how to perform regression on longitudinal data

[SPSSforMBA]

Dear Friends,

I am new to this forum. Sorry if I am posting a question which might have been already posted and answered.

I am performing statistical analysis for my MBA management project using SPSS.

In short, through this analysis, I am trying to determine the correlation of CreditCardFrauds (dependent var.) with GDP, NumberofCreditCards and InternetUsers (3 independent vars.) in 17 countries across 9 years.

With my initial reading of SPSS manuals and related text, I understood that due to the nature of the this data, the correct way to analyse this is using the MIXED-models from SPSS.

Problem I am facing is that when I Linear-Mixed-Models with Random effects explained in "IBM_SPSS_Advanced_Statistics.pdf", the SPSS never returns. I left it running for more than 6 hours, but it appeared to be hanging.

Can you please help me know if this is the correct model to use in SPSS for such kind of data? Does SPSS take long time to run analysis on such kind of data? (If required, I will share the command syntax that was run by SPSS and also a sample of my data).

Any help or guidance on this problem is highly appreciated. Thanks in advance for your time and guidance.

Best Regards.

[Gareth Edwards-7]

I am currently out of the office and will return Friday, Sept 20. Please e-mail Elizabeth McGaha at [hidden email] or Alicia Filson at [hidden email] if you need assistance.

[David Marso]

Yes, the syntax and a sample of the data would be useful.
Attach to this thread and maybe you'll get some takers.
Without this you are unlikely to receive much assistance if any!
Do not send it to my email!

[Rich Ulrich]

I don't know what is causing your problem.  The time is too long,
there ought to be a message, are you *sure*  that there isn't one?

Maybe, one message will say that you don't have enough data?
I will proceed on the basis of my guess at your modeling, but,
Yes, you probably need to post your basic syntax.

Look at what the problem is for a single country:
You have 9 years of data.  There is one DV (Fraud) and 5 IVs
(GDP; cards; net1 to net3).  Each of these numbers is highly
correlated with the previous year.

Can you make any simple sense from the data for one country?
 - I would suggest doing that, "making sense", for a few countries.
And work from there.
 - What are you hoping or expecting to learn? 
 - Do you expect to contrast countries, or collapse across them?

When you have a figure like GDP, sometimes it makes good
sense to convert *all*  the data to something that is per-capita.
Or convert all the other data to something that is per-GDP, and
drop GDP.

One direction of data reduction would be to find a formula
with two or maybe 3 parameters that describes each country; then
you perform regression with those three IVs, using the 17 countries.

--
Rich Ulrich

> Date: Thu, 13 Sep 2012 19:41:26 -0700

> From: [hidden email]
> Subject: how to perform regression on longitudinal data
> To: [hidden email]
>
> Dear Friends,
>
> I am new to this forum. Sorry if I am posting a question which might have
> been already posted and answered.
>
> I am performing statistical analysis for my MBA management project using
> SPSS.
>
> In short, through this analysis, I am trying to determine the correlation of
> CreditCardFrauds (dependent var.) with GDP, NumberofCreditCards and
> InternetUsers (3 independent vars.) in 17 countries across 9 years.
>
> With my initial reading of SPSS manuals and related text, I understood that
> due to the nature of the this data, the correct way to analyse this is using
> the MIXED-models from SPSS.
>
> Problem I am facing is that when I Linear-Mixed-Models with Random effects
> explained in "IBM_SPSS_Advanced_Statistics.pdf", the SPSS never returns. I
> left it running for more than 6 hours, but it appeared to be hanging.
>
> Can you please help me know if this is the correct model to use in SPSS for
> such kind of data? Does SPSS take long time to run analysis on such kind of
> data? (If required, I will share the command syntax that was run by SPSS and
> also a sample of my data).
>
> Any help or guidance on this problem is highly appreciated. Thanks in
> advance for your time and guidance.
>
> Best Regards.

...

[SPSSforMBA]

Hi All,

Really thanks for noting my request.

Attaching the screenshots of the command syntax and sample data set:
variables-and-data-for-mixed-model.png

mixed-model-regression-in-spss.png

As you can see, Im trying to hypothesize that fraud loss (in USD millions) is related to GDP/capita (USD), internet users (percentage of population) and number of credit cards (in millions). As these factors increase, there is a tendency for increase in fraud loss. Currently, this can be observed for most of the developed nations, like USA, Canada, Australia, Western European countries, SouthKorea and so on.

