class: center, middle, inverse, title-slide # Neue Statistik ### Statistik mit R
The R Bootcamp
### September 2020 --- layout: true <div class="my-footer"> <span style="text-align:center"> <span> <img src="https://raw.githubusercontent.com/therbootcamp/therbootcamp.github.io/master/_sessions/_image/by-sa.png" height=14 style="vertical-align: middle"/> </span> <a href="https://therbootcamp.github.io/"> <span style="padding-left:82px"> <font color="#7E7E7E"> www.therbootcamp.com </font> </span> </a> <a href="https://therbootcamp.github.io/"> <font color="#7E7E7E"> Statistik mit R | September 2020 </font> </a> </span> </div> --- .pull-left4[ # Neue = Alte Statistik <ul> <li class="m1"><span><high>Bessere Studien</high></span></li> <ul class="level"> <li><span>Informative designs</span></li> <li><span>Planung der Stichprobengrösse</span></li> <li><span>Registrierung</span></li> <li><span>Kein p-hacking</span></li> <li><span>Vollständiges berichten</span></li> <li><span>Replikation</span></li> </ul> <br> <li class="m2"><span><high>Bessere Statistik</high></span></li> <ul class="level"> <li><span>Unsicherheit berichten</span></li> <li><span><i>p</i> Herabsetzen</span></li> <li><span>Bayesianische Statistik</span></li> </ul> </ul> ] .pull-right5[ <br><br><br> <p align = "center"> <img src="image/blastfromthepast.jpg" height=500px style="border-radius:50px"><br> <font style="font-size:10px">from <a href="https://www.amazon.com/Blast-Past-Brendan-Fraser/dp/0780626494">amazon.com</a></font> </p> ] --- # Replikationskrise .pull-left4[ <ul> <li class="m1"><span>Eine Replikationskrise geht um in <high>diversen akademischen Feldern</high></span></li> <li class="m2"><span>In einer grossen Studie in der Psychologie konnten <high>nur 36% der Ergebnisse repliziert</high> werden.</span></li> <li class="m3"><span>Die niedrige Replikationsrate ist das Ergebnis von <high>Questionable Research Practices</high>.</span></li> </ul> <br><br> Siehe vergleichbare Studien in [**Medicine**](https://therbootcamp.github.io/SwR_2019Apr/_sessions/NewStats/literature/Ioannidis2005.pdf), [**Economics**](https://therbootcamp.github.io/SwR_2019Apr/_sessions/NewStats/literature/Camerer2016.pdf), [**Marketing**](https://therbootcamp.github.io/SwR_2019Apr/_sessions/NewStats/literature/Hunter2001.pdf), [**Social sciences**](https://therbootcamp.github.io/SwR_2019Apr/_sessions/NewStats/literature/Camerer2018.pdf). ] .pull-right5[ <p align = "center"> <img src="image/crisis.png" height=410px><br> <font style="font-size:10px">from <a href="https://ink.library.smu.edu.sg/cgi/viewcontent.cgi?article=6256&context=lkcsb_research">Open Science Collaboration</a></font> </p> ] --- # Zeichen guter Studien .pull-left4[ <ul> <li class="m1"><span>Informative Designs<br><font style="font-size:16px">Designe den bestmöglichen Test der Hypothese.</font></span></li> <li class="m2"><span><high>Stichprobenplanung</high><br><font style="font-size:16px">Wähle angemessenes <i>N</i> für klare Entscheidungen.</font></span></li> <li class="m3"><span>Registrierungen<br><font style="font-size:16px">Registriere deine Hypothesen vor der Datenerhebung.</font></span></li> <li class="m4"><span><high>Kein <i>p</i>-hacking</high><br><font style="font-size:16px">Teste geplante Analysen oder berichte sie als explorativ.