IPTW: Making Treatment Groups Comparable and Checking That It Worked

อ่านฉบับภาษาไทย (Thai version)

Abstract

Inverse probability of treatment weighting (IPTW) improves comparability between treatment groups. It changes how much each patient contributes to the analysis, without creating new patients. Each patient is weighted by one over the probability of the treatment actually received, given measured baseline characteristics; that probability is the propensity score. A treated patient with a score of 0.80 gets a weight of 1.25, while one with a score of 0.20 gets a weight of 5, because rare treatment assignment earns a larger weight. The result is a weighted pseudo-population in which treated and control patients look alike on what was measured. Before any outcome is analysed, three diagnostics show whether weighting worked: covariate balance, propensity-score overlap and the weight distribution. When balance fails, the treatment model is revised and the diagnostics are run again. This article works through each step with simple invented numbers and shows why balance and overlap answer different questions.

---

🔍 Click to enlarge Click anywhere to close

Two treatment groups that started out different

A team reviewing a hospital database compares patients who received Treatment A with patients who received Treatment B. The outcomes differ, but so do the patients. Those on Treatment A were older, more often had diabetes and had higher severity scores. Before anyone reads the outcome difference as a treatment effect, the team needs groups that start out alike.

This article follows one line of reasoning about inverse probability of treatment weighting (IPTW). It asks why weighting is needed, what a propensity score is and why a rare treatment earns a large weight. It then shows what IPTW can achieve, how to check that it worked and what to do when it did not.

Why do we need weighting?

In observational research, patients who receive different treatments are often systematically different at baseline. Patients receiving Treatment A may be older, have more severe disease or have more comorbidities than patients receiving Treatment B. A crude outcome comparison may therefore reflect two things at once:

$$\text{Observed difference} = \text{Treatment effect} + \text{Baseline differences}$$

Propensity-score methods try to improve comparability between groups. They use measured baseline characteristics to construct a more comparable weighted population [1].

What is a propensity score?

For treatment assignment, the propensity score is the probability of receiving the treatment given the measured baseline covariates [2]:

$$e(X) = P(A=1 \mid X)$$

Here $A$ is the treatment received, 1 for treated and 0 for control, and $X$ is the set of measured baseline covariates. In practice the score is often estimated with a logistic regression of $A$ on $X$. It answers one question about each patient:

Given this patient's baseline characteristics, how likely was this patient to receive the treatment?

Different patients normally have different propensity scores. The goal is not to make everybody's propensity score equal. Instead, the score is used to construct weights that make the treatment groups more comparable on measured baseline covariates [1].

Other uses of the score, such as matching, are covered in the propensity score guide.

IPTW: inverse probability of treatment weighting

The most familiar application of the propensity score is IPTW. IPTW asks one question about each patient:

How probable was the treatment that this patient actually received?

For an ATE-type analysis, a simplified form gives each patient one over that probability [3]. The ATE, or average treatment effect, compares everyone treated with no one treated in the whole population studied.

For treated patients:

$$w_i = \frac{1}{e(X_i)}$$

For control patients:

$$w_i = \frac{1}{1-e(X_i)}$$

For a control patient, $1-e(X_i)$ is the probability of the treatment actually received, which is no treatment. The same inverse-probability idea, applied to who enters a study rather than who receives a treatment, is taken up in part 4.

Why does IPTW work? Two treated patients

Consider two treated patients, with invented numbers. Both received the treatment, but their propensity scores differ.

1. Patient A: the score

   \[ PS = 0.80 \]

   Treatment was quite likely for someone with Patient A's characteristics.

2. Patient A: the weight

   \[ w = \frac{1}{0.80} = 1.25 \]

   Receiving the treatment was the expected case, so the weight is modest.

3. Patient B: the score

   \[ PS = 0.20 \]

   Treatment was relatively unlikely, yet the patient actually received treatment.

4. Patient B: the weight

   \[ w = \frac{1}{0.20} = 5 \]

   Receiving the treatment was the unusual case, so the weight is large.

