--- title: "Reproducible Statistical Inference" author: "Altman" date: "May 26, 2016" output: ioslides_presentation css: styles.css --- ## Reproducibility There are 3 basic concepts of reproducibility in research: 1. Reproducible experiments

2. Reproducible analysis

3. Reproducible inference ## Reproducibility There are 3 basic concepts of reproducibility in research: 1. Reproducible experiments Another lab can do the same experiment and obtain similar results. 2. Reproducible analysis Another analyst can redo the analysis with the same data and obtain identical results. 3. Reproducible inference After reproducing the experiment in another lab, similar scientific inferences will be made. ## Reproducibility Today: - How selection bias leads to irreproducibility - How simulation can help us assess statistical validitiy - How to set up synthetic data that mimics actual data - Using synthetic data to assess statistical significance ## Example S1 We have done an experiment in which we measured a phenotypic score on each sample (e.g. biopsy result) and also measured the gene expression for 1000 genes in the samples. Our objective is to determine which genes are associated with the phenotypic score, and to develop a prediction equation for the phenotype ```{r libraries, echo=FALSE, warning=FALSE,results="hide", message=FALSE} library(limma) ``` ```{r generateData,echo=FALSE, warning=FALSE,results="hide", message=FALSE} set.seed(7081991) y=rnorm(20) trt=as.integer(y>0) x=matrix(rnorm(20000,6,1.5),nrow=20) ``` ## Example S2 We pursue 2 analyses: 1. Correlation between gene expression and phenotypic score. 2. Division of the samples into "low" and "high" scores, followed by a t-test to determine if there is a difference in gene expression in the low and high score groups. ## Example S3 There are some genes with correlation as low as -0.6 or as high as 0.6 with the phenotype. These are good candidates for genes that are important to determining or predicting genotype. ```{r correlation, echo=FALSE} cors=cor(x,y) hist(cors,nclass=50,main="Correlation with Phenotype",xlab="Correlation") ``` ## Example S4 Lets select the 10 genes with the highest absolute correlation and see how well they predict phenotypic score. ```{r select10,echo=FALSE, warning=FALSE,results="hide", message=FALSE} ocor=order(abs(cors),decreasing=TRUE) ``` ```{r regression,echo=FALSE, warning=FALSE,results="hide", message=FALSE} pheno=y sigGenes=x[,ocor[1:10]] regfit=lm(pheno~sigGenes) ``` ```{r anovaTable} anov=anova(regfit) anov ``` Note that $R^2=$ `r round(anov$"Sum Sq"[1]/sum(anov$"Sum Sq"),2)` and the p-value is `r round(anov$Pr[1],6)` which are highly significant. ## Example S5 Now lets do an alternative analysis classifying the phenotype and "low" and "high" and using an eBayes t-test to distinguish between them. Note that this is a 2-sample t-test with 18 d.f. so the values greater than 2.1 (or less than -2.1) are significant at p<.05. There are several of these. ```{r ttests,echo=FALSE, warning=FALSE,results="hide", message=FALSE} design=model.matrix(~trt) fit.out=lmFit(t(x),design=design) efit.out=eBayes(fit.out) ``` ```{r tstat,echo=FALSE,fig.height=3.5} hist(efit.out$t[,2],nclass=20,xlab="t-values") ``` ## Example S6 We might also wonder if the most statistically significant genes in this analysis match the ones from the correlation analysis. ```{r both} plot(cors,efit.out$t[,2],xlab="Correlation",ylab="t-statistic") ``` We see that there is very high correlation between the correlation statistic and the t-statistic, so that almost the same gene set will be selected. ## Example S7 Naturally, we feel that we have done a great job of this experiment, but our collaborator wants to redo it to verify. We compare results: ## Example S8 ```{r regenerateData,echo=FALSE, warning=FALSE,results="hide", message=FALSE} y1=rnorm(20) trt1=as.integer(y1>0) x1=matrix(rnorm(20000,6,1.5),nrow=20) cors1=cor(x1,y1) ``` ```{r