Hypothesis Testing: Categorical Data Jin Xe Categorical Data > Dats classed into categories (ordered or unordered. Methods include 2. Two-Sample Tes for Sinomal Distribution Tete of Assocation and Homgeny Goodness. of Fit Test 4. The Kappa State = © invodueion a Sar © TwoSample Test for Binomial Proportions,a ST Two Sample pi vs po Breast Cancer Example Assume that A study on rik factors for Breast cancer (Bea) = Cazes: Bea women in selected hospitals = Contrals: Non-Bea women of comparable age who were in the hospital a the same time 5) ~-binlpism), 22 ~ bintpam) > The hypotheses of interest, Homan ow Kemer > Data > The test depends cn whether the samples ar independent or pled 2 Independent Samples |. normal: theory method or contingency table method i, Fishers exact text bares Samples (MeNemar's Tes) |. normal theary method [Age at fest bith Bea status | 23029 | torah peters came Normal Theory Approach + By lng hey (oe st > ee) a-n¥x(o-n(2+2)) ‘Pa = P(age at first birth > 30jcontrol) * we Hosp pa _ nn ~ Renin — Approximate Methods (Normal-Theory or Contingency-Table) = B= (m1 + a2)/(m +m) aR 9)> Sand ma -91> 8a ST Normal Theory Approach Normal Theory Approach > Tost statisti » For the Bca example, we can use the prop. test function in R >xsc(683, 1498); n*6(3220, 10245) Ddprop.teet(, a, alternative = e(*two.eided"), conf level = 0.95, correct » TRUE) + ahr + ph; contin cretion 2-aanple test for equality of proportions data: x out of a 7 Rapes Bi 4> non Fenquased = 77-8851, af = 4, porsiae < 2.2036 = Pralue siseraative hypotheasa: t4o sided po vae~ 2% P(2 > Somat 95 percent confadence aneereal 0,04999981 0, 08178846 Sample Size Determination Power Calculation ‘Suppose ie anticipate tat times as many people a ingroup 1 wil be enrled in group 2 = [a (eB)nne year] J ed were where ad Cees (mpi + mapa)/(m + ma) » 4.91 and gy ae as defined before For one sided test, replace a/2 with 0, For one sided ts, replace a2 witha.a Contingency Table Approach ~ Chi-square Test + Cardiovascular Disease: A study looked a the fects of Oral Contraceptive (OC) use on heart dense in women 40 to 48 years of ee Mi over 3-year period OC use ates Yer Te. Tear Teer 13 087 "5000 Non user | 7 9.093 | 10,000 Total | 20145980 [75.000 > The row totals/margins ar ied and known Chi-square Test Let Oy; be the observed frequency of cl (3). > Under Hp, the expected el frequencies, denoted by Ey. 20 Bu sng, Ba=m(1-7) By =mp and Ba =m ~ 9, > For lage sample (al iy > 5), under Hi, ey ew + When some Fi <5, we use Fisher's exact test ST Contingency Table Approach ~ Chi-square Test Mi over 3-year period Deus ates | Ver Wo Tear User i ra 7 Non vser_| On, On ma Total [ra ™a 7 P= POM over 3.yr periodlOC user) P= POM! over 3-y periodlon OC user) + Mas ps = pave th psf pa Ce ‘—_ a Contingency Table Approach » Reade pemze-nateix(c( 19,4907, 7,9983) rrow = 2, neal = 2, yrou = TRUE) >enteg.test(t22) continuity correction ata: 122 0.005626 squared = 7.8665, af = 1, prvaluea Fisher's Exact Test > A retoipectve dit stady on cardiovascular dase (CVD) among ren ages 50-54 n a speci county who ded ove 9 T-menth period. ‘Type of Dict Cause of Death | High Salt_Low Sat | Total Ton-OV 7 BS vo 5 o_| as Toad 7 33 |e » Hypothesis pr and Heep < pa, where pi =Pfhigi-salt dizt!non CV death) Plaigi-slt dit CV death) a The Urn (Hypergeometic) Model Color ofthe ball | Drawn Not drawn | Total She aes Xanlaend Red m=X mom tx] Total 7 ma [+e fora ST Fisher's Exact Test » Pevaie ~ the probability of geting 2 data as extreme as or more ‘extreme than the observed data if Tip was tue, ~ This probability can be computed under the assumption that cow realm taal margins are Gel + Tho Um (Hypergcometic) model: Suppose an um contains ns blue balls and ns re balls. and m, balls are drawn randomly. What is ‘the probability that actly ofthe Balls are Bue? Ce a Fisher's Exact Test Returning tothe CVD example, assume the row totals ae fed The ur model can be applied Type of Diet ‘Cause of Death [Tigh Satt_Low Sat [ Torr Non VO 73 m= vo 30 [ms Tor 73 Than the probability of getting a data as extreme as of more extreme than the abvered data fH pow = P(X $2) = PUY =0) + PX = 017 + 105 +.252 = ) + PIX =2)a Fisher's Exact Test > In general assume the table rearranged in sucha way that

