http://www.tableausoftware.com/sites/default/files/pages/tcc11-demystifyingstats.zip

               

              Scott

               

              Scott