Scribd Logo
Upload

Welcome to Scribd!

  • 19 views

    Department of Statistics: Course Stats 330

    Uploaded by

    PETER
    This document contains model answers for an assignment on log-linear models. It includes: 1) A data frame constructed from counts data on car accidents by ejection, seatbelt use, and injury type. 2) Fitting a log-linear model to the data and finding that a homogeneous association model provides a good fit with no ill-fitting cells. 3) Calculating that the odds ratio of a fatal vs. non-fatal injury given seatbelt use is 5.56, and this does not depend on ejection status. A 95% confidence interval of (5.01, 6.20) is also provided. 4) Analyzing monthly suicide count data by gender and

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd

    Department of Statistics: Course Stats 330

    Uploaded by

    PETER
    19 views4 pages
    This document contains model answers for an assignment on log-linear models. It includes: 1) A data frame constructed from counts data on car accidents by ejection, seatbelt use, and injury type. 2) Fitting a log-linear model to the data and finding that a homogeneous association model provides a good fit with no ill-fitting cells. 3) Calculating that the odds ratio of a fatal vs. non-fatal injury given seatbelt use is 5.56, and this does not depend on ejection status. A 95% confidence interval of (5.01, 6.20) is also provided. 4) Analyzing monthly suicide count data by gender and

    Original Description:

    07a5 Ans

    Original Title

    07a5 Ans

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    This document contains model answers for an assignment on log-linear models. It includes: 1) A data frame constructed from counts data on car accidents by ejection, seatbelt use, and injury type. 2) Fitting a log-linear model to the data and finding that a homogeneous association model provides a good fit with no ill-fitting cells. 3) Calculating that the odds ratio of a fatal vs. non-fatal injury given seatbelt use is 5.56, and this does not depend on ejection status. A 95% confidence interval of (5.01, 6.20) is also provided. 4) Analyzing monthly suicide count data by gender and

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    19 views4 pages

    Department of Statistics: Course Stats 330

    Uploaded by

    PETER
    This document contains model answers for an assignment on log-linear models. It includes: 1) A data frame constructed from counts data on car accidents by ejection, seatbelt use, and injury type. 2) Fitting a log-linear model to the data and finding that a homogeneous association model provides a good fit with no ill-fitting cells. 3) Calculating that the odds ratio of a fatal vs. non-fatal injury given seatbelt use is 5.56, and this does not depend on ejection status. A 95% confidence interval of (5.01, 6.20) is also provided. 4) Analyzing monthly suicide count data by gender and

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    You are on page 1of 4

    DepartmentofStatistics

    COURSESTATS330

    ModelanswersforAssignment5,2007

    Question1

    1. TypethedataintoRandconstructadataframesuitableforfittingloglinearmodels.
    Printoutthedataframe.[5marks]

    > counts=c(1105, 411111, 4624,157342, 14, 483, 497, 1008)


    > accident.df = data.frame(Counts=counts, expand.grid(Ejected = c("Yes",
    "No"), Seatbelt = c("Yes", "No"), Injury=c("Nonfatal", "Fatal")))
    > accident.df
    Counts Ejected Seatbelt Injury
    1 1105 Yes Yes Nonfatal
    2 411111 No Yes Nonfatal
    3 4624 Yes No Nonfatal
    4 157342 No No Nonfatal
    5 14 Yes Yes Fatal
    6 483 No Yes Fatal
    7 497 Yes No Fatal
    8 1008 No No Fatal

    2. Fitasuitableloglinearmodeltothedata.Arethereanycellsthatarenotfittedwellby
    themodel?[5marks]

    Ifwefitthesaturatedmodelanddoananova,weobtain

    > accident.glm = glm(Counts~Ejected*Seatbelt*Injury, data=accident.df,


    + family=poisson)
    > anova(accident.glm, test="Chisq")
    Analysis of Deviance Table
    Model: poisson, link: log
    Response: Counts
    Terms added sequentially (first to last)
    Df Deviance Resid. Df Resid. Dev P(>|Chi|)
    NULL 7 1624865
    Ejected 1 729871 6 894994 0
    Seatbelt 1 111458 5 783536 0
    Injury 1 772092 4 11444 0
    Ejected:Seatbelt 1 7877 3 3568 0
    Ejected:Injury 1 2423 2 1145 0
    Seatbelt:Injury 1 1142 1 3 2.741e-250
    Ejected:Seatbelt:Injury 1 3 0 -5.010e-11 9.115e-02

    1
    This suggests that the threeway interaction term is not significant, and that the
    homogeneous association model is appropriate for thesedata. Fitting this model, and using
    thesummaryfunction,weget

    > ham.glm = glm( counts~(Ejected+Seatbelt+Injury)^2, data=accident.df,


    + family=poisson)
    > summary(ham.glm)

    Call:
    glm(formula = counts ~ (Ejected + Seatbelt + Injury)^2, family = poisson,
    data = accident.df)

