Ashish -

I take it that you do not have any kind of regressor data that you could use for your finite population? If you had, say, some administrative data on all truckers, say miles driven, or loads delivered, and stratified by region, or some other appropriate category, you could use prediction. An example with one regressor for regression through the origin -which is common - is shown by following the attached link. You need a regressor that is correlated with you variable of interest. The sample size estimate is to be done for each question, and assumes continuous data. This may not work for you.

You need regressor data to do anything rigorous with nonprobability sampling. Otherwise, as Dean and others explain, you are not going to be able to infer well to your population.

Bias is a problem. But as in looking at response propensity groups when dealing with nonresponse in a more rigorous sampling scheme, the way you stratify your data not only reduces variance in the more rigorous study, but could reduce your bias here.

You could look at scatterplots to compare results from different questions, which may give you some insight. If you have too little data, you won't have much to examine. But if you try to collect too much, nonsampling error, such as measurement error, becomes more of a problem. The number of observations needed to get much insight depends upon data variability. You won't know that without prior information, such as a pilot study. In fact, the paper at the link below, and any estimate of sample size needs for a design-based (randomized) study, also require this kind of information on variability.

If you obtain a survey sampling book, you can find out more. In the case of Cochran, W.G.(1977), Sampling Techniques, 3rd ed., John Wiley & Sons, he mentions four ways of obtaining prior information on variance, but in his 'formula' for obtaining a simple random same size estimate, I note he uses the word "guess." :-)

Best wishes with your project.

Cheers - Jim 

Conference Paper Projected Variance for the Model-based Classical Ratio Estim...

As your mentors say, if you are using non.probability sampling, N is not important, beacause your sample cannot be statistically representative.  Some researchers suggest to use 90-100 people, beacause it's 3x the value where t student distribution level-off.  But you need to consider if your study population can give that value whitout a lot of extra effort

Hi Ashish,

Non-probability sampling - this sampling approach is most appropriate for qualitative research, as well as small exploratory quantitative studies. I assume that yours is a quantitative approach. Non-probability samples are useful when information on the total population is unknown or unavailable. Non-probability sampling strategy is less rigorous for quantitative research approaches than probability as, in such cases, it tends to produce less accurate and less representative samples — thus limiting the ability of the researcher to make generalisations about the findings at a population level. In order to make inferences about the population, the sample must have the same characteristics as the population. That is, the sample must be representative of the population. If the sample characteristics are not the same as the population, then a sampling error has occurred, which threatens the external validity of the study.

Convenience sampling (also called incidental sampling or accidental sampling) uses the most readily accessible and willing persons or objects as study participants. They are one of the easier populations to target - but often the least generalisable. The calculation for size, with a limited population, is often down to power calculations. 

Ashish -

I take it that you do not have any kind of regressor data that you could use for your finite population? If you had, say, some administrative data on all truckers, say miles driven, or loads delivered, and stratified by region, or some other appropriate category, you could use prediction. An example with one regressor for regression through the origin -which is common - is shown by following the attached link. You need a regressor that is correlated with you variable of interest. The sample size estimate is to be done for each question, and assumes continuous data. This may not work for you.

You need regressor data to do anything rigorous with nonprobability sampling. Otherwise, as Dean and others explain, you are not going to be able to infer well to your population.

Bias is a problem. But as in looking at response propensity groups when dealing with nonresponse in a more rigorous sampling scheme, the way you stratify your data not only reduces variance in the more rigorous study, but could reduce your bias here.

You could look at scatterplots to compare results from different questions, which may give you some insight. If you have too little data, you won't have much to examine. But if you try to collect too much, nonsampling error, such as measurement error, becomes more of a problem. The number of observations needed to get much insight depends upon data variability. You won't know that without prior information, such as a pilot study. In fact, the paper at the link below, and any estimate of sample size needs for a design-based (randomized) study, also require this kind of information on variability.

If you obtain a survey sampling book, you can find out more. In the case of Cochran, W.G.(1977), Sampling Techniques, 3rd ed., John Wiley & Sons, he mentions four ways of obtaining prior information on variance, but in his 'formula' for obtaining a simple random same size estimate, I note he uses the word "guess." :-)

Best wishes with your project.

Cheers - Jim 

Conference Paper Projected Variance for the Model-based Classical Ratio Estim...

