Penelitian Kuasieksperimental
Published on October 9, 2026
Feature | Detail |
|---|---|
Definition | A study design that evaluates the effect of an intervention without random assignment of participants to groups |
Core Distinction from RCT | No randomization; the investigator still controls the intervention but cannot control allocation |
Core Distinction from Observational | The investigator actively assigns or manipulates an intervention (unlike cohort or case-control studies where exposure occurs naturally) |
Why It Exists | Randomization is often unethical, impractical, or impossible in real-world clinical and public health settings |
Level of Evidence | Below RCTs but above observational studies in the evidence hierarchy |
Primary Weakness | Susceptible to confounding and selection bias because groups may differ at baseline |
Primary Strength | High external validity (generalizability) because it is conducted in real-world conditions |
Keywords in Vignettes | "Before and after," "policy change," "program implementation," "historical control," "non-equivalent control group," "no randomization was performed," "natural experiment" |
01Definition and Core Principles
A quasi-experimental study occupies the middle ground between a true experiment (randomized controlled trial) and a purely observational study. The investigator introduces or controls an intervention, which is what makes it "experimental." However, participants are not randomly allocated to the intervention or comparison group, which is what makes it "quasi."
The reason this design exists is pragmatic. Consider a hospital that implements a new hand hygiene protocol across all its wards. You cannot randomly assign some wards to wash hands and others not to; the intervention applies to everyone. But you can still measure infection rates before and after the protocol was introduced, or compare your hospital to a similar one that did not adopt the protocol. That is the essence of a quasi-experiment.
The critical concept for exam purposes is this: the absence of randomization means that any observed difference between groups could be due to the intervention, or it could be due to pre-existing differences between the groups (confounding). Every question about quasi-experimental validity will circle back to this principle.
02Types of Quasi-Experimental Designs
Classification Table
One-group pretest-posttest
Structure
O₁ → X → O₂
Control Group?
No
Key Feature
Measures the same group before and after the intervention
Main Threat
Maturation, history, regression to the mean
Posttest-only with nonequivalent groups
Structure
Group A: X → O₁ / Group B: → O₁
Control Group?
Yes (nonequivalent)
Key Feature
Compares an intervention group to a non-randomized comparison group, measured only after
Main Threat
Selection bias (no baseline data to verify equivalence)
Pretest-posttest with nonequivalent control group
Structure
Group A: O₁ → X → O₂ / Group B: O₁ → → O₂
Control Group?
Yes (nonequivalent)
Key Feature
Both groups measured before and after; only one receives the intervention
Main Threat
Selection bias, partially mitigated by baseline comparison
Interrupted time series (ITS)
Structure
O₁ O₂ O₃ → X → O₄ O₅ O₆
Control Group?
No (single series) or Yes (controlled ITS)
Key Feature
Multiple measurements before and after the intervention to detect a change in trend
Main Threat
History (concurrent external event)
Controlled interrupted time series
Structure
Same as ITS, but a parallel group without intervention is tracked simultaneously
Control Group?
Yes (nonequivalent)
Key Feature
Adding a control series to ITS strengthens causal inference
Main Threat
Reduced history threat; selection remains
Regression discontinuity
Structure
Assignment based on a cutoff score on a continuous variable
Control Group?
Yes (above/below cutoff)
Key Feature
Participants just above and just below the cutoff are compared
Main Threat
Requires a sharp, known cutoff; misspecification of the functional form
Natural experiment
Structure
An external event or policy creates de facto "treatment" and "control" groups
Control Group?
