{"id":34958,"date":"2019-10-31T22:01:31","date_gmt":"2019-10-31T19:01:31","guid":{"rendered":"https:\/\/prohoster.info\/blog\/kogda-stoit-proveryat-gipotezu-o-ne-menshej-effektivnosti\/"},"modified":"2019-10-31T22:01:31","modified_gmt":"2019-10-31T19:01:31","slug":"kogda-stoit-proveryat-gipotezu-o-ne-menshej-effektivnosti","status":"publish","type":"post","link":"https:\/\/prohoster.info\/cs\/blog\/news\/kogda-stoit-proveryat-gipotezu-o-ne-menshej-effektivnosti","title":{"rendered":"Kdy je dobr\u00e9 ov\u011b\u0159it hypot\u00e9zu o minim\u00e1ln\u011b stejn\u00e9 \u00fa\u010dinnosti?","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

\"Kdy
\n\u010cl\u00e1nek od t\u00fdmu Stitch Fix navrhuje pou\u017e\u00edt p\u0159\u00edstup klinick\u00fdch studi\u00ed s indik\u00e1torem nehor\u0161\u00edho v\u00fdsledku (non-inferiority trials) v marketingov\u00fdch a produktov\u00fdch A\/B testech. Tento p\u0159\u00edstup je skute\u010dn\u011b aplikovateln\u00fd, kdy\u017e testujeme nov\u00e9 \u0159e\u0161en\u00ed, kter\u00e9 m\u00e1 v\u00fdhody, kter\u00e9 testy nemohou zm\u011b\u0159it.<\/p>\n

Nejsnaz\u0161\u00ed p\u0159\u00edklad je sn\u00ed\u017een\u00ed n\u00e1klad\u016f. Nap\u0159\u00edklad automatizujeme proces prvn\u00ed lekce, ale nechceme p\u0159\u00edli\u0161 sn\u00ed\u017eit celkovou konverzi. Nebo testujeme zm\u011bny zam\u011b\u0159en\u00e9 na ur\u010dit\u00fd segment u\u017eivatel\u016f, p\u0159i\u010dem\u017e sledujeme, aby konverze v jin\u00fdch segmentech neklesla p\u0159\u00edli\u0161 (p\u0159i testov\u00e1n\u00ed n\u011bkolika hypot\u00e9z nezapom\u00edn\u00e1me na \u00fapravy).<\/p>\n

V\u00fdb\u011br spr\u00e1vn\u00e9 hranice nehor\u0161\u00edho v\u00fdsledku p\u0159id\u00e1v\u00e1 dal\u0161\u00ed obt\u00ed\u017ee p\u0159i navrhov\u00e1n\u00ed testu. Ot\u00e1zka, jak vyb\u00edrat \u0394, nen\u00ed v \u010dl\u00e1nku dostate\u010dn\u011b podrobn\u011b rozebr\u00e1na. Zd\u00e1 se, \u017ee tento v\u00fdb\u011br nen\u00ed zcela transparentn\u00ed ani v klinick\u00fdch studi\u00edch. P\u0159ehled<\/a><\/noindex> Podle medic\u00ednsk\u00fdch publikac\u00ed k t\u00e9matu non-inferiority pouze v polovin\u011b p\u0159\u00edpad\u016f b\u00fdv\u00e1 v\u00fdb\u011br hranice od\u016fvodn\u011bn, a \u010dasto jsou tato od\u016fvodn\u011bn\u00ed nejasn\u00e1 nebo nedostate\u010dn\u00e1.<\/p>\n

V ka\u017ed\u00e9m p\u0159\u00edpad\u011b se tento p\u0159\u00edstup zd\u00e1 zaj\u00edmav\u00fd, proto\u017ee sn\u00ed\u017een\u00edm pot\u0159ebn\u00e9 velikosti vzorku m\u016f\u017ee zrychlit testov\u00e1n\u00ed a tedy i rychlost rozhodov\u00e1n\u00ed. \u2014 Daria Mukhina, produktov\u00e1 analyti\u010dka mobiln\u00ed aplikace Skyeng.<\/i>
\n<\/a><\/noindex>
\nT\u00fdm Stitch Fix r\u00e1d testuje r\u016fzn\u00e9 v\u011bci. Cel\u00e1 technologick\u00e1 komunita obecn\u011b r\u00e1da prov\u00e1d\u00ed testy. Kter\u00e1 verze webu p\u0159itahuje v\u00edce u\u017eivatel\u016f \u2013 A nebo B? P\u0159in\u00e1\u0161\u00ed verze A doporu\u010dovac\u00ed model v\u00edc pen\u011bz ne\u017e verze B? T\u00e9m\u011b\u0159 v\u017edy p\u0159i ov\u011b\u0159ov\u00e1n\u00ed hypot\u00e9z pou\u017e\u00edv\u00e1me nejjednodu\u0161\u0161\u00ed p\u0159\u00edstup z z\u00e1kladn\u00edho kurzu statistiky: <\/p>\n

\"Kdy
\n
\nA\u010dkoli tento term\u00edn z\u0159\u00eddka pou\u017e\u00edv\u00e1me, tento typ testov\u00e1n\u00ed se naz\u00fdv\u00e1 \u201eov\u011b\u0159en\u00ed hypot\u00e9zy o p\u0159evaze\u201c. P\u0159i tomto p\u0159\u00edstupu p\u0159edpokl\u00e1d\u00e1me, \u017ee mezi dv\u011bma variantami nen\u00ed \u017e\u00e1dn\u00fd rozd\u00edl. Dr\u017e\u00edme se t\u00e9to my\u0161lenky a vzd\u00e1v\u00e1me se j\u00ed pouze tehdy, kdy\u017e jsou z\u00edskan\u00e1 data dostate\u010dn\u011b p\u0159esv\u011bd\u010div\u00e1, tedy ukazuj\u00ed, \u017ee jedna z variant (A nebo B) je lep\u0161\u00ed ne\u017e druh\u00e1.<\/p>\n

