Getting started with infosigasp

Raw, typed and clean data

read_infosiga() separates tabular import, type parsing and cleaning through the processing argument.

Every mode decodes Latin-1 to UTF-8, parses the CSV structure and combines the period files. Raw mode is therefore a lossless tabular representation rather than a byte-for-byte copy of the source archive.

Choose the mode according to the task: use "raw" to audit source values, "typed" to parse documented classes while retaining source labels and padding, and the default "clean" mode for most analysis. All modes retain import diagnostics in attr(x, "problems"); typed and clean modes replace values that fail their declared parser with NA, while raw mode preserves the field text.

raw <- read_infosiga("sinistros", processing = "raw")
typed <- read_infosiga("sinistros", processing = "typed")
clean <- read_infosiga("sinistros", processing = "clean")

Typed mode parses full dates to Date, times to hms, numeric fields to integer or double, and empty fields to NA. Identifiers remain character to preserve leading zeros and avoid integer overflow.

Clean mode then applies the following transformations:

• Year-month dates. The ano_mes_* columns (published as "YYYY/MM") become first-of-month Date values.
• Whitespace. Text columns lose their leading and trailing whitespace. The source pads some fields to a fixed width (nacionalidade ships as "BRASILEIRA "); untrimmed, those values break grouping and joins.
• Missing values. The "NAO DISPONIVEL" (“not available”) marker becomes NA (trimming runs first, so space-padded markers are caught).
• Ordered factors. Ordinal columns sort and plot in their natural order instead of alphabetically.
  • dia_da_semana: Domingo < … < Sábado (the Brazilian week starts on Sunday);
  • turno: MADRUGADA < MANHA < TARDE < NOITE;
  • gravidade_lesao (victims): LEVE < GRAVE < FATAL;
  • faixa_etaria_demografica / faixa_etaria_legal: age bands in order.
• Crash-type flags. The binary tp_sinistro_* columns ("S" / empty) become logical, so you can sum() or filter them directly. The categorical tp_sinistro_primario stays text.
• Counts. Missing qtd_* entries remain NA, including when another vehicle or injury-severity count is available for the same crash. Clean mode does not infer that an unreported count is zero.
• Numeric strings. tempo_sinistro_obito (days from crash to death) becomes integer, and numero_logradouro loses the spurious trailing ".0" the export appends ("193.0").
• Coordinates. Coordinate pairs outside the São Paulo state boundary with a 2 km buffer become NA. The buffer retains plausible near-border records, while the spatial test rejects extreme malformed values, (0, 0) placeholders and points elsewhere inside the state’s rectangular extent. No rows are dropped; typed and raw modes preserve the coordinates before this validation.

Before converting an ordinal, flag or integer column, clean mode validates its observed representations. If it finds an unexpected value, it preserves the entire column and warns; the class can therefore remain character after an upstream change.

These safeguards apply to closed-domain conversions, not to every parsed column. For example, a malformed numeric coordinate can become NA during the typed import and will be listed in attr(x, "problems").

sinistros <- read_infosiga("sinistros")
levels(sinistros$dia_da_semana)

dia_da_semana is an ordered factor, so a weekday tabulation comes out in calendar order rather than alphabetically.

table(sinistros$dia_da_semana)

Standardizing category labels

Clean mode does not harmonize nominal categories beyond its documented missing-value rule. Select only the harmonizations you need with standardize.

veiculos <- read_infosiga("veiculos", standardize = "cores")

sort(table(veiculos$cor_veiculo), decreasing = TRUE)

Vehicle colour is the clearest case. The source carries dozens of distinct values for about sixteen real colours, because two upstream systems coexist in every year, one upper-case and one title-cased, and the two disagree on gender agreement.

Note that toupper() alone will not merge Branca and Branco. Detailed liveries and multi-tone values remain unchanged because grouping them would be an analytical recode rather than label harmonization.

The remaining changes are listed below.

• "municipios". Municipality names take their official IBGE spelling, matched by cod_ibge; INFOSIGA administrative-region names receive consistent accents and case without changing their classification.
• "profissoes". Occupations are title-cased, which merges values that differ only by capitalisation. Known nonresponse markers—including variants of "NAO INFORMADA", "SEM INFORMACAO" and "DESCONHECIDO"—become NA.

Use standardize = c("municipios", "profissoes") to combine options, or standardize = "all" for every option applicable to the selected dataset. profissao keeps its unaccented spelling and occupations stay ungrouped because no authoritative list covers these free-text values. Standardization requires processing = "clean".

