1. PhasesThis project was carried out in four distinct phases:Data sourcing: Relevant datasets were identified and downloaded from the SIA’s online archive.Data ingestion and structuring: Annual datasets were ingested, structured, and compiled into a single dataset upon which queries could be run. Data was ingested using manual entry and Optical Character Recognition (OCR) utilising Anthropic’s Claude Sonnet large language model. All data was manually verified post-ingestion and transcription errors were rectified. Annual datasets were structured and normalised using the Pandas package for Python in Visual Studio Code (v1.109), where ‘Year’ and ‘Region’ headers were also added. The full dataset was compiled and formatted for MS Excel.Cleaning and EDA: All rows were manually checked against the original datasets for transcription accuracy, and recording errors were resolved. Exploratory Data Analysis (EDA) was then conducted in MS Excel and Jamovi (v2.7.13) to identify trends for further exploration.Analysis: Comprehensive analysis was conducted in MS Excel using Excel formulae, the Jamovi standard statistical package, and the Pandas package for Python in Visual Studio Code (v1.131). Further information, such as Company Size and Fee, was added during this phase. The results are presented alongside discussion of their significance in the full report.2. DatasetsAll quantitative data used in this study was downloaded from the publicly accessible archive hosted on the SIA’s website. Datasets are listed annually and include all PSCs registered with the SIA that year. All datasets (except 2024, discussed separately) disaggregate data by province and region, with rows organised under the following headings: index number, company name, licence number, licence class, and number of PSO permits issued. Datasets from 2015 to 2019 also record compliance status; for these years, only companies listed as Compliant were included. Datasets for 2015 to 2023 are published in December and are explicitly authorised by Paul Kingston Isari, the SIA Registrar, as formal public advertisements.The 2024 dataset is distinct, having been published online on 31 August 2024 as a print-out of the dataset rather than the usual formal public advertisement. It includes the additional headers “Registered Business Centre,” “Business Postal Address,” “Date of Expiry,” “Licence Holder,” and “Date of Issue.” This dataset may not include companies registered in 2024 on or after 1 September. The results in 2024 may therefore be undercounted compared to other years.3. Data DictionaryThe following headings were used in this analysis:Company Name: The registered business name of the firm offering security services. Used in this analysis to proxy ‘continuing companies’ (see Section 6a, Exact Name Matching).Licence Class: Which class of licence the company registered for, restricted to A, B, or C.Number of PSO Permits Issued: Also called “Guards,” “Guard Numbers,” etc. Used to calculate company size for analysis disaggregation.The following information was obtained from the datasets and listed in newly created columns:Province: The province in which the company was registered.Code: A three-letter code unique to each province. Useful during the EDA process, but fundamentally provides the same information as the “Province” heading.Region: The region containing the province in which the company was registered. The Autonomous Region of Bougainville (AROB) was listed separately as a region.The following information was calculated and added to the dataset in new columns:Company Size: Delimited by the number of PSO permits issued to the company, used to proxy PSO employment. Companies were designated ‘Small’ (< 51 PSOs), ‘Medium’ (51–200), or ‘Large’ (> 200). This column was useful in aggregating companies by size and calculating annual SIA revenue.Fee: Calculated using fee-schedule information published online by the SIA, including the increased fee applicable above the base rate and the 22 PGK fee per permit granted. The result is the total base fee payable by each company for a given year’s registration (exclusive of other possible fees, such as the 1,000 PGK late-registration fee), and was useful in analysing revenue. Calculated for each company using the following Excel formula (where Cell_A = number of PSOs and Cell_B = Licence Class):=LET(val, ‘Dataset’![Cell_A],baseFee, IF(val<51, 1100, IF(val<=201, 2750, 5500)),adjustedFee, IF(‘Dataset’![Cell_B] = A, baseFee*2, baseFee),adjustedFee + val * 22)The following information was not used:Licence Number: Some companies changed licence numbers year-to-year, while others retained theirs. No immediate use was identified for licence numbers in this analysis, and they were removed from the compiled dataset.Province Index Number: Index numbers were included for each province. No immediate use was identified for these numbers and they were removed from the compiled dataset.4. ProxiesThis analysis relies on two proxies: 1) the count of company licences issued by the SIA and recorded in the dataset as representative of the number of active registered companies, and 2) the count of PSO permits issued and recorded in the dataset for each company as representative of the number of PSOs employed by that company at that temporal interval (in this case, 1 year).The first proxy incorporates companies which may have paid licensing fees but are dormant or non-operational within the total count of registered (assumedly operational) companies. The