Ov\u011b\u0159en\u00ed hypot\u00e9zy o p\u0159evaze je vhodn\u00e9 pro \u0159e\u0161en\u00ed mnoha probl\u00e9m\u016f. Vyd\u00e1v\u00e1me verzi B doporu\u010duj\u00edc\u00edho modelu pouze tehdy, pokud je o\u010dividn\u011b lep\u0161\u00ed ne\u017e st\u00e1vaj\u00edc\u00ed verze A. Av\u0161ak v n\u011bkter\u00fdch p\u0159\u00edpadech tento p\u0159\u00edstup nefunguje tak dob\u0159e. Pod\u00edvejme se na n\u011bkolik p\u0159\u00edklad\u016f.<\/p>\n

1) Pou\u017e\u00edv\u00e1me extern\u00ed slu\u017ebu<\/b>, kter\u00e1 pom\u00e1h\u00e1 identifikovat fale\u0161n\u00e9 bankovn\u00ed karty. Na\u0161li jsme jinou slu\u017ebu, kter\u00e1 stoj\u00ed v\u00fdrazn\u011b m\u00e9n\u011b. Pokud levn\u011bj\u0161\u00ed slu\u017eba funguje stejn\u011b dob\u0159e jako ta, kterou aktu\u00e1ln\u011b pou\u017e\u00edv\u00e1me, vybereme si ji. Nen\u00ed nutn\u00e9, aby byla lep\u0161\u00ed ne\u017e ta, kterou nyn\u00ed pou\u017e\u00edv\u00e1me. <\/p>\n

2) Chceme se vzd\u00e1t zdroje dat<\/b> A a nahradit ho zdrojem dat B. Mohli bychom odlo\u017eit odchod od A, pokud B vykazuje velmi \u0161patn\u00e9 v\u00fdsledky, ale pokra\u010dov\u00e1n\u00ed v pou\u017e\u00edv\u00e1n\u00ed A nen\u00ed mo\u017en\u00e9. <\/p>\n

3) Cht\u011bli bychom p\u0159ej\u00edt od p\u0159\u00edstupu k modelov\u00e1n\u00ed<\/b>A k p\u0159\u00edstupu B ne proto, \u017ee o\u010dek\u00e1v\u00e1me lep\u0161\u00ed v\u00fdsledky od B, ale proto\u017ee n\u00e1m to poskytuje v\u011bt\u0161\u00ed operativn\u00ed flexibilitu. Nem\u00e1me d\u016fvod se domn\u00edvat, \u017ee B bude hor\u0161\u00ed, ale nep\u0159ejdeme na n\u011bj, pokud tomu tak bude.<\/p>\n

4) U\u010dinili jsme n\u011bkolik kvalitativn\u00edch zm\u011bn<\/b> v designu webov\u00e9 str\u00e1nky (verze B) a domn\u00edv\u00e1me se, \u017ee tato verze p\u0159evy\u0161uje verzi A. Neo\u010dek\u00e1v\u00e1me \u017e\u00e1dn\u00e9 zm\u011bny v konverzi nebo jak\u00fdchkoli kl\u00ed\u010dov\u00fdch ukazatel\u00edch v\u00fdkonnosti, podle kter\u00fdch obvykle hodnot\u00edme web. Ale v\u011b\u0159\u00edme, \u017ee existuj\u00ed v\u00fdhody v parametrech, kter\u00e9 jsou bu\u010f nem\u011b\u0159iteln\u00e9, nebo nem\u00e1me technologie k jejich m\u011b\u0159en\u00ed.<\/p>\n

Ve v\u0161ech t\u011bchto p\u0159\u00edpadech nen\u00ed zkoum\u00e1n\u00ed p\u0159evahy nejvhodn\u011bj\u0161\u00edm \u0159e\u0161en\u00edm. P\u0159esto v\u011bt\u0161ina odborn\u00edk\u016f v takov\u00fdch situac\u00edch pou\u017e\u00edv\u00e1 tento p\u0159\u00edstup jako v\u00fdchoz\u00ed. \u0158\u00e1dn\u011b prov\u00e1d\u00edme experiment, abychom spr\u00e1vn\u011b ur\u010dili velikost \u00fa\u010dinku. Kdyby bylo pravdou, \u017ee verze A a B funguj\u00ed velmi podobn\u011b, je pravd\u011bpodobn\u00e9, \u017ee bychom nedok\u00e1zali odm\u00edtnout nulovou hypot\u00e9zu. Doch\u00e1z\u00edme k z\u00e1v\u011bru, \u017ee A a B ve sv\u00e9 podstat\u011b funguj\u00ed stejn\u011b? Ne! Neschopnost odm\u00edtnout nulovou hypot\u00e9zu a p\u0159ijet\u00ed nulov\u00e9 hypot\u00e9zy nen\u00ed tot\u00e9\u017e.<\/p>\n

