Home

In medical research, a central objective is to determine whether specific risk factors or interventions have a meaningful impact on clinical outcomes or disease progression. The most rigorous approach to answering such causal questions is the randomized controlled trial (RCT), in which participants are prospectively assigned to interventions under controlled conditions.

Evidence Pyramid; from the Oxford Centre for Evidence-Based Medicine

RCTs are considered the gold standard because they:

• Directly address causal questions
• Provide well-defined causal estimands
• Establish a clear time zero (start of follow-up)
• Ensure structured reporting of drop-outs and missing data
• Minimize confounding through randomization

However, RCTs are not always feasible due to ethical, logistical, financial, or time constraints. Consequently, a substantial portion of clinical evidence is derived from observational studies, including cohort studies and registries, which may be collected prospectively or retrospectively.

Association vs. Causality

A critical distinction in observational research needs to be made between:

• Association: A statistical relationship between variables
• Causality: A direct effect of one variable on another, implying generalizability beyond the study population

The latter, causality, is very challenging in observational studies, as these studies lack randomization and are therefore vulnerable to multiple sources of bias, including selection bias, information bias, attrition bias, time bias and confounding (both measured and unmeasured). These biases complicate causal interpretation and limit the strength of conclusions that can be drawn from observational studies.

Target Trial Emulation: Causal inference from observational data?

Target trial emulation is one concept that has been develop to address this problem. The approach aims to approximate (emulate) a pragmatic randomized trial using observational data – thereby enabling causal inference from such studies.

Rather than analyzing observational data in an ad hoc manner, researchers explicitly define the protocol of a hypothetical RCT (the “Target Trial”) and then emulate it using the available data.

Key Steps in Target Trial Emulation

1. Define the target trial protocol
   
   Specify all elements of a hypothetical RCT, including:
   o Eligibility criteria
   o Treatment strategies (interventions and comparators)
   o Study procedures
   o Start of follow-up (time zero)
   o Outcomes of interest
   o Statistical analysis plan (pre-specified)
   
2. Emulate the trial in observational data
   
   o Identify individuals in the dataset who meet the eligibility criteria
   o Ensure data availability across all relevant time points
   o Align exposure, follow-up, and outcome definitions with the target trial
   
3. Analyze according to the predefined plan
   
   o Perform analyses strictly following the predefined protocol
   o Estimate effect measures such as risk ratios (RR) or hazard ratios (HR)

How to specify and emulate a target trial (from Hernán et al. 2016 and Hansford et al. 2023)

This approach aims to reduce selection bias (by clearly defining the study population in advance), minimize time-related biases (e.g., immortal time bias) through explicit definition of time zero, increase transparency and reproducibility (via a protocol-driven design) and enable more causal interpretation compared to standard observational approaches.

However, target trial emulation cannot fully replicate an RCT. For once, target trials will always remain pragmatic trials conducted in real-world conditions, and randomization or blinding cannot be emulated. Therefore, residual confounding, measurement and performance bias remain major concerns, and need to be handled using methods such as multivariable regression or sensitivity analyses (see our Stats Made Simple Episode on bias and confounding). Additionally, the approach depends strongly on data quality; poor or incomplete data can never be fully corrected statistically

Therefore, target trials are well-suited for pragmatic, real-world questions, where high-quality, large-scale observational data are available, but randomized trials are  currently not available or impractical.

To summarize, target trial emulation represents an interesting statistical framework for improving causal inference from observational research, and the method is increasingly used in medical research. By explicitly mimicking a randomized trials, it enhances methodological rigor and interpretability of observational studies. However, careful attention to confounding, bias, and data quality remains essential, and findings should be interpreted within these constraints.

REFERENCES

1.    Hansford HJ, Cashin AG, Jones MD, Swanson SA, Islam N, Dahabreh IJ, et al. Development of the TrAnsparent ReportinG of observational studies Emulating a Target trial (TARGET) guideline. BMJ Open. 2023 Sep 12;13(9):e074626.
2.    Cashin AG, Hansford HJ, Hernán MA, Swanson SA, Lee H, Jones MD, et al. Transparent Reporting of Observational Studies Emulating a Target Trial-The TARGET Statement. JAMA. 2025 Sep 23;334(12):1084–93.
3.    Young JG, Stensrud MJ, Tchetgen Tchetgen EJ, Hernán MA. A causal framework for classical statistical estimands in failure-time settings with competing events. Stat Med. 2020 Apr 15;39(8):1199–236.
4.    Hernán MA, Sauer BC, Hernández-Díaz S, Platt R, Shrier I. Specifying a target trial prevents immortal time bias and other self-inflicted injuries in observational analyses. J Clin Epidemiol. 2016 May 27;79:70–5.
5.    Hernán MA, Wang W, Leaf DE. Target trial emulation: A framework for causal inference from observational data. JAMA. 2022 Dec 27;328(24):2446–7.
6.    Hernán MA, Dahabreh IJ, Dickerman BA, Swanson SA. The target trial framework for causal inference from observational data: why and when is it helpful? Ann Intern Med. 2025 Mar;178(3):402–7.

