Categorical Data Analysis in Complex Surveysfaculty.nps.edu/rdfricke/OA4109/Lecture 9-3 -- Categorical Data... · Categorical Data Analysis in Complex Surveys" Professor Ron Fricker"

Categorical Data Analysis in Complex Surveys"

Professor Ron Fricker"Naval Postgraduate School"

Monterey, California"

2/22/13 1

Reading Assignment:"Lohr chpt. 10"

Goals for this Lecture"

•  What about categorical data analysis for complex designs?"

•  How do complex designs affect chi-square hypothesis tests?"

•  What are some methods to correct for the sampling design in order to correctly do the tests?"

2/22/13 2

What’s the Problem?"

•  Chi-square distribution of test statistic results from SRS assumption"

•  In complex surveys table counts unlikely to reflect relative frequencies of the categories in the population "–  Can’t just plug counts into standard X 2 calculations"–  Results in incorrect p-values"

2/22/13 3 3

Computer/Cable TV Example"

•  What if interviewed two individuals in each house and got same answers?"

2/22/13 4

Computer? Yes No

Cable? Yes 119 188 307 No 88 105 193

207 293 500

New data:

Original data:

Computer? Yes No

Cable? Yes 238 376 614 No 176 210 386

414 586 1000

2 4.562X =-value 0.03p =

2 2.281X =-value 0.13p =

4

Let’s Check the Calculations (1)"

•  Original data and chi-square results:"

2/22/13 5

Correct p-value

Let’s Check the Calculations (2)"

•  New data and chi-square results:"

2/22/13 6

p-value wrong!

After Specifying Sampling Design, R “survey” Package Gets It Right"

2/22/13 7

“cluster” identifies the households

p-value is right

Original data:" New data:"

New data correctly analyzed with cluster design accounted for:"

Effects of Stratified Sampling Design on Hypothesis Tests and CIs (1) "

•  If rows in contingency table correspond to strata, usual chi-square test of homogeneity fine"–  But may want to test association between other

(non-strata) factors"•  In general, stratification increases precision of

estimates"–  E.g., stratified sample of size n gives same

precision for estimating pij as a SRS of size n / dij, where dij is the design effect"

2/22/13 8 8

Effects of Stratified Sampling Design on Hypothesis Tests and CIs (2) "

•  Thus p-values for chi-square tests with stratification are conservative"–  E.g., actual p-value will be smaller than calculated

p-value"–  Means if null rejected, it is appropriate"–  However, could also fail to reject and miss a

significant result (“Type II error”)"•  If don’t reject but close, how to tell if null should

be rejected?"

2/22/13 9 9

Effects of Clustered Sampling Design on Hypothesis Tests and CIs"

•  Opposite effect from stratification"•  As we illustrated, p-values artificially low"

–  Means if fail to reject null, it is appropriate"•  However, if do reject null, how to tell if null

really should be rejected?"•  Clustering unaccounted for can result in

spurious “significant” results"–  I.e., more likely to commit “Type I” error"

2/22/13 10 10

Corrections to Chi-square Tests"

•  Wald tests"•  Bonferroni tests"•  Matching moments"•  Model-based methods"

2/22/13 11 11

Think of Problem in Terms of Cell Probabilities (1)"

•  Use sampling weights to estimate population quantity"

where"

•  Thus"

2/22/13 12

ˆk kij

k Sij

kk S

w yp

w∈

∈

=∑∑

1 if observation unit is in cell ( , )0 otherwise kij

k i jy ⎧

= ⎨⎩

sum of weights for observation units in cell ( , )ˆsum of weights for all observation units in sampleij

i jp =

12

Think of Problem in Terms of Cell Probabilities (2)"

•  So, using the , construct the table"

•  Can express the test statistics as"

""

2/22/13 13

ˆ ijp

13

2 222

All All Allcells cells cells

ˆ ˆ( ) ( )(observed expected)expected

ij ij ij ij

ij ij

np - np p - p - X nnp p

= = =∑ ∑ ∑

2

All Allcells cells

ˆobserved ˆ2 observed ln 2 lnexpected

ijij

ij

pG n p

p⎛ ⎞⎛ ⎞

= × = ⎜ ⎟⎜ ⎟ ⎜ ⎟⎝ ⎠ ⎝ ⎠∑ ∑

Wald Tests (1)"

•  For a 2x2 table, null hypothesis of independence is"

•  This is equivalent to testing"

•  Let"

2/22/13 14 14

, 1 , 2ij i jp p p i j+ += ≤ ≤

0 11 22 12 21

11 22 12 21

: 0: 0a

H p p p pH p p p p

− =− ≠

11 22 12 21ˆ ˆ ˆ ˆ ˆp p p pθ = −

Wald Tests (2)"