V\u00fdpo\u010dty objemu vzorku (kter\u00e9 jste, samoz\u0159ejm\u011b, provedli) se obvykle prov\u00e1d\u011bj\u00ed s p\u0159\u00edsn\u011bj\u0161\u00edmi mezemi pro chybu prvn\u00edho druhu (pravd\u011bpodobnost chybn\u00e9ho zam\u00edtnut\u00ed nulov\u00e9 hypot\u00e9zy, \u010dasto ozna\u010dovan\u00e1 jako alfa) ne\u017e pro chybu druh\u00e9ho druhu (pravd\u011bpodobnost neschopnosti zam\u00edtnout nulovou hypot\u00e9zu, za p\u0159edpokladu, \u017ee nulov\u00e1 hypot\u00e9za je chybn\u00e1, \u010dasto ozna\u010dovan\u00e1 jako beta). Typick\u00e1 hodnota pro alfa je 0,05, zat\u00edmco typick\u00e1 hodnota pro beta je 0,20, co\u017e odpov\u00edd\u00e1 statistick\u00e9 s\u00edle 0,80. To znamen\u00e1, \u017ee m\u016f\u017eeme neodhalit skute\u010dn\u00fd vliv velikosti, kterou jsme uvedli ve sv\u00fdch v\u00fdpo\u010dtech s\u00edly, s pravd\u011bpodobnost\u00ed 20 % a to je docela v\u00e1\u017en\u00fd nedostatek informac\u00ed. Jako p\u0159\u00edklad zva\u017eme n\u00e1sleduj\u00edc\u00ed hypot\u00e9zy:<\/p>\n

\"Kdy
\n
\nH0: m\u016fj batoh NEN\u00cd v m\u00e9 m\u00edstnosti (3)
\nH1: m\u016fj batoh je v m\u00e9 m\u00edstnosti (4)<\/i><\/p>\n

Pokud jsem prohledal svou m\u00edstnost a na\u0161el sv\u016fj batoh \u2014 skv\u011ble, mohu zam\u00edtnout nulovou hypot\u00e9zu. Ale pokud jsem prozkoumal m\u00edstnost a nena\u0161el sv\u016fj batoh (obr\u00e1zek 1), jak\u00fd z\u00e1v\u011br bych m\u011bl u\u010dinit? Jsem si jist\u00fd, \u017ee tam nen\u00ed? Dostate\u010dn\u011b d\u016fkladn\u011b jsem hledal? Co kdy\u017e jsem prohledal pouze 80 % m\u00edstnosti? U\u010dinit z\u00e1v\u011br, \u017ee batoh v m\u00edstnosti ur\u010dit\u011b nen\u00ed, by bylo un\u00e1hlen\u00e9 rozhodnut\u00ed. Nen\u00ed divu, \u017ee nem\u016f\u017eeme \u201ep\u0159ijmout nulovou hypot\u00e9zu\u201c.
\n\"Kdy
\nOblast, kterou jsme prohledali
\nNena\u0161li jsme batoh \u2014 m\u00e1me p\u0159ijmout nulovou hypot\u00e9zu?<\/i><\/p>\n

Obr\u00e1zek 1. Prohled\u00e1n\u00ed 80 % m\u00edstnosti je zhruba tot\u00e9\u017e jako prov\u00e9st studii se silou 80 %. Pokud jste nena\u0161li batoh po prozkoum\u00e1n\u00ed 80 % m\u00edstnosti, lze vyvodit z\u00e1v\u011br, \u017ee tam nen\u00ed?<\/i> <\/p>\n

Co by m\u011bl v t\u00e9to situaci ud\u011blat datov\u00fd odborn\u00edk? M\u016f\u017eete v\u00fdrazn\u011b zv\u00fd\u0161it s\u00edlu v\u00fdzkumu, ale pak budete pot\u0159ebovat vzorek mnohem v\u011bt\u0161\u00ed velikosti a v\u00fdsledek bude st\u00e1le neuspokojuj\u00edc\u00ed.<\/p>\n

Na\u0161t\u011bst\u00ed jsou podobn\u00e9 probl\u00e9my ji\u017e dlouho zkoum\u00e1ny ve sv\u011bt\u011b klinick\u00fdch studi\u00ed. P\u0159\u00edpravek B je levn\u011bj\u0161\u00ed ne\u017e p\u0159\u00edpravek A; o\u010dek\u00e1v\u00e1 se, \u017ee p\u0159\u00edpravek B bude m\u00edt m\u00e9n\u011b vedlej\u0161\u00edch \u00fa\u010dink\u016f ne\u017e p\u0159\u00edpravek A; p\u0159\u00edpravek B se snadn\u011bji p\u0159epravuje, proto\u017ee ho nen\u00ed t\u0159eba skladovat v chladni\u010dce, zat\u00edmco p\u0159\u00edpravek A ano. Ov\u011b\u0159\u00edme hypot\u00e9zu o nejm\u00e9n\u011b stejn\u00e9 \u00fa\u010dinnosti. To je pot\u0159eba, abychom prok\u00e1zali, \u017ee verze B je stejn\u011b dobr\u00e1 jako verze A \u2013 alespo\u0148 v ur\u010dit\u00e9m p\u0159edem stanoven\u00e9m limitu \u201enejm\u00e9n\u011b stejn\u00e9 \u00fa\u010dinnosti\u201c, \u0394. O chv\u00edli pozd\u011bji se podrobn\u011bji pod\u00edv\u00e1me na to, jak tento limit stanovit. Ale nyn\u00ed p\u0159edpokl\u00e1dejme, \u017ee se jedn\u00e1 o minim\u00e1ln\u00ed rozd\u00edl, kter\u00fd je prakticky v\u00fdznamn\u00fd (v kontextu klinick\u00fdch zkou\u0161ek se tomu obvykle \u0159\u00edk\u00e1 klinick\u00e1 v\u00fdznamnost).<\/p>\n

Hypot\u00e9zy o nejm\u00e9n\u011b stejn\u00e9 \u00fa\u010dinnosti p\u0159evrac\u00ed v\u0161e naruby:<\/p>\n

