It seems that Respondent Driven Sampling would be appropriate. 

First try to collect the emails of academic staff either by requesting the college administration stating the research objectives to them or by some other source to known to you. Use the cross section approach. First collects the data as far as possible using your email approach. then  stratify the data as per your hypothesis and then attempt the analysis 

Maksym -

The following article seems to indicate that though respondent driven sampling could be good in many large studies to find rare parts of a population, I don't think it would be ideal for your application:

 http://www.ncbi.nlm.nih.gov/pmc/articles/PMC1705515/

"Variance Estimation, Design Effects, and Sample Size Calculations for Respondent-Driven Sampling"

In your case, you know the population, but the pending expected inability to obtain all email addresses for those staff that will be selected appears to be equivalent to an expected nonresponse problem.

You might try something straightforward.  If you draw a simple random sample across all 133 universities without regard to university, so that you have a one-stage simple random sample with 'staff member' as your unit, and thus I expect that N, the population size, is in the thousands, then you could draw n out of that one 'bucket' of N staff members. (For greater accuracy, and/or more information, you may want to stratify or categorize by staff-type of some kind.)  Then you would expect that for larger universities, you would have more staff members selected, and it might cut down on the number of places you need to contact for help in obtaining email addresses than you might think.  However, you still need to obtain a list of all names.  It is common to have frame errors, such as a slightly incorrect such list, but you will hopefully come close.

Your eventual nonresponse problem from no email address might then be treated as "missing at random," and you might draw a replacement from that same university without a great deal of bias to your results. 

If you are looking at continuous data, or yes/no data for proportions, then you can find information (even 'formulas') for sample size needs determinations - which will yield a different result for each variable/question, so you may need to look at which question is most critical - in a survey sampling book such as Cochran, W.G(1977), Sampling Techniques, 3rd ed., John Wiley & Sons.  There are many other such texts. 

Regardless of what kind of data you are examining, sample size needs are based on the accuracy you want to obtain.  Sample size 'formulas' generally assume no bias, and are based on reducing variance of an estimated mean, by increasing the sample size. This can be a problem if your sample size becomes large and it might increase nonsampling error and/or be too costly or take too long. Also, bias from the replacement of missing data can be a problem when it was not really missing at random.  For continuous data, if you replace a missing at random case with an 'average' number, say from the same university, or all universities, then you will artificially reduce your estimated variance, and thus think the result is better than it is, unless you look into more advanced techniques.  (Drawing a replacement from the same university, or taking an 'average' from it, may help ensure that assuming missing at random is OK, even though you were using a simple random sample across all UK staff.) 

Even for qualitative data, I think you still want to avoid bias, and have a large enough sample size to cover the variability of the population.  In any case, if you can avoid bias as noted above (also perhaps research "response propensity" groups which is basically the same idea), and collect enough data to account for variability in the population, hopefully without over-taxing your resources, then you may do fine.  For qualitative data, perhaps you could account for variability by sampling until you have no new 'surprises' - but I worked with continuous data and don't have any other suggestions for qualitative data.  

I also did not work with likert-scales, and they would not be in typical survey sampling books of which I am aware, but I suppose that they would be a variation on yes/no data and covered somewhere.

You might consider starting with a small pilot study.

Best wishes - Jim

Eric

Thank you.

The argument for Snowball Sampling would be that academics=researchers therefore it is least likely that they will pass the link to non-UK-based academics, which may eliminate one of the main disadvantages of SS, even though I am not sure how to prove it. However, if for instance, I will email an academic of my choice in every university in the UK and ask them to share the link with a colleague from OTHER department, it is still non-probability sampling as it will be the academic's own choice on whom to sent the link to. It seems obvious that the academic would send the link to his/her friend to make sure they would complete the questionnaire. Please feel free to critique this point. Thank you.

Ramnath, thanks for your response.

Would it reduce the 'validity' or 'representativeness' of my sample frame to be used either for Random or Stratified Sampling, if I could collect 90% ,I assume, of emails of each university?

Thank you.

Perhaps you should read

 http://www.ncbi.nlm.nih.gov/pmc/articles/PMC1705515/

"Variance Estimation, Design Effects, and Sample Size Calculations for Respondent-Driven Sampling," 

and see what it says about sample sizes.

Jim

Thank you very much for your exhaustive response.

I have finally read the article as got to a computer. I found it not easy, this is not my field. However, what I have gathered is that if I double my sample size to more than 200 which is quite possible and much easier than to obtaining all names/emails, then I could employ Respondent Driven Sampling. Please correct me if I got it wrong.

What remains unclear is that if I go for Stratified Random Sampling having an incomplete list of names (approx.90%) plus an expected nonresponse problem, would the sampling still be representative. I have no knowledge on it.

I think I should add some more information about the study (quantitative). It aims to measure the factors that most influence the utilisation of IT by academic staff during education process. There are 4 Independent Variables (factors that influence the usage of IT), 4 Moderating Variables (age, gender, experience, voluntariness of use) and one Dependent Variable (Use Behaviour).

Thank your for your help.

Maksym -

Whether you use 100 in a simple random sample, or 200 in a response driven sample, to perhaps obtain about the same variances (or I suppose 'variabilities' for qualitative data? - not my area), you still do not know if the variance is acceptable, or perhaps better than needed.  (My guess on response driven sampling is that it also may still retain appreciable bias without a rather large sample.) 

I think a pilot study may be in order. You could first decide which method you would rather try, and do a small effort to see what obstacles you might encounter. 

By the way, I worked on official surveys where establishments had to respond, though there were a few problems.  But from what I have heard, a 90% response rate on a survey of your type (not establishments, and not mandatory) is likely considered quite good.  For response driven sampling, you could add to bias by unknown nonresponse issues, I would imagine. 

But whatever method you'd like to try, for quantitative data, you should report standard errors and all metadata (saying how you collected your data, your analyses, problems encountered, methodology in general).   Reporting metadata would be in order for any kind of survey for any kind of data. 

A great deal may also depend upon the kind of data you collect.  (By the way, sample size "calculators" you see on the internet generally fail to tell you that they are only for yes/no data, simple random sampling, and generally for a worst case regarding standard deviation, and without a finite population correction factor, which could sometimes mean they would tell you to use a sample size that is bigger than the population size.) 

I suppose either method might work for you, but a pilot study might possibly be very helpful. 

Cheers - Jim

Dear Maksym

Systematic sampling is a more commonly employed method for sampling unknown population size, I hope the attached file will be useful for you.

Good Luck

you should use stratified sampling 

No doubt you should go for stratified random sampling, by doing stratified random sampling things become simpler for you. Make strata according to the type of university i.e. may be some are central, some are state and some are deemed university.

Then you decide the number of university to select from each strata by PPS. And after that you select your desirable sample size from that.

In this your sample will be representative and provide a good unbiased result.

Thanks

I hope it helps.