•  Then for large samples, under the null"

•  follows an approximately standard normal distribution "

•  Equivalently, follows a chi-square distribution with 1 degree of freedom"

•  Must estimate the variance appropriately"

2/22/13 15 15

( )ˆ ˆV̂θ θ

( )2ˆ ˆV̂θ θ

( )ˆV θ

Example: Survey of Youth in Custody"

•  1987 survey of incarcerated youth"–  Sample of n=2,588 juveniles and young adults in long-term, state-

operated juvenile institutions"–  Interviewed about family background, previous criminal history, and

drug and alcohol use"–  Selected variables are contained in the syc data frame"

"

2/22/13 16

Example: Survey of Youth in Custody (2)"

•  Is there an association between:"–  “Was anyone in your family ever incarcerated?”"–  “Have you ever been put on probation or sent to a

correctional institution for a violent offense?”"•  Table with sum of weights:""

2/22/13 17 17


Incorrect analyses"

2/22/13 18

Raw counts: Do not appropriately reflect population distribution

Weighted counts: Without adjustment, sample size overinflated

18


•  Results in the following estimated proportions:"

•  Test statistic:"•  How to estimate the variance? "

2/22/13 19 19

11 22 12 21ˆ ˆ ˆ ˆ ˆ 0.0053p p p pθ = − =


•  Use resampling method:"•  Thus, the standard error

of is"•  So the test statistic is"

•  p-value:"•  Result: No evidence of association "

2/22/13 20

0.0158 7 0.006=θ̂

( )ˆ 0.0053 0.89

0.0060ˆV̂t θ

θ= = =

6Pr( ) 2 Pr( 0.89) 0.41T t Tν => = × > =

Example: Survey of Youth in Custody (6)"•  Doing the calculations in R:"

–  Results consistent with book – but we can’t account for the stratification and fpc…

2/22/13 21 21

Better p-value?

Wald Tests for Larger Tables"

•  Let"•  Hypotheses are"

•  Wald test statistic is where is the estimated covariance matrix "

•  Problem is, need a large number of PSUs to estimate covariance matrix"–  E.g., 4x4 table results in 9x9 covariance matrix

that requires estimation of 45 variance/covariances"

2/22/13 22

11 12 ( 1)( 1), ,...,T

r cθ θ θ − −⎡ ⎤= ⎣ ⎦θ

0 :: for one or more cellsa

HH

=≠θ 0θ 0

2 1ˆ ˆ ˆV̂( )TWX

−= θ θ θ ˆV̂( )θ

Bonferroni Tests (1)"

•  Alternative to Wald test"•  Idea is to separately (and conservatively) test

each "•  Test each of m=(r-1)(c-1) tests separately at

/m significance level"•  Reject null that variables are independent if

any of the m separate tests reject"

2/22/13 23

, 1 1, 1 1ij i r j cθ ≤ ≤ − ≤ ≤ −

α

Bonferroni Tests (2)"

•  Specifically, reject if"

"for any i and j, where is the appropriate degrees of freedom"–  Resampling: #resample groups – 1"–  Another method: #PSUs – #strata"

•  Lohr says method works well in practice"

2/22/13 24

θ̂ ij V̂ θ̂ ij( ) > tα /2m,ν

0 :H =θ 0

ν


•  Is there a relationship between age and whether a youth was sent to an institution for a violent offense?"

2/22/13 25


•  Hypotheses are"

•  What happens if clustering ignored?"–  With n=2,621, we have"

"which gives an (incorrect) p-value of ~ 0"•  Compare to a Bonferroni test…"

2/22/13 26

0 11 11 1 1

12 12 1 2

: 00

H p p pp p p

θθ

+ +

+ +

= − == − =

22 32

1 1

ˆ ˆ ˆ( )34

ˆ ˆij i j

i j i j

p - p pX n

p p+ +

= = + +

= =∑∑


•  For these data, and"•  Using resampling,

we get the table:"•  And from this,"

•  Thus "

2/22/13 27

11ˆ 0.013θ = 12

ˆ 0.0119θ =

( )11ˆs.e. 0.0074,θ =

( )12ˆs.e. 0.0035θ =

θ̂11

s.e. θ̂11( ) = 1.8, θ̂12

s.e. θ̂12( ) = 3.4 0.05/ 2 2, 6 2.97t ν× = =and"

Reject null (more appropriately)

Matching Moments (1) "

•  As before, X 2 test statistics from complex survey don’t follow distribution"

•  Idea is to use first and second moments to adjust test statistic"–  I.e., perhaps multiples of X 2 test statistics

approximately follow distribution"•  Developed by Rao and Scott (1981, 1984)""