Yes (naturally occurring)
Key Feature
The investigator exploits an event outside their control (e.g., a new law in one province but not another)
Main Threat
Selection and unmeasured confounders; depends on the assumption that the event was "as if random"
Design Type | Structure | Control Group? | Key Feature | Main Threat |
|---|---|---|---|---|
One-group pretest-posttest | O₁ → X → O₂ | No | Measures the same group before and after the intervention | Maturation, history, regression to the mean |
Posttest-only with nonequivalent groups | Group A: X → O₁ / Group B: → O₁ | Yes (nonequivalent) | Compares an intervention group to a non-randomized comparison group, measured only after | Selection bias (no baseline data to verify equivalence) |
Pretest-posttest with nonequivalent control group | Group A: O₁ → X → O₂ / Group B: O₁ → → O₂ | Yes (nonequivalent) | Both groups measured before and after; only one receives the intervention | Selection bias, partially mitigated by baseline comparison |
Interrupted time series (ITS) | O₁ O₂ O₃ → X → O₄ O₅ O₆ | No (single series) or Yes (controlled ITS) | Multiple measurements before and after the intervention to detect a change in trend | History (concurrent external event) |
Controlled interrupted time series | Same as ITS, but a parallel group without intervention is tracked simultaneously | Yes (nonequivalent) | Adding a control series to ITS strengthens causal inference | Reduced history threat; selection remains |
Regression discontinuity | Assignment based on a cutoff score on a continuous variable | Yes (above/below cutoff) | Participants just above and just below the cutoff are compared | Requires a sharp, known cutoff; misspecification of the functional form |
Natural experiment | An external event or policy creates de facto "treatment" and "control" groups | Yes (naturally occurring) | The investigator exploits an event outside their control (e.g., a new law in one province but not another) | Selection and unmeasured confounders; depends on the assumption that the event was "as if random" |
Notation Key
O = Observation (measurement)
X = Intervention
Subscript numbers indicate the sequence of measurements
One-Group Pretest-Posttest
This is the simplest and weakest quasi-experimental design. A single group is measured, receives an intervention, and is measured again. If the outcome improves, the investigator attributes it to the intervention. The problem is that many other things could explain the change: natural disease progression (maturation), an unrelated event occurring at the same time (history), or the tendency for extreme initial values to move toward the average on repeat measurement (regression to the mean). Exam questions will describe this design and ask you to identify why the conclusion is unreliable.
Pretest-Posttest with Nonequivalent Control Group
This is the most commonly tested quasi-experimental design. Two groups are compared, both measured before and after, but only one receives the intervention. The word "nonequivalent" is the exam signal: the groups were not formed by randomization, so they may differ in ways that affect the outcome. For example, a teaching hospital implements a new antibiotic stewardship program while a community hospital serves as the comparison. The teaching hospital may already have different prescribing patterns, patient acuity, and staffing levels. Even if both groups have baseline data, selection bias remains the dominant threat because the groups may differ on unmeasured variables.
Interrupted Time Series (ITS)
This design collects multiple data points over time, both before and after an intervention is introduced at a known point. It is considered one of the strongest quasi-experimental designs because the repeated measurements allow you to establish a pre-intervention trend and then detect whether the intervention changed the level (an immediate jump or drop) or the slope (a change in the rate of increase or decrease) of the outcome. A single pretest-posttest design cannot distinguish a true intervention effect from an ongoing trend; ITS can. The main vulnerability is history: if another event coincides with the intervention, you cannot separate their effects. Adding a control series (controlled ITS) addresses this.
Regression Discontinuity
This design is less commonly tested but increasingly recognized. Assignment to intervention is based on whether a participant's score on a continuous variable falls above or below a predetermined cutoff. For example, patients with a blood pressure above 140 mmHg receive a new drug, while those below do not. The analysis compares outcomes for patients just above and just below the cutoff, under the assumption that these individuals are essentially similar. When the cutoff rule is strictly applied and the analysis is correctly specified, this design can provide causal estimates comparable to an RCT for the population near the cutoff.
Natural Experiment
The investigator does not assign the intervention at all. Instead, an external event (a policy change, a natural disaster, a legislative difference between jurisdictions) creates groups that differ in their exposure. The investigator then analyzes the resulting data as if it were an experiment. The validity of a natural experiment depends entirely on whether the "assignment" by the external event was unrelated to the outcome of interest. Exam vignettes will describe a policy that took effect in one region and ask what study design is being used.
03Threats to Internal Validity
Understanding threats to validity is non-negotiable for this topic. Exam questions frequently present a quasi-experimental scenario and ask you to identify the most likely source of bias.
