In the applications above  we reasoned that  it was unlikely to be many if any repeats on the same person as these were sample surveys of a few thousand people from each of a set of  countries, If you have some idea of the sampling fraction and frame you could make these sorts of argument.  Clearly the more time varying the dependent outcome the more like independent  samples you have.

In practice I would probably pool them and put in a dummy for each cross-section and through interactions between predictors and these dummies see the extent to which the relationship with the outcome changes through time.

The problem with treating repeated measures as if they were independent data points is that you underestimate the variance and you get too narrow confidence intervals. Do you have a guess of the percentage of repeated responders? If it is low, the underestimation of variance will be low and combining data may be acceptable. If you do not know, it would be safer, in my opinion, to run three separte models.

It all depends what you are repeatedly measuring - eg if there are higher level units - eg counties or countries that you know to be repeatedly measured - you can treat them as a level in a multilevel model. 

Have a look at structures and classifications here

http://www.bristol.ac.uk/cmm/software/mlwin/mlwin-resources.html#mlmbkgrd

you may also be interested in a colleague's paper 

https://www.researchgate.net/publication/259638850_Two_Multilevel_Modeling_Techniques_for_Analyzing_Comparative_Longitudinal_Survey_Datasets

and one of mine that analyses repeated cross-sections in this way  

https://www.researchgate.net/publication/270103441_Social_Policy_and_Economic_Development_in_Europe_Investigating_the_Macro_Determinants_of_Individual_Health_and_Well-being_in_a_Multilevel_Multivariate_Analysis_of_Thirty_European_Nations

Article Two Multilevel Modeling Techniques for Analyzing Comparative...

Conference Paper Social Policy and Economic Development in Europe: Investigat...

Thank you so much~ we cannot tell the proportion of repeated responders from what we have so far.

I was wondering, even if the three survey contain some proportions of same respondents, but the information of these respondents changes with time (some characteristics of the individual are different at different point of time), will it matter if I pooling them together and run one model on it?  

Good question, me too can I get the answer?

In the applications above  we reasoned that  it was unlikely to be many if any repeats on the same person as these were sample surveys of a few thousand people from each of a set of  countries, If you have some idea of the sampling fraction and frame you could make these sorts of argument.  Clearly the more time varying the dependent outcome the more like independent  samples you have.

In practice I would probably pool them and put in a dummy for each cross-section and through interactions between predictors and these dummies see the extent to which the relationship with the outcome changes through time.

Thank you, Professor Kelvyn Jones! Your answer helps!

The reason I want to pooling the three surveys is to capture the vehicle price variation among the three years. I cannot extract the price information due to lack of vehicle purchase year data from only one survey. So I plan to generate the average vehicle price of the three years separately and give the three years a categorical variable (as low, middle and high price). In this way, I might be able to get the relationship between price (of course some other variables are at issue here) and vehicle ownership? I am not quite sure about this. Pls kindly advice.  

I am sorry but I do not understand what you are trying to do and what information that you have. Can you please be clearer eg what does generate mean? what does give the three years a categorical variable mean?; where does ownership fit in?

Sorry.

Much appreciated for your help :)

So what I want to do is to examine the influence of car price (not the vehicle market price, but the money you need to pay for the right to own a vehicle regulated by local law) on car ownership. However, the survey does not report the car price but only car ownership (how many car does the person has). In my case, car price depends on which year the vehicle was purchased. However, we do not have the year information from the survey itself. Hence, to capture the variation of car price among vehicles, we are considering to pool the three surveys (2004, 2008 & 2012), and get the average car price for each year. And then assign the three average prices to each vehicle, which means that we will have three values for car price in the model.

Hope it clears...

If I understand you correctly,  you wish to model a binary variable (whether a subject in a survey owns a car). Your explanatory variable is the "car price", a tax which depends on the year of purchase, but you don't have this information. If you are using the results of a survey to answer a question for which it was not originally designed, I think you will have to accept that it might not be possible. You may have to find another source of information, or even change your question.

Hi, Miland Joshi, your understanding is right! And I agree with you, if the survey is not designed for that purpose, maybe I should not manage it in that way. Thanks for your suggestions.

I too agree with this - in comparison to these problems your potentially repeated measures issue is as nothing!

What was the research question? Is it about individual change or change in the marketplace? Questions about unemployment over time, for example, use both single sample approaches and population studies to look at change over time. If you our looking at the latter, then the methods of Economists or market researchers might be helpful. It Sounds like your study fits with market research, not growth curve models unless you see markets as organisms. Outside my expertise.

o.o

Grace, you said this in one of your earlier responses:

So I plan to generate the average vehicle price of the three years separately and give the three years a categorical variable (as low, middle and high price). In this way, I might be able to get the relationship between price (of course some other variables are at issue here) and vehicle ownership?

   Saying back what you said here, you want to know how the price of a vehicle impacts vehicle ownership in the United States.  But you have no idea how much any individual person paid for his/her car, so you are planning to aggregate the cost of all vehicles in the US at each time point of the survey and pretend that is the cost that every person in that time-point was looking at spending for a car.

   Based on statista.com, the average price of new and used vehicles bought or leased in the US in 1990 was $8691, in 2000 was $12469, and in 2010 was $13105.  But the range of the price of a new car in 1990 in the US was over $50,000, and was higher in 2000 and 2010.  Your theory in addressing this question in this way is that one number, the average, is a valid number for everyone considering a car purchase in that year.

   This is a classic level-of-modeling error - you are attaching the information of a group (the US) to the persons within the group as though everyone experiences it in the same way.  You might be able to get by with it in a dissertation but no journal will publish it.

So, a different question to you:  Does your survey data include the individual's income?

    If so, you can recast your model as the relationship between income and car ownership within the changing context of how much vehicle costs have changed.  In this type of analysis, the average cost of a vehicle is a context condition, not a predictor.

   If I have misunderstood your question, I apologize.  Just trying to wrap my head around what you are doing.

Thank you so much, Julia! You addressed my question! What bothers me so far is whether it is methodologically right to use the average price as an indicator for car ownership. I know it is a super coarse way.

Luckily, the survey includes individual's income. You said we can treat the average cost of a vehicle as a context condition. Do you mean we exclude the vehicle price from our model but consider it as a factor when interpreting the results? Thanks again...

Actually, you would use it as a categorical variable, and see if the relationship between income and car ownership differs significantly in each condition group.  You can do this without pooling the data, just comparing the coefficient from each logistic regression.  Or, you can pool the data and test it as a grouping condition.

Julia, Many thanks for your advices. I got your point. :)