Data Frame Exercises in R: 17 Real-World Practice Problems

Explanation: data.frame() takes equal-length vectors as named arguments and stacks them column-wise. Since R 4.0 character vectors stay as character by default; before that, stringsAsFactors = TRUE was the silent default and a long-standing source of subtle bugs. The integer sequence 101:104 keeps integer storage rather than being coerced to double, which matters if you later join on it.

Explanation: A data frame is internally a list of equal-length columns, so sapply() walks each column and applies class(). The named character vector tells you immediately which columns are numeric, factor, character, or Date. For columns with multiple classes (such as POSIXct, POSIXt) prefer vapply(mtcars, function(x) class(x)[1], character(1)) so the result stays a flat character vector instead of collapsing into a matrix.

Explanation: Wrapping a column subset in lapply() and reassigning back is the canonical base-R idiom for coercing several columns at once. Constructing the frame with data.frame(..., stringsAsFactors = TRUE) would also have converted sku and wasted memory on a factor with one level per row. For larger frames you can stay base-R with Filter(is.character, df) to discover candidate columns and then loop on the survivors.

Explanation: All three forms return the same atomic numeric vector for a single column. Dollar syntax $ is shortest but performs partial matching on column names, which is a footgun inside scripts; [[]] and [, ] accept a column name as a string and pair cleanly with column names stored in variables. Use [[]] when the column name lives in a variable; pass drop = FALSE if you must keep the result as a one-column data frame instead of a vector.

Explanation: head() and tail() both preserve original row names, so the resulting rbind carries 1, 2, 3 and 148, 149, 150 in the row name slot. The Species column reveal makes it clear the rows are sorted by Species, a thing many EDA pipelines miss when they only call head() and assume the bottom of the frame looks the same. Treat this one-liner as a default sanity audit for any new data frame.

Explanation: table() returns a table object (a thin wrapper over a named integer matrix) which prints as a named row of counts. Because Species is a factor with three levels, the counts respect the factor level order rather than alphabetical order or first-appearance order. To turn the result into a tidy data frame for plotting, wrap in as.data.frame(table(...)); to convert counts into proportions, wrap in prop.table().

Explanation: The bare-bracket [rows, cols] form takes a logical vector along the row axis. mtcars$mpg > 25 evaluates to TRUE or FALSE per row, the & operator combines element-wise with the cylinder check, and the empty column slot keeps every column. Forgetting the trailing comma drops you to list-of-columns mode and silently returns a different shape, a classic R beginner trap to watch for.

Explanation: Negative indexing by column name does not exist in base R directly. You could write mtcars[, -which(names(mtcars) %in% c("vs","am"))] but it is noisy and breaks if a target name is missing. Using setdiff() on names is clearer and survives both reordering and missing keys gracefully. For a single drop, mtcars$vs <- NULL works too, but it mutates in place which is usually not what you want in an analysis pipeline.

Explanation: subset() uses non-standard evaluation so you can reference column names without quoting, which is great for interactive use and a trap inside functions (R CMD check complains about no visible binding). For programmatic filtering prefer airquality[airquality$Month == 5, c("Ozone","Wind")]. The behaviour around NA in the condition is identical to bracket subsetting: rows where the condition is NA are dropped, not silently kept.

Explanation: Assigning a vector to a new dollar-name slot is the simplest column-add path: R recycles the vector length and validates it matches nrow(). Doing this on a copy (ex_4_1 <- mtcars) keeps the original frame untouched, which matters in long notebooks where downstream cells may still rely on the base column set. As a functional alternative, transform(mtcars, kpl = round(mpg * 0.425144, 2)) produces the same shape in one expression.

Explanation: cut() accepts breakpoints as a numeric vector and returns an ordered factor with one level per interval. Using -Inf and Inf as the outer bounds is cleaner than min(wt) - 1 and max(wt) + 1 and immune to new data falling outside the historical range. The default is right-closed intervals; pass right = FALSE if your domain prefers left-closed bins (financial price buckets and tax brackets typically want this).

Explanation: The trick is the targeted left-hand side: ex_4_3$Ozone[is.na(...)] <- value writes only into the NA positions and leaves the observed values untouched. Always pass na.rm = TRUE to median() or you compute NA and silently overwrite every reading with NA. Median imputation is a defensible quick fix; for a real air-quality pipeline you would want temporal interpolation or a model-based imputation that respects seasonality and weather covariates.

Explanation: order() returns the row indices that would sort the inputs; passing them as the row argument to [ , ] reorders the data frame. The leading minus negates numeric vectors to give descending order. For character or factor columns where minus does not work, use order(x, decreasing = TRUE) instead, but note that flag applies to ALL keys, so for mixed-direction sorts on non-numeric keys, convert with -as.integer(factor) or use dplyr::arrange() with desc().

Explanation: unique() on a data frame returns first occurrences as full rows, keeping the original row names so you can trace each pair back to the car that introduced it. Eight pairs out of a possible 12 (gear in 3-5, cyl in 4-8) tells you something structural: every gear count appears with every cylinder count except 3-gear-and-4-cyl is missing. For deduping by a key in much larger frames, df[!duplicated(df[, keys]), ] is more memory-efficient than unique() on a wide frame.

Explanation: duplicated() returns FALSE for the first occurrence of each unique row and TRUE for every later copy, so it preserves the original row count. If you want both copies flagged (so the original also reads TRUE), call duplicated(audit) | duplicated(audit, fromLast = TRUE). To drop duplicates outright use audit[!duplicated(audit), ]. Use duplicated(audit[, c("user","action")]) to ignore noisy columns like millisecond timestamps when judging logical equivalence between rows.

Explanation: rbind.data.frame matches columns by name rather than position, so the East frame's flipped column order is harmless as long as the names and types match. If a column exists in one frame but not the other, rbind() errors loudly rather than silently filling NAs. For unioning frames with mismatched columns (where you want NA fill for missing columns), use dplyr::bind_rows() or data.table::rbindlist(fill = TRUE) instead of base rbind.

Explanation: merge() defaults to an inner join, returning only rows where the key matches in both inputs. Order 1005 for customer 4 disappears because that customer is missing from the customers table. Switch to all.x = TRUE for a left join (NA columns from the right side appear for unmatched orders), or all = TRUE for a full outer join. For million-row frames, data.table or dplyr::inner_join() are substantially faster than base merge, but the join semantics are identical.