Today’s program Herwart / Axel: Kiva intro (the Galak et al. paper) Follow-up questions Non-response (and respondent list) Multi-level models in Stata.

Today’s program

• Herwart / Axel: Kiva intro (the Galak et al. paper)

• Follow-up questions

• Non-response (and respondent list)

• Multi-level models in Stata

Advanced Methods and Models in Behavioral Research – 2012

Your follow-up questions

See Galak_etal_follow_ups.rtf


Issues

• Consider doable in short vs long term

• Consider doable with the same or similar data, and on the same or another site

• Consider topic (it should not be Kiva)

• Not just ... “this might also be interesting”

Typical paper follow-ups

• Paper is wrong• There is an alternative explanation for the analytical

results• Paper’s conclusion might be dependent on

design/measurement/analysis. Redo with different kind of design/measurement/analysis

• Paper is right, but conclusions limited to a given time or place or context

• Paper argues that X’s of a given kind are important you check several other X’s of that kind, or you check several X’s of a different kind

• Relative importance of (kinds of) X’s• A connection XY is given with a “theory” behind it. You

check that theory behind it in more detail (typically with another kind of design/measurement/analysis).

Finding other papers on the topic

• Look at the references in the original paper

• Search for Kiva related papers (either in Google scholar or directly in Web of Science / Scopus ...)

• Search for more general key-words: “micro-financing” & “decision-making”

• Other literature: try “matching” (e.g. Literature on online dating), other sites with similar setup (e.g. eBay), persuasion and trust, charitable giving ...

New analysis/paper: conjoint experiment

• Tests importance of stimuli directly (allows comparison of importance of different stimuli + choice of stimuli)(also note the analysis on magnitude of donation in Galak et al., and their complicated analysis of the similarity argument)

• Population consists largely of non-donors• No real money involved

• ...

Non-response analysis

• Not all of the ones invited are going to participate

• Think about selective non-response: some (kinds of) individuals might be less likely to participate.

How might that influence the results?

sample


Back to the (multi-level) statistics...


MULTI – LEVEL ANALYSIS


Multi-level models or ...

dealing with clustered data.One solution: the variance component model

• Bayesian hierarchical models • mixed models (in SPSS)• hierarchical linear models • random effects models • random coefficient models • subject specific models • variance component models • variance heterogeneity models


Clustered data / multi-level models

• Pupils within schools (within regions within countries)

• Firms within regions (or sectors)

• Vignettes within persons


Two issues with clustered data

• Your estimates will (in all likelihood) be too precise: you find effects that do not exist in the population

[further explanation on blackboard]

• You will want to distinguish between effects within clusters and effects between clusters

[see next two slides]


On individual vs aggregate data

For instance: X = introvert X = age of McDonald’s employee Y = school results Y = like the manager


Had we only known, that the data are clustered!

So the effect of an X within clusters can be different from the effect between clusters!

Using the school example: lines represent schools. And within schools the effect of being introvert is positive!


MAIN MESSAGES

Be able to recognize clustered data and deal with it appropriately (how you do that will follow)

Distinguish two kinds of effects: those at the "micro-level" (within clusters) vs those at the aggregate level (between clusters)

(and ... do not test a micro-hypothesis with aggregate data)


A toy example – two schools, two pupils

Overall mean(0)

Two schools each with two pupils. We first calculate the means.

Overall mean= (3+2+(-1)+(-4))/4=0

3

2

-1

-4

exam

sco

re

School 2School 1

(taken from Rasbash)


Now the variance

Overall mean(0)

3

2

-1

-4

exam

sco

re

School 2School 1

The total variance is the sum of the squares of the departures of the observations around mean divided by the sample size (4) =

(9+4+1+16)/4=7.5


The variance of the school means around the overall mean

3

2

-1

-4

exam

sco

re

School 2School 1

Overall mean(0)

2.5

-2.5

The variance of the school means around the overall mean=

(2.52+(-2.5)2)/2=6.25 (total variance was 7.5)


The variance of the pupils scores around their school’s mean

3

2

-1

-4

exam

sco

re

School 2School 1

2.5

-2.5

The variance of the pupils scores around their school’s mean=

((3-2.5)2 + (2-2.5)2 + (-1-(-2.5))2 + (-4-(-2.5))2 )/4 =1.25


-> So you can partition the variance in individual level and school level

How much of the variability in pupil attainment is attributable to factors at the school and how much to factors at the pupil level?

