0
$\begingroup$

I will be collecting data from three three different groups. I will be interviewing women (probabilistic sample) and then asking them to provide a name to build the other two groups (convenience sample). I would like to know what is the best approach to analyze all this data?

My study is about: Breastfeeding support in the workplace. I already define three groups to obtain data from: 1. working mothers; 2. managers; 3. other employees.
Working mothers sample will be obtained from a database provided by a HR department in the academic setting I am working at. Then I will ask the mothers in the sample to recommend their immediate manager name and also the closest employee working with her (example: sharing office). Here comes my firts question: What is this type of sampling? Then I will be sending a separate self-administered questionnaire to the groups. A different type for women. Managers and colleagues will be receiving an almost identical one. Finally I want to know how to overlap the results in one analysis (if possible).

Hypothesis: Is workplace support influencing the breastfeeding intention and also the breastfeeding duration?

What is the best approach to integrate the three sets of data into one analysis? Is the proposed sampling ok?

$\endgroup$
9
  • 4
    $\begingroup$ Lots more information is needed. $\endgroup$
    – Glen
    Jun 8, 2012 at 3:48
  • 3
    $\begingroup$ This question is too vague. Please, provide more context about your design and the kind of data you expect to get. $\endgroup$
    – chl
    Jun 8, 2012 at 6:36
  • $\begingroup$ I don't think this is vague. He is applying snowball sampling and wants to know if he can still do ordinary survey sampling statistical inference given that some subjects are gotten by referral without randomization. I don't think the downvotes are justified. $\endgroup$ Jun 8, 2012 at 15:26
  • 1
    $\begingroup$ Is the sampling method ok? Which is the best type of analysis that I can conduct based on my hypothesis? How can I integrate all three sets of dat in one analysis. $\endgroup$ Jun 8, 2012 at 15:29
  • $\begingroup$ The update to the question suggests now that "snowball sampling" might not be involved here, @Michael, because the referrals appear to be to different populations and apparently (it's still a little vague) they will receive different questionnaires. Conceivably, then, the population consists of all 3-tuples (working women on maternity leave in 2011, manager, employee); sampling is random for the first component, but non-random for the second two. But I could be misinterpreting--which is why the question could still use some improvement. $\endgroup$
    – whuber
    Jun 8, 2012 at 15:58

2 Answers 2

4
$\begingroup$

At first glance this looks like snowball sampling, but snowball sampling usually proceeds to get more referalls from the refered subjects.

This could also be considered cluster sampling where a cluster is defined by a working mother, her manager, and her closest co-worker. You are sampling an entire cluster by choosing the mother. So you could use techniques based on cluster sampling. Or it could even be that each triplet is really your observational unit (if you are more interested in relationships between the 3 people than something like the average age of the managers) and could be treated as a simple random sample (but the data is the triplet).

The problem comes in that it is unlikely that the population of interest matches the population that you are sampling from. Any employees who are not the closest coworker to a mother have 0 probability of being sampled, any manager who does not supervise a mother has 0 probability of being sampled, and any manager who supervises more than 1 mother has a higher chance of being sampled than those that only supervise 1. The last part could be handled by weighting, but the under coverage in the other 2 could seriously bias any results.

$\endgroup$
2
  • $\begingroup$ Thank you so much Greg. That was very useful. What can you recommend me to do to avoid the under coverage. Should I use other sampling method for the managers and coworkers? I am really interested about the relationships and not much in the description of the managers and coworkers characteristics. $\endgroup$ Jun 8, 2012 at 20:22
  • $\begingroup$ As @Peter says in his response, if your focus is really on the relationships, then your sample is probably fine. In that case you are not interested in managers that don't supervise working mothers, so that undercoverage is not a problem (as long as you focus on the relationship and don't branch into manager characteristics). For the co-workers just make sure you have a clear way to define how you are getting them (or you could get a list of candidates and randomly select from that list). $\endgroup$
    – Greg Snow
    Jun 11, 2012 at 15:50
3
$\begingroup$

I think the sampling method looks completely appropriate, but that the discussion of "snowball sampling" is not helpful.

You really only have one sample - working mothers. They should be your unit of analysis. All the randomness comes from how they were selected. Among the attributes of working mothers are their manager and their closest peer, and the answers those people make to questions. But any statistical inference you do about those people should be based on the fact that originally it was the working mothers who were sampled.

For example, it would be legitimate to draw inferences to the total population of working mothers such as

  • "X% of working mothers did Y"
  • "X% of working mothers' managers thought Z"
  • "X% of working mothers had a closest colleague who did W"

But you couldn't draw any conclusions about managers in general, or workers in general, without a lot more information and some complex weighting systems. The issue is that managers can only be included in your sample if they manage a working mother; and the more working mothers they manage, the more likely they are to be in your sample; hence you would need to control for this factor if you want to infer about managers in general.

In summary - make sure whatever analysis you do has working mothers as the unit of analysis, and you will be ok.

$\endgroup$
10
  • $\begingroup$ She can analyze and summarize the data from the managers and coworkers but cannot infer anything about a more general population of coworkers and managers other than to the subset that is connected to working mothers. Initially I thought she was using the initial list to find more respondents. But in reality for each working mother selected there is a supervisor and coworker that she is connected to. $\endgroup$ Jun 8, 2012 at 21:14
  • $\begingroup$ Dear Peter: Thank you. Your comment clear my path. For writing the paper I just have to mention the mothers as the unit of analysis. Also, since I am trying to "measure" how the workplace support influences the initiation and duration of breastfeeding, what is the best analysis to perform with the mother set of data and with the data from managers and coworkers. Do I have to run independent test or I can integrate the three sets? $\endgroup$ Jun 8, 2012 at 21:31
  • $\begingroup$ @MichaelChernick - yes, good comment, that is what I was trying to convey. $\endgroup$ Jun 8, 2012 at 21:44
  • $\begingroup$ @WendyAlfaro - I'm not sure I understand the question, but I would definitely try to integrate the three sets. So analysis of co-workers and of managers only takes place explicitly linked to the working mothers who were originally sampled. $\endgroup$ Jun 8, 2012 at 21:45
  • $\begingroup$ Ok. Thank you to both of you. I have a clearer understanding now. But what type of analysis should I run after collecting the data in a integrative way (multivariate, regression, etc.). Should the questionnaires be exactly alike to be able to run the analysis? $\endgroup$ Jun 8, 2012 at 21:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.