</font></span></li> <li class="m5"><span>Vollständige Berichterstattung<br><font style="font-size:16px">Berichte alle durchgeführten Analysen.</font></span></li> <li class="m6"><span>Replikation<br><font style="font-size:16px">Wiederhole deine Studie.</font></span></li> </ul> ] .pull-right5[ <p align = "center"> <img src="image/filedrawer_bw.png"><br> <font style="font-size:10px">adapted from <a href="https://www.someecards.com/usercards/viewcard/MjAxMy0zOGE1Mzg3YjY4YzE1NTA0/?tagSlug=birthday">Someecards.com</a></font> </p> ] --- # Stichprobenplanung .pull-left4[ <ul> <li class="m1"><span>Die Grösse der Stichprobe sollte die nötige <high>Trennschärfe für das Detektieren eines Effekts ermöglichen</high></high></span></li> <li class="m2"><span>Grössere Stichproben → <high>schmalere Stichprobenverteilungen</high> → höhere Trennschärfe.</span></li> </ul> <p align = "center"> <img src="image/error_table.png" width=285px><br> </p> ] .pull-right5[ <p align = "center"> <img src="image/power.png" height=420px><br> </p> ] --- # Stichprobenplanung .pull-left4[ <ul> <li class="m1"><span>Die Grösse der Stichprobe sollte die nötige <high>Trennschärfe für das Detektieren eines Effekts ermöglichen</high></high></span></li> <li class="m2"><span>Grössere Stichproben → <high>schmalere Stichprobenverteilungen</high> → höhere Trennschärfe.</span></li> </ul> <p align = "center"> <img src="image/error_table.png" width=285px><br> </p> ] .pull-right5[ <p align = "center"> <img src="image/power_large.png" height=420px><br> </p> ] --- # Stichprobenplanung in R .pull-left4[ <ul> <li class="m1"><span>Die Grösse der Stichprobe sollte die nötige <high>Trennschärfe für das Detektieren eines Effekts ermöglichen</high>.</span></li> <li class="m2"><span>Grössere Stichproben → <high>schmalere Stichprobenverteilungen</high> → höhere Trennschärfe.</span></li> </ul> <p align = "center"> <img src="image/error_table.png" width=285px><br> </p> ] .pull-right5[ ```r # Lade pwr Paket library(pwr) # N für einen grossen Effekt im t-test pwr.t.test(sig.level = .05, power = .95, d = .2) # grosser Effekt ``` ``` ## ## Two-sample t test power calculation ## ## n = 650.7 ## d = 0.2 ## sig.level = 0.05 ## power = 0.95 ## alternative = two.sided ## ## NOTE: n is number in *each* group ``` ] --- # Stichprobenplanung in R .pull-left4[ <ul> <li class="m1"><span>Die Grösse der Stichprobe sollte die nötige <high>Trennschärfe für das Detektieren eines Effekts ermöglichen</high>.</span></li> <li class="m2"><span>Grössere Stichproben → <high>schmalere Stichprobenverteilungen</high> → höhere Trennschärfe.</span></li> </ul> <p align = "center"> <img src="image/error_table.png" width=285px><br> </p> ] .pull-right5[ ```r # Lade pwr Paket library(pwr) # N für einen grossen Effekt im t-test pwr.t.test(sig.level = .05, power = .95, d = .8) # grosser Effekt ``` ``` ## ## Two-sample t test power calculation ## ## n = 41.59 ## d = 0.8 ## sig.level = 0.05 ## power = 0.95 ## alternative = two.sided ## ## NOTE: n is number in *each* group ``` ] --- .pull-left2[ # <i>p</i>-hacking <high><i>“If you torture the data long enough, it will confess.”