This course presents research conducted to increase our understanding of how data collection decisions affect survey errors. This is not a “how–to-do-it” course on data collection, but instead reviews the literature on survey design decisions and data quality in order to sensitize learners to how alternative survey designs might impact the data obtained from those surveys. The course reviews a range of survey data collection methods that are both interview-based (face-to-face and telephone) and self-administered (paper questionnaires that are mailed and those that are implemented online, i.e. as web surveys). Mixed mode designs are also covered as well as several hybrid modes for collecting sensitive information e.g., self-administering the sensitive questions in what is otherwise a face-to-face interview. The course also covers newer methods such as mobile web and SMS (text message) interviews, and examines alternative data sources such as social media. It concentrates on the impact these techniques have on the quality of survey data, including error from measurement, nonresponse, and coverage, and assesses the tradeoffs between these error sources when researchers choose a mode or survey design.
3.1.3 Proxy Responding
Loading...
来自 University of Michigan 的课程
Data Collection: Online, Telephone and Face-to-face
123 个评分
从本节课中
Module 3: Interviewers and Interviewing
In this lesson, we explore the various roles interviewers take on beside asking questions and collecting answers, as well as some of the different approaches to interviewing that have been proposed and how they affect the accuracy of responses. By the end of Lesson 3, you will appreciate the benefits and costs of collecting data in interviews and will be able to contrast them with the costs and benefits of self-administration.
与讲师见面
Frederick Conrad, Ph.D.Research Professor, Survey Methodology
Institute for Social Research
We've been talking about interviewer's non interviewing tasks.
Generally, conducted prior to the interview and
just been discussing selecting a house hold member to be the respondent.
Related selection issue concerns the use of proxy respondents for
the selected respondent.
So if the selection procedure identifies a household member who's not
available at the moment.
Some surveys or survey organizations will allow the interviewer
to conduct the interview with somebody who is available serving as
a proxy that is answering as if they were the selected respondent.
Other surveys of the organizations don't allow proxy respondent,
the general thinking is the data quality is worse but
there are many well-known organizations that do use proxies.
There are several advantages to collecting information from proxies.
It takes less time and therefore less expense if proxies are used when
selected respondents are not available.
For example, the current population survey, the main government labor force
survey in the US reduces its cost by about 12% by using proxies.
And use of proxies results in higher response rates simply because
if a selected respondent is not available
this doesn't lower the response rate if a proxy is available instead.
As I said, the general belief is that self reports, that is,
reports of respondents about themselves are better than proxy reports.
The response process is thought to differ,
probably does differ between proxies and respondents answering about themselves.
The general idea is that proxies are more likely to rely on.
General knowledge about the selected respondent
as opposed to on the basis of actual events.
So people are better at answering questions in terms of their actual
behavior when the questions are about themselves than about somebody else.
We tend to use general knowledge or
it has in this literature been called symantec knowledge or
dispositional knowledge about the intended or target respondent.
As opposed to answering about one's self in terms of the episodes or
events that have gone on in one's life.
This can lead to differences, proxies are more likely to use estimation than recall,
that is they're based on the general dispositions.
The general characteristics of the person they're answering about,
they're going to estimate the answers to questions requiring a number.
More likely to do this than to recall actual behaviors,
because they're not aware of the actual behaviors.
I mean really, nobody knows as much about anyone else as they know about themselves.
The dating information, when events are dated,
did an event occur within the reference period or not.
Less likely to be as accurate when a proxy is
thinking about the date of an event that
somebody else engaged in then in the event that they themselves have engaged in.
Self respondents, on the other hand, are more likely to use direct retrieval just
to recall the information necessary to answer the question.
So the answer's more likely to be based on actual events than general knowledge.
And whether one is answering about oneself or
a different household member is likely to affect the motivation to work hard in
providing the most accurate answer possible.
Proxy reporters may be less motivated
because they're not answering about themselves and it may be harder.
They have to use these estimation processes rather than just
directly retrieving information.
Proxies may be less subject to social desirability issues, so
let's say smoking is stigmatized for one household member.
If that person answers a Let say himself, are you smoking?
Or how many cigarettes have you smoked this week?
There might be a tendency to shade the truth to present
one self in a better light.
But if someone if asked, is your husband smoking or
how many cigarettes did your husband smoked last week?
That may not be as stigmatized or embarrassing for
the spouse answering about the husband.
And so, the answer actually may be more truthful from
a proxy when the topics are socially undesirable.
The accuracy of the proxy reports is likely to be affected by the amount
of communication between the selected or target respondent.
And the proxy, so, family members probably know more about each other,
and therefore proxies answering about family members probably likely
to be more accurate than, say.
Roommates or other kinds of living arrangements where
the people are not as well connected and
the level of communication is not as great as it is among family members.
A review paper by Jeffrey Moore in the 1980s
investigated the really assumption that self respondents, respondents
answering about themselves, provide better quality data than proxy respondents.
And he actually found very little evidence for this in his review of the literature.
He did point out that the literature then was really quite
flawed because number of the studies comparing quality of
answered between proxies and self respondents were observational.