Selection bias
Definition
Groups differ at baseline due to non-random assignment
Which Designs Are Most Vulnerable?
All quasi-experimental designs, especially posttest-only with nonequivalent groups
Maturation
Definition
Participants change naturally over time (aging, healing, learning)
Which Designs Are Most Vulnerable?
One-group pretest-posttest
History
Definition
An external event co-occurs with the intervention
Which Designs Are Most Vulnerable?
One-group pretest-posttest, ITS without control
Regression to the mean
Definition
Extreme baseline values tend to normalize on remeasurement
Which Designs Are Most Vulnerable?
One-group pretest-posttest, any design that selects participants based on extreme values
Testing effect
Definition
Taking a pretest influences performance on the posttest
Which Designs Are Most Vulnerable?
Designs with repeated testing
Instrumentation
Definition
The measurement tool or method changes between pretest and posttest
Which Designs Are Most Vulnerable?
Any design with before-after measurement
Attrition (differential dropout)
Definition
Participants leave the study at different rates between groups
Which Designs Are Most Vulnerable?
Pretest-posttest with nonequivalent control group
Diffusion of treatment
Definition
The control group gains access to the intervention
Which Designs Are Most Vulnerable?
Nonequivalent control group designs where groups are geographically close
Threat | Definition | Which Designs Are Most Vulnerable? |
|---|---|---|
Selection bias | Groups differ at baseline due to non-random assignment | All quasi-experimental designs, especially posttest-only with nonequivalent groups |
Maturation | Participants change naturally over time (aging, healing, learning) | One-group pretest-posttest |
History | An external event co-occurs with the intervention | One-group pretest-posttest, ITS without control |
Regression to the mean | Extreme baseline values tend to normalize on remeasurement | One-group pretest-posttest, any design that selects participants based on extreme values |
Testing effect | Taking a pretest influences performance on the posttest | Designs with repeated testing |
Instrumentation | The measurement tool or method changes between pretest and posttest | Any design with before-after measurement |
Attrition (differential dropout) | Participants leave the study at different rates between groups | Pretest-posttest with nonequivalent control group |
Diffusion of treatment | The control group gains access to the intervention | Nonequivalent control group designs where groups are geographically close |
Selection bias is the single most important threat to internalize. Every quasi-experimental question on the exam can be connected to this concept. When randomization is absent, you cannot guarantee that the groups are comparable. Any difference in outcome may reflect pre-existing differences rather than the intervention itself.
04Methods to Strengthen Quasi-Experimental Designs
Because randomization is absent, investigators use statistical and design strategies to approximate its effect. Exam questions may ask which technique reduces confounding in a quasi-experimental study.
Propensity score matching
How It Works
Estimates the probability of receiving the intervention based on observed covariates, then matches participants with similar scores across groups
Limitation
Only adjusts for measured confounders
Difference-in-differences (DiD)
How It Works
Compares the change over time in the intervention group to the change over time in the control group, removing time-invariant confounders
Limitation
Assumes parallel trends (both groups would have followed the same trajectory without the intervention)
Instrumental variable analysis
How It Works
Uses a variable that affects exposure but not the outcome (except through the exposure) to estimate a causal effect
Limitation
Valid instruments are rare and difficult to verify
Multivariable regression adjustment
How It Works
Statistically controls for known confounders
Limitation
Cannot adjust for unmeasured confounders
Adding a control group
How It Works
Upgrades a single-group design to a nonequivalent control group design
Limitation
Control group may still differ systematically
Increasing data points (time series)
How It Works
Converts a simple before-after into an ITS design
Limitation
Requires routine, consistent data collection over a long period
Strategy | How It Works | Limitation |
|---|---|---|
Propensity score matching | Estimates the probability of receiving the intervention based on observed covariates, then matches participants with similar scores across groups | Only adjusts for measured confounders |
Difference-in-differences (DiD) | Compares the change over time in the intervention group to the change over time in the control group, removing time-invariant confounders | Assumes parallel trends (both groups would have followed the same trajectory without the intervention) |
Instrumental variable analysis | Uses a variable that affects exposure but not the outcome (except through the exposure) to estimate a causal effect | Valid instruments are rare and difficult to verify |
Multivariable regression adjustment | Statistically controls for known confounders | Cannot adjust for unmeasured confounders |
Adding a control group | Upgrades a single-group design to a nonequivalent control group design | Control group may still differ systematically |
Increasing data points (time series) | Converts a simple before-after into an ITS design | Requires routine, consistent data collection over a long period |
The principle tested most often: no statistical method can fully substitute for randomization, because all adjustment techniques only address confounders that the investigator has identified and measured. This is the fundamental limitation that separates quasi-experiments from RCTs.
