Missing samples (observations) in time series data

Question

[prathyushareddy87]

Hi all,

I am trying to find the differential expressed genes for a dataset where the samples are treated with a drug at three different time-points(baseline, 16 weeks, 52 weeks). There are no controls in the study. I am trying to use the limma in R and analyze this as they are paired samples. I am using F statistic to do this. I have one challenge here. Not all time points equal samples. Few observations are missing. For instance, I have baseline and week 16 and 52 week is missing. The easiest way to handle the missing observations I thought was to take the samples which has all three time points. But I think in statistical point of view this might not be correct way. Can anybody suggest what method to use to handle this issue. I am reading lot of papers. I came across imputation technique. I am not sure whether or not the technique would be applicable to this scenario.

I am attaching the sample data here

Accession   Title   time    timepoints  subjectid
GSM2352693  SUBJ.1720, SLE, baseline    baseline    3   SUBJ.1720
GSM2352694  SUBJ.1720, SLE, week16  week 16 3   SUBJ.1720
GSM2352695  SUBJ.1720, SLE, week52  week 52 3   SUBJ.1720
GSM2352696  SUBJ.0003, SLE, baseline    baseline    3   SUBJ.0003
GSM2352697  SUBJ.0003, SLE, week16  week 16 3   SUBJ.0003
GSM2352698  SUBJ.0003, SLE, week52  week 52 3   SUBJ.0003
GSM2352699  SUBJ.0065, SLE, baseline    baseline    2   SUBJ.0065
GSM2352700  SUBJ.0065, SLE, week52  week 52 2   SUBJ.0065
GSM2352701  SUBJ.1587, SLE, baseline    baseline    3   SUBJ.1587
GSM2352702  SUBJ.1587, SLE, week16  week 16 3   SUBJ.1587
GSM2352703  SUBJ.1587, SLE, week52  week 52 3   SUBJ.1587
GSM2352704  SUBJ.1028, SLE, baseline    baseline    3   SUBJ.1028
GSM2352705  SUBJ.1028, SLE, week16  week 16 3   SUBJ.1028
GSM2352706  SUBJ.1028, SLE, week52  week 52 3   SUBJ.1028
GSM2352707  SUBJ.0901, SLE, baseline    baseline    3   SUBJ.0901
GSM2352708  SUBJ.0901, SLE, week16  week 16 3   SUBJ.0901
GSM2352709  SUBJ.0901, SLE, week52  week 52 3   SUBJ.0901
GSM2352710  SUBJ.1544, SLE, baseline    baseline    3   SUBJ.1544
GSM2352711  SUBJ.1544, SLE, week16  week 16 3   SUBJ.1544
GSM2352712  SUBJ.1544, SLE, week52  week 52 3   SUBJ.1544
GSM2352713  SUBJ.0200, SLE, baseline    baseline    3   SUBJ.0200
GSM2352714  SUBJ.0200, SLE, week16  week 16 3   SUBJ.0200
GSM2352715  SUBJ.0200, SLE, week52  week 52 3   SUBJ.0200
GSM2352716  SUBJ.0032, SLE, baseline    baseline    3   SUBJ.0032
GSM2352717  SUBJ.0032, SLE, week16  week 16 3   SUBJ.0032
GSM2352718  SUBJ.0032, SLE, week52  week 52 3   SUBJ.0032
GSM2352719  SUBJ.1545, SLE, week16  week 16 2   SUBJ.1545
GSM2352720  SUBJ.1545, SLE, week52  week 52 2   SUBJ.1545

Th R code I am trying to use is

library(limma)
# limma
design <- model.matrix(~0 + eset_filtered$cohort)
colnames(design) <- levels(eset_filtered$cohort)
fit <- lmFit(eset_filtered, design)
contrast.matrix <- makeContrasts("week16-baseline","week52-baseline",levels = design) #name
fit2 <- contrasts.fit(fit, contrast.matrix)
fit2 <- eBayes(fit2)
top.active <- topTable(fit2, adjust="BH", n=nrow(eset_filtered))

Any suggestions would be really appreciated. Thanks in advance.

Answer 1

[Kevin Blighe]

The easiest way to handle the missing observations I thought was to take the samples which has all three time points. But I think in statistical point of view this might not be correct way.

Why do you say that? - it is actually the desirable situation to have data at all time-points for all samples. The problem is that you reduce your statistical power, which is perhaps what you meant (?).

Given the data that you've got, I would not restrict myself to just a single type of analysis (I've analysed, in the past, various datasets like you've got):

• What changes from Baseline to 16 weeks? (simple comparison in Limma)
• What changes from 16 weeks to 52 weeks? (simple comparison in Limma)
• What changes from Baseline to 52 weeks? (simple comparison in Limma)
• What differs between all time-points (ANOVA and F-test)?

I would also do paired tests in the form of the Wilcoxon signed-rank test between each time-point on a pairwise basis (3 comparisons).

Further, if you have drug response data: Given the genes that differ statistically significantly between Baseline and 16 or 52 weeks, can the levels of these genes in just the baseline samples predict drug response / non-response at 16 or 52 weeks? This could be answered via the construction of a logistic regression model with y as DrugResponse (at 16 or 52 weeks) and the x predictors as the baseline expression of the key genes. Then, once you select best predictors, perform ROC analysis on the final model.

Kevin