Summary of suggestions received for Trend Analysis - Anova One-way with intervals that are not equal

[giorgio de marchis]

Hi everybody!

 

Here there is a summary of the suggestions I have received to solve the problem of the Trend Analysis with different intervals:

 

1.- Marija suggested the way to do it using One-way: “the value of X that SPSS Oneway uses when it calculates polynomial contrasts is the value that you assign to your group codes. For example, if your X values are 1, 3 and 10, you should assign values of 1, 3 and 10 for the grouping variable (the variable that defines groups)”. This seems confirmed by the text I have found in the Syntax Reference Guide: “ONEWAY computes the sums of squares for each order polynomial from weighted polynomial contrasts, using the category of the independent variable as the metric. These contrasts are orthogonal.”

I have pointed out to Marija that the graph comes out from One-way is the same when I use equal intervals (e.g. 1;2;3;..) than when I use different ones (e.g. 1.2; 3.5; 8; …). She suggested that maybe there is a bug in the plot Oneway produces. Actually, I have not been able to get a good plot from SPSS even with graphics, and I had to do it with Excel in order to get the different intervals on the X axis (Is this a new problem for the discussion group?)

 

2.- Another simple solution is the one offered by Martin. He suggested the Cochran-Armitage Trend Test. It is done from a cross-tabulation, so it is very simple to do. It seems that SPSS doesn’t do it directly but Martin has told me that Marta has written a syntax to make it:

http://www.kingdouglas.com/SPSS/DiverseCultures/Marta/Code/Armitage-Cochran%20Test%20for%20Trend.TXT

XLSTAT and SAS can do it.

In my case, the problem was that this statistic just measures linearity.

 

3.- After reading my imprecise message William suggested some bibliography if I was working with Longitudinal Data and growth curves. I was not, but the suggestion was very interesting. At the end I have dropped this line because, as pointed out by Paul, I am not working with intra subjects design.

 

4.- Paul also suggested an excellent solution: Curvilinear Regression. It allows to analyze many different curves.

 

5.- Anyd told me to calculate the polynomial coefficients. It is also a good way to solve the problem, but a little bit more complicated.

 

6.- Michael suggested to estimate for monotonic trends while preserving the test of quadratics. Keppel & Wickens discuss the details in Chapter 5. 

Michael, kindly, summarized the procedure to me, but I have found it too complicated.

 

Thank you everybody again for your help.

 

Giorgio

 

[SPSS Support]

With regard to the issue in number 1 of the plots in ONEWAY not reflecting distances among numeric codes for the factor, this is indeed the way it is, and it's not a bug, as the design never included such a feature. The same is true for the profile plots in the GLM and UNIANOVA procedures. Although you can fit polynomials and user-defined contrasts in these procedures using various options, these are always done within the context of models in which factors are treated as nominal.

 

You can get a plot of the means by group using unequal spacing using the Chart Builder if you have the variable type for the factor defined as Scale. You can do this in the Data Editor, or using the VARIABLE TYPE command, or temporarily in the dialog box for the Chart Builder by right-clicking on the variable in the source list, then changing the type in the popup dialog that comes up. Make a line chart, specifying mean of the variable you put on the Y axis. The spacings should reflect the numeric values of the variable on the X axis. If you want point markers, after creating the graph, activate it in the Chart Editor, and specify Elements>Add Markers.

 

David Nichols

---

From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of giorgio de marchis
Sent: Wednesday, May 20, 2009 4:10 AM
To: [hidden email]
Subject: [SPSSX-L] Summary of suggestions received for Trend Analysis - Anova One-way with intervals that are not equal

Hi everybody!

 

Here there is a summary of the suggestions I have received to solve the problem of the Trend Analysis with different intervals:

 

1.- Marija suggested the way to do it using One-way: “the value of X that SPSS Oneway uses when it calculates polynomial contrasts is the value that you assign to your group codes. For example, if your X values are 1, 3 and 10, you should assign values of 1, 3 and 10 for the grouping variable (the variable that defines groups)”. This seems confirmed by the text I have found in the Syntax Reference Guide: “ONEWAY computes the sums of squares for each order polynomial from weighted polynomial contrasts, using the category of the independent variable as the metric. These contrasts are orthogonal.”

I have pointed out to Marija that the graph comes out from One-way is the same when I use equal intervals (e.g. 1;2;3;..) than when I use different ones (e.g. 1.2; 3.5; 8; …). She suggested that maybe there is a bug in the plot Oneway produces. Actually, I have not been able to get a good plot from SPSS even with graphics, and I had to do it with Excel in order to get the different intervals on the X axis (Is this a new problem for the discussion group?)

 

2.- Another simple solution is the one offered by Martin. He suggested the Cochran-Armitage Trend Test. It is done from a cross-tabulation, so it is very simple to do. It seems that SPSS doesn’t do it directly but Martin has told me that Marta has written a syntax to make it:

http://www.kingdouglas.com/SPSS/DiverseCultures/Marta/Code/Armitage-Cochran%20Test%20for%20Trend.TXT

XLSTAT and SAS can do it.

In my case, the problem was that this statistic just measures linearity.

 

3.- After reading my imprecise message William suggested some bibliography if I was working with Longitudinal Data and growth curves. I was not, but the suggestion was very interesting. At the end I have dropped this line because, as pointed out by Paul, I am not working with intra subjects design.

 

4.- Paul also suggested an excellent solution: Curvilinear Regression. It allows to analyze many different curves.

 

5.- Anyd told me to calculate the polynomial coefficients. It is also a good way to solve the problem, but a little bit more complicated.

 

6.- Michael suggested to estimate for monotonic trends while preserving the test of quadratics. Keppel & Wickens discuss the details in Chapter 5. 

Michael, kindly, summarized the procedure to me, but I have found it too complicated.

 

Thank you everybody again for your help.

 

Giorgio