Finding the most 'important' variable

[gheelengooi]

I am trying to analyze a junk of landslide data (landslide @ different locations) to see which factor has the most effect on landslide, that includes slope angle, rainfall intensity etc etc, which sums up to 7 to 8 of them.

I am wondering which kind of analysis I should adapt in order to do that. I am thinking of using multiple regression analysis and sensitivity analysis but I am still trying to figure it out.

Any good suggestion for that?

(I use Matlab and Excel)

 

[Studiot]

Since there are several very different mechanisms causing landslips you need to break down your data under these headings first.
Otherwise the effect of the variables you mention will not be directly comparable and any data analysis meaningless.

 

[gheelengooi]

Thanks Studiot!

Yeap I already broken them down to landslip, debris flow, landslide etc according to Takahashi's definition.

I just want to figure out how to do the analysis

 

[Studiot]

Well your opening post suggested formal statistical analysis.

For this you require a (or several) hypothesis to test.

 

[Mech_Engineer]

It seems to me the most sensitive variable will have the steepest slope (partial derivative)...

 

[brewnog]

Why not try an ANOVA or some such?

 

[Danger]

This is something that weirds me out as an uneducated guy with a scientific mind. How can any factor be dismissed? For instance, slope and ground viscosity will effect the speed of decent, but other things such as geological formations, winds, etc. can alter the vector and hence velocity of the sliding bits. How can one determine which is most important?

 

[brewnog]

Danger, you can use some clever statistical tools which (assuming you have enough data) will show you the strength of the effect of each variable, and the effect of the variables you didn't think of. Providing the amount of variation you have measured is mostly explained by the factors included in your model, then hey presto.

 

[Danger]

Okay, got it. That same question has nagged at me in several different fields of study. Thanks, Brewski.

 

[brewnog]

Okay, got it. That same question has nagged at me in several different fields of study. Thanks, Brewski.

It's a big "aha!" moment when you see it work. Next time you're at a computer with a statistician get him to show you, it'll blow your mind.

 

[Danger]

Next time you're at a computer with a statistician get him to show you, it'll blow your mind.

I'll have to take your word for it; I'm homebound with a couple of G3's and my beloved MacBook. PF is my only life.

On the other hand, if I snork down my morning allotment of inhalers right now, I can probably make it all the way down the block to the beer store and back.

 

[Studiot]

some clever statistical tools

IMHO such things have no place in geotechnical engineering.

Since the next elections are a few years off we will have to wait for such an opening.

 

[gheelengooi]

Thank you guys! I will finish the analysis in the coming two days. I will try multiple regression with sensitivity analysis and also ANOVA.

And as for Danger:-

This is something that weirds me out as an uneducated guy with a scientific mind. How can any factor be dismissed? For instance, slope and ground viscosity will effect the speed of decent, but other things such as geological formations, winds, etc. can alter the vector and hence velocity of the sliding bits. How can one determine which is most important?

You're right, but not totally. In geotechnical engineering, one can categorize the mechanism, geological formation etc etc but then you would have too few cases for each group because there are so many unknown things that we cannot measure affecting flow landslide.

So instead of going 'academic', a lot of times we would perform in-house experiments by building flume but we need some statistics to support why we build the experiments that way.

That's why we're looking into the field data and try to see which variable is the most 'important' so that we can plan our experiments and see what insights we can gain.

 

[Danger]

So if I'm interpreting that properly, you are essentially eliminating the middle-man and directly correlating your own theories with your own empirical data? Direct feedback loop?

 

[gheelengooi]

So if I'm interpreting that properly, you are essentially eliminating the middle-man and directly correlating your own theories with your own empirical data? Direct feedback loop?

Well urm...I don't quite get you. The flow landslide data are collected from major landslide events in the region I am doing research, which is Hong Kong. They are not empirical data, right?

 

[Danger]

Well urm...I don't quite get you. The flow landslide data are collected from major landslide events in the region I am doing research, which is Hong Kong. They are not empirical data, right?

Perhaps I misunderstand the term. To me, that would be the empirical data, as in something that is observed from a physical event. Do you not then correlate that with your theoretical models to determine their veracity?
Sorry, man... I'm a bit lost. You might have to "dummy it down" a tad more for me to catch the reality.