Types of studies

Given the infeasibility of conducting randomised studies (especially with randomisation of the individual participant) in this area, we considered eligible for inclusion in our review all randomised as well as non‐randomised studies of the following types.

• Randomised controlled trials (RCTs), defined as studies in which participants are randomly allocated into groups to receive an intervention. Identical treatment is provided to all groups with the exception of the intervention the research is designed to study (Hammond 2015).

• Cluster‐randomised trials (CRTs); that is, RCTs that involve groups of participants, as opposed to individuals, as the unit of randomisation. Comparisons are then made between these clusters rather than between individuals (Kaura 2015).

• Interrupted time‐series (ITS) studies, where observations of a group are taken repeatedly over time and used to establish an underlying trend, which is then interrupted by an intervention. The analysis of ITS data provides statistical evidence about whether changes in the trend represent real increases or decreases (Bernal 2017).

• Controlled before‐after (CBA) studies, in which outcomes of interest are measured in both intervention and control groups before and after an intervention has been performed (EPOC 2017a).

In order to be inclusive and to facilitate capturing all relevant data, we considered studies published in the peer‐reviewed literature as well as unpublished data from clinical trial registries to be eligible for inclusion.

While the highest‐quality (lowest risk of bias) results could be expected from RCTs, we reasoned that restricting the review to these trials only might neglect evidence in the form of studies without individual participant randomisation. Indeed, for RDAT, as previously mentioned, randomisation of the individual participant may not be practical. Therefore, we made the decision to include other study designs (CRTs, ITS studies and CBA studies), which are easier to conduct in an occupational health setting, even though they are more prone to bias. Such studies could provide important, practically relevant information on the effectiveness of RDAT. We have taken the higher risk of bias inherent in these study designs into account in our analysis and conclusions.

Types of participants

We considered studies conducted on adult workers in any occupation as eligible for inclusion, with the exception of commercial drivers, as this is the subject of another Cochrane Review (Cashman 2009).

Types of interventions

Studies that evaluated the effectiveness of workplace RDAT according to the following criteria were eligible for inclusion.

• Randomness: each worker studied had to have an equal likelihood of being chosen for testing, and with the choice made through a probabilistic method (e.g. using a random‐number table, a computer‐generated list of random numbers, or drawing numbers out of a hat to select from a list of consecutively numbered names).

• Substances tested: random testing in the study was performed to detect:

  • alcohol;

  • illicit or prescription substances other than alcohol; or

  • both alcohol and illicit or prescription substances.

• Frequency of testing: random testing in the study was performed on at least a quarterly basis or more frequently, with scheduled retesting at regular intervals as a planned part of the program.

• Proportion of workers tested: random testing in the study could be performed on any proportion of the worker population, as long as test results from at least 10 subjects were described in each contributing study.

• Setting: the study could be conducted in any work‐related setting, excluding those where the tested employees were commercial drivers (as this is covered in another Cochrane Review, Cashman 2009).

We considered studies that investigated co‐interventions to be eligible for inclusion, provided that they were implemented in both the RDAT and comparator arms.

Types of outcome measures

Studies that evaluated the effectiveness of workplace RDAT by using any suitable measures of any of the outcomes below were eligible for inclusion.

Electronic searches

We conducted a systematic literature search to identify all published and unpublished studies that could be considered eligible for inclusion in this review. We adapted the search strategy we developed for MEDLINE (Appendix 1) for use in the other electronic databases. We did not restrict the search by date or language of publication.

We searched the following databases and registers to identify potential studies.

• Cochrane Central Register of Controlled Trials (CENTRAL; 2020, Issue 11) in the Cochrane Library (searched 4 November 2020).

• MEDLINE Ovid (1946 to 30 October 2020) (searched 1 November 2020 (Appendix 1).

• PsycINFO Ovid (1806 to October week 4 2020) (searched 2 November 2020) (Appendix 2).

• Embase Ovid (1974 to 29 October 2020) (searched 4 November 2020) (Appendix 3).

• Google Scholar (https://scholar.google.ca/; searched 3 November 2020).

• Occupational Safety and Health (OSH) Update (19th century to 24 May 2019). We previously had access to this database through the Cochrane Work Editorial Group, but did not have such access for our search update in November 2020.

• OSH References Collection (19th century to October 2020).

• ClinicalTrials.gov (www.ClinicalTrials.gov; searched 3 November 2020).