Written by Victoria Konzett, EMEUNET Newsletter Sub-Committee member

Statistical tests are at the heart of clinical research, yet many of us have wondered at some point: should I use a parametric or a non-parametric test? The answer is not always straightforward and depends on the nature of the data, the sample size, and the assumptions each test requires. This short overview aims to clarify the key differences and help you choose the right tool for your data

What makes a test “parametric”?

Parametric tests assume that your data follow a known distribution, most commonly the normal (Gaussian) distribution, the familiar bell-shaped curve [1]. They also assume homogeneity of variance and rely on estimating population parameters, such as the mean and standard deviation. Some, such as the classical Student’s t-test and ANOVA, also assume homogeneity of variance, though not all parametric tests share this requirement.

Common examples include the Student’s t-test, ANOVA, and the Pearson correlation coefficient. Because they use all the information in the continuous data, parametric tests tend to have greater statistical power, meaning they are more likely to detect a true effect when one exists, provided their assumptions are met.

When to go non-parametric?

Non-parametric tests, sometimes called “distribution-free” tests, do not require the data to follow a specific distribution [2]. Instead, they typically work with ranks or medians rather than raw values. This makes them particularly useful when:

• The data are not normally distributed, and the sample size is too small for the central limit theorem to apply.
• The outcome is measured on an ordinal scale (e.g., a visual analog scale or pain rated as mild, moderate, severe).
• There are significant outliers that could distort the mean.

Well-known equivalents include the Mann-Whitney U test, the Wilcoxon signed-rank test, and the Kruskal-Wallis test.

How to choose?

A common misconception is that non-parametric tests are always “safer”. While they make fewer assumptions, using a non-parametric test on truly normally distributed data sacrifices statistical power and may result in a failure to detect a real difference [3].

Recent evidence and expert consensus suggest that non-parametric tests might actually be overused in medical research [4]. Parametric tests are often remarkably robust to mild violations of normality, particularly with larger sample sizes. On the flip side, if your data show extreme skewness, non-parametric tests can actually be substantially more powerful than parametric ones.

Practical tips

Before defaulting to a test, always start by exploring your data visually (using histograms or Q-Q plots). You can complement this with formal normality tests (like Shapiro-Wilk) , but remember that these can be overly sensitive in large samples and underpowered in small ones [1,2] (Figure 1).

If your data deviate from normality, consider data transformation (e.g., log transformation) to see if you can safely use a more powerful parametric test before switching to non-parametric methods.

Figure 1. Quick decision algorithm for choosing between parametric and non-parametric tests (diagram realized with Rstudio version 4.4.2 and Microsoft Powerpoint).

Take-home message

Neither parametric nor non-parametric tests are inherently better. The right choice depends on your data. Always start by exploring your data visually, check the underlying assumptions, and choose accordingly. When in doubt, consulting a biostatistician early in your study design can save a lot of trouble later.

References and further reading

1. Vetter TR. Fundamentals of Research Data and Variables: The Devil Is in the Details. Anesth Analg. 2017 Oct;125(4):1375-1380. 
2. le Cessie S, Goeman JJ, Dekkers OM. Who is afraid of non-normal data? Choosing between parametric and non-parametric tests. Eur J Endocrinol. 2020 Feb;182(2):E1-E3.
3. Bensken WP, Ho VP, Pieracci FM. Basic Introduction to Statistics in Medicine, Part 2: Comparing Data. Surg Infect (Larchmt). 2021 Aug;22(6):597-603.
4. Bridge PD, Sawilowsky SS. Increasing physicians’ awareness of the impact of statistics on research outcomes: comparative power of the t-test and Wilcoxon rank-sum test in small samples applied research. J Clin Epidemiol. 1999;52(3):229–235.

Written by Elvis Hysa, EMEUNET Newsletter Sub-Committee member