</i></high> Ronald Coase <br><br><br><br><br><br><br><br><br><br><br><br> <p align="right"><font style="font-size:10px">from <a href="https://fivethirtyeight.com/features/science-isnt-broken/#part1">fivethirtyeight.com</a></font> </p> ] .pull-right75[ <br> <iframe src="https://projects.fivethirtyeight.com/p-hacking/index.html?initialWidth=1024&childId=phacking&parentTitle=Science%20Isn%E2%80%99t%20Broken%20%7C%20FiveThirtyEight&parentUrl=https%3A%2F%2Ffivethirtyeight.com%2Ffeatures%2Fscience-isnt-broken%2F" width="100%" scrolling="yes" marginheight="0" frameborder="0" height="550px"></iframe> ] --- .pull-left45[ # <i>p</i>-hacking <ul> <li class="m1"><span><i>p</i>-hacking bezeichnet des Missbrauch von Datenanalyse, um <high>Effekte zu zeigen wo es eigentlich keine gibt</high>.</span></li><br> <li class="m2"><span><high>Optionales Stoppen</high><br><br>Stoppen der Datenanalyse sobald Signifikanz erreicht wurde.</span></li><br> <li class="m3"><span><high>HARKing</high><br><br>Hypothesizing After the Data are Known. Geschieht wenn nicht-geplante Analysen als geplant dargestellt werden.</span></li> </ul> ] .pull-right45[ <br> <p align = "center"> <img src="image/P-hacking.png" height=550px><br> </p> ] --- # Dos and Don'ts .pull-left5[ <p align="center"> <img src="image/exploration.png" width = 440px> <h style="font-size:60px;position:absolute;bottom:390px;left:168px;color:white">Exploration</h><br> <font style="font-size:10px">Columbus looking for India, from <a href="https://www.history.com/topics/exploration">history.com</a></font> <br><br> <font style="font-size:28px"><i>Do</i></font><br><br2> Alles<br><bt> <br><br><br><br><font style="font-size:28px"><i>Don't</i></font><br><br2> Die Ergebnisse als konfirmatorisch präsentieren.<br><br> </font> </p> ] .pull-right5[ <p align="center"> <img src="image/confirmation.png" width = 440px> <h style="font-size:60px;position:absolute;bottom:390px;left:675px;color:white">Konfirmation</h><br> <font style="font-size:10px">Villemard vision for 2000, from <a href="https://www.sadanduseless.com/world-in-2000/">sadanduseless.com</a></font> <br><br> <font style="font-size:28px"><i>Do</i></font><br><br2> Stelle Vorhersagen auf.<br> Plane die Stichprobengrösse.<br> Lege den Analyseplan fest.<br> Registriere.<br> <br><font style="font-size:28px"><i>Don't</i></font><br><br2> Führe nicht-geplante Analysen durch.<br><br> </font> </p> </p> ] --- .pull-left45[ # Bessere Statistik <ul> <li class="m1"><span>Probleme mit <i>p</i>-Werten.</span></li> <ul> <li><span><high>Schwer zu interpretieren</high> als Mass für Evidenz.</span></li> <li><span>Dienen einer <high>arbiträren Binarisierung</high>.</span></li> </ul><br> <li class="m2"><span>"Neue" Statistik.</span></li> <ul> <li><span>Fokus auf <high>Schätzung</high>.</span></li> <li><span>Kommunikation der <high>Unsicherheit</high>.</span></li> </ul> </ul> <br><br><br> `$$Confidence\,Interval\,(CI) = \bar{x} \pm t_{1-\alpha} \sigma_{\bar{x}}$$` ] .pull-right45[ <br> <p align = "center"> <img src="image/confint.png" height=550px><br> </p> ] --- .pull-left45[ # Konfidenzinterval <ul> <li class="m1"><span>Konfidenzintervalle sind <high>rearrangierte Signfikanztests</high>.