collapse (version 1.2.1)

qsu: Fast (Grouped, Weighted) Summary Statistics for Cross-Sectional and Panel-Data

Description

qsu, shorthand for quick-summary, is an extremely fast summary command inspired by the (xt)summarize command in the STATA statistical software.

It computes a set of 7 statistics (nobs, mean, sd, min, max, skewness and kurtosis) using a numerically stable one-pass method generalized from Welford's Algorithm. Statistics can be computed weighted, by groups, and also within-and between entities (for panel-data, see Details).

Usage

# S3 method for default
qsu(x, g = NULL, pid = NULL, w = NULL, higher = FALSE, array = TRUE, …)
# S3 method for matrix
qsu(x, g = NULL, pid = NULL, w = NULL, higher = FALSE, array = TRUE, …)
# S3 method for data.frame
qsu(x, by = NULL, pid = NULL, w = NULL, cols = NULL,
    higher = FALSE, array = TRUE, vlabels = FALSE,…)
# Methods for compatibility with plm:
# S3 method for pseries
qsu(x, g = NULL, w = NULL, effect = 1L, higher = FALSE, array = TRUE, …)
# S3 method for pdata.frame
qsu(x, by = NULL, w = NULL, cols = NULL, effect = 1L,
    higher = FALSE, array = TRUE, vlabels = FALSE, …)

# S3 method for qsu
print(x, digits = 2, nonsci.digits = 9, na.print = "-",
      return = FALSE, print.gap = 2, …)

Arguments

x

a numeric vector, matrix, data.frame, panel-series (plm::pseries) or panel-data.frame (plm::pdata.frame)

g

a factor, GRP object, atomic vector (internally converted to factor) or a list of vectors / factors (internally converted to a GRP object) used to group x.

by

(p)data.frame method: Same as g, but also allows one- or two-sided formulas i.e. ~ group1 + group2 or var1 + var2 ~ group1 + group2. See Examples.

pid

same input as g/by: Specify a panel-identifier to also compute statistics on between- and within- transformed data. data.frame method also supports one- or two-sided formulas. Transformations are taken independently from grouping with g/by (grouped statistics are computed on the transformed data). However, passing any LHS variables to pid will overwrite any LHS variables passed to by.

w

a vector of (non-negative) weights. Adding weights will compute the weighted mean, sd, skewness and kurtosis, and transform the data using weighted individual means if pid is used.

cols

select columns to summarize using column names, indices or a function (i.e. is.numeric). Two-sided formulas passed to by or pid overwrite cols.

higher

logical. Add higher moments (skewness and kurtosis).

array

logical. If computations have more than 2 dimensions (up to a maximum of 4D: variables, statistics, groups and panel-decomposition) output to array, else output (nested) list of matrices.

vlabels

logical. Use variable labels in the summary. See vlabels.

effect

plm methods: Select which panel identifier should be used for between and within transformations of the data. 1L means first variable in the plm::index, 2L the second etc.. More than one variable can be supplied.

…

arguments to be passed to or from other methods.

digits

the number of digits to print after the comma/dot.

nonsci.digits

the number of digits to print before resorting to scientific notation (default is to print out numbers with up to 9 digits and print larger numbers scientifically).

na.print

character string to substitute for missing values.

return

logical. Don't print but instead return the formatted object.

print.gap

integer. Spacing between printed columns. Passed to print.default.

Value

A matrix, array or list of matrices of summary statistics. All matrices and arrays have a class 'qsu' and a class 'table' attached, responding i.e. to print.qsu and aperm.table...

Details

The algorithm used to compute statistics is well described here (see sections Welford's online algorithm, Weighted incremental algorithm and Higher-order statistics. Skewness and kurtosis are calculated as described in Higher-order statistics and are mathematically identical to those implemented in the moments package. Just note that qsu computes the kurtosis (like momens::kurtosis), not the excess-kurtosis (= kurtosis - 3) defined in Higher-order statistics. The Weighted incremental algorithm described can easily be generalized to higher-order statistics).

Grouped computations specified with g/by are carried out extremely efficiently as in fsum (in a single pass, without splitting the data).

If pid is used, qsu performs a panel-decomposition of each variable and computes 3 sets of statistics: Statistics computed on the 'Overall' (raw) data, statistics computed on the 'Between' - transformed (pid - averaged) data, and statistics computed on the 'Within' - transformed (pid - demeaned) data.

More formally, let x (bold) be a panel vector of data for N individuals indexed by i, recorded for T periods, indexed by t. xit then denotes a single data-point belonging to individual i in time-period t (t/T must not represent time). Then xi. denotes the average of all values for individual i (averaged over t), and by extension xN. is the vector (length N) of such averages for all individuals. If no groups are supplied to g/by, the 'Between' statistics are computed on xN., the vector of individual averages. (This means that for a non-balanced panel or in the presence of missing values, the 'Overall' mean computed on x can be slightly different than the 'Between' mean computed on xN.). If groups are supplied to g/by, xN. is expanded to the vector xi. (length N x T) by replacing each value xit in x with xi., while preserving missing values in x. Grouped Between-statistics are then computed on xi., with the only difference that the number ob observations ('Between-N') reported for each group is the number of distinct non-missing values of xi. in each group (not the total number of non-missing values of xi. in each group, which is already reported in 'Overall-N').