Joining to other Brazilian data

Always join on cod_ibge, never on the name. IBGE and INFOSIGA-SP spell nine municipalities differently. Eight contain an apostrophe that INFOSIGA-SP renders as a space ("Santa Bárbara d'Oeste" against "SANTA BARBARA D OESTE"), and IBGE’s "São Luiz do Paraitinga" appears as "SAO LUIS DO PARAITINGA". A name join loses all nine silently. cod_ibge is also the standard Brazilian municipality key, so it joins INFOSIGA-SP to census and population data. Those are the denominators that turn crash counts into crash rates.

Inspecting the structure without downloading

The package ships a small sample of each dataset so you can inspect the columns without any network access.

sample_path <- system.file("extdata", "sinistros_sample.csv", package = "infosigasp")
sample <- readr::read_delim(
  sample_path,
  delim = ";",
  show_col_types = FALSE
)
dim(sample)
#> [1] 100  48
names(sample)
#>  [1] "id_sinistro"                     "tipo_registro"                  
#>  [3] "data_sinistro"                   "ano_sinistro"                   
#>  [5] "mes_sinistro"                    "dia_sinistro"                   
#>  [7] "hora_sinistro"                   "ano_mes_sinistro"               
#>  [9] "dia_da_semana"                   "turno"                          
#> [11] "logradouro"                      "numero_logradouro"              
#> [13] "tipo_via"                        "tipo_local"                     
#> [15] "latitude"                        "longitude"                      
#> [17] "cod_ibge"                        "municipio"                      
#> [19] "regiao_administrativa"           "administracao"                  
#> [21] "conservacao"                     "circunscricao"                  
#> [23] "tp_sinistro_primario"            "qtd_pedestre"                   
#> [25] "qtd_bicicleta"                   "qtd_motocicleta"                
#> [27] "qtd_automovel"                   "qtd_onibus"                     
#> [29] "qtd_caminhao"                    "qtd_veic_outros"                
#> [31] "qtd_veic_nao_disponivel"         "qtd_gravidade_fatal"            
#> [33] "qtd_gravidade_grave"             "qtd_gravidade_leve"             
#> [35] "qtd_gravidade_ileso"             "qtd_gravidade_nao_disponivel"   
#> [37] "tp_sinistro_atropelamento"       "tp_sinistro_colisao_frontal"    
#> [39] "tp_sinistro_colisao_traseira"    "tp_sinistro_colisao_lateral"    
#> [41] "tp_sinistro_colisao_transversal" "tp_sinistro_colisao_outros"     
#> [43] "tp_sinistro_choque"              "tp_sinistro_capotamento"        
#> [45] "tp_sinistro_engavetamento"       "tp_sinistro_tombamento"         
#> [47] "tp_sinistro_outros"              "tp_sinistro_nao_disponivel"

2015–2018 covers fatal crashes only

Non-fatal records begin in 2019. Before that, every row is a fatal crash. Counting crashes per year over the whole series therefore produces a roughly 20-fold jump in 2019 that reflects only the expansion of data collection.

Those counts come from a 2026 release and shift as DETRAN-SP revises the data. The break in 2019 does not move. The same break appears in pessoas, where the LEVE and GRAVE injury levels simply do not occur before 2019.

So for any trend that reaches back before 2019, restrict to fatalities.

library(dplyr)

read_infosiga("sinistros") |>
  filter(tipo_registro == "SINISTRO FATAL") |>
  count(ano_sinistro)

Otherwise, start the series in 2019 by filtering ano_sinistro >= 2019 after importing the data.

A third of sinistros rows are notifications

tipo_registro distinguishes confirmed crashes from "NOTIFICACAO", a reported event not yet confirmed as a crash. Notifications are about a third of all rows, so treating every row as a crash overstates the total considerably.

read_infosiga("sinistros") |>
  count(tipo_registro, sort = TRUE)

DETRAN-SP reclassifies records as it validates them, so the newest months carry an unusually high share of notifications. The last month or two of any release is also incomplete. Drop the tail of the series before reading anything into a recent trend.

Smaller caveats

• tempo_sinistro_obito is capped at 30 days, the standard convention for attributing a death to a crash. Fatality counts here are 30-day counts.
• The qtd_* vehicle columns disagree with the veiculos row count for a small share of crashes. Pick one definition and say which. (The qtd_gravidade_* columns do match pessoas exactly.)
• Coordinate availability varies by year. It is highest through the middle of the series and lower at both ends, so a mapped subset is not a uniform time sample.
• About a third of crashes have no pessoas row. Use a left join to keep them.