write_dta support for long string (strL)?

[deribo]

I need to process a .json containing logdata into a .dta file. The file itself contains a handful of string variables exceeding 5000 and more characters.

As Stata (since Stata 13) supports long string variables (strL variable) and haven() supports long string at least in write_sav() I wanted to ask if it is in any way possible to write strL-variables with haven (haven't found anything on this)? If not if there's support for this planned in the future?

Bests

library(haven)

longFun <- function(n) {
  do.call(paste0, replicate(5000, sample(LETTERS, n, TRUE), FALSE))
}

longString <- data.frame(V1 = longFun(1), stringsAsFactors = F)
                         
write_dta(longString, paste0(tempdir(), "\\longString.dta"))
#> Error in write_dta_(data, normalizePath(path, mustWork = FALSE), version = stata_file_format(version)): Writing failure: A provided string value was longer than the available storage size of the specified column.

[evanmiller]

strLs are supported in the underlying ReadStat library (as READSTAT_TYPE_STRING_REF) - it will be up to the haven authors to implement support on the R side.

[hadley]

@evanmiller what would support look like? We'd need to detect if a string was longer than some threshold, and if so, change variable types?

[evanmiller]

@hadley Probably something like that. The relevant threshold depends on the file version

#define DTA_OLD_MAX_WIDTH    128
#define DTA_111_MAX_WIDTH    244
#define DTA_117_MAX_WIDTH   2045

Then use

readstat_string_ref_t *readstat_add_string_ref(readstat_writer_t *writer,
    const char *string);

readstat_error_t readstat_insert_string_ref(readstat_writer_t *writer,
    const readstat_variable_t *variable, readstat_string_ref_t *ref);

If a string is repeated you can re-use the reference, so some kind of hash table of string refs might make sense (but it's not necessary).

[hadley]

Makes sense; I'm unlikely to do this in the near term, but it makes sense to do in the future.

[hadley]

Is there a reason to not use strL variables by default? (assuming version > 13)

[hadley]

@evanmiller should I need to do more than this for it to work?

// when creating variable
readstat_add_variable(writer_, name, READSTAT_TYPE_STRING_REF, 0);

// when inserting values
readstat_insert_string_ref(writer_, var, readstat_add_string_ref(writer_, val));

It writes fine, but when I read I get "Invalid file, or file has unsupported features."

[evanmiller]

@hadley I'd need to see the whole program; don't forget the begin row / end row stuff and also check the return value of readstat_insert_string_ref.

[hadley]

You can see my efforts so far in https://github.com/tidyverse/haven/pull/584/files (ignore the unordered_map stuff; I'm not using that yet). Basic reprex showing that short strings work:

library(haven)

long_string <- function(n, m) {
  do.call("paste0", replicate(m, sample(LETTERS, n, TRUE), simplify = FALSE))
}

df1 <- data.frame(x = long_string(10, 3000), stringsAsFactors = FALSE)
df2 <- data.frame(x = long_string(10, 2000), stringsAsFactors = FALSE)

path <- tempfile()
write_dta(df1, path); read_dta(path)
#> Error: Failed to parse /private/tmp/RtmpbcBlqL/file22a13813e594: Invalid file, or file has unsupported features.
write_dta(df2, path); read_dta(path)
#> # A tibble: 10 x 1
#>    x                                                                            
#>    <chr>                                                                        
#>  1 EHKANXLOEWRBYWATCSFBTVNYPEGACKDXOINTAPZXEDQIFBWVPTRHRGFLLHPSZXJSYVQQJMEQEKVE…
#>  2 BAPXNRUZDAIBWWGZBRNTVYZAGVYPZKHENVLXELZHIRODCGMJWYTHICPQSQUPCECJEYUYZKJDUYCU…
#>  3 IYQKLASDVPJYJUCGVEFKRXYCKCIAPXYTUASZACLJAMJQDDSBJNWGWGAHIYAGSJYLSUWCMOXTXPOS…
#>  4 HQTJUVYZQILDACSAREXAHIJSPBHTUKQDRYEWUGCMWVCOCVKIZYCGUZAJOWZCXYWLQTZOLAGNKQRW…
#>  5 QZCREAKKWDQGQRJPZBNNHXHYQVYNIINLOBOVQPEVXERRAZOHRQJKNJPTSWOASIGTKJXQIWFJWKWJ…
#>  6 WAWYXUGPVHCYKVDBKGXSSEMASLOZRQKQWQCASOLONEOZBPPBZKPYMUECSSWFLFUFFYNDAOUOAGTN…
#>  7 EQOWRNDEBBSXZBSBREUNJWIOZFIERHFUWXGFRJOTHMYWIKXQWAXQQGTKOAZCNJTOQAAFKQXWLQDX…
#>  8 GHJDZYJXMSUXCOOBKGXMZNGOPRBQZNJROXRPZOCVQCHFEAHTVBLUSBNICZLMLDHYOHMSZKECWNRN…
#>  9 BDRJURSKGNAKWFFFAHAWPBSPVQSAQVPOWHYSREZVJIAYWUSZYRUSNLRPZNNWEVGTSTTXAEXZSRFY…
#> 10 ESCPCHPKHJWWWICVJUCAMKKHJPOUQFEOFDIKATKBYNGSXFRPDTERQEYOGNMSIJWBXULVXBMDFLOP…

^{Created on 2021-04-10 by the reprex package (v2.0.0)}

[arthurgailes]

Is there a reason to not use strL variables by default? (assuming version > 13)

For short strings, strL can consume upwards of 10x as much memory. In addition, they throw errors in some Stata commands, e.g. merge.