Entering panel-data (cross-sectional time-series data) into SPSS for regression

[fazzmo]

I am kind of desperate as I dont know how to enter pooled data (I have data about 42 countries over 7 years per variable) into SPSS and then run a simple regression.

So far SPSS does not assign the observations to a certain country and year and rather treats each variable as independent. How do I tell SPSS that I have 42 observations (one for each country) for 7 years per variable and not 294 (7*42) observations?

This is how I arranged it so far:



Do I need to create some Interaction variables (Time*Variable)?

Once I manage to enter this correctly, 4 independent variables (Ease, BusFree, PropRight, Corrup) make up one latent variable which is supposed to have an effect on one dependent variable. I will use SPSS Amos to perform a confirmatory factor analysis and to run a regression, if that's possible.

I would appreciate your help very much as I've been trying to figure this out for days :/ I just need a simple way to get some legitimate results, it does not have to be highly scientific.

Thank you so much in advance.

[Bruce Weaver]

With 42 countries, I would guess you want to treat country as random rather than fixed.  Assuming your "simple regression" has a continuous outcome variable, take a look at the MIXED procedure, which allows you to run random intercept & random coefficient models.  If you are unfamiliar with these types of models, I recommend Jos Twisk's introductory book as a good starting point.  One of the chapters includes examples of how to run the models he describes in various packages, including SPSS.

http://books.google.ca/books/about/Applied_Multilevel_Analysis.html?id=N5nCQgAACAAJ&redir_esc=y

HTH.

fazzmo wrote

I am kind of desperate as I dont know how to enter pooled data (I have data about 42 countries over 7 years per variable) into SPSS and then run a simple regression.

So far SPSS does not assign the observations to a certain country and year and rather treats each variable as independent. How do I tell SPSS that I have 42 observations (one for each country) for 7 years per variable and not 294 (7*42) observations?

This is how I arranged it so far:



Do I need to create some Interaction variables (Time*Variable)?

Once I manage to enter this correctly, 4 independent variables (Ease, BusFree, PropRight, Corrup) make up one latent variable which is supposed to have an effect on one dependent variable. I will use SPSS Amos to perform a confirmatory factor analysis and to run a regression, if that's possible.

I would appreciate your help very much as I've been trying to figure this out for days :/ I just need a simple way to get some legitimate results, it does not have to be highly scientific.

Thank you so much in advance.

[fazzmo]

Thank you so much for clarifying. My outcome variable is indeed continuous.

So there is no way to take all the data into account at once? For instance, by creating interaction-terms?

I spent quite some time on gathering this data, would be too bad if a lot of these efforts have been in vain.

[Bruce Weaver]

I don't know what you mean by "take all the data into account at once".  But models run via MIXED certainly can include interaction terms.  See the Twisk book I suggested, for example.  I would not advise plunging into these models before doing a little background reading.  GIGO is likely if you do.

fazzmo wrote

Thank you so much for clarifying. My outcome variable is indeed continuous.

So there is no way to take all the data into account at once? For instance, by creating interaction-terms?

I spent quite some time on gathering this data, would be too bad if a lot of these efforts have been in vain.

[David Greenberg]

You have entered the data in what is called the "long form." For many
kinds of statistical analyses, this is a good form; for others,you
will want data in the "wide form." That form would have one row for
each country, and would treat each year's observations as distinct
variables. The problem if you want to "run a simple regression" is
that doing so implicitly treats each observation as independent, when
observations of the same country at different time points are not
independent. There are ways of taking this into account. You need to
read up on panel data methodology to find out what they are and how to
implement them. David Greenberg, Sociology Department, New York
University

On Sat, Dec 1, 2012 at 7:33 AM, fazzmo <[hidden email]> wrote:

> I am kind of desperate as I dont know how to enter pooled data (I have data
> about 42 countries over 7 years per variable) into SPSS and then run a
> simple regression.
>
> So far SPSS does not assign the observations to a certain country and year
> and rather treats each variable as independent. How do I tell SPSS that I
> have 42 observations (one for each country) for 7 years per variable and not
> 294 (7*42) observations?
>
> This is how I arranged it so far:
>
> <http://spssx-discussion.1045642.n5.nabble.com/file/n5716565/Dataset.jpg>
>
> Do I need to create some Interaction variables (Time*Variable)?
>
> Once I manage to enter this correctly, 4 independent variables (Ease,
> BusFree, PropRight, Corrup) make up one latent variable which is supposed to
> have an effect on one dependent variable. I will use SPSS Amos to perform a
> confirmatory factor analysis and to run a regression, if that's possible.
>
> I would appreciate your help very much as I've been trying to figure this
> out for days :/ I just need a simple way to get some legitimate results, it
> does not have to be highly scientific.
>
> Thank you so much in advance.
>
>
>
> --
> View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Entering-panel-data-cross-sectional-time-series-data-into-SPSS-for-regression-tp5716565.html
> Sent from the SPSSX Discussion mailing list archive at Nabble.com.
>
> =====================
> To manage your subscription to SPSSX-L, send a message to
> [hidden email] (not to SPSSX-L), with no body text except the
> command. To leave the list, send the command
> SIGNOFF SPSSX-L
> For a list of commands to manage subscriptions, send the command
> INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[MR]

Hi Team,

I have simple query but it is turning out to be very complex for me. I am using AMOS to run SEM model.

