Maitiro ekushandisa Propensity Score Matching: A Danho-ne-Nhanho Gwaro

Munguva ichangopfuura ndanga ndichishanda pakuedza kwakavakirwa paPropensity Score Matching uye ndichiri kutsvaga ruzivo ndakasangana nekushaikwa kwezvinhu pamusoro penyaya. Zvizhinji zvezvinyorwa zvandakawana zviri pamusoro pekubudirira kweiyo nzira uye hazvina kunyatsotsanangurwa maererano nedzidziso. Naizvozvo, ndafunga kugovera newe gwara rakazara pane Propensity Score Matching chimiro nematanho ayo.

Chii chinonzi Propensity Score Matching uye nei uchiishandisa?

"Propensity mamakirwo ekufananidza kunosanganisira kuumba seti dzakafanana dzezvidzidzo zvakarapwa uye zvisina kurapwa zvinogovana kukosha kwakafanana kweiyo propensity mamaki. Kana sampu inofananidzwa yaumbwa, mhedzisiro yekurapa inogona kufungidzirwa nekuenzanisa zvakananga mhedzisiro. "

Tsanangudzo iyi yakatanga kupihwa naRosenbaum PR, Rubin DB muchinyorwa "Kuongorora kunzwisiswa kune isina kucherechedzwa bhinari covariate muchidzidzo chekucherechedza nemhedzisiro" ye1983.

Kuzvitaura zviri nyore, iyi imwe yekuwedzera A/B bvunzo nzira inoshandiswa kana sampuro randomisation ikasashanda . Propensity mamakisi (mukana wekupihwa kuboka rekuyedza) reboka rekurapa rinoverengerwa kune wese mushandisi uyezve mushandisi anofananidzwa nemumwe mushandisi zvichibva pane zvakaitika kare zvekushandiswa kwechigadzirwa kuumba boka rinodzora. Mushure mezvo, mhedzisiro yemapoka maviri inofananidzwa uchishandisa bvunzo yenhamba uye chiitiko chekuyedza chinoyerwa.

Asi sei uchishandisa nzira yakaoma yekutsvaga boka rekutonga kana A / B chikuva ichigona kuzviita panzvimbo? Mune zvimwe zviitiko hazvigoneke kushandisa A/B papuratifomu ine yakavakirwa-mukati kupatsanura basa. Heano maitiro anogona kuitika:

1. Zvinodhura uye zvinopedza nguva kuti bhizinesi riite bvunzo yeA/B nekuda kweakasiyana siyana anoshanda basa uye kutaurirana (hongu, zvinogona kuitika).
2. Kambani inogona kushaya purofiti ichiri kuyedza kuita mari kana kushambadzira-ane chekuita nemabasa / hurongwa kana ikaita yekirasi bvunzo / yekudzora kuyedza.
3. Kazhinji hazvigoneke kuve nesampling isina kurongeka mukuedza pasina Indaneti inodiwa pabvunzo yeA/B.
4. Izvo zvakare hazvishande kuisa bvunzo yeA/B mukuedza kwepamhepo kana chinhu chakayedzwa chisina kugovaniswa.

Ndakanga ndine nyaya yechina mukuita kwangu uye zvakaitika ndichishanda ne e-commerce chigadzirwa. Chikwata chechigadzirwa chaigadzirira kuyedza basa rekupa mabhonasi kune vashandisi mushure mekuisa yekutanga odha. Dambudziko nderekuti basa racho rakanga richishanda kwete kune vese vashandisi vaiisa yekutanga odha. Mimwe mamiriro ezvinhu, akadai sekukosha kwehurongwa, nezvimwewo, zvaifanira kuzadzikiswa. Muchiitiko ichi, zviri kupfuura miganhu yeA/B bvunzo papuratifomu kupatsanura traffic pakati pebvunzo nemapoka ekudzora. Heino chikonzero nei Propensity Score Matching yaive sarudzo.

Framework yePropensity Score Matching

A c yakakwana sisitimu yakavakirwa pachinyorwa " Propensity mamaki inofananidzwa neR: yakajairwa nzira uye maficha matsva "uye ine matanho mashanu (Mufananidzo 2).

Danho rekutanga nderekuunganidza iyo data painofungidzirwa chibodzwa chechimiro uye mushandisi anofananidzwa anowanikwa.

Danho rechipiri ndere kufungidzira zvibodzwa uchishandisa nzira, senge logistic regression, uye kudzidzisa pane dataset kufanotaura kana mushandisi achapihwa kuboka rebvunzo. Kune wese mushandisi, modhi yakadzidziswa inoburitsa mukana wekuve muboka rebvunzo.

Nhanho yechitatu inoreva kuenzanirana kwakavakirwa pane propensity mamaki, apo nzira dzakasiyana dzekufananidza dzinoyedzwa, semuvakidzani wepedyo.

Muchinhanho chechina, kuenzana kwema covariates pakati pekurapa uye mapoka ekudzora anotariswa nekuverenga kuenzanisa nhamba uye kugadzira zvirongwa. Chiyero chisina kunaka chinoratidza kuti modhi inofungidzira propensity mamakisi inoda kutsanangurwa.