05Comparison with Other Study Designs
This table is designed to help you answer questions that present a study scenario and ask you to classify the design.
Investigator assigns the intervention?
True Experiment (RCT)
Yes
Quasi-Experiment
Yes
Observational (Cohort)
No
Observational (Case-Control)
No
Random allocation?
True Experiment (RCT)
Yes
Quasi-Experiment
No
Observational (Cohort)
No
Observational (Case-Control)
No
Control group?
True Experiment (RCT)
Yes (randomized)
Quasi-Experiment
Sometimes (nonequivalent)
Observational (Cohort)
Yes (unexposed)
Observational (Case-Control)
Yes (controls)
Direction of inquiry
True Experiment (RCT)
Prospective
Quasi-Experiment
Prospective or retrospective
Observational (Cohort)
Prospective (or retrospective)
Observational (Case-Control)
Retrospective
Primary measure of association
True Experiment (RCT)
Risk ratio, NNT
Quasi-Experiment
Varies
Observational (Cohort)
Risk ratio, hazard ratio
Observational (Case-Control)
Odds ratio
Strength of causal inference
True Experiment (RCT)
Highest
Quasi-Experiment
Moderate
Observational (Cohort)
Lower
Observational (Case-Control)
Lower
Selection bias risk
True Experiment (RCT)
Low (randomization)
Quasi-Experiment
High
Observational (Cohort)
Moderate
Observational (Case-Control)
High
When to use
True Experiment (RCT)
Intervention can be randomized
Quasi-Experiment
Randomization unethical or impractical
Observational (Cohort)
Exposure occurs naturally
Observational (Case-Control)
Rare diseases, case identification first
Feature | True Experiment (RCT) | Quasi-Experiment | Observational (Cohort) | Observational (Case-Control) |
|---|---|---|---|---|
Investigator assigns the intervention? | Yes | Yes | No | No |
Random allocation? | Yes | No | No | No |
Control group? | Yes (randomized) | Sometimes (nonequivalent) | Yes (unexposed) | Yes (controls) |
Direction of inquiry | Prospective | Prospective or retrospective | Prospective (or retrospective) | Retrospective |
Primary measure of association | Risk ratio, NNT | Varies | Risk ratio, hazard ratio | Odds ratio |
Strength of causal inference | Highest | Moderate | Lower | Lower |
Selection bias risk | Low (randomization) | High | Moderate | High |
When to use | Intervention can be randomized | Randomization unethical or impractical | Exposure occurs naturally | Rare diseases, case identification first |
The discriminating question on the exam is always: Did the investigator assign the intervention? If yes, it is experimental or quasi-experimental. Then ask: Was there randomization? If yes, it is a true experiment. If no, it is quasi-experimental. If the investigator did not assign the intervention and simply observed what happened, it is observational.
06Differential Diagnosis of Study Designs (Distractors)
Prospective cohort study
why it's confusing
Both follow groups over time and compare outcomes
Key Discriminator
In a cohort study, the investigator does not assign the exposure. The groups are defined by naturally occurring exposures. In a quasi-experiment, the investigator introduces or controls the intervention.
Randomized controlled trial
why it's confusing
Both involve an investigator-assigned intervention and a comparison group
Key Discriminator
The RCT uses random allocation. If the vignette states that participants were randomly assigned, it is an RCT, not a quasi-experiment.