Issue: I want to apply gender weight as my data is not aligned with population distribution.
Model: AMOS Structural Equation modeling
Variable: Gender weight - male 1.3; female .7.

In AMOS there is no place where I can use case weights. I realize I need to generate covariance matrix but how would i do it based on just one weight variable? Are there any other alternatives?

Any help is greatly appreciated.

Manmit

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Ryan]

Manmit,

 

What type of model are you fitting, and for what purpose? Are you certain you want to weight cases based on the population distribution? Often, within a structural equation modeling framework, one wants to demonstrate specific type(s) of equivalence (factorial, latent means, causal structures) depending on the model and intent, across populations of interest (e.g., males versus females).

Ryan

On Mon, Dec 3, 2012 at 8:50 PM, MR <[hidden email]> wrote:

Hi Team,

I have simple query but it is turning out to be very complex for me. I am using AMOS to run SEM model.

Issue: I want to apply gender weight as my data is not aligned with population distribution.
Model: AMOS Structural Equation modeling
Variable: Gender weight - male 1.3; female .7.

In AMOS there is no place where I can use case weights. I realize I need to generate covariance matrix but how would i do it based on just one weight variable? Are there any other alternatives?

Any help is greatly appreciated.

Manmit

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[MR]

Hi Ryan,

I am building customer satisfaction>loyalty model on attitudinal factors. Unfortunately we have to work with client data which is skewed towards females, incorrectly (data capture issue). 

Additionally, we also have to weight the data by spend variable from transaction data. I tried AMOS manual and user guide but could not find any reference on how to do this. 

M 

On 2012-12-03, at 9:48 PM, R B wrote:

Manmit,

 

What type of model are you fitting, and for what purpose? Are you certain you want to weight cases based on the population distribution? Often, within a structural equation modeling framework, one wants to demonstrate specific type(s) of equivalence (factorial, latent means, causal structures) depending on the model and intent, across populations of interest (e.g., males versus females).

Ryan

On Mon, Dec 3, 2012 at 8:50 PM, MR <[hidden email]> wrote:

Hi Team,

I have simple query but it is turning out to be very complex for me. I am using AMOS to run SEM model.

Issue: I want to apply gender weight as my data is not aligned with population distribution.
Model: AMOS Structural Equation modeling
Variable: Gender weight - male 1.3; female .7.

In AMOS there is no place where I can use case weights. I realize I need to generate covariance matrix but how would i do it based on just one weight variable? Are there any other alternatives?

Any help is greatly appreciated.

Manmit

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Ryan]

M,

I still do not see why you require a weighting scheme. If I were tasked to build and assess the fit of a path/causal model linking customer satisfaction (a unidimensional construct measured using multiple manifest variables?) to customer loyalty (a unidimensional construct measured using multiple manifest variables?), and I was concerned about gender differences, I would assess model fit for females, and test for model equivalence on data collected from males. 

I rarely use weights. The last time I applied weights was in an attempt to equalize non-randomized groups (hoping that the variables, e.g., gender, could NOT predict the condition to which they were assigned), so I'm not the best person to address your concern.

I can appreciate how large national surveys apply weights to ensure that the rates (whatever they may be) are based on a sample that is representative of the larger population. Having said that, I am NOT a survey methodologist, and so my understanding of such methods is admittedly superficial. 

At any rate, back to your model, I would be concerned that my causal model is equivalent across client populations, leading me to the approach suggested previously. 

Finally, yes, perhaps spending/transactions should be incorporated into the model, but I would take a different tact given the lens through which I view model building and testing in structural equation modeling. 

Perhaps someone with more experience with applying weights in structural equation models can chime in at this point. 

Ryan

On Dec 3, 2012, at 9:52 PM, MR <[hidden email]> wrote:

Hi Ryan,

I am building customer satisfaction>loyalty model on attitudinal factors. Unfortunately we have to work with client data which is skewed towards females, incorrectly (data capture issue). 

Additionally, we also have to weight the data by spend variable from transaction data. I tried AMOS manual and user guide but could not find any reference on how to do this. 

M 

On 2012-12-03, at 9:48 PM, R B wrote:

Manmit,

 

What type of model are you fitting, and for what purpose? Are you certain you want to weight cases based on the population distribution? Often, within a structural equation modeling framework, one wants to demonstrate specific type(s) of equivalence (factorial, latent means, causal structures) depending on the model and intent, across populations of interest (e.g., males versus females).

Ryan

On Mon, Dec 3, 2012 at 8:50 PM, MR <[hidden email]> wrote:

Hi Team,

I have simple query but it is turning out to be very complex for me. I am using AMOS to run SEM model.

Issue: I want to apply gender weight as my data is not aligned with population distribution.
Model: AMOS Structural Equation modeling
Variable: Gender weight - male 1.3; female .7.