\"Kdy
\n
\nNyn\u00ed m\u00edsto toho, abychom p\u0159edpokl\u00e1dali, \u017ee mezi nimi nen\u00ed rozd\u00edl, p\u0159edpokl\u00e1d\u00e1me, \u017ee verze B je hor\u0161\u00ed ne\u017e verze A, a budeme se t\u00e9to hypot\u00e9zy dr\u017eet, dokud neprok\u00e1\u017eeme, \u017ee tomu tak nen\u00ed. To je p\u0159esn\u011b ten okam\u017eik, kdy m\u00e1 smysl pou\u017e\u00edt jednostrann\u00e9 testov\u00e1n\u00ed hypot\u00e9zy! V praxi to lze ud\u011blat vytvo\u0159en\u00edm d\u016fv\u011bryhodn\u00e9ho intervalu a ur\u010den\u00edm, zda je interval skute\u010dn\u011b v\u011bt\u0161\u00ed ne\u017e \u0394 (obr\u00e1zek 2).
\n\"Kdy
\n<\/p>\n

V\u00fdb\u011br \u0394<\/h3>\n

\nJak spr\u00e1vn\u011b zvolit \u0394? Proces v\u00fdb\u011bru \u0394 zahrnuje statistick\u00e9 od\u016fvodn\u011bn\u00ed a odborn\u00e9 posouzen\u00ed. Ve sv\u011bt\u011b klinick\u00fdch studi\u00ed existuj\u00ed normativn\u00ed doporu\u010den\u00ed, podle kter\u00fdch by delta m\u011bla p\u0159edstavovat nejmen\u0161\u00ed klinicky v\u00fdznamn\u00fd rozd\u00edl \u2013 tedy takov\u00fd, kter\u00fd bude m\u00edt v\u00fdznam v praxi. Zde je cit\u00e1t z evropsk\u00e9ho pokynu, kter\u00fd v\u00e1m m\u016f\u017ee pomoci ov\u011b\u0159it: \u201ePokud byla rozd\u00edl zvolena spr\u00e1vn\u011b, d\u016fv\u011bryhodn\u00fd interval, kter\u00fd le\u017e\u00ed zcela mezi \u2013\u2206 a 0\u2026, st\u00e1le posta\u010duje k prok\u00e1z\u00e1n\u00ed nejm\u00e9n\u011b stejn\u00e9 \u00fa\u010dinnosti. Pokud se tento v\u00fdsledek nezd\u00e1 p\u0159ijateln\u00fd, znamen\u00e1 to, \u017ee \u2206 nebyla spr\u00e1vn\u011b zvolena.\u201c<\/p>\n

Delta rozhodn\u011b nesm\u00ed p\u0159es\u00e1hnout velikost efektu verze A ve srovn\u00e1n\u00ed s prav\u00fdm kontroln\u00edm skupinou (placebo \/ absence l\u00e9\u010dby), proto\u017ee to n\u00e1s vede k z\u00e1v\u011bru, \u017ee verze B je hor\u0161\u00ed ne\u017e prav\u00fd kontrol a z\u00e1rove\u0148 vykazuje \u201enejm\u00e9n\u011b stejnou \u00fa\u010dinnost\u201c. P\u0159edpokl\u00e1dejme, \u017ee kdy\u017e byla p\u0159edstavena verze A, na jej\u00edm m\u00edst\u011b byla verze 0 nebo funkce v\u016fbec neexistovala (viz obr\u00e1zek 3). <\/p>\n

Na z\u00e1klad\u011b v\u00fdsledk\u016f testov\u00e1n\u00ed hypot\u00e9zy o nad\u0159azenosti byla zji\u0161t\u011bna velikost efektu E (tj. p\u0159edpokl\u00e1dan\u011b \u03bc^A\u2212\u03bc^0=E). Nyn\u00ed je A na\u0161\u00edm nov\u00fdm standardem a chceme se ujistit, \u017ee B nen\u00ed hor\u0161\u00ed ne\u017e A. Dal\u0161\u00ed zp\u016fsob, jak zapsat \u03bcB\u2212\u03bcA\u2264\u2212\u0394 (nulov\u00e1 hypot\u00e9za) \u2014 \u03bcB\u2264\u03bcA\u2212\u0394. Pokud p\u0159edpokl\u00e1d\u00e1me, \u017ee \u010din\u00ed alespo\u0148 E, pak \u03bcB \u2264 \u03bcA\u2212E \u2264 placebo. Nyn\u00ed vid\u00edme, \u017ee na\u0161e hodnota pro \u03bcB zcela p\u0159ekra\u010duje \u03bcA\u2212E, co\u017e t\u00edm p\u00e1dem zcela vyvrac\u00ed nulovou hypot\u00e9zu a umo\u017e\u0148uje n\u00e1m z\u00e1v\u011br, \u017ee B nen\u00ed hor\u0161\u00ed ne\u017e A, ale z\u00e1rove\u0148 \u03bcB m\u016f\u017ee b\u00fdt \u2264 \u03bc placeba, co\u017e nen\u00ed to, co pot\u0159ebujeme. (obr\u00e1zek 3).<\/p>\n

\"Kdy
\nObr\u00e1zek 3. Uk\u00e1zka rizik volby hranice nejmen\u0161\u00ed \u00fa\u010dinnosti. Pokud je limit p\u0159\u00edli\u0161 vysok\u00fd, m\u016f\u017eeme doj\u00edt k z\u00e1v\u011bru, \u017ee B nen\u00ed hor\u0161\u00ed ne\u017e A, ale z\u00e1rove\u0148 se neli\u0161\u00ed od placeba. Nebudeme m\u011bnit l\u00e9k, kter\u00fd je zjevn\u011b \u00fa\u010dinn\u011bj\u0161\u00ed ne\u017e placebo (A), za l\u00e9k, kter\u00fd m\u00e1 stejnou \u00fa\u010dinnost jako placebo. <\/i><\/p>\n