So, I collected data for such 17 countries over the period of 9 years and expect to see the similar trend across these countries.

As you say, I may have to convert the factors to either per-capita or per-GDP or so.
But, since I did not know this till now, that's why I wanted to know from SPSS community whether I am performing this regression correctly or not.

Please guide me what correction or changes I need to do to this model and/or data for get the results from SPSS within reasonable about of time.

Many thanks in advance for your time and guidance.

Best Regards,
bganesh

[Peter Løvgreen (101plg)]

Jeg er på tjenesterejse.
Jeg er tilbage torsdag d. 20. september 2012.
Med venlig hilsen Peter Løvgreen

=====================
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

[Maguin, Eugene]

Bganesh,

You are running an extraordinarily complex model with just a spit of data--153 cases. Every term is a random effect including all interactions. And you have a repeated factor in there as well. Furthermore, you are treating continuous variables as 'BY' variables. That is crazy. I'm kind of speechless (well, actually not, I just don't want to say what I'm thinking). Actually, I'll say this. Somebody has done you an extreme disservice in their teaching you about data analysis.

Why haven't you started with a much simpler model? Say, this model.
Mixed fraudloss with year gdp issuers ccqty/fixed= year gdp issuers ccqty/
    Random=intercept | subject(country) covtype(un).

But, you could do something even simpler. You have two factors, basically, country and year. The most basic model is

Regression dependent=fraudloss/enter gdp issuers ccqty.

*  you have to build some country contrast terms. Let these be c1 to c16.
Regression dependent=fraudloss/enter year c1 to c16 gdp issuers ccqty.

Sort cases by country.
Split file by country.
Regression dependent=fraudloss/enter year gdp issuers ccqty.
Split file off.

You'll see how much the coefficients differ across country

And you could turn this around

*  Using the country contrast terms.
Sort cases by year.
Split file by year.
Regression dependent=fraudloss/enter c1 to c16 gdp issuers ccqty.
Split file off.

Gene Maguin




-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of SPSSforMBA
Sent: Friday, September 14, 2012 10:31 AM
To: [hidden email]
Subject: Re: how to perform regression on longitudinal data

Hi All,

Really thanks for noting my request.

Attaching the screenshots of the command syntax and sample data set:
http://spssx-discussion.1045642.n5.nabble.com/file/n5715086/variables-and-data-for-mixed-model.png
variables-and-data-for-mixed-model.png
http://spssx-discussion.1045642.n5.nabble.com/file/n5715086/mixed-model-regression-in-spss.png
mixed-model-regression-in-spss.png

As you can see, Im trying to hypothesize that fraud loss (in USD millions) is related to GDP/capita (USD), internet users (percentage of population) and number of credit cards (in millions). As these factors increase, there is a tendency for increase in fraud loss. Currently, this can be observed for most of the developed nations, like USA, Canada, Australia, Western European countries, SouthKorea and so on.

So, I collected data for such 17 countries over the period of 9 years and expect to see the similar trend across these countries.

As you say, I may have to convert the factors to either per-capita or per-GDP or so.
But, since I did not know this till now, that's why I wanted to know from SPSS community whether I am performing this regression correctly or not.

Please guide me what correction or changes I need to do to this model and/or data for get the results from SPSS within reasonable about of time.

Many thanks in advance for your time and guidance.

Best Regards,
bganesh




--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/how-to-perform-regression-on-longitudinal-data-tp5715082p5715086.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

[David Marso]

1.  Copy Paste so we don't have to type?
As well as the transformation Rich suggested, perhaps leave out the interactions for the first pass and see what happens.  Add them later?
--

SPSSforMBA wrote

Hi All,

Really thanks for noting my request.

Attaching the screenshots of the command syntax and sample data set:
variables-and-data-for-mixed-model.png

mixed-model-regression-in-spss.png

As you can see, Im trying to hypothesize that fraud loss (in USD millions) is related to GDP/capita (USD), internet users (percentage of population) and number of credit cards (in millions). As these factors increase, there is a tendency for increase in fraud loss. Currently, this can be observed for most of the developed nations, like USA, Canada, Australia, Western European countries, SouthKorea and so on.

