Does adding audits produce less biased data? My team made a survey about sexual behaviors among college students in China, and the result looks unreasonable1, so we suspect that many of the participants aren't serious when filling out the questionnaires.
I'm considering adding audits in the questionnaires. For example,

On average, how often do you have sex with others?
  
  
*
  
*never
  
*twice a year or less
  
*more than twice a year, less than or equal to once a month
  
*more than once a month, less than or equal to once a week
  
*more than once a week, less than or equal to once a day
  
*more than once a day, less than or equal to once an hour
  
*more than once an hour
  

Since it's practically impossible for one to have sex more than once an hour, if a participant chose more than once an hour, we'll mark their response as invalid and ignore it when performing analysis.
Or maybe a more generic one like:

Monnow Bridge (Welsh: Pont Trefynwy) in Monmouth, Wales, is the only
  remaining fortified river bridge in Great Britain with its gate tower
  standing on the bridge. It crosses the River Monnow 500 metres (1,600
  ft) above its confluence with the River Wye. Please choose Disagree below.
  
  
*
  
*Agree strongly
  
*Agree
  
*Disagree
  
*Disagree strongly
  

If a participate chose anything other than Disagree, it indicates that they wasn't paying attention when taking the survey, and thus their response would be filtered out.
My question is: does adding such audits produce less biased data? if so, will it have a large impact? also, will it introduce extra noise?
1: I'm not going to disclose the accurate figures, but it looks like "0% of the participants had sex between once a week and once a day, while ~5% of them had sex more than once a day". Also note that the Chinese society is quite conservative.
 A: Extra noise...no. Throwing away data from people who clearly aren't taking the survey seriously certainly won't make things any worse.
I can envision some very strange circumstances where this could add bias, so it's impossible to say that it wouldn't. For example, if people who answer questions unrealistically make up a disproportionate number of those who would answer a specific way if they were answering truthfully, then you might be biasing against this response. However, given that the alternative is to accept answers that are almost definitely nonsensical, there was likely much more bias to begin with before you discarded these answers. So you can expect to reduce the bias, even if you can't guarantee it.
However, the fact that you've collected data, examined the results, and then decided to re-run the experiment can be a source of bias. Exactly what distribution of responses would you have needed to believe that your survey was valid? If you think there's no conceivable way that 5% of the participants could respond a certain way, you're introducing a prior assumption that may not be valid. It would be best to include the audit questions the first time, rather than adding them in after you've already reviewed the first round of responses.
A: Unless the objective of your survey is finding out about lying behaviours, adding any "audits" or lie detectors is a really bad idea. It is okay to ask knowledge questions, but do not encourage lying behaviour.
Avoid formulating weak questions, even if they are meant to be "audits". In your second example, the attentive respondent will follow exactly your advice and disagree because that's what you've asked for. In the first example, the last option does not make any sense at all, for that would leave no time for anything but ...
In addition, you receive a "controlled" garbage response, now what?


*

*ignore the response in your study: very dangerous strategy. As soon as the proportion becomes significant, the study is no less flawed than without the "audit".

*impute it with a more plausible value, sure, but how? If you don't know what causes the bad item response, it is very difficult to impute it without introducing another bias;

*disclose it in your study: the preferable option, but hopefully the phenomenon is rare enough to avoid the readers from concluding the problem was the survey itself;


Therefore, the only viable solution consists in studying your topic well and eventually come up with a set of good questions. Test the questionnaire before-hand on a few subjects (e.g. a sample of 10-20) and then learn from that test run to revise the questionnaire accordingly. Seek feedback from peers who ran similar surveys as well. Finally, run your full data collection.
Should any weird significant response phenomenon arise, you should always ask yourself:


*

*what characterizes those people who provided the bad answers? If there's a pattern, you can resort to imputation;

*could the formulation of the question and the response items be improved?

*is the topic overly sensitive? May this lead a certain subsample not to respond at all to either question or survey?

*is the questionnaire too long? 

*is the survey addressed to the right audience? how to ensure that those who respond are actually from the population you are studying?

*who did not respond to the survey and what are the potential reasons?

*is the sample design and the collection mode appropriate?


Never assume that the questionnaire and the survey design would not introduce bias.