Volba \u03b1 <\/h3>\n

\nP\u0159ejd\u011bme k volb\u011b \u03b1. M\u016f\u017eete pou\u017e\u00edt standardn\u00ed hodnotu \u03b1 = 0,05, ale to nen\u00ed zcela f\u00e9r. Nap\u0159\u00edklad kdy\u017e nakupujete n\u011bco online a pou\u017eijete najednou n\u011bkolik slevov\u00fdch k\u00f3d\u016f, i kdy\u017e by se nem\u011bly s\u010d\u00edtat \u2014 prost\u011b se v\u00fdvoj\u00e1\u0159i dopustili chyby a V\u00e1m to pro\u0161lo. Podle pravidel by m\u011bla b\u00fdt hodnota \u03b1 rovna polovin\u011b hodnoty \u03b1, kter\u00e1 se pou\u017e\u00edv\u00e1 p\u0159i testov\u00e1n\u00ed hypot\u00e9zy o nad\u0159azenosti, tj. 0,05 \/ 2 = 0,025.<\/p>\n

Velikost vzorku <\/h3>\n

\nJak odhadnout velikost vzorku? Pokud se domn\u00edv\u00e1te, \u017ee skute\u010dn\u00fd rozd\u00edl pr\u016fm\u011br\u016f mezi A a B je 0, pak v\u00fdpo\u010det velikosti vzorku bude stejn\u00fd jako p\u0159i testov\u00e1n\u00ed hypot\u00e9zy o nad\u0159azenosti, krom\u011b toho, \u017ee nahrad\u00edte velikost efektu limitem nejmen\u0161\u00ed \u00fa\u010dinnosti, za p\u0159edpokladu, \u017ee pou\u017e\u00edv\u00e1te \u03b1 nejmen\u0161\u00ed \u00fa\u010dinnost = 1\/2 \u03b1 nad\u0159azenost <\/b>(\u03b1neinvazivita=1\/2\u03b1superiorita). Pokud m\u00e1te d\u016fvod se domn\u00edvat, \u017ee varianta B m\u016f\u017ee b\u00fdt m\u00edrn\u011b hor\u0161\u00ed ne\u017e varianta A, ale chcete dok\u00e1zat, \u017ee nen\u00ed hor\u0161\u00ed v\u00edce ne\u017e \u0394, m\u00e1te \u0161t\u011bst\u00ed! Ve skute\u010dnosti to zmen\u0161uje velikost va\u0161eho vzorku, proto\u017ee je snaz\u0161\u00ed prok\u00e1zat, \u017ee B je hor\u0161\u00ed ne\u017e A, pokud skute\u010dn\u011b v\u011b\u0159\u00edte, \u017ee je trochu hor\u0161\u00ed a ne rovnocenn\u00fd.<\/p>\n

P\u0159\u00edklad s rozhodnut\u00edm <\/h3>\n

\nP\u0159edpokl\u00e1dejme, \u017ee chcete p\u0159ej\u00edt na verzi B za p\u0159edpokladu, \u017ee je hor\u0161\u00ed ne\u017e verze A maxim\u00e1ln\u011b o 0,1 bodu na 5bodov\u00e9 \u0161k\u00e1le spokojenosti z\u00e1kazn\u00edk\u016f\u2026 P\u0159istupme k t\u00e9to \u00faloze pomoc\u00ed hypot\u00e9zy o superiorit\u011b.<\/p>\n

Pro testov\u00e1n\u00ed hypot\u00e9zy o superiorit\u011b bychom po\u010d\u00edtali velikost vzorku n\u00e1sleduj\u00edc\u00edm zp\u016fsobem:<\/p>\n

\"Kdy
\n
\nTo znamen\u00e1, \u017ee pokud m\u00e1te ve skupin\u011b 2103 pozorov\u00e1n\u00ed, m\u016f\u017eete si b\u00fdt s 90 % jist\u00ed, \u017ee odhal\u00edte efekt velikosti 0,10 nebo vy\u0161\u0161\u00ed. Ale pokud je hodnota 0,10 pro v\u00e1s p\u0159\u00edli\u0161 velk\u00e1, mo\u017en\u00e1 nebude st\u00e1t za to testovat hypot\u00e9zu o superiorit\u011b pro tuto hodnotu. Mo\u017en\u00e1 se pro spolehlivost rozhodnete prov\u00e9st studii pro men\u0161\u00ed velikost efektu, nap\u0159\u00edklad 0,05. V takov\u00e9m p\u0159\u00edpad\u011b budete pot\u0159ebovat 8407 pozorov\u00e1n\u00ed, co\u017e znamen\u00e1, \u017ee se velikost vzorku t\u00e9m\u011b\u0159 \u010dty\u0159ikr\u00e1t zv\u00fd\u0161\u00ed. Ale co kdy\u017e se budeme dr\u017eet na\u0161\u00ed p\u016fvodn\u00ed velikosti vzorku, ale zv\u00fd\u0161\u00edme s\u00edlu na 0,99, abychom si byli jisti, \u017ee dostaneme pozitivn\u00ed v\u00fdsledek? V takov\u00e9m p\u0159\u00edpad\u011b n pro jednu skupinu bude 3676, co\u017e u\u017e je lep\u0161\u00ed, ale zvy\u0161uje velikost vzorku o v\u00edce ne\u017e 50 %. A jako v\u00fdsledek stejn\u011b nebudeme schopni vyvr\u00e1tit nulovou hypot\u00e9zu, co\u017e nep\u0159inese odpov\u011b\u010f na na\u0161i ot\u00e1zku.<\/p>\n

