Hi, if you enrolled all accessible (and participating) persons during a defined period in time, you applied consecutive sampling (or at least convenience sampling, if you too narrowly restricted the events at which you recruited your participants). Get more information on this webpage: https://explorable.com/non-probability-sampling

Regards

Hi, if you enrolled all accessible (and participating) persons during a defined period in time, you applied consecutive sampling (or at least convenience sampling, if you too narrowly restricted the events at which you recruited your participants). Get more information on this webpage: https://explorable.com/non-probability-sampling

Regards

Then I can not do inferential statistic which get from non probable sampling! right?

You are correct that you do not have a probability basis for using inferential statistics. Effectively, you appear to have a census of results from something that could be described as a single event or case, as in "case study." Tables of summary statistics could be part of an effective presentation of your research. Means, medians, standard deviations, and frequencies are all potentially useful summary statistics. There are usually a few different ways to describe studies correctly. The main thing is to describe the study in sufficient detail for your peers to understand what you have done.

http://psychology.about.com/od/types-of-research-studies/

Hi, I would say yes and no regarding inferential statistics. Of course, you cannot refer to the underlying total population, since your sample is not randomly pulled from it (selection bias because of your personal selection of events). However, your sample can be representative for people participating in these events. So, if you apply inferential statistics, you have to set up your statistical models using sound cofactors (as f.e. age, sex, education, etc.) with a theoretical background, you have to disclose how sampling was applied and to what extend generalizations regarding the underlying total population can be derived or not from your results. (see f. e. http://www.aapor.org/AM/Template.cfm?Section=Reports1&Template=/CM/ContentDisplay.cfm&ContentID=6210 ).

In fact, a plethora of clinical studies was done enrolling f.e. patients consecutively frequenting a hospital ward. Of course, this selection bias has to be disclosed. However, since their sample might be somehow representative of all patients frequenting this special ward, valid conclusions can be drawn.

Helmuth,

If Marjan has a basis for using inferential statistics, I haven't seen it presented here. There are two related rationales for the use of inferential statistics. The first is random sampling. The second is randomization of treatments to subjects where the subjects have not been selected at random (experiments). Most of the clinical studies you reference are randomized. The others get to present their arguments as to why they think inferential statistics are appropriate and I might even be understanding in some cases. If it's not physically possible to randomize then you do what you can and explain why you did it. However, please understand that if you didn't sample at random and you didn't randomize then there is no statistical theory to support your use of inferential statistics.

Chuck

Dear Chuck,

thank you for your remarks. Your are of course right, from a theoretical point of view. Nevertheless, inferential statistics are permanently applied on non-probabilistic samples in clinical science and the results are published. Of course, if you aim to compare diagnostic or therapeutic options to a control group in clinical trials, you have to randomly allocate your participants to the groups. In contrast, in various cohort studies, you follow up a distinct group of persons (f.e. a consecutive sample) longitudinally over a certain period, subgroup them according to genotypes or phenotypes (not randomly) and compare incidences between these groups by means of inferential statistics. Systematic (random) sampling (f.e. enroll every 5th of consecutive patients) is often not applied, since they - especially when studying rare diseases - would not be able to reach a sufficient sample size.

Of course, the results can not be generalized to the total population, so the target population, to which the sample is representative, has to be found and described (f.e. elderly, middle european people visiting a university hospital's angiological outpatient ward) and this could be difficult. On the other hand, the bias can be decreased by multicentric studies or the found biomarkers can be evaluated in the course of replication studies enrolling independent samples and subsequent meta-analyses. However, all of this has to be disclosed.

Regards, Helmuth

Regarding Marjan's design: There is not enough information... what is the aim of your study? Which hypotheses do you want to test?

The three types of non-random samplings are:

1. A Systematic Sample is a sample obtained by selecting every nth name in a population

The sampling interval is the distance in the list between each of the individuals selected for the sample

The sampling ratio is the population that is selected for the sample

2. A Convenience Sample is any group of individuals that is conveniently available to be studied

Are not considered representative of the population and should be avoided, if possible

3. A Purposive Sample is a sample selected because the individuals have special qualifications of some sort, or because of prior evidence of representation

Personal judgment is used for selection purposes

A major disadvantage is that the researcher’s judgment could be in error.

In my opinion yours should be Convenience Sample.

Perhaps a convenience sample, from which any conclusions would be rather subjective. But after reading Marjan's question above, I'd say hopefully close to a census for each event, and perhaps comparisons could be drawn between events, though I don't know that I would call them inferential. So then I saw this from Chuck, above, which I agree sounds like good, practical advice. Chuck said: " Effectively, you appear to have a census of results from something that could be described as a single event or case, as in 'case study.' Tables of summary statistics could be part of an effective presentation of your research." -

Usually in a census, when there is nonresponse, there may be something one can do such as follow-ups on nonresponders, to see if you can get an idea as to how the nonresponders mean answers might differ from the ones who responded right away. But since these 'events' are over, that does not sound practical, unless there is a list of attendees, but my guess is that there wasn't. Sometimes data are arranged in "response propensity groups" so that each group can be weighted upward accordingly. But again, there may not be sufficient information for that. So I think that caveats should be written to help readers of your findings, Marjan, to know how close each event may have been to being censused, and one can then see descriptive statistics for each. Also, as I noted in your question on nonprobability sampling, within each set of questions, you might regress one response against another to find any relationships there. If you are lucky enough to find some direct relationships as simple as a ratio (y = bx + e, where e is a heteroscedastic estimated residual - not hard to look at - say a classical ratio estimator, where you can get a standard error for b, using statistical software, or even using an excel spreadsheet and standard error of b from various texts, or say page 10 from a paper in my area with "optional" and "weekly surveys" in the title, or the one starting with "properties" talking about weighted least squares regression), then you may find some useful results that way.

Statistical software

PS - If you use statistical software for regression, to compare responses to various questions, to see how they relate within each questionaire, the default is always OLS (ie, equally weighted data), but if you get a nice linear relationship between two questions, through the origin, which often happens with survey data, OLS is not logical. I'm sure various software can handle this, but I only know how SAS does this. You could use SAS PROC REG, and put in NOINT, and the regression weight w = 1/x to obtain a classical ratio estimator. I'm not aware of the software to which you might have access, but it would likely be similar, I think.

Marjan -

In your question, you have this: "We have this categorization: probable and non-probable sampling methods. The first includes simple random sampling, stratified, cluster and multiple stage sampling. The second one includes convenience sampling, and snowball. Which one is mine?" - As Chuck indicated, you may not want to refer to this as a sample at all. It may be more like a census of each event, with nonresponse. The less nonresponse the better. However, some "censuses" do have quite a bit of it.

OK. One more. :-)

Marjan, I see you have statistical software experience on your profile. You could use scatterplots to show relationships between questions for each group. And for a between events comparison, you could use histograms, comparing percents. Graphics are always good. Hopefully you have access to software with good graphics. There should be some good possibilities, such as the two examples just given.

Eish Marjan, hope everything goes well. I've just learnt that poor sampling can destroy the whole good study. Im in a similar kind of predicament since I need to collect data from two institutions and from there I am looking at students doing a certain qualification and doing a certain subject. Now, I am told that my study is purposive! which is not always good. Now, I know the main question to ask ... Why did I choose students doing students doing these subjects I chose and why not others. The answer there is exactly what need to be understood. Importantly, we need to know this information before we embark on the data collection exercise.