</span></li> </ul> <table style="cellspacing:0; cellpadding:0; border:none; padding-top:10px"> <col width="30%"> <col width="70%"> <tr> <td bgcolor="white"> t-test </td> <td bgcolor="white" style="padding-top:15px;padding-bottom:15px"> <img src="image/form0.png" height=74px><br> </td> </tr> <tr> <td bgcolor="white"> Step 1 </td> <td bgcolor="white" style="padding-top:15px;padding-bottom:15px"> <img src="image/form1.png" height=34px><br> </td> </tr> <tr> <td bgcolor="white"> Step2 </td> <td bgcolor="white" style="padding-top:15px;padding-bottom:15px;bgcolor:white"> <img src="image/form2.png" height=34px><br> </td> </tr> <tr> <td bgcolor="white"> CI </td> <td bgcolor="white" style="padding-top:15px;padding-bottom:15px"> <img src="image/form3a.png" height=27px> <img src="image/form3b.png" height=27px> </td> </tr> </table> ] .pull-right45[ <br> <p align = "center"> <img src="image/confint.png" height=550px><br> </p> ] --- # Konfidenzinterval .pull-left45[ <ul> <li class="m1"><span>Konfidenzintervalle sind <high>rearrangierte Signfikanztests</high>.</span></li> </ul> <table style="cellspacing:0; cellpadding:0; border:none; padding-top:10px"> <col width="30%"> <col width="70%"> <tr> <td bgcolor="white"> t-test </td> <td bgcolor="white" style="padding-top:15px;padding-bottom:15px"> <img src="image/form0.png" height=74px><br> </td> </tr> <tr> <td bgcolor="white"> Step 1 </td> <td bgcolor="white" style="padding-top:15px;padding-bottom:15px"> <img src="image/form1.png" height=34px><br> </td> </tr> <tr> <td bgcolor="white"> Step2 </td> <td bgcolor="white" style="padding-top:15px;padding-bottom:15px;bgcolor:white"> <img src="image/form2.png" height=34px><br> </td> </tr> <tr> <td bgcolor="white"> CI </td> <td bgcolor="white" style="padding-top:15px;padding-bottom:15px"> <img src="image/form3a.png" height=27px> <img src="image/form3b.png" height=27px> </td> </tr> </table> ] .pull-right5[ ```r # Simulierte Daten t_1 <- rnorm(100, 3, 1) t_2 <- rnorm(100, 4, 1) # Standard Konfidenzinterval N <- length(t_1) Delta <- mean(t_1 - t_2) SE <- sd(t_1 - t_2) / sqrt(N) Delta + SE * qt(.95, N - 1) * c(-1, 1) ``` ``` ## [1] -1.1171 -0.6333 ``` ] --- # Konfidenzinterval .pull-left45[ <ul> <li class="m1"><span>Konfidenzintervalle sind <high>rearrangierte Signfikanztests</high>.</span></li> </ul> <table style="cellspacing:0; cellpadding:0; border:none; padding-top:10px"> <col width="30%"> <col width="70%"> <tr> <td bgcolor="white"> t-test </td> <td bgcolor="white" style="padding-top:15px;padding-bottom:15px"> <img src="image/form0.png" height=74px><br> </td> </tr> <tr> <td bgcolor="white"> Step 1 </td> <td bgcolor="white" style="padding-top:15px;padding-bottom:15px"> <img src="image/form1.png" height=34px><br> </td> </tr> <tr> <td bgcolor="white"> Step2 </td> <td bgcolor="white" style="padding-top:15px;padding-bottom:15px;bgcolor:white"> <img src="image/form2.png" height=34px><br> </td> </tr> <tr> <td bgcolor="white"> CI </td> <td bgcolor="white" style="padding-top:15px;padding-bottom:15px"> <img src="image/form3a.png" height=27px> <img src="image/form3b.png" height=27px> </td> </tr> </table> ] .pull-right5[ ```r # Bootstrap-Konfidenzinterval bf <- function(x,ind) { sum(x[ind])/length(x[ind]) } boot_res <- boot(t_1 - t_2, bf, 1000) boot.ci(boot_res) ``` ``` ## BOOTSTRAP CONFIDENCE INTERVAL CALCULATIONS ## Based on 1000 bootstrap replicates ## ## CALL : ## boot.ci(boot.out = boot_res) ## ## Intervals : ## Level Normal Basic ## 95% (-1.1635, -0.5846 ) (-1.1667, -0.5772 ) ``` ] --- # Effektstärke .pull-left45[ <ul> <li class="m1"><span>Wichtiger als die Signifikanz, ist ob ein <high>Effekt von bedeutsamer Grösse</high> ist.