'Within' statistics are always computed on the vector x - xi. + x.., where x.. is simply the 'Overall' mean computed from x, which is added back to preserve the level of the data. The 'Within' mean computed on this data will always be identical to the 'Overall' mean. In the summary output, qsu reports not 'N', which would be identical to the 'Overall-N', but 'T', the average number of time-periods of data available for each individual obtained as 'T' = 'Overall-N / 'Between-N'. See Examples.

Apart from 'N/T' and the extrema, the standard-deviations ('SD') computed on between- and within- transformed data are extremely valuable because they indicate how much of the variation in a panel-variable is between-individuals and how much of the variation is within-individuals (over time). At the extremes, variables that have common values across individuals (such as the time-variable 't' in a balanced panel), can readily be identified as individual-invariant because the 'Between-SD' on this variable is 0 and the 'Within-SD' is equal to the 'Overall-SD'. Analogous, time-invariant individual characteristics (such as the individual-id 'i') have a 0 'Within-SD' and a 'Between-SD' equal to the 'Overall-SD'.

qsu comes with it's own print method which by default writes out up to 9 digits at 2 decimal places. Larger numbers are printed in scientific format. for numbers between 7 and 9 digits, a comma ',' is placed after the 6th digit to designate the millions. Missing values are printed using '-'.

See Also

descr, pwcor, Fast Statistical Functions, Collapse Overview

Examples

# NOT RUN {
## World Development Panel Data
# Simple Summaries -------------------------
qsu(wlddev)                                 # Simple summary
qsu(wlddev, vlabels = TRUE)                 # Display variable labels
qsu(wlddev, higher = TRUE)                  # Add skewness and kurtosis

# Grouped Summaries ------------------------
qsu(wlddev, ~ region, vlabels = TRUE)       # Statistics by World Bank Region
qsu(wlddev, PCGDP + LIFEEX ~ income)        # Summarize GDP per Capita and Life Expectancy by
stats <- qsu(wlddev, ~ region + income,     # World Bank Income Level
             cols = 9:10, higher = TRUE)    # Same variables, by both region and income
aperm(stats)                                # A different perspective on the same stats

# Panel-Data Summaries ---------------------
qsu(wlddev, pid = ~ iso3c, vlabels = TRUE)   # Adding between and within countries statistics
# -> They show amongst other things that year and decade are individual-invariant,
# that we have GINI-data on only 161 countries, with only 8.42 observations per country on average,
# and that GDP, LIFEEX and GINI vary more between-countries, but ODA received varies more within
# countries over time.

# Using plm:
pwlddev <- plm::pdata.frame(wlddev,         # Creating a Panel-Data Frame frame from this data
                  index = c("iso3c","year"))
qsu(pwlddev)                                # Summary for pdata.frame -> qsu(wlddev, pid = ~ iso3c)
qsu(pwlddev$PCGDP)                          # Default summary for Panel-Series (class pseries)
qsu(G(pwlddev$PCGDP))                       # Summarizing GDP growth, see also ?G

# Grouped Panel-Data Summaries -------------
qsu(wlddev, ~ region, ~ iso3c, cols = 9:12) # Panel-Statistics by region
psr <- qsu(pwlddev, ~ region, cols = 9:12)  # Same on plm pdata.frame
psr                                         # -> Gives a 4D array
print.qsu(psr[,"N/T",,])                    # Checking out the number of observations:
# In North america we only have 3 countries, for the GINI we only have 3.91 observations on average
# for 45 Sub-Saharan-African countries, etc...
print.qsu(psr[,"SD",,])                     # Considering only standard deviations
# -> In all regions variations in inequality (GINI) between countries are greater than variations
# in inequality within countries. The opposite is true for Life-Expectancy in all regions apart
# from Europe, etc...
psrl <- qsu(wlddev, ~ region, ~ iso3c,      # Same, but output as nested list
            cols = 9:12, array = FALSE)
psrl                                        # We can use unlist2d to create a tidy data.frame
head(unlist2d(psrl, c("Variable","Trans"),
                    row.names = "Region"))

# Weighted Summaries -----------------------
n <- nrow(wlddev)
weights <- abs(rnorm(n))                    # Generate random weights
qsu(wlddev, w = weights, higher = TRUE)     # Computed weighted mean, SD, skewness and kurtosis
weightsNA <- weights                        # Weights may contain missing values... inserting 1000
weightsNA[sample.int(n, 1000)] <- NA
qsu(wlddev, w = weightsNA, higher = TRUE)   # But now these values are removed from all variables

# Grouped and panel-summaries can also be weighted in the same manor
# }