Before-and-after study (uncontrolled)
why it's confusing
This is actually a type of quasi-experimental design, but students may classify it as observational
Key Discriminator
The presence of an investigator-introduced intervention makes it quasi-experimental, even without a control group.
Cross-sectional study
why it's confusing
A quasi-experiment measured at a single point in time (posttest only) may look cross-sectional
Key Discriminator
Cross-sectional studies measure exposure and outcome simultaneously without an intervention. A posttest-only quasi-experiment still has an intervention that was deliberately applied.
Ecological study
why it's confusing
Both may use population-level data (e.g., comparing regions)
Key Discriminator
Ecological studies analyze aggregate data without individual-level exposure-outcome linkage. A quasi-experiment at the population level (e.g., ITS of hospital-wide infection rates) still has a defined intervention point and measures the effect on the same population.
Natural experiment vs. observational study
why it's confusing
A natural experiment has no investigator-assigned intervention, which sounds observational
Key Discriminator
The distinction is in the analytical framework: a natural experiment analyzes the external event as if it were an assigned intervention, using quasi-experimental methods. A cohort study of the same event would simply compare exposed and unexposed groups without the quasi-experimental framing.
Distractor Design | why it's confusing | Key Discriminator |
|---|---|---|
Prospective cohort study | Both follow groups over time and compare outcomes | In a cohort study, the investigator does not assign the exposure. The groups are defined by naturally occurring exposures. In a quasi-experiment, the investigator introduces or controls the intervention. |
Randomized controlled trial | Both involve an investigator-assigned intervention and a comparison group | The RCT uses random allocation. If the vignette states that participants were randomly assigned, it is an RCT, not a quasi-experiment. |
Before-and-after study (uncontrolled) | This is actually a type of quasi-experimental design, but students may classify it as observational | The presence of an investigator-introduced intervention makes it quasi-experimental, even without a control group. |
Cross-sectional study | A quasi-experiment measured at a single point in time (posttest only) may look cross-sectional | Cross-sectional studies measure exposure and outcome simultaneously without an intervention. A posttest-only quasi-experiment still has an intervention that was deliberately applied. |
Ecological study | Both may use population-level data (e.g., comparing regions) | Ecological studies analyze aggregate data without individual-level exposure-outcome linkage. A quasi-experiment at the population level (e.g., ITS of hospital-wide infection rates) still has a defined intervention point and measures the effect on the same population. |
Natural experiment vs. observational study | A natural experiment has no investigator-assigned intervention, which sounds observational | The distinction is in the analytical framework: a natural experiment analyzes the external event as if it were an assigned intervention, using quasi-experimental methods. A cohort study of the same event would simply compare exposed and unexposed groups without the quasi-experimental framing. |
07Traps and High-Yield Pearls
The most common way students lose points on quasi-experimental design questions is by failing to identify the absence of randomization as the central weakness. A vignette will describe a study where a hospital implemented a new protocol and infection rates dropped. The student, seeing a clear before-and-after effect, selects the answer choice that affirms a causal relationship. The correct answer almost always states that the conclusion is limited by confounding due to lack of randomization or by a named threat to internal validity such as history or maturation.
A second common trap involves confusing quasi-experimental with observational designs. The key is always whether the investigator controlled the intervention. If a vignette says "the hospital administration decided to implement a new protocol," that is an investigator-assigned intervention and the study is quasi-experimental, even if no one in the vignette is wearing a lab coat and calling it a study.
A third testing pattern involves interrupted time series. The exam may present a graph showing data points over several years with a vertical line marking a policy change. Students who recognize this as ITS will correctly identify that the design controls for pre-existing trends, making it stronger than a simple before-after comparison.
Finally, remember the hierarchy: when a vignette asks "what is the best study design to evaluate this intervention?" and randomization is described as unethical or infeasible, the answer is a quasi-experimental design, not an observational one. The quasi-experiment is the strongest design available when an RCT cannot be performed. The exam tests whether you understand that feasibility and ethics constrain design choice, and that the quasi-experiment is the appropriate compromise, not a flawed shortcut.