</span></li> <li class="m2"><span>Je nach Situation, werden <high>verschiedene Effektstärkemasse</high> herangezogen.</span></li> </ul> <br> <table style="cellspacing:0; cellpadding:0; border:none; padding-top:10px"> <col width="50%"> <col width="50%"> <tr> <td bgcolor="white"> <b>Situation</b> </td> <td bgcolor="white"> <b>Mass</b> </td> </tr> <tr> <td bgcolor="white"> Mittelwertsvergleiche </td> <td bgcolor="white"> Cohen's <i>d</i><br>Hedge's <i>g</i><br>η<sup>2</sup><br> </td> </tr> <tr> <td bgcolor="white"> Korrelation, Regression </td> <td bgcolor="white"> <i>r</i><br>R<sup>2</sup><br>f<sup>2</sup> </td> </tr> </table> ] .pull-right5[ <br> <p align = "center"> <img src="image/effectsize.png" height=400px><br> </p> ] --- # Effektstärke .pull-left45[ <ul> <li class="m1"><span>Wichtiger als die Signifikanz, ist ob ein <high>Effekt von bedeutsamer Grösse</high> ist.</span></li> <li class="m2"><span>Je nach Situation, werden <high>verschiedene Effektstärkemasse</high> herangezogen.</span></li> </ul> <br> <table style="cellspacing:0; cellpadding:0; border:none; padding-top:10px"> <col width="50%"> <col width="50%"> <tr> <td bgcolor="white"> <b>Situation</b> </td> <td bgcolor="white"> <b>Mass</b> </td> </tr> <tr> <td bgcolor="white"> Mittelwertsvergleiche </td> <td bgcolor="white"> Cohen's <i>d</i><br>Hedge's <i>g</i><br>η<sup>2</sup><br> </td> </tr> <tr> <td bgcolor="white"> Korrelation, Regression </td> <td bgcolor="white"> <i>r</i><br>R<sup>2</sup><br>f<sup>2</sup> </td> </tr> </table> ] .pull-right5[ ```r # Bootstrap-Konfidenzinterval für Cohen's d bf <- function(x,ind) { mean(x[ind])/sd(x[ind]) } boot_res <- boot(t_1 - t_2, bf, 1000) boot.ci(boot_res) ``` ``` ## BOOTSTRAP CONFIDENCE INTERVAL CALCULATIONS ## Based on 1000 bootstrap replicates ## ## CALL : ## boot.ci(boot.out = boot_res) ## ## Intervals : ## Level Normal Basic ## 95% (-0.8148, -0.3748 ) (-0.8126, -0.3770 ) ``` ] --- .pull-left45[ # Bayesianische Statistik <ul> <li class="m1"><span>Ergänzt klassische (frequentistische) Statistik um eine Verteilung der <high>Grundwahrscheinlichkeit der Hypothesen (Prior)</high>.</span></li> <li class="m2"><span>Erlaubt die Berechnung eines <high>wahren <i>p</i>-Werts</high>.</span></li> </ul> <p align = "center" style="padding-top:20px"> <img src="image/bayes_form.png"><br> </p> ] .pull-right45[ <br> <p align = "center"> <img src="image/bayes.png" height=520px><br> <font style="font-size:10px">from <a = href="https://xkcd.com/1132/">xkcd.com</a></font> </p> ] --- .pull-left45[ # Bayesianische Statistik <ul> <li class="m1"><span>Bayesianische Statistik war <high>lange unpopular</high>.</span></li> <ul> <li><span>Es ist <high>komputational anspruchsvoll</high> <mono>p(D)</mono> zu berechnen.</span></li> <li><span>Das Festlegen der Prior ist <high>notwendigerweise subjektiv</high>.</span></li> </ul> <li class="m2"><span>Allerdings hat sich dies geändert.</span></li> <ul> <li><span>Schnellere Computer/Algorithmen erlauben eine Effiziente Schätzung von <mono>p(D)</mono>.</span></li> <li><span>Einsicht, dass Statistik <high>subjektiv</high> ist.