For the loversidad aterative Maps

For uppersided stenative Hm. > pe povalue = PIX a) = PIX =a) + P(X = att) t+ POX =H) > For the o-sdedsleratve Hpi # ps prvalue = 2 x min|PUX Evaluate hypergeometric elstibuton in using phyper,r2.}. Ce ae a Paired Samples ~ Example Compare two chemotherapy regimens for Bea patients afer mastectomy ~ Patients ate asigned to pairs matched on age (withia§ years) land clinical conditions an fllned fr 8 years “Treatment & Treatment A | Sunive Die | Total Sunive | 510 16] 505 Die Be poe eel Toral [BIS 105 [BET re » Assuming independent samples (WRONG assumption), prop. teat (e(626, 616) -2+e(621 621)? ppvalue= 0.441 = no significant diference between the two frestrents ST Fisher's Exact Test + Reade dezze-mutrine(2,25,5,90),ar0¥ © 2, acel = 2, byroe Deinker- eat (220) ‘nteravivehypotbeia: tee ode ratio fa no onl 1 Ce a Paired Samples ~ McNemar's Test A concordant pr fs 9 matched par whose outcomes ar the same + A discordant pie i a matched pair whose outcomes ae diferent = Type A discordant pst: the treatment A member ofthe par has the event but treatment Bamber doesnot = Type B discordant pai siniarly dened + The concordant pats provide no information about siferences between treatmentsa Paired Samples ~ McNemar's Test vletp (a discordant par is of type 8), = Ifthe teatments are equally fictive then we expet an equal umber of type A and type 8 dscordant pairs. Thos we wish to test iysy 2 ww Hy #12 = Let np and ng be the number af discordant and type A dscordant pairs, respectively The tes statistic of the McNemr's text is (ra=nel= 0? x This is ecacly the sign test! a a R codes ~ McNemar’s Test > For large sample (nip > 20), we eam apply the norma method prop. test (ara Aara.D, pe05,correct=THUE, siternativer'tyo.sidee") » For small samples (roy < 20), we can use binon-tostGern_A,2°8,D,pr0.5, altornativen"tvo. sided") Example: Assess the statistical sgnifeance ofthe chemo data = prop.test -> pualue 0.0291 ~ inom. teat = pralve = 0.0266 ST R codes ~ McNemar’s Test > table tenenar. vest (table) Mefonar's Chi-equared teat with continuity correction Nekonar's chivaquared = 4.7619, éf = 1, prvalu 0.0201 Ce a Paired Samples ~ McNemar's Test + The sample sre need to conduct # two-sided test with significance level and power I~ 3 is 42 VATE) “Wpnlpa — 057 ster 9p tn 94 se popin o erns e + Pome aot Powe Plas en on tps 05am provided ny > 20 and mppall~ pa) > 5.© 8 «© Contingency Taber Test for association ~ Contingency table The data canbe graphically displayed as > Are the two stacked bar significantly diferent? > Lower anes appesr more prevalent inthe controls ad higher ages in TT Test for association ~ Contingency table Consider the following data from a case-control ca study Age at fest bieth status [290205242529 594335 | Toxo Gse | 320 — 106 01463200 | 3D ont | 42244322693 1092 __406_ | 10245 “oral |r 638 3008 1585 625 | Ta40 » Let pi, pa sro and ps be the probally dstribution of age at fst birth © 20,2024, 21-20, 30-24 ang > Xo given that 9 person is + We wish tote if the same probability dstbution holds fr cases as a TT The Null Hypothesis ~ Contingency table “The null hypothess can be expressed 35 > Ha:mo association batmona the case-control stu and age ot ft bin » Hy no ssscietion betwsen A and B in the follwing x C table + Data is callcted in an 12 contingeney table with Oy, observation inthe (,3)0h eal ¢ = Tyee =k. Let be the ow sum and my, bethe column Sum. Let n = $n, be the fatal numberof subjects 1 c —_ {| A |‘The Chi-square Test ~ Contingency table Test for association ~ Contingency table Under Ho, > Entering the Frequencies 35 2 matrie By >érage-natr(e(820, 1206, 2011 483, 220,1422, 4492, "2699, 1092,406) row = 2, neo) = 6, byrow = TRUE) > Under Hp, the test statistic hs a Ch square distribution with degree >jire-chieg. teat (retreg) ee nee 2 (Og- Fu? a data: freq. > Valid when no more that 20% of the expected frequencies are <5 [1,] 416.5793 1348,263 933.5967 371.8604 149.7007 ee ee ee Ce _— —cn — Chi-Square Test for