extent to which these companies may be present within the dataset is unknown and requires qualitative examination of a significant industry sample to estimate. No such work has been published. However, economic factors make a scenario where many firms are dormant unlikely. PNG’s private security industry is fiercely competitive. Wages are low, margins are thin, and the majority of operators are competing in the same size/class niche. If province information recorded at registration can be relied upon, a substantial number are also co-located in the same cities. In this context, the author assesses that the proportion of firms willing to pay the comparatively high costs of registration without conducting business is likely low. This proportion is very likely even lower for companies renewing their registrations. For this economic reason, the author assesses that the trends identified using the count of company licences issued by the SIA as a proxy for the number of active registered companies are still reliable indicators.Equally, the second proxy may incorporate PSOs who are non-active in their roles within the count of permit-carrying (assumedly operational) PSOs. This is less likely than for the first proxy, again for economic reasons. PSOs are required to have salary and superannuation contributions paid by their employers. Permits are also specific to each PSO, meaning a company cannot purchase permits in bulk from the SIA and issue them to PSOs when they are hired. This is done to record information about training and work history, likely to aid in the SIA’s compliance-checking. Inactive PSOs are therefore an ongoing cost to companies. For this reason, and given the PNG context stated above, the author assesses that the trends identified using the count of PSO permits issued and recorded in the dataset for each company as representative of the number of PSOs employed by that company at that temporal interval are still reliable indicators.The SIA provides clear and simple steps for de-registering a security company and revoking PSO permits on its website. The SIA also implores interested parties to ensure permit details are accurately recorded to avoid risking fines.5. ErrorRecording errors were identified during initial data ingestion and structuring, during the EDA, and during compilation. Common errors identified were missing index numbers, mistyped or misprinted data (e.g., numeric columns containing letters), and province/region misspellings. These were manually corrected and verified whenever possible. Resolved errors in the original datasets numbered approximately 5–6 per annual dataset.Licence numbers adhered to a common format, which in some cases was broken and likely indicative of a recording or printing error. Given their limited utility in this analysis and little prospect of manual verification, these were removed from the compiled dataset.Possible spelling errors were identified in company names, including misspellings of city and province names and common terms associated with the private security industry. These were left unchanged in all instances, as such errors may be intentional branding choices by business owners; it would not be possible to verify intentionality versus recording, entry, or printing error for all possible misspellings. Possible business-name errors were identified in approximately 1 per cent of all rows. Where business names were unreadable, which predominantly affected the 2016 dataset owing to binder holes in the scanned paper records, “[ILLEGIBLE]” was added to the company name. This affected fewer than 10 rows and had a negligible impact on analyses.Six of the 5,276 total rows were missing data on licence class or number of PSOs. These were removed from the compiled data; this likely had a negligible impact on analyses.Summary information recorded in the 2015–2019 datasets giving registered-company and PSO-permit totals at the provincial, regional, and national levels was found to be erroneous on several occasions. Errors in summary information were generally associated with recording errors in the relevant section, and some summaries excluded entries without explanation, undercounting totals. Regional summaries may have been calculated from aggregated provincial summaries containing errors, with these errors observed to carry through, in some cases, to the national summary. For this reason, no summary data was used in this analysis; all calculations were made directly from the compiled dataset. This, alongside the aforementioned deletions of rows with missing critical data, explains the small discrepancy between totals given in the official summaries and those presented in this analysis. The disparity is not significant enough to cast doubt on the SIA’s data collection.6a. Exact Name MatchingCompany names were matched across two-year intervals as a proxy for companies choosing to renew their registration and therefore continuing their business operations. This proxy may include companies providing services for a single, recurring short-term contract (e.g., for an annual event). It will also exclude companies renewing their registration under a different business name (hereafter, ‘rebranding’), and cannot detect potential matches where a name has been erroneously entered. Despite these caveats, the trends identified by this matching process lend support to its utility as a quantitative method in this case.Matches were identified by a script written by the author in