In compliance or regulatory monitoring, where for example one tests whether concentrations of pollutants exceed regulatory standards or not, the validity of the result (conclusion of hypothesis testing) is of primary interest, it may be argued that design-based methods are more appropriate( Brus and Knotters, 2008). Is it okay also if model based approach is used, is there room for that?
For compliance/regulatory monitoring, the standards and approaches applied are either identified or expectation that standard methods used. If you have some non-conventional approach, best to get it approved by regulatory agency first, make the case to them relative to the specific circumstance. They are only ones with authority to make this type decision, and it is apt to be specific to the circumstances and model precision relative to subject area as to whether modeling would suffice for any, partial or all monitoring compliance.
Thanks for the useful advice. I have followed the standard methods for analysis of water quality data.
I am using model based approach because I have space-time data for temporal trends and mapping quantities. Does that mean I can't, in addition to that, compare my data to the already existing water quality threshold values (Standards) provided by either EPA or WHO?
I am asking because it has been stated that design based methods are more appropriate for testing whether concentrations of pollutants exceed regulatory standards or not.
Bernice -
This is not my field, so it is good that William noted that the allowed procedures may be dictated. I had not thought of that. -- My work in modeling was model-based sampling and prediction for official statistics (electric sales and revenue totals, etc.).
Just some general thoughts which may or may not apply:
If you can use modeling, I assume that means regression modeling, and that you have regressor data on the whole population. You then need to be sure to stratify your population such that each set (stratum or subpopulation) is modeled appropriately by just one regression equation each. Such model-based application groups may overlap the actually published data groups.
Hypothesis tests are nebulous and generally misused and misunderstood. Note that an isolated p-value is basically meaningless. There is no universally acceptable p-value threshold. This depends upon sample size, standard deviation, and other context. If you can use confidence or prediction intervals, they are generally more practically interpretable.
Best wishes - Jim