In terms of our toy example we can now say

6.25/7.5= 82% of the total variation of pupils attainment is attributable to school level factors

1.25/7.5= 18% of the total variation of pupils attainment is attributable to pupil level factors

And this is important; we want to know how

to explain (in this example)

school attainment,and appararently thedifferences are at theschool level more than

the pupil level


Standard multiple regression won't do

Y D1 D2 D3 D4 D5 id …

+4 -1 -1 0 1 0 1

-3 1 1 1 0 -1 1

+2 0 0 1 0 -1 2

0 1 0 -1 1 0 2

+1 … … … … … 3

+2 … … … … … 3

-3 … … … … … 4

+4 … … … … … 4

… … … … … … …

So you can use all the data and just run a multiple regression, but then you disregard the clustering effect, which gives uncorrect confidence intervals (and cannot distinguish between effects at the cluster vs at the school level)

Possible solution (but not so good) You can aggregate within clusters, and then run a multiple regression on the aggregate data. Two problems: no individual level testing possible + you get less data points.

So what can we do?


Multi-level models

The usual multiple regression model assumes

... with the subscript "i" defined at the case-level.

... and the epsilons independently distributed with covariance matrix I.

With clustered data, you know these assumptions are not met.


Solution 1: add dummy-variables per cluster

• So just multiple regression, but with as many dummy variables as you have clusters (minus 1)

... where, in this example, there are j+1 clusters.

IF the clustering is (largely) due to differences in the intercept between persons, this might work.

BUT if there are only a handful of cases per person, this necessitates a huge number of extra variables


Solution 2: split your micro-level X-vars

Say you have:

then create:

and add both as predictors (instead of x1)

Make sure that you understand what

is happening here,and why it is of use.


Solution 3: the variance component model

In the variance component model, we split the randomness

in a "personal part" and a "rest part"


Now: how do you do this in Stata?

<See Stata demo> [note to CS: use age and schooling as examples to split at restaurant level]

relevant commandsxtset and xtregbys <varA>: egen <meanvarB> = mean(<varB>)gen dvarB = <varB> - <meanvarB>

convenience commandstab <var>, gen() droporder desedit sum


Up next

• How do we run the "Solution 1”, "Solution 2”, and “Solution 3” analysis and compare which works best? What about assumption checking?

• Random intercept we now saw, but how about random slopes?


When you have multi-level data (2 levels)1. If applicable: consider whether using separate dummies per

group might help (use only when this does not create a lot of dummies)

2. Run an empty mixed model (i.e., just the constant included) in Stata. Look at the level on which most of the variance resides.

3. If applicable: divide micro-variables in "group mean" variables and "difference from group mean" variables.

4. Re-run your mixed model with these variables included (as you would a multiple regression analysis)

5. (and note: use regression diagnostics secretly, to find outliers and such)

To Do

• Make an invitation list using the Excel template that I will send to you later this week (don’t invite anyone just yet!)

• Make sure you understand the multi-level concept with random intercepts (that is: c0 varies per cluster), and know how to do it in Stata

• Try the assignment on the website. Next week we will work on that data in class. Check out the practice data “motoroccasion8March2012.dta” on the website as well. It’s practice data.



Data: TVSFP on influencing behavior


Online already (though not visible)

motoroccasion8March2012.dta

Today’s program Herwart / Axel: Kiva intro (the Galak et al. paper) Follow-up questions Non-response (and respondent list) Multi-level models in Stata.

Documents

behavioral research

interesting advanced

survey9advanced methods

rtfadvanced methods

multilevel statistics

given kind

nonresponse analysisnot

selective nonresponse