</span></li> <li><span>Fokus auf <high>ignorante Priors</high>.</span></li> </ul> </ul> ] .pull-right45[ <br> <p align = "center"> <img src="image/bayes.png" height=520px><br> <font style="font-size:10px">from <a = href="https://xkcd.com/1132/">xkcd.com</a></font> </p> ] --- # Bayesian statistics .pull-left45[ <ul> <li class="m1"><span>Die Pakete <a href="https://cran.r-project.org/package=rstanarm/"><mono><b>rstanarm</b></mono></a> and <a href="https://cran.r-project.org/package=BayesFactor/"><mono><b>BayesFactor</b></mono></a> machen Bayesianische Analysen einfach.</span></li> <ul> <br> <table style="cellspacing:0; cellpadding:0; border:none; padding-bottom:30px"> <col width="25%"> <col width="20%"> <col width="55%"> <tr> <td bgcolor="white"> <b>Function</b> </td> <td bgcolor="white"> <b>Package</b> </td> <td bgcolor="white"> <b>Description</b> </td> </tr> <tr> <td bgcolor="white"> <mono>stan_glm</mono>, <mono>stan_glmer</mono> </td> <td bgcolor="white"> <mono>rstanarm</mono> </td> <td bgcolor="white"> Bayesian (mixed) regression </td> </tr> <tr> <td bgcolor="white"> <mono>ttestBF</mono>, <mono>anovaBF</mono> </td> <td bgcolor="white"> <mono>BayesFactor</mono> </td> <td bgcolor="white"> Standard h-tests </td> </tr> <tr> <td bgcolor="white"> <mono>lmBF</mono> </td> <td bgcolor="white"> <mono>BayesFactor</mono> </td> <td bgcolor="white"> Bayesian (mixed) regressions </td> </tr> </table> ] .pull-right5[ ``` ## stan_glm ## family: gaussian [identity] ## formula: income ~ height ## observations: 300 ## predictors: 2 ## ------ ## Median MAD_SD ## (Intercept) 9226.6 2134.4 ## height -10.3 12.4 ## ## Auxiliary parameter(s): ## Median MAD_SD ## sigma 2634.2 107.9 ## ## ------ ## * For help interpreting the printed output see ?print.stanreg ## * For info on the priors used see ?prior_summary.stanreg ``` ] --- # Bayesian statistics .pull-left45[ <ul> <li class="m1"><span>Die Pakete <a href="https://cran.r-project.org/package=rstanarm/"><mono><b>rstanarm</b></mono></a> and <a href="https://cran.r-project.org/package=BayesFactor/"><mono><b>BayesFactor</b></mono></a> machen Bayesianische Analysen einfach.</span></li> <ul> <br> <table style="cellspacing:0; cellpadding:0; border:none; padding-bottom:30px"> <col width="25%"> <col width="20%"> <col width="55%"> <tr> <td bgcolor="white"> <b>Function</b> </td> <td bgcolor="white"> <b>Package</b> </td> <td bgcolor="white"> <b>Description</b> </td> </tr> <tr> <td bgcolor="white"> <mono>stan_glm</mono>, <mono>stan_glmer</mono> </td> <td bgcolor="white"> <mono>rstanarm</mono> </td> <td bgcolor="white"> Bayesian (mixed) regression </td> </tr> <tr> <td bgcolor="white"> <mono>ttestBF</mono>, <mono>anovaBF</mono> </td> <td bgcolor="white"> <mono>BayesFactor</mono> </td> <td bgcolor="white"> Standard h-tests </td> </tr> <tr> <td bgcolor="white"> <mono>lmBF</mono> </td> <td bgcolor="white"> <mono>BayesFactor</mono> </td> <td bgcolor="white"> Bayesian (mixed) regressions </td> </tr> </table> ] .pull-right5[ ``` ## Bayes factor analysis ## -------------- ## [1] height : 0.1813 ±0% ## ## Against denominator: ## Intercept only ## --- ## Bayes factor type: BFlinearModel, JZS ``` ] --- class: middle, center <h1><a href="https://therbootcamp.github.io/SmR_2020Mai/_sessions/NewStats/NewStats_practical.html">Practical</a></h1>