Every time you are going to analyse your data you should do correlation analysis before the regression one because if there are no any dependencies in the data the models of any type will be absolutely useless as they could predict something nonexistent. In correlation analysis you could discover which factors have significant effect to your response variables and which of regression functions are better to use for these interdependence's description (I mean some nonlinear tranformations which could improve the tightness of interdependence, and to increase the correlation coefficient's significance).

Vladislav is right, you should use even more "primitive" analyses (like looking at diagrams, trends) before.

This is slightly off topic ( but I agree with everyone's answer ) but I would look at the work of Marc Nerlove. He's done a ton of econometric modelling related to argiculture and that area has subtleties because a lot of farmers base their supply on expectations. So, expectations formation comes into play in a big way. Good luck.

correlation shows the degree of association between two different variables ie whether positive or negative, therefore doing this prior to regression analysis paves the way to select variables that define the relation well

if data is not stationary, shall I use detrended or first differenced data for correlation analysis ?

You should first think over how the functional relations between the variables should be in principle. For linear relations, I would take absolute differences, for multiplications/divisions or exponentiel functions relative ones (percentage changes, dln). For a mix of functions, it is not as easy, but mostly one can find linear approximations for the relation of differences of both sorts.

Anton makes a great point in that any kind of functional relationship most likely depends heavily on the response. Some responses may be assumed to occur instantaneously. Other might be assumed to occur with a delay. Others may be assumed to occur over time and are called  dynamic as opposed to static.. Some may  be assumed to have a delay and also assumed to occur over time.

These issues have significant implications on the functional form. If you think it's instantaneous and static, ( by this  I mean no delays and no dynamics }, start with linear regression which is the most straightforward.

                                                                                                                   Mark

Yes. it  tells about the strength of association between response variable  and regressors. Secondly, correlation analysis also show the association among the explanatory variable which help in the identification of multicollinearity problems. The correlation coefficient between regressors is 0.95 or above give clear indication of multicollinearity in the estimating model if the variables taking in the same position in the estimating equation

regards

irshad

Mark,

Even if one can assume or theory says that the relation is instantaneous and static, one has to choose which functional relations seem realistic, because that influences the (quasi-correct) transromation of the variables. If the variables have a trend; I recommend a regression (which itself is, of course, linear) in differences. Which differences is dependent on the assumed relation, in most cases, I think, relative differences (% or dln) are the best choice. For high collinearities (see Sayeds answer) one could never find the "correct" coefficients of the correlated variables (but maybe a common coefficient of a composed variable) without additional information.

HI Anton: Agreed but I think we can't say too much unless we know what the response and the predictors are.

What aspect of agriculture are you attempting as dependent variable? Is it the output in tonnage? Is it the revenue? Is it quality of agricultural output? 

Yes. It would be necessary to conduct correlation analysis to know the level(s) at which your variables are stationary. Cheers

First, it depends on the type of data you are working with. You talk of stationarity or non stationarity of data when you are working with a secondary data. The need for stationary time series data is to avoid running a spurious regression. So if you are working with a time series data, you must first ensure that your data is stationary. Then you can run a regression without necessarily conducting correlation test. But if you are using cross-sectional data, you don't need to find out whether you data is staionary either at level form or first difference. Then you can carry out correlation test to ascertain the relationship between variables and conduct regression analysis to determine the influence of selected variables(Independent variables) on a particular variable(dependent).

Ali is right. Theory must be first. It need not be textbook (or other published) theory (which may not exist for your problem). It could also be your own model. It should be elaborated as far as possible (relevant variables, functional form of the relations, even the approximate values of the coefficients), then looking for the adequate data, before the statistical tests.