Result: Patient B is relatively unusual among treated patients and therefore receives more statistical weight: 5, against 1.25 for Patient A.

A weighted pseudo-population

This process constructs a weighted pseudo-population. IPTW using the propensity score is specifically intended to create a weighted sample in which measured baseline covariates are balanced between treated and control groups [4].

What can IPTW achieve?

The major ability of IPTW is easy to state and easy to see in a baseline table.

Before weighting

After weighting

What the two tables show

Weighting has made the measured baseline distributions much more similar. Mean age moves from 72 against 60 to 65.8 against 65.5. Diabetes moves from 60% against 35% to 47% against 46%, and the severity score from 8 against 4 to 6.1 against 6.0.

This improved comparability is the central causal purpose of propensity-score adjustment [1]. It holds for the characteristics that were measured and used to build the weights.

How do we know IPTW worked?

After creating the weights, we perform weighting diagnostics [4]. They sit between the weights and the outcome, so the outcome is analysed only after the diagnostics have been read.

A. Covariate balance

We examine standardised mean differences (SMDs). An SMD is the difference in a covariate's mean between the groups, divided by a pooled standard deviation, so covariates on different scales can be compared [5]. Small absolute SMDs suggest that measured baseline covariates are similar between the weighted groups.

A commonly used practical benchmark is approximately:

$$|SMD| < 0.10$$

The purpose, however, is to assess the overall pattern of balance rather than mechanically pass or fail a single number [4, 5]. Part 2 looks at balance checks in more detail.

B. Propensity-score overlap

We examine whether treated and control patients occupy reasonably overlapping regions of the propensity-score distribution. Conceptually, good overlap looks like this:

Good overlap

Control:      ███████████████
Treated:        ███████████████

Good overlap means there are patients with similar baseline treatment probabilities who received different treatments.

C. Weight distribution

Extremely large weights can indicate propensity scores close to 0 or 1. For example, a control patient with a propensity score of 0.99 receives:

$$w = \frac{1}{1-0.99} = 100$$

Such an observation can strongly influence the weighted analysis [4]. Part 3 looks at what can be done about extreme weights.

What if IPTW does not achieve balance?

Suppose that after weighting the SMD for severity is still large:

$$SMD_{\text{severity}} = 0.32$$

This tells us that the original weighting procedure has not adequately balanced severity. The correct response is not to interpret the outcome immediately.

Revising the treatment model

Suppose the original treatment model was a logistic regression on age and severity only.

1. The original model

   \[ \text{logit}[P(A=1)] = \beta_0 + \beta_1\,\text{Age} + \beta_2\,\text{Severity} \]

   This model lets the log odds of treatment change in a straight line with severity.

2. Suspect a nonlinear association

   The association between severity and treatment may be nonlinear, which one straight-line severity term cannot capture.

3. Revise the specification

   \[ \text{Severity}^2 \quad \text{or a spline, or} \quad \text{Age} \times \text{Severity} \]

   Add a squared severity term, a spline (a smooth curve built from joined pieces), or a clinically justified interaction such as age by severity.

4. Refit and check again

   \[ \text{New PS} \rightarrow \text{New weights} \rightarrow \text{Check balance again} \]

   The revised model gives new propensity scores, the new scores give new weights, and balance is checked again before any outcome is analysed.

Result: Flexible PS specifications can improve covariate balance when a simpler treatment model was inadequate [4].

Balance and overlap are different concepts

Balance and overlap should not be treated as synonyms. They ask different questions and are examined with different tools.

What each concept asks

Common misreadings and their fixes

• "Weighting adds patients to the groups."

  Weighting changes how much each individual contributes to the analysis. It does not physically create new patients.

  Fix: Read a weight as how much a patient counts in the pseudo-population, never as extra patients.

• "The propensity score should end up the same for everybody."

  Different patients normally have different propensity scores, and that is expected.

  Fix: Use the score to build weights, then judge the weighted groups on their measured baseline covariates.

• "An SMD below 0.10 is a pass and anything above it is a fail."

  The 0.10 benchmark is approximate and is used as a practical guide.