Trend in Binomial Proportions: Chi-Square Test for Trend: R-Codes + Previous, we show some reatonshp between breast cancer and age st fit birth, but i doesnot eur spctealy about the nature of We can us the prop tend test command of R fer this purpose the relations. prop.trend.test(s, m, score) > In porticuar, we notice an incessng tend in the proportion of Ate ak beast cance n each stcedng cole + eis tr with he mumber of succes i ach group Thus we ith ots + isa vector wth the sme ize in ech ou Figs Tht taeda he's ee ee tet Bs ee increasing or decreasing function of the S, » Hf there are ke groups, then x, n, and score should be of length ke > Reference: Masel, AE. (1961). Analyzing qualtatve dataChi-Square Test for Trend: R-Codes cts20, 2208, 1021, 483, 220) > Ee CLITHD, $898, 390%, 1585, 625) eet) Seend. eee, Chi-Square Goodness-of-Fit Test » In Chapters 6,7 and 8, we assumed that the data came from an underying probably model + In this section, we consider 2 general metho of testing for the oodnese off of probably model © Cri Seuare Goodness of Fit Test Chi-Square Goodness-of-Fit Test ample: The frequency dsribution of diastolic blood presse (OBP) of 11,736 alts ages 30.59 in Est Boston, given belon Group Obsereé_ Expectee (evil) Frequency _Frequengy =o) >i0,<00 Sa<7 Sm<80 San, <00 3,100 S o0,< 10 30 We with to test itis data came from a narml distribution,ED TT Chi-Square Goodness-of Fit Test Cont'd, Chi-Square Goodness-of- Fit Test Cont'd. The agreement betwen the observed and expected fequences can Recall that the normal stributon i indexed by two parameter, be quantes using ando? » We estimate these parameters fom the dats by and S?, the sample mean and vaiance where is the total rumber of lasses an isthe number of Th expaced eae of he ith es i the god porinete inte neta hd fo be cated Banx Pex Ba where [x and U; are the loner and upper elas limits ofthe ith dss. pralue = P63. 14 > Xap + nour ample we take 4 = ~2o and Ua =, + Ue this stony 0 mae than (1/5) ofthe expec eqenis pee and neath oe —_—— —— ee Chi-Square Goodness-of-Fit Test: Example Chi-Square Goodness-of-Fit Test: R-Codes + frampe Ase th goodness ft fhe nal rin tthe Dera Le eta se(60, 210,10) + Te spleen ad standard cea ong fo the eee eon, 30) iad ea ee prob<-pnorm(U, 80,68, 12)-pnorn(L,80.68,12) + To empath potable ofthe dass, we ue rm Teervedce (6790, 3132, 4504, 604,219,680, 251 Se SComtsteerved Spec ape (60 X < m0) ~ pore, 80.68 12) —porm( 7, 8085 12) Toda Coenres™ expecta) 2/opected) fe the fens bt ot) par tr iagentonereed) 1 2 farthest cls, P(X < 60) = paoen(0,8068, 12). eee eee + Since we estimating 2 pater and we hove eight dass inthe feguny abe a oe ee are© The Kappa Statistic Measuring the Strength of Association Survey 2 Survey 1 | Yes _No | Tota! Yes [mn mia No__| maim | ma ic a C be response categorie. + Leto; = ny/ and by = m/e the probabilities of response centegary forthe frst and second survey, Let ny = (ns +na)/n be observed concordance rate > The expected concordance rate fe pe = EL a TT Measuring the Strength of Association + We at inerested in the degree of association between to categorical vanabies When the same variables messured more than once, the dagree of association between the multiple measurements tls us about the reproducbilty ofthe survey (also known 3 relly). > Another name fer such association is concordance a there i ratural paring relsorship among the observed valves, Ce TT ‘The Kappa Statistic > A measure of reproccbilty of = suey is =P T= Fr The quanti «interpreted 2s ollons ‘> 076 -+ Exealnt Reproducbilny O4-¢ 20.75» Good Reproducity Om 04+ Marginal Repoducbilsy + Reade: nibrery (ert) Bibrary(wea) c-tablowwacrix(e(ayc,by4), nrow appaCe.tabie)co ‘The Kappa Statistic * Amount of best consumption reported by S87 female Amencan nurses at too diferent survey, Survey 2 Survey 1_| < 1 swings > L servings | Tota Zisevinaginee] 13 7 BE 5 iservngiek| 69 20 ‘oo Tat 2 Br 7 ‘The Kappa Statistic Compute she Kappa statistic and assess its statistical sigiicance > asbesey(eri) > Rbrary(oe) > Rappataatrix(e(136, 69, 92, 240), nrov = 2)) vale Ae Pele) ‘netgnted 013782 0.04085 9.361 8, 7054-2