Muchishanu chekupedzisira nhanho, mhedzisiro yebvunzo inofungidzirwa kushandisa inofananidzwa data uye bvunzo yehuwandu inoitwa.

Data Collection

Iyi nhanho ndeyekuunganidza zvinodiwa zvinosiyana, covariates uye confounders. Covariate (X) ishanduko yakazvimirira inogona kukanganisa mhedzisiro yekuyedza (Y), asi iyo isiri yekufarira zvakananga. Confounder chimwe chinhu chisiri icho chiri kudzidzwa chinobatanidzwa zvese nekugoverwa kuboka rebvunzo (W) uye nemhedzisiro yekuyedza (Y).

Girafu iri pazasi inoratidza hukama hwezvakasiyana. X ndeye covariate, W chiratidzo chekupihwa kurapwa, uye Y ndiyo mhedzisiro. Girafu riri kuruboshwe rinoratidza hukama hwe confounder uye iyo iri kurudyi inoratidza yakazvimirira kubatana kwecovariate kune mhedzisiro yekuedza (Y) uye yekuyedza kugoverwa kweboka (W).

Iwe unogona kubvunza kuti mangani emhando dzandinofanira kusarudza? Mhinduro iri nyore - zvakanyanya, zvirinani kuitira kuti uwane fungidziro yepamusoro yezvabuda uye kuderedza kurerekera pakudzidza . Uye pano ndiri kutaura nezve nhamba huru se 20-50 kana kutopfuura.

Propensity Score Estimation

Kuenderera mberi kune nhanho inotevera, inodiwa kuunganidza data uye kuseta mureza wekuve weboka rekurapa. Vamwe vese vashandisi vanogona kuumba boka rekutonga. Mushure mezvo propensity mamakisi inofungidzirwa uchishandisa nzira dzakasiyana, senge logistic regression kana masango asina kurongeka.

Zvizhinji zvezvinyorwa zvandakaverenga zvinokurudzira kuomerera kune logistic regression uye kusashandisa mamwe mamodheru akaomarara sezvo kurongeka kwepamusoro hakusi crucia l. Asi, kubudirira kwekufananidza nzira inotarisana nekururama.

Mushure mekusarudza nzira, muenzaniso wekufungidzira unodzidziswa pane data uchishandisa covariates yakasarudzwa kufanotaura kana mushandisi ari weboka rekuyedza. Chekupedzisira, iyo modhi inofanotaura kune wega mushandisi, uye propensity mamaki, mukana wekuve muboka rebvunzo, inoverengerwa. Panyaya yemasoftwares, muPython unogona kushandisa chero raibhurari yekufungidzira kutanga kubva kune basic scikit-dzidza uye kutamira kuMuporofita.

Data Matching

Chiitiko chinotevera ndechekushandisa nzira yekufananidza kutsvaga mushandisi anoenderana nemushandisi kubva muboka rekuyedza. Nokudaro, boka rekutonga rinoumbwa.

Pane nzira dzakasiyana dzekufananidza dzekusarudza kubva, semuenzaniso kunyatsofananidza kana Mahalanobis chinhambwe chinofananidzwa. Muchikamu chino ini ndichanyanya kukurukura nzira yakajairika yekufananidza muvakidzani wepedyo uye nekusiyana kwayo.

Muvakidzani wepedyo anofananidza (NNM) inoumbwa nezvikamu zviviri. Kutanga, iyo algorithm inotora vashandisi, mumwe nemumwe kubva kuboka rekurapa, mune yakatarwa. Zvadaro, kune wega wega mushandisi weboka rekuyedza, iyo algorithm inowana mushandisi muboka rekutonga ane iri pedyo propensity mamakisi. Aya matanho anodzokororwa kusvika pasina vashandisi vasara muyedzo kana mapoka ekutonga. MuPython, kune chaiwo maraibhurari ePSM sePyTorch, Psmpy , causallib . Kana kuti iwe unogara uchigona kunamatira kune chero yekare raibhurari ine anofananidza algorithms.

Pane zvakare sarudzo yekusarudza modhi yeNNM ine kana isina caliper. A caliper inoisa muganho wepamusoro wechinhambwe chezvibodzwa zvepropensity mupeya inofananidzwa. Nekudaro, mushandisi wega wega anogona chete kuenzaniswa nevashandisi veiyo propensity mamaki mukati mechikamu chidiki. Kana vashandisi vanokodzera vasingakwanisi kufananidzwa, mushandisi anoraswa.

Sei ndichifanira kushandisa caliper? Zvinokurudzirwa kuishandisa kana chinhambwe chezvibodzwa muzvikamu zvakafananidzwa chingave chakakura. Paunenge uchifunga nezve saizi yecaliper, funga zvinotevera: kana kuenzanisa kuita kusingagutsi, kuenzanisa kunogona kuitwa neyakaomesesa caliper uye kana kuenzanisa kuchibudirira asi huwandu hwemapeya akafananidzwa idiki, caliper inogona kukwidziridzwa ( https:/ /www.ncbi.nlm.nih.gov/pmc/articles/PMC8246231/ ).

