Re: st: Poll of polls

[Richard Ohrvall <richard.ohrvall@gmail.com>]

Casey, Clive, Eric and others,

thank you for your suggestions and comments.

More information about the data was asked for. The response variable
is voting intention shares for political parties. I am looking at
Swedish data, we have an election in September this year. This means
that there are many political parties in the data. The number of polls
per month vary over time, they get more frequent when we get closer to
the election. At this point in time, they are roughly one poll per
pollster and there are 7-8 pollsters. However, two of them are using
web panels, so one could argue that they should not be included. The
remaining are using RDD. We do not have robocalls, at least not yet.
With one exception they are using fairly similar wording in the
questions used.

The different pollsters have different number of respondents, so that
is one parameter to take into the calculation.

If you have some further thoughts, please let me know. If someone
wants to look at the data, I could clean it up and send it to you.

All the best,
Richard


2010/6/23 Clive Nicholas <clivelists@googlemail.com>
>
> Richard Ohrvall wrote:
>
> > I am sorry if this is a bit too unspecific, but I am currently looking
> > into what is usually called "poll of polls", i.e. techniques to take
> > estimates from different opinion polls and estimating time series. I
> > know, it is outside of established statistical theory, but I am
> > playing around with political opinion polls to look at different
> > methods to achieve time series. Some of the issues are a) how to
> > handle "house effects", i.e. that different pollsters systematically
> > diverge from others, b) how to smooth the data over time, e.g. some
> > sort of moving average. So, my questions are 1) if any of you have
> > seen anything done on this using Stata? 2) Do you have any
> > ideas/suggestions about the best way to tackle this (e.g. if -lowess-
> > is a path worth exploring)?
>
> [...]
>
> You should be clearer on what your response variable is. Is it voting
> intention shares for political parties? Is it party-leader popularity
> ratings? Are they measured monthly, quarterly or annually? Either way,
> these are measured on the 0-100 scale and can pose their own problems
> if their limits are reached. Analysed on their own on a house-by-house
> basis would, in my view, call for normalizing the scales and
> regressing them on your independent variables in a fractional logit or
> probit model, viz:
>
> glm y x, family(binomial) link(logit) robust
>
> However:
>
> (a) if your intention is to model 'polling house' effects on opinion
> shares across all polling houses and points of time simultaneously,
> then this would call for a pooled TSCS (or CSTS) approach with the
> polling houses included as fixed-effect dummies, so look at the family
> of -xt- models and pick the one most suitable for your data;
>
> (b) download Nick Cox and Kit Baum's -mvsumm- from SSC for creating
> moving-average variables;
>
> and
>
> (c) -lowess- could be used to smooth time-series graphs, but also
> check out Cox and Baum's -tsgraph- (SSC) or -help xtline-.
>
> (Much) more details on your data, which reads very interestingly,
> would help here.
>
> --
> Clive Nicholas
>
> [Please DO NOT mail me personally here, but at
> <clivenicholas@hotmail.com>. Please respond to contributions I make in
> a list thread here. Thanks!]
>
> "My colleagues in the social sciences talk a great deal about
> methodology. I prefer to call it style." -- Freeman J. Dyson.
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/