  Fix: Assess the overall pattern of balance across the covariates rather than mechanically passing or failing a single number.

• "Balance is poor, but the outcome analysis can go ahead anyway."

  With an SMD for severity of 0.32 after weighting, severity has not been adequately balanced, so the outcome comparison would still carry baseline differences.

  Fix: Revise the model or weighting strategy, for example with a squared term, a spline or an interaction, and recheck the diagnostics before interpreting the outcome.

• "Good balance means good overlap."

  Balance asks whether covariates are similarly distributed after weighting. Overlap asks whether comparable treated and control patients exist at all.

  Fix: Check both: SMDs and balance plots for balance, propensity-score distributions for overlap.

Glossary

baseline covariates (ตัวแปรร่วม ณ จุดเริ่มต้น)

Patient characteristics measured before treatment that may make the groups differ.

propensity score (คะแนนแนวโน้มการได้รับการรักษา)

The probability of receiving the treatment given the measured baseline covariates, e(X) = P(A=1 given X).

IPTW (การถ่วงน้ำหนักด้วยส่วนกลับของความน่าจะเป็นในการได้รับการรักษา)

Inverse probability of treatment weighting: each patient is weighted by one over the probability of the treatment actually received.

ATE (ผลเฉลี่ยของการรักษาในประชากรทั้งหมด)

Average treatment effect: everyone treated versus no one treated, in the whole population studied.

pseudo-population (ประชากรเสมือน)

The weighted population in which measured baseline covariates are balanced between treated and control groups.

standardised mean difference (ผลต่างค่าเฉลี่ยมาตรฐาน)

The difference in a covariate's mean between groups divided by a pooled standard deviation, used to judge balance.

covariate balance (ความสมดุลของตัวแปรร่วม)

Whether baseline covariates are similarly distributed between the weighted treatment groups.

propensity-score overlap (การทับซ้อนของคะแนนแนวโน้ม)

Whether treated and control patients with comparable propensity scores actually exist.

spline (สไปลน์)

A smooth curve built from joined pieces, used to let a covariate enter a model in a nonlinear way.

References

1. Desai RJ, Franklin JM. Alternative approaches for confounding adjustment in observational studies using weighting based on the propensity score: a primer for practitioners. BMJ. 2019;367:l5657. doi:10.1136/bmj.l5657 https://doi.org/10.1136/bmj.l5657
2. Rosenbaum PR, Rubin DB. The central role of the propensity score in observational studies for causal effects. Biometrika. 1983;70(1):41-55. doi:10.1093/biomet/70.1.41 https://doi.org/10.1093/biomet/70.1.41
3. Austin PC. An introduction to propensity score methods for reducing the effects of confounding in observational studies. Multivariate Behav Res. 2011;46(3):399-424. doi:10.1080/00273171.2011.568786 https://doi.org/10.1080/00273171.2011.568786
4. Austin PC, Stuart EA. Moving towards best practice when using inverse probability of treatment weighting (IPTW) using the propensity score to estimate causal treatment effects in observational studies. Stat Med. 2015;34(28):3661-3679. doi:10.1002/sim.6607 https://doi.org/10.1002/sim.6607
5. Austin PC. Balance diagnostics for comparing the distribution of baseline covariates between treatment groups in propensity-score matched samples. Stat Med. 2009;28(25):3083-3107. doi:10.1002/sim.3697 https://doi.org/10.1002/sim.3697

Key takeaways

• Weighting changes how much each patient contributes to the analysis; it does not create new patients.
• The propensity score is used to build weights that make the groups comparable, not to make everybody's score equal.
• IPTW weights each patient by one over the probability of the treatment received, so rare treatment assignment earns a larger weight: 1.25 for a score of 0.80, 5 for a score of 0.20.
• IPTW's major ability is to improve comparability between treatment groups, checked by covariate balance, propensity-score overlap and the weight distribution before the outcome is analysed.
• If balance fails, revise the model or weighting strategy and recheck; balance and overlap are different concepts, and both are needed.

Related in the wiki: [[propensity-score-guide]]