In AMOS there is no place where I can use case weights. I realize I need to generate covariance matrix but how would i do it based on just one weight variable? Are there any other alternatives?

Any help is greatly appreciated.

Manmit

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Mircea Comsa]

Hi,

One possibility is to use as data imput a variance-covariance matrix.
Weight the data (in spss or ...) then compute the var-cov matrix. You can
use the syntax:

CORRELATIONS
  /VARIABLES= x1 x2 x3 ...
  /PRINT=TWOTAIL SIG
  /STATISTICS DESCRIPTIVES
  /MISSING=LISTWISE
  /matrix out ("C:\...\cov_matrix.sav").

All the best,
Mircea

> Hi Team,
>
> I have simple query but it is turning out to be very complex for me. I am
> using AMOS to run SEM model.
>
> Issue: I want to apply gender weight as my data is not aligned with
> population distribution.
> Model: AMOS Structural Equation modeling
> Variable: Gender weight - male 1.3; female .7.
>
> In AMOS there is no place where I can use case weights. I realize I need
> to generate covariance matrix but how would i do it based on just one
> weight variable? Are there any other alternatives?
>
> Any help is greatly appreciated.
>
> Manmit
>
> =====================
> To manage your subscription to SPSSX-L, send a message to
> [hidden email] (not to SPSSX-L), with no body text except the
> command. To leave the list, send the command
> SIGNOFF SPSSX-L
> For a list of commands to manage subscriptions, send the command
> INFO REFCARD
>

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Molte]

Fazzmo: Have you become any wiser in regards to your inquiry?

I'm in a similar position as you and came across this thread while trying to figure out how to analyze a cross-sectional-time-series data set in SPSS for my bachelor’s thesis.

I'm interested in the potential effect a dichotomous variable on a continuous dependent variable in 47 U.S. states over a period of 17 years (really 16, the first will be discarded since I'm using lagged variables) I will also control for state and time specific fixed effects. Most interesting now is however to find out how to do a simple regression.

I would really appreciate if anyone could provide "tutorial help"?

Best regards,

[fazzmo]

First of all, thanks for your advice, David.

I transformed the data into a wide format to take account of each variable per year.

I read about transition models and (cross-)lagged panel-data and I think this is the right way.
However, my model is complicated, that it seems very difficult to put into practice.

To put it simple, 4 items for which I collected data for 7 years (2005-2011) and for 42 countries constitute one latent variable. In an attempt to incorporate all the causalities, I understand that it should look like this for "one" latent variable.



However, I do not know whether this is any right, nor do I know how to go on, as I have four latent variables and these four latent variables are supposed to have an effect on a certain dependent variable.

[Maguin, Eugene]

You're at a kind of methodological crossroads. As others have pointed out a cross-sectional time series model has a panel model structure. You have an N of 42. You assert that the four items form a latent variable. The normal way to test that proposition, as well as the hypothesized structural relationships, is through a latent variable structural equation model (SEM), which can be implemented in AMOS, as well as other software. It can not be done in spss.  However (and an extremely important however), with a sample size of 42, you can't do it. Why? Your covariance matrix is 28x28. 406 non redundant elements. Your model shown in your figure estimates 28 residual item variances, 21 (first order) factor loadings, 7 (second order) factor loadings, 7 residual factor variances, 6 regression coefficients, at least 7 residual factor covariances and 1 factor variance. 77 thus far. It's very likely that you'll need residual covariances between, at a minimum, adjacent corresponding it!
 ems. That's 24+77=101. The degrees of freedom will be 305, if I've counted and added right. The default estimation will be by maximum likelihood (ML), not least squares as spss regression uses, and my understanding is that ML requires more cases than parameters being estimated.  Some programs allow other types of estimation models; Amos may do so but I don't know.

Gene Maguin








-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of fazzmo
Sent: Wednesday, December 05, 2012 10:32 AM
To: [hidden email]
Subject: Re: Entering panel-data (cross-sectional time-series data) into SPSS for regression

First of all, thanks for your advice, David.

I transformed the data into a wide format to take account of each variable per year.

I read about transition models and (cross-)lagged panel-data and I think this is the right way.
However, my model is complicated, that it seems very difficult to put into practice.

To put it simple, 4 items for which I collected data for 7 years (2005-2011) and for 42 countries constitute one latent variable. In an attempt to incorporate all the causalities, I understand that it should look like this for "one" latent variable.

<http://spssx-discussion.1045642.n5.nabble.com/file/n5716684/Dataset2.jpg>

However, I do not know whether this is any right, nor do I know how to go on, as I have four latent variables and these four latent variables are supposed to have an effect on a certain dependent variable.



--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Entering-panel-data-cross-sectional-time-series-data-into-SPSS-for-regression-tp5716565p5716684.html
Sent from the SPSSX Discussion mailing list archive at Nabble.com.

=====================
To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[fazzmo]

Thank you Gene,

your statement that this is most likely not going to work is very helpful. Especially in light of my rather basic understanding of all this (I am business student trying to finish his Master Thesis ;)). Looks like I should talk to my supervisor about this. Thank you, once again.

Nils