We searched the World Health Organization International Clinical Trials Registry Platform (www.who.int/ictrp/en) in May 2019. Due to high traffic related to COVID‐19, this site was not searchable despite multiple attempts when we updated our search in November 2020, and so we searched the ISRCTN Registry and EU Clinical Trials Register (searched 4 November 2020) instead.

Searching other resources

We searched the reference lists of original studies and review articles for additional references and continued this in an iterative fashion until no new studies were identified. We also searched the publication history of frequently cited authors known to us.

Selection of studies

We used Covidence 2020 to assist with study selection. We conducted the selection of eligible studies in two stages. First, two review authors (out of CE, TJ, MM, GW, DS, DD) independently screened the titles and abstracts of the systematic search results to identify studies for inclusion. Each reviewer assessed studies as potentially eligible or ineligible. We excluded, as ineligible, studies that clearly did not fulfil our inclusion criteria or definitely fulfilled one or more exclusion criteria. At the second stage, we retrieved the full‐text publications of the potentially eligible studies and two review authors (out of CE, TJ, DK, GW, DS) independently assessed these and screened studies for final inclusion. We recorded reasons for exclusion of studies assessed as full‐texts in Table 2. We resolved any disagreement through discussion or, if required, we consulted another review author (SS).

Data extraction and management

We used a data extraction form for study characteristics and outcomes. Two review authors (TJ and DK) independently extracted study characteristics from the included study. Disagreements were resolved by consensus, or if required, by consulting another review author (SS). Where available, we extracted the following study characteristics.

• Study information: study design; study duration; and study location.

• Participant information: number of participants; number of withdrawals; reason for withdrawals; sex or gender; mean age or age range; race or ethnicity; occupation(s); employer or company; and study inclusion/exclusion criteria.

• Interventions: screening methods; whether screening was for alcohol, non‐alcoholic substances, or both; frequency of screening; number of employees tested; and total number in testing pool.

• Outcomes: fatal injuries; non‐fatal injuries; non‐injury accidents; rate of positive results found by RDAT; absenteeism; and adverse events.

We noted in the 'Characteristics of included studies' table if outcome data were not reported in a usable way.

We took occupations specifically identified for the study population and coded these occupations into the 10 broad occupational structure categories and the unit group codes of Canada's National Occupational Classification (Statistics Canada and ESDC 2018). We also reported the industry employing the study population as a code in the North American Industry Classification System (NAICS) (Statistics Canada 2012). Two review authors (TJ, GW) independently performed and reconciled this coding, resolving disagreements by consensus or by involving another review author (SS).

We did not enter any data into the Data and analysis section of Review Manager 5.4 (RevMan 2020). Had data been available, one review author (TJ or GW) would have entered data into Review Manager 5.4. A second review author (CE) would have confirmed that the data were entered correctly by spot‐checking study characteristics for accuracy against the study reports, and comparing the data presented in the systematic review with those in the study reports. Had we decided to include studies published in languages in which our author team is not proficient, we would have arranged for a native speaker or a sufficiently proficient translator to complete a data extraction form for us.

Assessment of risk of bias in included studies

Two review authors (CE, GW) independently assessed the risk of bias in the included study. We resolved any disagreements by discussion or by involving another author (SS).

For the ITS study we included, we used the risk of bias criteria developed by the Cochrane Effective Practice and Organisation of Care (EPOC) Group (EPOC 2017b). Thus, we graded each potential risk of bias as 'high', 'low', or 'unclear' in each of the following domains.

• Was the intervention independent of other changes?

• Was the shape of the intervention effect prespecified?

• Was the intervention unlikely to affect data collection?

• Was knowledge of the allocated interventions adequately prevented during the study?

• Were incomplete outcome data adequately addressed?

• Was the study free from selective outcome reporting?

• Was the study free from other risks of bias?

Had it been necessary, we would have assessed the risk of bias in RCTs and CRTs according to the following standard domains, grading each potential risk of bias as 'high', 'low', or 'unclear'.

• Random sequence generation

• Allocation concealment

• Blinding of participants and personnel

• Blinding of outcome assessment

• Incomplete outcome data

• Selective outcome reporting

• Other biases

Had it been necessary, we would have assessed CBA studies using the risk of bias criteria as given in Sterne 2016, grading each potential risk as 'low', 'moderate', 'serious', 'critical', or 'no information', in each of the following domains.