Python, using the Pandas package, run in Visual Studio Code. Company names were trimmed of the strings “limited” and “Ltd,” all bracketed text (including “(PNG)” and “[ILLEGIBLE]”), and whitespace. This standardisation increased match fidelity without introducing false positives, as evidenced in manual testing. Year-list sets were then filtered by licence class, number of PSOs, or both, before being matched, with results printed to a table. The script identified total matches across both continuous and discontinuous selectors (e.g., a company moving from Small to Large, or Class B to Class A); however, filtered matches were included only if continuous (e.g., remaining Small and Class A across both years). A separate script identified these ‘missing’ discontinuous matches: companies whose names matched across both years but whose licence class or size changed.The results demonstrate that, over a two-year period, companies changing licence class or size is highly unusual. A small but clear trend was visible in companies upgrading from Class B to Class A in a given year. The identification of this trend strengthens the case for this method’s use in detecting potential rebranding: rebranding-driven changes in licence class are assumed to affect transitions uniformly (A–B, A–C, B–A, B–C, C–A, and C–B are all treated as equally likely), and a company seeking to capitalise on a class change by rebranding would most plausibly do so when moving into the highest, most expensive tier, Class A. That a trend in B–A transitions is nonetheless evident is further evidence supporting the validity of this method and its results in this analysis.False rebranding caused by data errors (e.g., misspellings) was likewise assumed to be equally likely to affect every recorded business name. Because large companies are comparatively few, their match percentages may be disproportionately affected by recording error; as industry leaders, these companies are also critical to the analysis. Special attention was therefore paid to manually verifying large-company business names, which was feasible given their comparative prominence in the sector and their advertising presence.6b. Temporal IntervalsName matching was carried out for nine distinct two-year intervals present in the compiled dataset: 2015–2016, 2016–2017, 2017–2018, 2018–2019, 2019–2020, 2020–2021, 2021–2022, 2022–2023, and 2023–2024. Further scope exists to analyse this dataset at other intervals (e.g., 3, 4, or 5 years), which could better articulate, and potentially forecast, company survivorship over different periods since 2015. Given the already broad suite of methods deployed in this analysis, however, temporal-interval analysis was deemed out of scope.The results of this method should be considered indicative, providing a sound lowest-possible estimate of companies surviving each two-year interval. While scope plausibly exists for false negatives caused by company rebranding and data-recording error, extremely limited scope exists for false positives.6c. False PositivesA small but real possibility exists that this method could record false positive results; for example, if a business closes and another, separate business takes on the previously registered name the following year. This may be more likely for generic business names, or where an existing name is sold or used under agreement. No examples of this were identified during the EDA, nor when reviewing previous work on PNG’s private security industry.7. Logistic Regression Model for Company Renewal RatesTo test whether company registration-renewal rates changed over 2015–2024, and whether this trend differed by company size, renewal outcomes were modelled using logistic regression. For each year (2016–2024; 2015 was excluded, as no prior-year renewal base exists for it) and company-size category (Small, Medium, Large), the number of prior-year-registered companies eligible to renew and the number that renewed were treated as binomial counts:renewals ~ Binomial(n, p),n = prior-year registrationslogit(p) = β0 + β1(year) + β2(size) + β3(year×size)Small companies were specified as the reference category, such that coefficients for Medium and Large companies are expressed relative to Small companies. The year×size interaction term tests whether the temporal trend in renewal rates differs across company sizes.Coefficients were estimated by maximum likelihood (iteratively reweighted least squares). Model fit was assessed via the ratio of the Pearson chi-squared statistic to residual degrees of freedom, with values substantially above 1 indicating overdispersion relative to the binomial variance assumption. The initial model returned a dispersion ratio of approximately 2.5. Inspection of residuals by year and company size traced this overdispersion to three large deviations: Small companies in 2017, and Small companies in 2020 and 2021 (with 2020–2021 renewal rates falling well below model predictions, possibly reflecting pandemic-related disruption that disproportionately affected small companies). Statistical significance of individual coefficients was assessed at α = 0.05.Addendum: Logistic Regression ResultsThe following tables report the full output of the logistic regression model described in Section 7 of the Methodology annex (renewal rates by company size, 2016–2024). No correction for overdispersion or anomalous years has been applied; the model is reported as fitted.
Click Here For The Original Source
——————————————————–
