Dear Sayantani,

No (If you ask in one word), Statistically we won't encourage regression analysis on weighted data (until unless its not in mannered linear form/or functions as linear transformation, & that simply have no significance).

However It's a topic of discussion.

While classical statisticians won't recommend regression on weighted data, it may mislead the coefficient& inference (because you are weighting data as per you customised objective and that will create a kinda biases so randomness of your dataset will no longer be valid in this scenario).

I'd rather prefer data as it is (conditioned on some necessary cleaning).

I came across an old paper sometimes ago and that suggest to not to use weighting in regression analysis.

However there are few(probably in Bioststs), who do use weighted input data while doing analysis/modelling.

My point of discussion is that we choose simple random sampling to avoid biasness and assign equal probabilities to all the units in frame (also for fulfilling the assumption of randomness), once you assign the weights, it won't be be random/unbias anymore & hence the statistical validity of resultant/output of statistical analysis would be tough to establish.

Do let me know in case of any queries.

Regards,Abhay

No, i agree with Abhay

A sincere thanks to Abhay and Abiodun.

In that case, the sampling design has no role to play. Is it possible to incorporate the sampling design anyhow?

Hello Sayantani,

If a sample was selected by using a complex sampling design (stratification, clustering, and weights) you have to include the design variables in your analysis.

The importance of adding strata and cluster design variables into the analysis is to compute correctly the standard errors of your estimates. Regarding the weights, you can compare your results with and without weights. If you observe differences, this would be an indication that the unweighted estimates are biased. Therefore, you should have to compute weighted estimates.

Roberto

Thank you Roberto.

Mainly I use data from demographic surveys which adopt complex sampling designs, often multistage ones.

Both sampling weights, from a sampling design, and regression weights, to account for heteroscedasticity, are useful. 

For a simple example of using the design, consider the generalized ratio estimator (Thompson(2012),  p. 102). 

References:

Särndal, CE, Swensson, B. and Wretman, J. (1992), Model Assisted Survey Sampling, Springer-Verlang.  

Thompson, S.K.(2012), Sampling, 3rd ed, John Wiley & Sons.

https://www.researchgate.net/publication/319434665_Comparison_of_Model-Based_to_Design-Based_Ratio_Estimators

thank you sir

Shall I use sample weight as Gross State Domestic Product (GSDP) for multiple regression Child Mortality model? However results came is very impressive, does it make sense?