Muhammad - I think that in the 1940s, there was a debate between purposive sampling and random sampling for official statistics. Apparently some thought that it would be more "representative" for an expert to pick a sample, purposively, according to his/her expertise. However, random sampling was developed from which inferences could be made regarding the population from which the sample was drawn, and standard errors could be estimated. I am not aware of any distinction between purposive and nonprobability, though someone else may know.
A famous - or infamous - example might be the telephone survey that predicted that Dewey would defeat Roosevelt for president of the US. That was before many people in the USA had telephones, so the sample - perhaps a quota or convenience sample? - contained more Republican responders than Democrats, as Republicans were more likely to have telephones.
There appear to be many types of purposive samples of which I am barely aware, but they all have potential bias due to lack of randomization. However, there is an important exception. If regressor data are available on the entire population, then model-based estimations can be made. These are predictions. This is not forecasting, but relies on a sample, possibly purposive, that relates to the regressor data. From this relationship/model, the non-sampled data are predicted (estimated).
In the case of the model-based classical ratio estimator (CRE), variances can be obtained when a regressor (x) datum is available for each sample (y) datum, and only the sum of all the x data are available for the population. An early example of the CRE was when LaPlace estimated the population of France, using birth records for regressor data, I think. I don't think he realized that he could have estimated a standard error on his result.
A modern use of prediction is when, in official statistics, a census, on industrial output, for example, may be done infrequently (say annually), and a sample of the larger plants may be done more frequently, say monthly. The regressor data may be the same data elements/variables of interest, from the census as are being estimated from the samples. Or, the regressor data could be an administratively known quantity, such as the nameplate capacity of the equipment.
So there is randomization, for probability/design-based sampling and estimation, and there is model-based estimation that can be done on any sample, though one needs to be careful to stratify and apply regression models only to data that should be modeled together, to avoid bias.
A random sample can, by bad luck, be nearly all large or nearly all small members of the population, and badly over- or underestimate population totals, respectively. That is why model-assisted design-based sampling and estimation can be so useful.
Statistically speaking, there are two sampling methods i.e. 1) Probabilistic method, and 2) Non-Probabilistic method.
In probabilistic, that's where each subject has equal chance of being selected unlike the case in non-probabilistic method.
Each of the sampling method has got several sampling techniques; e.g. Purposive sampling is one of the sampling techniques under non-probabilistic sampling method.
Please have a look the link below for more information;