Balance Diagnostics

Munguva iyi inotariswa kana covariates ebvunzo uye akafananidzwa ekudzora mapoka akaenzana, nekudaro, inoti kana mutambo wakarurama.

Pane nzira nhatu dzekuenzanisa diagnostics:

-Nhamba dzinotsanangura: yakamisikidzwa zvinoreva musiyano (SMD) kana musiyano reshiyo (VR)

- nhamba miedzo

- kuona: qq-plot, histogram kana rudo chiitiko

Muchinyorwa ini ndinonyanya kuisa pfungwa pane yekutanga uye yechitatu sarudzo.

Chekutanga, ngatikurukurei yakamisikidzwa kureva musiyano uye musiyano reshiyo. Ndeapi maitiro anoratidza kuti covariate yakaenzana? Ini ndinokurudzira kuti kukosha kweSMD kuri pazasi 0.1 Panyaya yeVR, kukosha kuri pedyo ne1.0 kunoratidza chiyero .

Chechipiri, maererano nemaitiro ekuona, imwe yenhamba inotsanangura iri pamusoro inoverengerwa kune yega yega covariate uye inoratidzwa nemifananidzo. Ini pachangu ndinofarira chirongwa cherudo sezvo vese covariates vanogona kuiswa mugirafu rimwe uye covariates pamberi uye mushure mekufananidza zvinogona kuenzaniswa zviri nyore. Ndinoisa muenzaniso wegirafu pazasi.

Ko kana ma covariates achiri asina kugadzikana mushure mekufananidza? Kuenzanisira, yakajairwa mutsauko (SMD) yecovariates frequency yekutenga uye AOV inotenderedza 0.5, iyo iri pamusoro inodiwa 0.1. Zvinoreva kuti ma covariate haana kuenzana uye kudzokorora kunodiwa.

Pane nzira ina dzekuita patsva kuenzanisa:

1. Wedzera covariates itsva

2. Ingoshandura nzira yekufananidza sezvo pane zvakawanda zvavo

3. Batanidza Propensity Score Matching nechaiyo yekufananidza nzira

4. Wedzera saizi yemuenzaniso

Estimate of Treatment Effects

Chekupedzisira, tave kusvika padanho rekupedzisira apo mhedzisiro yekuyedza inofungidzirwa. Kune kunyanya matatu marudzi efungidziro yemhedzisiro: avhareji yekurapa maitiro (ATE), avhareji kurapwa maitiro pane akarapwa (ATT), uye avhareji kurapwa maitiro pakutonga (ATC). Chaizvoizvo, ATE musiyano wakaverengerwa mune kiyi metric pakati pebvunzo nemapoka ekudzora (zvakafanana kuyera metric hombe muyedzo yeA/B). Inoverengerwa senzira yekurapa maitiro, ATE = avg (Y1 - Y1) sezvakaratidzwa pazasi mumufananidzo.

Nepo ATT neATC iri avhareji yekurapa maitiro ebvunzo uye yekudzora boka, zvichiteerana. Yese inzira dzakananga uye dzinonzwisisika dzekufungidzira.

ATE ndiyo yakajairika mhando uye inoshandiswa kana kudzora uye mapoka makuru metric akaenzaniswa uye akaedzwa maitiro anoyerwa. Nepo ATT neATC zvichisarudzwa kana absolute metrics achidikanwa kuboka rega rega. Pakupedzisira, kuongororwa kwenhamba kwakakodzera kunoitwa kutarisa kukosha kwehuwandu hwemhedzisiro.

Kuganhurirwa kwePropensity Score Matching

Mushure mekutsanangurwa kwakadzama kweiyo Propensity Score Matching nzira, ingave nguva yekutanga kuishandisa mubasa rako, asi pane zvimwe zvinogumira zvinofanirwa kutariswa.

1. Bootstrap haikurudzirwe kuti ishandiswe nePropensity Score Matching sezvo ichiwedzera musiyano. ( https://economics.mit.edu/sites/default/files/publications/ON THE FAILURE OF THE BOOTSTRAP FOR.pdf )

 2. Stable unit treatment value assumption (SUTVA) principle must be met. 3. Propensity Score Matching implies using two machine learning algorithms (one for propensity score calculations and the second one for matching), which can be a pricy method to use for a company. On that account, it's advisable to negotiate with your team on A/B test conduction. 4. Finally, as discussed above, a big number of covariates are suggested to be used in the models. Thus, it requires a high-powered machine(-s) to calculate the results of the models. Again, it's a costly method to implement.

Nekudaro, kana zvichikwanisika kuita Propensity Score Matching , zviite uye usazeze kuwedzera ruzivo rwako uye ruzivo runoshanda. Rombo rakanaka nemiedzo yako yeramangwana uye kuwana muchina kudzidza