So, I collected data for such 17 countries over the period of 9 years and expect to see the similar trend across these countries.

As you say, I may have to convert the factors to either per-capita or per-GDP or so.
But, since I did not know this till now, that's why I wanted to know from SPSS community whether I am performing this regression correctly or not.

Please guide me what correction or changes I need to do to this model and/or data for get the results from SPSS within reasonable about of time.

Many thanks in advance for your time and guidance.

Best Regards,
bganesh

[Hector Maletta]

Totally agree with Eugene's recommendations. Start with an extremely simple
model, then add more effects only if needed.
Moreover, any set of data points could be PERFECTLY predicted by a
sufficiently complex model, but it would not be generalizable to other sets
of data points, because much of the observed variation is essentially
"noise". A simpler model is usually more robust. And then there is the
matter of the number of cases: one needs many cases to test a complex model,
and you have  a very small number of cases for such a large number of
effects.
To reinforce Eugene's points: you do not need 16 country dummies: group the
countries according to some principle (per capita GDP, some other financial
indicator, world region, etc.) to make a more robust contrast between TYPES
of countries. If it shows something, you may go into more detail (although
you may need more cases to do so).

Hector

-----Mensaje original-----
De: SPSSX(r) Discussion [mailto:[hidden email]] En nombre de
Maguin, Eugene
Enviado el: Friday, September 14, 2012 12:24
Para: [hidden email]
Asunto: Re: how to perform regression on longitudinal data

Bganesh,

You are running an extraordinarily complex model with just a spit of
data--153 cases. Every term is a random effect including all interactions.
And you have a repeated factor in there as well. Furthermore, you are
treating continuous variables as 'BY' variables. That is crazy. I'm kind of
speechless (well, actually not, I just don't want to say what I'm thinking).
Actually, I'll say this. Somebody has done you an extreme disservice in
their teaching you about data analysis.

Why haven't you started with a much simpler model? Say, this model.
Mixed fraudloss with year gdp issuers ccqty/fixed= year gdp issuers ccqty/
    Random=intercept | subject(country) covtype(un).

But, you could do something even simpler. You have two factors, basically,
country and year. The most basic model is

Regression dependent=fraudloss/enter gdp issuers ccqty.

*  you have to build some country contrast terms. Let these be c1 to c16.
Regression dependent=fraudloss/enter year c1 to c16 gdp issuers ccqty.

Sort cases by country.
Split file by country.
Regression dependent=fraudloss/enter year gdp issuers ccqty.
Split file off.

You'll see how much the coefficients differ across country

And you could turn this around

*  Using the country contrast terms.
Sort cases by year.
Split file by year.
Regression dependent=fraudloss/enter c1 to c16 gdp issuers ccqty.
Split file off.

Gene Maguin




-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
SPSSforMBA
Sent: Friday, September 14, 2012 10:31 AM
To: [hidden email]
Subject: Re: how to perform regression on longitudinal data

Hi All,

Really thanks for noting my request.

Attaching the screenshots of the command syntax and sample data set:
http://spssx-discussion.1045642.n5.nabble.com/file/n5715086/variables-and-da
ta-for-mixed-model.png
variables-and-data-for-mixed-model.png
http://spssx-discussion.1045642.n5.nabble.com/file/n5715086/mixed-model-regr
ession-in-spss.png
mixed-model-regression-in-spss.png

As you can see, Im trying to hypothesize that fraud loss (in USD millions)
is related to GDP/capita (USD), internet users (percentage of population)
and number of credit cards (in millions). As these factors increase, there
is a tendency for increase in fraud loss. Currently, this can be observed
for most of the developed nations, like USA, Canada, Australia, Western
European countries, SouthKorea and so on.

So, I collected data for such 17 countries over the period of 9 years and
expect to see the similar trend across these countries.

As you say, I may have to convert the factors to either per-capita or
per-GDP or so.
But, since I did not know this till now, that's why I wanted to know from
SPSS community whether I am performing this regression correctly or not.

Please guide me what correction or changes I need to do to this model and/or
data for get the results from SPSS within reasonable about of time.

Many thanks in advance for your time and guidance.

Best Regards,
bganesh




--
View this message in context:
http://spssx-discussion.1045642.n5.nabble.com/how-to-perform-regression-on-l
ongitudinal-data-tp5715082p5715086.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

=====================
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