They weren't experimentally, they weren't experiments so the respondents
were not randomly assigned to be self-respondents or proxy respondents.
Which is really what you'd need to evaluate the quality of the data from
the proxies.
Instead, what he observed was that what appears to be a difference in quality
may be a difference in fact, so ideally, the contacted household member.
in a methodological study would be randomly assigned to be a self or
proxy reporter.
But in practice, this can never be the case because proxies are used
when the selected respondent is not available.
And so most of the studies that more was revealing, were of this type,
they were observational studies, that is, they weren't conducted as experiments.
The finding that he focused on came from health studies,
in which proxies were shown to report fewer health events that is for
example going to the doctor or the emergency room.
For the household member they're reporting about or
that person is that person is not available.
Then they are to report about themselves and the idea was well the assumption had
been this indicates the proxies are not at giving us complete in thorough answers
about the unavailable but, it selected respondent as they are about themselves.
But, more observe that well, actually
if the selected respondent is not available they're not home.
A person who is available and can serve as a proxy is home.
And someone who's home is more likely to be home for
health reasons and be not as healthy as the person who's not available,
who's not at home, who may well be at work.
And so, while it appears that proxies may be say not giving a complete or
a thorough answers for the intended respondent as they can for themselves.
This is not necessarily the case it could just be that there are real differences
between the health proxies and the health of the people they're reporting about.
He also observed that most of these studies which were not done for
methodological purposes, had no validation data.
That is, no way to verify the accuracy of answers or
directly assess the quality of answers.
And so, they're generally not conclusive and so,
at least what he's suggesting is that proxy data might not be
any worse than the data from respondents reporting about themselves.
And he really argued for the need to do controlled experiments where household
members randomly assigned to be proxies or to report about themselves.
More recently,
some of the kinds of experimental studies more call for has been done.
So a study by Schwarz and Wellens is a good example of this.
It was based on the distinction between in social psychology between observers and
actors.
The idea that if observers are asked to describe actors behavior.
They'll rely more on the other person's underlying dispositions or
traits than actual situations or events, but
when responding about themselves, it'll be the opposite though.
They'll think more about situations or
events similar to what we've discussed about general knowledge versus episodes.
They built on the idea by Jones and Nisbett that actors has richer knowledge
of their own behavior and they tell a coherent story.
That is to find causal explanations for the events in their lives.
Based on this idea, Schwartz and Wellens predicted that
if participants are randomly assigned to be proxies or
to report about themselves that the self-reports.
Should reflect inconsistencies in behaviors apart from the usual
routines because somebody answering about their own life is going to be aware
of those departures from the norm, from regular activities.
Whereas, answering about someone else sort of requires answering on the basis of,
well they typically do this.
So I'm going to say that I'm going to base my answer on what I know about
the routines.
So to test this idea that there be a more inconsistency or
more variation in the answers from self reporters than from proxies.
They had 24 couples report either their own or
their partners typical behavioral frequency like how many times did you or
did your partner checkout library books from the university library.
Have them rate either their own or their partner's typical behavior or
frequency as well as the highest and
lowest weekly frequency of that behavior during the preceding month.
So the idea is, by looking at the range of the typical frequency,
they could see if self reporters actually have a larger range of highest and
lowest than proxy reporters.
And that's really what they found here, so here's the library example in the top row.
The range between the highest and lowest frequencies is larger for
self-reporters than for proxies.
Same for snacks, same for being late.
So basically, the difference between highest and lowest for
all these behaviors is Is larger for self reporters than for proxies.
And the idea, again, is just that proxies are not going to
be aware of the exceptions to the routine the way self reporters are.
The proxies are not with the people they're reporting about all day every day,
even though they're members of the same couple.
And so, have to rely on routines rather than actual behavior.
There is some indication that there are different thought processes or
different strategies use to answer questions by proxies versus
self reporters which could have a bearing on data quality.
So to summarize, what we've been talking about in the last
few segments actually the interviewers non-interview tasks.
Interviewers can affect for better or worse all of the sources of survey here.
Sampling, coverage, nonresponse and measurement.
When it comes to recruitment which has direct bearing on response rates and
nonresponse error potentially Interviewers vary in there success and
the manner of speech may matter.
When it comes to within household sampling the different methods that interviewers
use vary in their explicitness about the probability of selection.
Remember the birthday methods really can't tell
us what the probability of selections.
We don't know anything about the size of the household.
Their intrusiveness, the kitsch method required in enumerating all the males and
all the females and ordering them from oldest to youngest.
And the efficiency of the method, so for
example the kitsch method is probably the least efficient.
And with respect to self proxy issues the response processes used
by self reporters and proxy reporters seem to differ.
And cost non-response and measurement error all need to be taken
into account when deciding whether or not to use proxies in place of self reporters.
What we'll turn to next is the actual interviewing techniques that
interviewers use.
So, once they've completed all these kind of pre-interview,
non-interviewing activities.
It's actually time to conduct the interview and we'll look at some of
the different interviewing techniques and their consequences for data quality next.