Recorrelation, echo=FALSE,fig.height=3,results="hide"} min(c(cors,cors1)) max(c(cors,cors1)) hist(cors,main="Our Correlations",xlab="Correlation",breaks=seq(-0.725,0.8,0.025)) hist(cors1,main="Collaborator's Correlations",xlab="Correlation",breaks=seq(-0.725,0.8,0.025)) ``` ## Example S9 Of course we don't expect to obtain the exact same correlations, but the pattern is fairly similar. What about the regression on the top 10 genes? ## Example S10 ```{r Reregression,echo=FALSE, warning=FALSE,results="hide", message=FALSE} ocor1=order(abs(cors1),decreasing=TRUE) phenoC=y1 sigGenesC=x1[,ocor1[1:10]] regfit1=lm(phenoC~sigGenesC) ``` **Our ANOVA table:** ```{r anovaTable0, echo=FALSE} anov=round(anov,5) anov[is.na(anov)]="" as.data.frame(anov) ``` **Collaborator's ANOVA table:** ```{r ReanovaTable,,echo=FALSE, warning=FALSE,results="hide", message=FALSE} anov1=round(anova(regfit1),5) anov1[is.na(anov1)]="" ``` ```{r anovaTable1, echo=FALSE} as.data.frame(anov1) ``` These are quite similar. ## Example S11 Similarly if we compute the 2-sample t-tests: ```{r Rettests,echo=FALSE, warning=FALSE,results="hide", message=FALSE} design1=model.matrix(~trt1) fit1.out=lmFit(t(x1),design=design1) efit1.out=eBayes(fit1.out) ``` ```{r Retstat, fig.height=2.9,echo=F} #min(efit.out$t[,2],efit1.out$t[,2]) #max(efit.out$t[,2],efit1.out$t[,2]) hist(efit.out$t[,2], breaks=seq(-4.5,3.6,.15),main="Our t-tests",xlab="t-values") hist(efit1.out$t[,2],breaks=seq(-4.5,3.6,.15),main="Collaborator's t-tests",xlab="t-values") ``` ## Example S12 The correspondance looks equally good if we consider a heatmap of gene expression and other typical measures. But there is a problem - Lets look e.g. at the genes with |cor|>.5 or p<0.05 for our study and our collaborator's study. ## Example S13 Here are the genes with correlation less than -0.5 or greater than 0.5. ```{r intersectionr, echo=FALSE} vennDiagram(vennCounts(cbind(abs(cors)>0.5,abs(cors1)>0.5)),names=c("Ours","Collab"),main="High Correlation Genes") ``` ## Example S14 Here are the genes with t less than -2.1 or greater than 2.1. ```{r intersectiont, echo=FALSE} vennDiagram(vennCounts(cbind(abs(efit.out$t[,2])>2.1,abs(efit1.out$t[,2])>2.1)),names=c("Ours","Collab"),main="Diff Expressed Genes") ``` ## Example S15 Even though overall our collaborator's results seem similar to ours, the resulting gene lists are very different. What went wrong? ## Example S16 The two "studies" cited here are "in silico" studies. For each, I generated 20 "phenotypic scores" and then independently generated 1000 Normally distributed gene expression values for each "sample". All the gene expression values are independent of the phenotypic scores and of each other. So why do we get such similar results for the 2 "studies"? ## Example S17 Firstly lets look at the correspondance between the actual results - e.g. the correlation of the 2 sets of correlations is `r round(cor(cors,cors1),2)`. ```{r corCor, echo=FALSE, fig.height=5} par(omi=c(.1,.1,0,.1)) plot(cors,cors1,xlab="Our Correlations",ylab="Collaborator's Correlations") ``` ## Example S18 Similarly, we can look at the correspondance between the t-values which have correlation `r round(cor(efit.out$t[,2],efit1.out$t[,2]),2)`. ```{r corT, echo=FALSE, fig.height=5} par(omi=c(.1,.1,0,.1)) plot(efit.out$t[,2],efit1.out$t[,2],xlab="Our t-values",ylab="Collaborator's t-values") ``` ## What is going on? Even though the data were generated at random, we appeared to obtain significant (even highly significant) results. As well, the magnitude of the correlations, t-tests, regression R^2^ etc. were very concordant between the 2 totally independent experiments. To understand whether outcomes of our **biological** experiments have a **biological** interpretation, we need to understand the behavior of our statistical methods when the data are random.