Co kdy\u017e m\u00edsto toho otestujeme hypot\u00e9zu o ne men\u0161\u00ed efektivit\u011b? <\/p>\n

\"Kdy
\n
\nVelikost vzorku se bude po\u010d\u00edtat podle stejn\u00e9ho vzorce, s v\u00fdjimkou jmenovatele.
\nRozd\u00edly od vzorce pou\u017e\u00edvan\u00e9ho p\u0159i testov\u00e1n\u00ed hypot\u00e9zy o superiorit\u011b jsou n\u00e1sleduj\u00edc\u00ed:<\/p>\n

\u2014 Z1\u2212\u03b1\/2 je nahrazen Z1\u2212\u03b1, ale pokud to d\u011bl\u00e1te podle pravidel, nahrazuje \u03b1 = 0,05 \u03b1 = 0,025, tedy je to stejn\u00e9 \u010d\u00edslo (1,96)<\/p>\n

\u2014 v jmenovateli se objevuje (\u03bcB\u2212\u03bcA)<\/p>\n

\u2014 \u03b8 (velikost efektu) je nahrazeno \u0394 (hranice ne men\u0161\u00ed efektivity)<\/p>\n

Pokud p\u0159edpokl\u00e1d\u00e1me, \u017ee \u00b5B = \u00b5A, pak (\u00b5B \u2212 \u00b5A) = 0 a v\u00fdpo\u010det velikosti vzorku pro limitn\u00ed efektivitu je pr\u00e1v\u011b to, co bychom z\u00edskali p\u0159i v\u00fdpo\u010dtu p\u0159evahy pro velikost efektu 0,1, skv\u011bl\u00e9! M\u016f\u017eeme prov\u00e9st studii stejn\u00e9 velikosti s r\u016fzn\u00fdmi hypot\u00e9zami a jin\u00fdm p\u0159\u00edstupem k z\u00e1v\u011br\u016fm, a dostaneme odpov\u011b\u010f na ot\u00e1zku, na kterou chceme opravdu odpov\u011bd\u011bt.<\/p>\n

Nyn\u00ed p\u0159edpokl\u00e1dejme, \u017ee skute\u010dn\u011b nepova\u017eujeme, \u017ee \u00b5B = \u00b5A, a
\nmysl\u00edme si, \u017ee \u00b5B je trochu hor\u0161\u00ed, mo\u017en\u00e1 o 0,01 jednotky. To zvy\u0161uje n\u00e1\u0161 jmenovatel a sni\u017euje velikost vzorku pro skupinu na 1737.<\/p>\n

Co se stane, pokud je verze B ve skute\u010dnosti lep\u0161\u00ed ne\u017e verze A? Vyvrac\u00edme nulovou hypot\u00e9zu, \u017ee B je hor\u0161\u00ed ne\u017e A, o v\u00edce ne\u017e \u0394 a p\u0159ijmeme alternativn\u00ed hypot\u00e9zu, \u017ee B, pokud je hor\u0161\u00ed, nen\u00ed hor\u0161\u00ed ne\u017e \u0394, a m\u016f\u017ee b\u00fdt lep\u0161\u00ed. Zkuste toto z\u00e1v\u011br zahrnout do cross-funk\u010dn\u00ed prezentace a pod\u00edvejte se, co z toho vzejde (opravdu, zkuste to). V situaci, kdy je t\u0159eba orientovat se na perspektivu, nikdo nechce souhlasit s \u201ehor\u0161\u00ed nen\u00ed v\u00edce ne\u017e \u0394 a m\u016f\u017ee b\u00fdt lep\u0161\u00ed\u201c.<\/p>\n

V tomto p\u0159\u00edpad\u011b m\u016f\u017eeme prov\u00e9st studii, kter\u00e1 se velmi stru\u010dn\u011b naz\u00fdv\u00e1 \u201eov\u011b\u0159en\u00ed hypot\u00e9zy, \u017ee jedna z variant p\u0159evy\u0161uje druhou nebo j\u00ed ustupuje\u201c. Pou\u017e\u00edvaj\u00ed se dva soubory hypot\u00e9z:<\/p>\n

Prvn\u00ed soubor (stejn\u011b jako p\u0159i ov\u011b\u0159ov\u00e1n\u00ed hypot\u00e9zy o nedostate\u010dn\u00e9 efektivit\u011b): <\/p>\n

\"Kdy
\n
\nDruh\u00fd soubor (stejn\u011b jako p\u0159i ov\u011b\u0159ov\u00e1n\u00ed hypot\u00e9zy o p\u0159evaze):<\/p>\n

\"Kdy
\n
\nDruhou hypot\u00e9zu testujeme pouze v p\u0159\u00edpad\u011b, \u017ee prvn\u00ed byla zam\u00edtnuta. P\u0159i sekven\u010dn\u00edm testov\u00e1n\u00ed udr\u017eujeme celkovou \u00farove\u0148 chyb prvn\u00edho druhu (\u03b1). V praxi to lze dos\u00e1hnout t\u00edm, \u017ee vytvo\u0159\u00edme 95% konfidenci\u00e1ln\u00ed interval pro rozd\u00edl mezi pr\u016fm\u011bry a ov\u011b\u0159\u00edme, zda cel\u00fd interval p\u0159ekra\u010duje -\u0394. Pokud interval nep\u0159esahuje -\u0394, nem\u016f\u017eeme zam\u00edtnout nulovou hodnotu a p\u0159estaneme. Pokud cel\u00fd interval skute\u010dn\u011b p\u0159esahuje \u2212\u0394, budeme pokra\u010dovat a zjist\u00edme, obsahuje-li interval 0. <\/p>\n

