3

I've got three data frames, dfLON, dfMOS, and dfATA. Each features the same variables: y is a continuous variable, and a, b, and c are binary categorical, and there is also some NA.

I'd like to build separate linear regression models, one for each data set.

With my current code I've managed to make a list of data frames and pass it into lm(). But is there are more concise way to view the results than eg fitdfLON <- DfList[[1]]? I've provided three data frames in this example but I actually have ~25 so I'd have to type it 25 times!

Any help would be much appreciated.

Starting point (dfs):

dfLON <- data.frame(y=c(1.23,2.32,3.21,2.43),a=c(1,NA,1,2),b=c(1,1,2,2),c=c(2,1,2,1))
dfMOS <- data.frame(y=c(4.56,6.54,4.43,5.78),a=c(2,1,2,1),b=c(2,1,1,2),c=c(1,2,1,2))
dfATA <- data.frame(y=c(1.22,6.54,3.23,4.23),a=c(2,2,2,1),b=c(1,2,1,2),c=c(1,NA,1,2))

Current code: 

Mylm <- function(df){
 fit <- lm(y ~ a + b + c, data=df)
  return(fit)
}
DfList <- lapply(list(dfLON, dfMOS, dfATA), Mylm)

fitdfLON <- DfList[[1]]
fitdfMOS <- DfList[[2]]
fitdfATA <- DfList[[3]]
Claus Wilke
  • 16,992
  • 7
  • 53
  • 104
LLL
  • 723
  • 1
  • 9
  • 27
  • Simply use a named list instead of separate objects: `DfList <- setNames(lapply(list(dfLON, dfMOS, dfATA), Mylm), c("fitdfLON", "fitdfMOS", fitdfATA"));` Then reference items: `DfList$fitdfLON; DfList$fitdfMOS; DfList$fitdfATA` – Parfait Nov 28 '17 at 18:55

3 Answers3

1

If the names of the data.frame have a common pattern, you can use a combination of mget and ls to extract them and run lm using lapply

fit = lapply(mget(ls(pattern = "^df[A-Z]{3}")), function(x) lm(y ~ a + b + c, data = x))
fit$dfATA

#Call:
#lm(formula = y ~ a + b + c, data = x)

#Coefficients:
#(Intercept)            a            b            c  
#      6.235       -2.005           NA           NA

If you just want coefficients for all, you could do

do.call(rbind,
        lapply(X = mget(ls(pattern = "^df[A-Z]{3}")),
               FUN = function(x) lm(formula = y ~ a + b + c, data = x)[[1]]))
#      (Intercept)      a      b  c
#dfATA      6.2350 -2.005     NA NA
#dfLON      0.0300 -0.780  1.980 NA
#dfMOS      8.2975 -1.665 -0.315 NA

Instead of ls(pattern = "df[A-Z]{3}"), you could also just provide a vector that has the names of all the data.frame

d.b
  • 32,245
  • 6
  • 36
  • 77
1

Whenever you're running models on many different datasets, it makes sense to tidy them using the broom library. This produces a clean data frame for each model, which you can then output or use in downstream analyses.

Simplest example:

library(broom)

Mylm <- function(df){
  fit <- lm(y ~ a + b + c, data=df)
  tidy(fit) # tidy the fit object
}

list(dfLON, dfMOS, dfATA) %>% lapply(Mylm)

#[[1]]
#         term estimate std.error statistic p.value
#1 (Intercept)     0.03       NaN       NaN     NaN
#2           a    -0.78       NaN       NaN     NaN
#3           b     1.98       NaN       NaN     NaN
#
#[[2]]
#         term estimate std.error  statistic    p.value
#1 (Intercept)   8.2975  0.969855  8.5554025 0.07407531
#2           a  -1.6650  0.445000 -3.7415730 0.16626155
#3           b  -0.3150  0.445000 -0.7078652 0.60785169
#
#[[3]]
#         term estimate std.error statistic   p.value
#1 (Intercept)    6.235  3.015000  2.067993 0.2867398
#2           a   -2.005  1.740711 -1.151828 0.4551559

And you can now combine this with the map_dfr() function from purrr to combine everything into one combined data frame:

library(purrr)

# note the named list entries; these will go into the "model" column
# without them, you'd just get a model number
list("LON" = dfLON, "MOS" = dfMOS, "ATA" = dfATA) %>% 
  map_dfr(Mylm, .id = "model")

#  model        term estimate std.error  statistic    p.value
#1   LON (Intercept)   0.0300       NaN        NaN        NaN
#2   LON           a  -0.7800       NaN        NaN        NaN
#3   LON           b   1.9800       NaN        NaN        NaN
#4   MOS (Intercept)   8.2975  0.969855  8.5554025 0.07407531
#5   MOS           a  -1.6650  0.445000 -3.7415730 0.16626155
#6   MOS           b  -0.3150  0.445000 -0.7078652 0.60785169
#7   ATA (Intercept)   6.2350  3.015000  2.0679934 0.28673976
#8   ATA           a  -2.0050  1.740711 -1.1518281 0.45515586

And to make things more compact, you can define the function on the fly inside map_dfr. Seems appropriate when all you're doing is fit a linear model.

list("LON" = dfLON, "MOS" = dfMOS, "ATA" = dfATA) %>% 
  map_dfr(~ tidy(lm(y ~ a + b + c, data = .)),
          .id = "model")

#  model        term estimate std.error  statistic    p.value
#1   LON (Intercept)   0.0300       NaN        NaN        NaN
#2   LON           a  -0.7800       NaN        NaN        NaN
#3   LON           b   1.9800       NaN        NaN        NaN
#4   MOS (Intercept)   8.2975  0.969855  8.5554025 0.07407531
#5   MOS           a  -1.6650  0.445000 -3.7415730 0.16626155
#6   MOS           b  -0.3150  0.445000 -0.7078652 0.60785169
#7   ATA (Intercept)   6.2350  3.015000  2.0679934 0.28673976
#8   ATA           a  -2.0050  1.740711 -1.1518281 0.45515586
Claus Wilke
  • 16,992
  • 7
  • 53
  • 104
0
#make a list of all the dataframes
df = list(dfATA = dfATA, dfLON =dfLON, dfMOS = dfMOS)

#fitting the model
lmr = lapply(df, function(x){
  lmr = lm(x$y ~ x$a + x$b+ x$c, x)
})

#Get coefficients for each model
coefficients = lapply(lmr, function(x) x[["coefficients"]])
coefficients = unlist(coefficients)
Nad Pat
  • 3,129
  • 3
  • 10
  • 20