Hello Hadar,

It would be helpful to have some additional information about the variable(s) or measure(s) you're working with in order to offer some constructive advice. As well, what is the specific purpose of this inquiry? Is it undertaken with the intention of showing differences, equivalence, consistency of relative values, identical distributions, or something else?

Do note that correlations can be high between two sets of scores and yet the scores may have wildly displaced mean/median values. So, part of the issue would have to be, what specific information would convince you (and/or others) that the two data sources were similar (or, different, if that's the concern)?

Good luck with your work.

Mr Yudkin:

For comparing lab and field data, we can use t test or z test if lab and field data consist on one treatment, respectively. If lab or field data are determined greater than one treatments then we can use ANOVA.

Hypotheses t test, z test or ANOVA:

H0: there are not difference among treatments

H1: there are difference at least one pair of treatments.

Depending on if you have multiple pairwise comparisons, standard statistical testing (Z, t, or F Tests) can become rather cumbersome. You can also use Nelson & Matejcik's "multiple comparison of the best" (MCB) procedure which is a derivation of a two-step Bonferroni comparison test. I use this method when comparing simulation results from multiple systems, as it allows me to generate multiple 100(1-a)% confidence intervals based off the difference from the best response observed. See my attached notes starting on page 27 if you have any questions.

Agree with Respected Nicholas Shallcross

Hadar -

If you have tried linear regression, then you should be able to plot your data on a scatterplot.

I'm not sure how you did this. Was there one regressor/predictor/independent variable, lab result, and the dependent variable was field result? Thus you were predicting field results from lab results? Whatever you did, you should be able to plot your data to see what the relationship looks like. (Note: A polynomial regression is still technically a "linear regression," even though curved.) You could also look at a "graphical residual analysis" to check fit. Also, "cross-validation" is used to check for overfitting to your particular sample, and may help to see if your relationship holds up here. An intercept term would likely be problematic. It would indicate a systematic, constant offset. Particularly if your sample size is small, using an intercept term when it is likely actually zero would not be helpful. Plotting the data, as well as subject matter knowledge, may be helpful when deciding whether or not an intercept term is justified.

Cheers - Jim

Thank you everyone for your advice.

I actually want to see the consistency of relative values from lab and field data. There are 2 variables (comparison between x and y). Although I have different treatments in the experiments I want to check all treatments in terms of those two variables.