    Deviance Residuals:
    1 2 3 4 5 6 7 8
    0.20704 -0.01071 -0.10095 0.01731 -1.59987 0.31400 0.30951 -0.21583

    Coefficients:
    Estimate Std. Error z value Pr(>|z|)
    (Intercept) 7.00137 0.02992 233.99 <2e-16 ***
    EjectedNo 5.92527 0.02996 197.76 <2e-16 ***
    SeatbeltNo 1.43913 0.03321 43.33 <2e-16 ***
    InjuryFatal -3.96315 0.06944 -57.07 <2e-16 ***
    EjectedNo:SeatbeltNo -2.39964 0.03334 -71.97 <2e-16 ***
    EjectedNo:InjuryFatal -2.79779 0.05526 -50.63 <2e-16 ***
    SeatbeltNo:InjuryFatal 1.71732 0.05402 31.79 <2e-16 ***
    ---
    Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

    (Dispersion parameter for poisson family taken to be 1)

    Null deviance: 1.6249e+06 on 7 degrees of freedom


    Residual deviance: 2.8540e+00 on 1 degrees of freedom
    AIC: 93.853

    Number of Fisher Scoring iterations: 3

    NoneoftheDevianceresidualsarelarge,indicatingthatallcellsarereasonablywellfitted.

    3. Compute the conditional odds ratio between seatbelt use and injury, conditional on
    Ejected.Dotheseoddsratiosdependonbeingejectedfromthecar?Whateffectdoes
    seatbeltusehaveonthetypeofinjury?Alsocomputea95%confidenceintervalforthe
    conditionaloddsratio.[10marks]

    ThehomogeneousassociationmodelimpliesthattheoddsratiosbetweenInjuryandSeatbelt
    in the separate tables for Eject = yes and Eject=No are the same, ant that the logodds
    ratioisthecoefficientfortheSeatbeltInjuryinteraction.Theestimatedcoefficientis1.71732,
    so the odds ratio in both tables is exp(1.71732) = 5.569582. Thus, the odds of an accident
    being fatal are about 5 and a half times more if a seat belt is not worn. A 95% confidence
    intervalisobtainedby

    2
    > exp(confint(ham.glm))
    Waiting for profiling to be done...
    2.5 % 97.5 %
    (Intercept) 1.034985e+03 1.163800e+03
    EjectedNo 3.532271e+02 3.972509e+02
    SeatbeltNo 3.952975e+00 4.502698e+00
    InjuryFatal 1.656784e-02 2.175183e-02
    EjectedNo:SeatbeltNo 8.497287e-02 9.683767e-02
    EjectedNo:InjuryFatal 5.472342e-02 6.796101e-02
    SeatbeltNo:InjuryFatal 5.013605e+00 6.196228e+00

    i.e. (5.013,6.196) to 3 dp.

    Q2.Thefollowingcodemakesthedataframeandplotstherelativefrequenciesofthe
    variousmonthsseparatelyformalesandfemales:

    counts = scan()
    3755 3251 3777 3706 3717 3660 3669 3626 3481 3590 3650 3392
    1362 1244 1496 1452 1448 1376 1370 1301 1337 1351 1416 1226
    heights = matrix(counts, 2,12, byrow=T)
    relheights = heights/apply(heights, 1, sum)
    barplot(relheights, beside=T, legend.text = T, ylim=c(0,0.12))
    0.12

    Male
    Female
    0.10
    0.08
    0.06
    0.04
    0.02
    0.00

    Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

    3
    Weseefromthegraphthatthedistributionoverthemonthsissimilarformalesand
    females.However,themonthsseemtodiffer:thereisadipinDecemberandFebrurary,
    andapeakinMarch.

    Aretheseobserveddifferencesduetochance?Wecanseeifthedistributionsarethe
    sameformalesandfemalesbytestingifgenderandmonthareindependent,or,
    alternativelyiftheinteractionsarezerointhePoissonmodel.Weget

    > suicide.df = data.frame(counts=counts, expand.grid(month= c("Jan",
    "Feb", "Mar", "Apr", "May", "Jun", "Jul",
    + "Aug", "Sep", "Oct", "Nov", "Dec"), sex=c("Male","Female")))
    >
    > anova(glm(counts~month*sex, data=suicide.df, family=poisson),
    test="Chisq")
    Analysis of Deviance Table
    Model: poisson, link: log
    Response: counts
    Terms added sequentially (first to last)
    Df Deviance Resid. Df Resid. Dev P(>|Chi|)
    NULL 23 12704.5
    month 11 118.4 12 12586.2 3.886e-20
    sex 1 12574.2 11 12.0 0.0
    month:sex 11 12.0 0 4.441e-14 0.4

    Thepvaluecorrespondingtoadevianceof12.0in11dfis0.3636indicatingthatthe
    distributionsarethesameforbothsexes.However,themodelwithgenderalonehasa
    devianceof130.35on22df,correspondingtoapvalueof0.000,sothatthereisa
    significantdifferencebetweenthemonths.

    You might also like