P-value - Parametric Vs Non-parametric

[mss90]

Greetings

Is it possible to compare the p-values derived from non-parametric tests to parametric tests? For example, lets say dataset 1: Copper concentration in spoil is not normally distributed, so I have to use a non-parametric test on it and get a p-value. Dataset 2: copper concentration in plants happen to be normally distributed so I use a parametric test and derive a p-value. In my discussion, can I compare the two p-values despite the fact the methods used to derive them are quite different?

Thanks

Marius

 

[Miner]

I would say no. P-values are always in terms of a specified null hypothesis. Since the two null hypotheses are different, the p-values could not be directly compared.

 

[ondansetron]

I would say no. P-values are always in terms of a specified null hypothesis. Since the two null hypotheses are different, the p-values could not be directly compared.

I believe that it p-values are typically not comparable, as a general rule. I had read this from a pretty credible source, but I can't find it at the moment.

Question for the OP: what would you hope to gain from comparing p-values? There may be another way to learn the information you hoped to obtain.

 

[Dason]

Even if you're using the same test... What do you mean by comparing the p values?

 

[hlsmith]

So these are two independent samples? What are the sample sizes? What tests are you actually running? List what variables are in which test.


Sidenote, even if you can compare the two samples, you wouldn't want to compare the p-values, but the effect estimates and precesion. P-values are misleading if the two samples are different sizes, etc.


Second, you can also for the ease of the paper's write-up, just run all non-parametric tests. Also can you transform data to normal and run parametric on both.


In general, please provide more information. Thanks!

 

[mss90]

So these are two independent samples? What are the sample sizes? What tests are you actually running? List what variables are in which test.


Sidenote, even if you can compare the two samples, you wouldn't want to compare the p-values, but the effect estimates and precesion. P-values are misleading if the two samples are different sizes, etc.


Second, you can also for the ease of the paper's write-up, just run all non-parametric tests. Also can you transform data to normal and run parametric on both.

In general, please provide more information. Thanks!

First time im doing stats like this in a long time so I am a bit rusty. However, things seems to be clearing up. In terms of comparing p-values, I guess I was thinking of something like this; The p-value is < 0.05 for copper concentrations in the soil, hence there is no significant differnace in copper concentrations in the spoil and the vegetated areas, however the p-value for EDTA Cu extractable is >0.05. This suggests that the EDTA concentration of Cu is significantly different in the spoil compared to the vegetated areas whereas the total Cu extractable is not.. Something like that? Em I making any sense?


Yeah, independent. n = 18-30 depending on the dataset where each data set is split into 3-5 groups with significance lvl of 0.05.
One way ANOVA w/Tukey and KW (using Pairwise comparisons in SPSS to decipher which groups are significantly different) on continous data such as weight, concentration, % etc.

How would it work to run non-para on normally distributed data? Is it possible to transform not normally distributed data to normally distributed data?

Cheers