Ashish - 

About 20 months ago, I heard a presentation by Mike Brick on nonprobability sampling. (I've told Mike that with a name like his, he should have been a private detective - from the 1940s or 50s.)   It was the president's invited seminar for the Washington Statistical Society (a chapter of the ASA). He noted a report at the time, and I thought it might be of interest to you. I just found it on the internet, and I am attaching a link below. 

Jim  

https://www.aapor.org/AAPORKentico/AAPOR_Main/media/MainSiteFiles/NPS_TF_Report_Final_7_revised_FNL_6_22_13.pdf

First of all thanku all for giving me the answers , appreciate your co-operation

1.Luis Campo - thanks fr ur answer as you have mentioned "Some researchers suggest to use 90-100 people, beacause it's 3x the value where t student distribution level-off. But you need to consider if your study population can give that value whitout a lot of extra effort" -  in regard to your answer i dont understand what you mean by " 3x the value where t student distribution level-off."   Further " need to consider if your study population can give that value whitout a lot of extra effort"  - here i am not able to understand  " that value"  and last  an you please elaborate " wwithout a lot of extra-effort"  please if possible cite with examples thnku vry much.

2.James R Knaub - Thanks Sir for your answer .... sir i have done my pilot study taking 49 truck drivers  .  i have taken  their occupational details like the distance travelled by them , time of rest  in between the journey, number of round trips per month  and their socio demographic profiles, i have taken into account detaiked sexual history , Condom use history and pattern use  their knowledge and attitude and perception regarding  condom . further konwledge of hiv transmission . My dependant variable are Unsafe/Safe sex ; Condom use ; HIV Transmission knowledge  and rest our the independant factors ....So what do you further suggest i should do. Now in this case how will scatter plot help?Thankyou for your response and pdf attachment.

3.Patrick Druggan - Sir do you have this book in a pdf format if u do plz do attach it . Thankyou

Lastly , appeciate you all for your co-operation

Ashish -

I worked with surveys of establishments (skewed data) rather than people, fewer variables, and all my data were continuous. I'm not certain all of your data are continuous. I'm not an expert in your area, so i could say something nonsensical. But it sounds like you have a regression model going here already, so with regard to scatterplots, you might, if possible, make a number of plots of your y values against one regressor at a time and against combinations of regressors, to see if that let's you know which regressor or combination of regressors seem most useful. You could gain some insight. Also, there are other methods discussed in other questions on ResearchGate for looking at the importance of various regressors in multiple regression.

If you want to conjecture to truckers outside of your sample, and do not have a randomized sample, you have to use a model where the regressor data are known for the entire population, and you have to be able to categorize your data such that only data that model together well go into a given model for a given category, where the regression coefficients apply for that category. At least, that is the way i have worked. Otherwise results could be seriously biased. This sounds like a tough job here.

Since your pilot study is only n = 49, you do not have a lot of data to stratify into different categories.

Here is a thought: If you could identify two or maybe three groups of truckers, hopefully not too unevenly split regarding the data you have collected, that you think may respond quite differently, you could apply your model to each group and see how the estimated regression coefficients differ. You could also see about plotting y vs predicted y in each of the two or three groups. The one with the bigger estimated residuals would be the more variable group that would need a larger sample size in your follow-on work.

Scatterplots of y vs predicted y for continuous data would be a linear regression, through the origin, with an estimated slope of 1, and heteroscedastic variance.  Your actual model, however, could be decidedly nonlinear. 

Remember that if you overextend your resources, you may increase measurement and other nonsampling error.

I hope that is useful for you.

Cheers - Jim   

James R Knaub - thankyou sir ....you are right most of my data is discrete data and not continuous so  in this case you mean to say regression coefficient fits the best .....further can multiple logistic regression be used? thankyou sir

Ashish - I think that whatever modeling works for your data, that it still seems like a good idea to try splitting it into groups that represent truckers you would think may respond differently. Note, however, that even if it turns out that the data should be modeled together, in one group, you will still obtain somewhat different results from using any subset of your data, as opposed to another one. But if there really should be different groupings, you had best find them. Otherwise, your sample may be overly representative of one group, and underrepresentive of another. This also might mean your data split may have to be quite 'uneven' - but hopefully you have some data in your pilot study for each such group. - Jim 

Thankyou sir