2/22/13 28

2( 1)( 1)r cχ − −

2χ

Matching Moments (2)"

•  Won’t go into detail here – see Lohr if interested"

•  First-order correction: "–  Estimate expected value of X2 as well as various

design effects"–  Use to calculate a new test statistic ( )"–  Compare to an F distribution to calculate p-values"

•  Second-order correction:"–  Estimate first and second moments of X 2, as well

as various design effects"–  Use to calculate a new test statistic ( )"–  Compare to a distribution to calculate p-values"

2/22/13 29 29

2FX

2SX

2χ

Back to Survey of Youth in Custody"

•  Table below shows estimated design effects"

•  Several are large, resulting from clustering within facilities"–  Some facilities are just for nonviolent offenders

and others just for violent"

2/22/13 30

Survey of Youth Study (cont’d)"

•  Just doing a first-order correction"•  Turns out that"•  Appropriate reference distribution is an"•  Thus, the p-value is"

•  Conclusion: Reject the null "–  Consistent with Bonferroni result"–  p-value may be a bit too small (because of large

design effect differences) "

2/22/13 31

2 16.2FX =2,12F

2,12Pr( 16.2 / 2) 0.006F > =

31

Model-based Methods"

•  Rather than use covariance estimates to adjust chi-squared tests, use model"

•  Lohr describes model for cluster sample with two observations per cluster"–  Model assumes design effects are the same in

each cell and margin"•  Well just illustrate with example from Lohr"

2/22/13 32

Example: Exploring association Between Gender and Schizophrenia (1)"

•  Data: 71 hospitalized pairs of siblings"–  Schizophrenia tends to run in families"–  Thus, there is dependence (clustering) among

siblings"•  Tabularization of the

142 individuals:"–  Assuming independence,

giving a p-value of 0.00002 –  But clearly not correct: p-value too small "

2/22/13 33 33

2 17.89X =

Example: Association Between Gender and Schizophrenia (2)"

•  Issue: clustering was ignored"–  In 52 of 71 sibling pairs, either both have

schizophrenia or both do not"

2/22/13 34


•  Model the clustering:"–  Let qij be the probability a pair falls in the (i,j) cell"

•  E.g., q11 is the probability that both siblings are schizophrenic and male"

–  Then model the qijs by"

"""where a is clustering effect and qi is probability an individual is in class i (i = SM, SF, NM, NF)"

•  If a=0 then individuals are independent, and regular chi-square test applies "

2/22/13 35 35

2(1 ) if (1 ) if i i

iji j

aq a q i jq

a q q i j⎧ + − =⎪= ⎨ − ≠⎪⎩


•  If a=1 then siblings perfectly correlated and regular chi-square test applies but with X 2/2 (since each pair is one piece of information)"

•  For 0 < a < 1, if the model is true, under the null X 2/(1+a) follows a distribution"

•  Fitting model via maximum likelihood, we get and estimate probabilities:"–  Table does not exhibit significant lack of fit from previous

table of showing clustering"–  Test statistic for independence is X 2/(1+0.3006)=13.76 which

gives a p-value of 0.0002 – reject "

2/22/13 36

2( 1)( 1)r cχ − −

ˆ 0.3006a =

Other Software: Stata"

•  Stata 9: SVY procedures support both one-way and two-way tables"–  svy: tab [var]"–  svy: tab [var1] [var2]"

•  Need to order manuals to see which methods used"

•  See www.stata.com/stata9/svy.html for more detail"

2/22/13 37

Other Software: SAS"

•  SAS v8 has some procedures for complex survey analysis (PROC SURVEYMEANS PROC SURVEYREG), but no PROCs for categorical data analysis"–  PROC FREQ and PROC CATMOD would

incorrectly estimate standard errors"•  SAS v9.1 has PROC SURVEYFREQ for

categorical data analysis"–  Can specify various Wald and Rao-Scott tests "

•  See http://support.sas.com/onlinedoc/913/docMainpage.jsp"

2/22/13 38

And Still Other Software…"

•  …designed for complex survey analysis include SUDAAN and WestVar"–  Don’t know if they can do these calculations or not"–  See their documentation ""

ü JMP: Cannot do appropriate calculations"

2/22/13 39

What We Have Just Learned"

•  Gained some insight into "–  What to do about categorical data analysis for

complex designs"–  How complex designs affect chi-square hypothesis

tests"•  Learned about some methods to correct for

the sampling design in chi-square tests"

2/22/13 40

Categorical Data Analysis in Complex Surveysfaculty.nps.edu/rdfricke/OA4109/Lecture 9-3 -- Categorical Data... · Categorical Data Analysis in Complex Surveys" Professor Ron Fricker"

Documents