• Bias due to confounding

• Bias in selection of participants into the study

• Bias in classification of interventions

• Bias due to deviations from intended interventions

• Bias due to missing data

• Bias in measurement of outcomes

• Bias in selection of the reported result

Potential confounding domains that we anticipated would be relevant to included studies were socioeconomic status, age, sex, and tobacco smoking.

Measures of treatment effect

Had we found suitable data, we would have entered the outcome data for each study into the data tables into Review Manager 5.4 (RevMan 2020), in order to calculate the treatment effects.

For the one included ITS study, we extracted and re‐analysed their data to assess differences in average percent of positive tests (level) and in the time trend (slope). To do this, we followed the procedure described in Ramsay 2003. We performed multiple checks for auto‐correlation and used the most appropriate regression model to assess pre‐ and post‐intervention level and time‐trends. Each data point was defined as an annual period, starting with period 1. Periods 1, 2, and 3 were pre‐intervention, and periods 4, 5, 6, 7, and 8 were post‐intervention.

The general model tested was:

Y = β₀ + β₁t + β₂I(t ≥ 4) + β₃tI(t ≥ 4) + ε,ε~N(0,σ)

Period 4 is the beginning of the intervention, and the indicator variable I(・) is one when the condition in the parenthesis is true, and zero otherwise. The parameters have the following interpretation:

β₀ is the pre‐intervention intercept

β₁ is the slope of the regression pre‐intervention

β₂ is the level change pre‐ and post‐intervention

β₃ is the change in slope post‐intervention

We performed the statistical analysis in Stata 15 for Windows (Stata 2017).

We allowed for the possibility that the errors were auto‐correlated, and in that case, we would have used the Prais‐Winston first‐order auto‐correlation version of generalised least squares (GLS) regression estimation, otherwise reverting to ordinary least squares. We defined a change in level as an immediate shift in the regression line post‐intervention, and a change in slope post‐intervention representing a gradual response to the intervention.

Unit of analysis issues

Had we included studies that employed a cluster‐randomised design and reported sufficient data to be included in the meta‐analysis, but did not make an allowance for the design effect, we would have calculated the design effect based on a fairly large assumed intracluster correlation of 0.10. We based this assumption of 0.10 being a realistic estimate on an analogy with studies on implementation research (Campbell 2001). We would have followed the methods described in the Cochrane Handbook for Systematic Reviews of Interventions for the calculations (Higgins 2020).

In the re‐analysis that we did on data reported in the included study, given that these data were reported on an annual basis, the unit is the proportion of positive outcome randomised drug testing.

Dealing with missing data

We would have employed a conservative approach for dealing with missing data, preferring baseline observation carried forward over last observation carried forward, if both were reported. If it had been possible and appropriate to calculate values for missing data from other statistics reported in studies, we would have done so. If such computation were not possible, we would have contacted authors to request additional data and reported which study analyses made use of unpublished data.

Assessment of heterogeneity

We would have used the I² statistic to assess statistical heterogeneity among the studies in each meta‐analysis. We would have discussed any substantial statistical or clinical heterogeneity.

We would have considered the following three substance groupings as being clinically heterogeneous.

• Alcohol only

• Illicit or prescription substances but no alcohol

• Alcohol and any illicit or prescription substances

We would have provided further detail about the planned assessment of clinical heterogeneity in the section 'Subgroup analysis and investigation of heterogeneity'.

Assessment of reporting biases

Had we been able to pool data from more than 10 studies in any single meta‐analysis, we would have created and examined a funnel plot to explore possible reporting biases.

Data synthesis

As we included only one study, we could not pool data nor perform a meta‐analysis.

Subgroup analysis and investigation of heterogeneity

We planned to conduct subgroup analyses if there were sufficient data. We would have examined the effects of RDAT according to the presence or absence of four elements:

• safety‐sensitive work;

• manual labour;

• testing for cannabinoids; and

• testing for opioids.

We would have treated studies of different designs separately; that is, we would not have combined data from the four eligible study designs. We would have discussed the implications of 'high' or 'low' risks of bias, inherent in these study designs and specific to the studies.

Further, we would have pooled outcome data for three comparable time points, defining short‐term follow‐up to be up to and including one month, medium‐term to be greater than one month and less than one year, and long‐term as one year or greater.

Sensitivity analysis

We would have performed a sensitivity analysis to assess the robustness of our conclusions by omitting studies with a 'high' risk of bias, if studies with both 'low' and 'high' risks of bias had been present.

Summary of findings and assessment of the certainty of the evidence

We constructed a 'Summary of findings' table that presents our GRADE judgements concerning the certainty of the evidence.