Existuje je\u0161t\u011b jeden typ v\u00fdzkum\u016f, kter\u00fd jsme neprobrali \u2013 studie ekvivalence. <\/p>\n

V\u00fdzkumy tohoto typu mohou b\u00fdt nahrazeny studiemi pro ov\u011b\u0159en\u00ed hypot\u00e9zy o nezmen\u0161en\u00e9 \u00fa\u010dinnosti a naopak, ale maj\u00ed jedno d\u016fle\u017eit\u00e9 odli\u0161en\u00ed. Zkou\u0161ka pro ov\u011b\u0159en\u00ed hypot\u00e9zy o nezmen\u0161en\u00e9 \u00fa\u010dinnosti m\u00e1 za c\u00edl prok\u00e1zat, \u017ee varianta B je alespo\u0148 tak dobr\u00e1 jako A. A studie ekvivalence m\u00e1 za c\u00edl prok\u00e1zat, \u017ee varianta B je alespo\u0148 tak dobr\u00e1 jako A, a varianta A je tak\u00e9 tak dobr\u00e1 jako B, co\u017e je slo\u017eit\u011bj\u0161\u00ed. V podstat\u011b se sna\u017e\u00edme ur\u010dit, jestli le\u017e\u00ed cel\u00fd interval spolehlivosti pro rozd\u00edl pr\u016fm\u011br\u016f mezi \u2212\u0394 a \u0394. Takov\u00e9 studie vy\u017eaduj\u00ed v\u011bt\u0161\u00ed velikost vzorku a prov\u00e1d\u011bj\u00ed se m\u00e9n\u011b \u010dasto. Proto p\u0159\u00ed\u0161t\u011b, kdy\u017e budete prov\u00e1d\u011bt studii, jej\u00ed\u017e hlavn\u00edm c\u00edlem je ujistit se, \u017ee nov\u00e1 verze nen\u00ed hor\u0161\u00ed, nesouhlas\u00edte s \"nemo\u017enost\u00ed vyvr\u00e1tit nulovou hypot\u00e9zu\". Pokud chcete ov\u011b\u0159it skute\u010dn\u011b d\u016fle\u017eitou hypot\u00e9zu, zva\u017ete r\u016fzn\u00e9 mo\u017enosti.
\n
Zdroj: habr.com<\/a><\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"

\u0421\u0442\u0430\u0442\u044c\u044f \u043e\u0442 \u043a\u043e\u043c\u0430\u043d\u0434\u044b Stitch Fix \u043f\u0440\u0435\u0434\u043b\u0430\u0433\u0430\u0435\u0442 \u0438\u0441\u043f\u043e\u043b\u044c\u0437\u043e\u0432\u0430\u0442\u044c \u043f\u043e\u0434\u0445\u043e\u0434 \u043a\u043b\u0438\u043d\u0438\u0447\u0435\u0441\u043a\u0438\u0445 \u0438\u0441\u0441\u043b\u0435\u0434\u043e\u0432\u0430\u043d\u0438\u0439 \u043d\u0435 \u043c\u0435\u043d\u044c\u0448\u0435\u0439 \u044d\u0444\u0444\u0435\u043a\u0442\u0438\u0432\u043d\u043e\u0441\u0442\u0438 (non-inferiority trials) \u0432 \u043c\u0430\u0440\u043a\u0435\u0442\u0438\u043d\u0433\u043e\u0432\u044b\u0445 \u0438 \u043f\u0440\u043e\u0434\u0443\u043a\u0442\u043e\u0432\u044b\u0445 A\/B \u0442\u0435\u0441\u0442\u0430\u0445. \u0422\u0430\u043a\u043e\u0439 \u043f\u043e\u0434\u0445\u043e\u0434 \u0434\u0435\u0439\u0441\u0442\u0432\u0438\u0442\u0435\u043b\u044c\u043d\u043e \u043f\u0440\u0438\u043c\u0435\u043d\u0438\u043c, \u043a\u043e\u0433\u0434\u0430 \u043c\u044b \u0442\u0435\u0441\u0442\u0438\u0440\u0443\u0435\u043c \u043d\u043e\u0432\u043e\u0435 \u0440\u0435\u0448\u0435\u043d\u0438\u0435, \u0438\u043c\u0435\u044e\u0449\u0435\u0435 \u043f\u0440\u0435\u0438\u043c\u0443\u0449\u0435\u0441\u0442\u0432\u0430, \u043d\u0435\u0438\u0437\u043c\u0435\u0440\u044f\u0435\u043c\u044b\u0435 \u0442\u0435\u0441\u0442\u0430\u043c\u0438. \u0421\u0430\u043c\u044b\u0439 \u043f\u0440\u043e\u0441\u0442\u043e\u0439 \u043f\u0440\u0438\u043c\u0435\u0440 \u2014 \u0441\u043d\u0438\u0436\u0435\u043d\u0438\u0435 \u043a\u043e\u0441\u0442\u043e\u0432. \u041d\u0430\u043f\u0440\u0438\u043c\u0435\u0440, \u0430\u0432\u0442\u043e\u043c\u0430\u0442\u0438\u0437\u0438\u0440\u0443\u0435\u043c \u043f\u0440\u043e\u0446\u0435\u0441\u0441 \u043d\u0430\u0437\u043d\u0430\u0447\u0435\u043d\u0438\u044f \u043f\u0435\u0440\u0432\u043e\u0433\u043e \u0443\u0440\u043e\u043a\u0430, \u043d\u043e \u043d\u0435 \u0445\u043e\u0442\u0438\u043c \u0441\u0438\u043b\u044c\u043d\u043e \u0443\u0440\u043e\u043d\u0438\u0442\u044c \u0441\u043a\u0432\u043e\u0437\u043d\u0443\u044e \u043a\u043e\u043d\u0432\u0435\u0440\u0441\u0438\u044e. \u0418\u043b\u0438 \u0442\u0435\u0441\u0442\u0438\u0440\u0443\u0435\u043c […]<\/p>\n","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"author":1,"featured_media":26311,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[702],"tags":[],"class_list":["post-34958","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news"],"aioseo_notices":[],"aioseo_head":"\n\t\t\n\t\n\t\n\t\n\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\n","aioseo_head_json":{"title":"\ud83e\udd47Kdy prov\u00e1d\u011bt ov\u011b\u0159en\u00ed hypot\u00e9zy o nezmen\u0161en\u00e9 \u00fa\u010dinnosti? | ProHoster","description":"\u010cl\u00e1nek od t\u00fdmu Stitch Fix.","canonical_url":"https:\/\/prohoster.info\/cs\/blog\/news\/kogda-stoit-proveryat-gipotezu-o-ne-menshej-effektivnosti","robots":"max-image-preview:large","keywords":"","webmasterTools":{"miscellaneous":""},"schema":null,"og:locale":"cs_CZ","og:site_name":"ProHoster | \u041a\u0443\u043f\u0438\u0442\u044c \u043d\u0430\u0434\u0435\u0436\u043d\u044b\u0439 \u0445\u043e\u0441\u0442\u0438\u043d\u0433 \u0434\u043b\u044f \u0441\u0430\u0439\u0442\u043e\u0432 \u0441 \u0437\u0430\u0449\u0438\u0442\u043e\u0439 \u043e\u0442 DDoS, VPS VDS \u0441\u0435\u0440\u0432\u0435\u0440\u044b","og:type":"article","og:title":"\ud83e\udd47\u041a\u043e\u0433\u0434\u0430 \u0441\u0442\u043e\u0438\u0442 \u043f\u0440\u043e\u0432\u0435\u0440\u044f\u0442\u044c \u0433\u0438\u043f\u043e\u0442\u0435\u0437\u0443 \u043e \u043d\u0435 \u043c\u0435\u043d\u044c\u0448\u0435\u0439 \u044d\u0444\u0444\u0435\u043a\u0442\u0438\u0432\u043d\u043e\u0441\u0442\u0438? | ProHoster","og:description":"\u0421\u0442\u0430\u0442\u044c\u044f \u043e\u0442 \u043a\u043e\u043c\u0430\u043d\u0434\u044b Stitch Fix.","og:url":"https:\/\/prohoster.info\/cs\/blog\/news\/kogda-stoit-proveryat-gipotezu-o-ne-menshej-effektivnosti","og:image":"https:\/\/prohoster.info\/wp-content\/uploads\/2021\/11\/logo-350.jpg","og:image:secure_url":"https:\/\/prohoster.info\/wp-content\/uploads\/2021\/11\/logo-350.jpg","og:image:width":350,"og:image:height":350,"article:published_time":"2019-10-31T19:01:31+00:00","article:modified_time":"2019-10-31T19:01:31+00:00","article:publisher":"https:\/\/www.facebook.com\/prohoster","article:author":"https:\/\/www.facebook.com\/prohoster"},"aioseo_meta_data":{"post_id":"34958","title":null,"description":null,"keywords":null,"keyphrases":null,"primary_term":null,"canonical_url":null,"og_title":null,"og_description":null,"og_object_type":"default","og_image_type":"default","og_image_url":null,"og_image_width":null,"og_image_height":null,"og_image_custom_url":null,"og_image_custom_fields":null,"og_video":null,"og_custom_url":null,"og_article_section":null,"og_article_tags":null,"twitter_use_og":false,"twitter_card":"default","twitter_image_type":"default","twitter_image_url":null,"twitter_image_custom_url":null,"twitter_image_custom_fields":null,"twitter_title":null,"twitter_description":null,"schema":{"blockGraphs":[],"customGraphs":[],"default":{"data":{"Article":[],"Course":[],"Dataset":[],"FAQPage":[],"Movie":[],"Person":[],"Product":[],"ProductReview":[],"Car":[],"Recipe":[],"Service":[],"SoftwareApplication":[],"WebPage":[]},"graphName":"","isEnabled":true},"graphs":[]},"schema_type":null,"schema_type_options":null,"pillar_content":false,"robots_default":true,"robots_noindex":false,"robots_noarchive":false,"robots_nosnippet":false,"robots_nofollow":false,"robots_noimageindex":false,"robots_noodp":false,"robots_notranslate":false,"robots_max_snippet":null,"robots_max_videopreview":null,"robots_max_imagepreview":"large","priority":null,"frequency":null,"local_seo":null,"seo_analyzer_scan_date":"2026-01-21 21:15:19","breadcrumb_settings":null,"limit_modified_date":false,"reviewed_by":null,"ai":null,"created":"2021-03-01 02:12:27","updated":"2026-01-21 21:15:19","focus_keyword":null,"additional_keywords":null,"truseo_locale":null},"gt_translate_keys":[{"key":"link","format":"url"}],"_links":{"self":[{"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/posts\/34958","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/comments?post=34958"}],"version-history":[{"count":0,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/posts\/34958\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/media\/26311"}],"wp:attachment":[{"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/media?parent=34958"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/categories?post=34958"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/tags?post=34958"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}