Correlation vs. Causation and Confounding Variables Questions

Recognize that correlation (statistical relationship between variables) doesn't imply causation (direct cause-and-effect relationship). Identify confounding variables that might explain an observed correlation. For example, summer ice cream sales and crime rates both increase but neither causes the other—warm weather is the confounder. Practice identifying lurking variables in business scenarios.

EasyTechnical

89 practiced

Define lurking variable in the context of data analysis. Provide two industry examples (one e-commerce, one operations) where a lurking variable might mislead decision makers, and outline quick exploratory data analysis steps you would run to surface such lurking variables.

MediumTechnical

90 practiced

Explain the difference-in-differences (DiD) identification strategy. State the parallel trends assumption and describe at least two empirical checks or falsification tests you would run to assess whether the assumption is plausible. Also describe a policy scenario in business where DiD would be appropriate.

HardTechnical

74 practiced

Discuss how measurement error in confounders or the treatment variable biases causal effect estimates. Explain attenuation bias, classical versus systematic measurement error, and propose at least three methods to mitigate bias including validation data, instrumental variables for measurement, and regression calibration.

Sample Answer

Measurement error in confounders or the treatment distorts estimated causal effects because models use noisy proxies rather than true variables, producing biased and inconsistent estimates.

Classical vs systematic error- Classical (random) error: observed X* = X + u with u independent of X and other variables. For a continuous treatment in a linear model, classical measurement error in the treatment causes attenuation (bias toward zero) of the treatment coefficient: E[β̂] ≈ β · Var(X)/Var(X*). For confounders, classical error leads to residual confounding because the adjustment is incomplete, often biasing the treatment effect toward the crude (biased) association.- Systematic (nonclassical) error: errors correlated with X, treatment, outcome, or covariates (e.g., differential misreporting). Bias direction is unpredictable and can be severe, including sign reversal.

Attenuation bias- Intuition: measurement error inflates denominator variance (noise), shrinking the slope estimate. In multivariable causal settings, attenuation of confounders can produce under-adjustment and either exaggerate or attenuate the treatment effect depending on correlations.

Mitigation methods1) Validation data (gold standard subset): measure true variable X in a subset. Use that to estimate error model and correct estimates (e.g., two-sample correction, measurement-error-corrected likelihood). Requires representative validation and sufficient size.2) Instrumental variables for measurement: find Z correlated with true X but independent of measurement error and outcome except through X (classic IV assumptions). Use IV to recover consistent estimates even when X* is noisy. Careful: IV must be valid for the true variable, not the noisy proxy.3) Regression calibration: model E[X | X*, W] using validation or external info and replace X by its prediction in the outcome model. Works well for small-to-moderate error and linear/logistic outcomes; standard errors must account for calibration uncertainty.Additional techniques: SIMEX (simulation-extrapolation), multiple imputation for measurement error, Bayesian hierarchical measurement-error models, and sensitivity analysis to quantify robustness.

Practical notes- Always characterize error type (classical vs differential), use validation data where possible, report corrected estimates and uncertainty, and perform sensitivity analyses because nonclassical errors often cannot be fully corrected without strong assumptions.

HardTechnical

144 practiced

Implement propensity score weighting in Python to estimate ATE and ATT. Write code to fit a logistic regression propensity model, compute stabilized inverse probability weights, check covariate balance after weighting, and estimate treatment effect with robust standard errors or bootstrap. Use scikit-learn and statsmodels.

Sample Answer

Approach: Fit a logistic propensity model (scikit-learn), compute stabilized inverse probability weights (IPW) for ATE and ATT, check covariate balance via weighted standardized mean differences (SMD), estimate treatment effect with weighted OLS using statsmodels with robust SEs and optionally bootstrap CIs.

python

import numpy as np
import pandas as pd
from sklearn.linear_model import LogisticRegression
import statsmodels.api as sm
from statsmodels.stats.weightstats import DescrStatsW
from sklearn.preprocessing import StandardScaler
from sklearn.utils import resample

# Example function implementing the pipeline
def propensity_weighting(df, treatment_col, outcome_col, covariates,
                         estimand='ATE', bootstrap_iters=0, random_state=0):
    np.random.seed(random_state)
    X = df[covariates].astype(float)
    y = df[treatment_col].astype(int)
    # optional: scale covariates for better logistic behavior
    scaler = StandardScaler()
    Xs = scaler.fit_transform(X)

    # fit propensity model
    model = LogisticRegression(solver='lbfgs', max_iter=200)
    model.fit(Xs, y)
    ps = model.predict_proba(Xs)[:, 1]
    eps = 1e-6
    ps = np.clip(ps, eps, 1 - eps)

    # stabilized weights
    p_t = y.mean()
    if estimand == 'ATE':
        sw = np.where(y == 1, p_t / ps, (1 - p_t) / (1 - ps))
    elif estimand == 'ATT':
        # For ATT, treated units get weight 1, controls get ps/(1-ps) * (1-p_t)/p_t
        sw = np.where(y == 1, 1.0, (ps / (1 - ps)) * ((1 - p_t) / p_t))
    else:
        raise ValueError("estimand must be 'ATE' or 'ATT'")

    df2 = df.copy()
    df2['ps'] = ps
    df2['sw'] = sw

    # balance: weighted standardized mean differences
    def weighted_mean_std(series, w):
        d = DescrStatsW(series, weights=w, ddof=0)
        return d.mean, d.std

    balance = {}
    for col in covariates:
        m1, s1 = weighted_mean_std(df2.loc[df2[treatment_col]==1, col], df2.loc[df2[treatment_col]==1, 'sw'])
        m0, s0 = weighted_mean_std(df2.loc[df2[treatment_col]==0, col], df2.loc[df2[treatment_col]==0, 'sw'])
        pooled_sd = np.sqrt((s1**2 + s0**2) / 2)
        smd = (m1 - m0) / (pooled_sd + 1e-8)
        balance[col] = smd

    # outcome regression with weights (weighted least squares)
    X_out = sm.add_constant(df2[treatment_col])
    y_out = df2[outcome_col]
    wls_model = sm.WLS(y_out, X_out, weights=df2['sw'])
    res = wls_model.fit(cov_type='HC3')  # robust SE
    ate_est = res.params[treatment_col]
    se = res.bse[treatment_col]

    # optional bootstrap for CI
    if bootstrap_iters and bootstrap_iters > 0:
        boot_est = []
        n = len(df2)
        for i in range(bootstrap_iters):
            b = resample(df2, replace=True, n_samples=n)
            # refit propensity in bootstrap sample for proper variability
            Xb = scaler.fit_transform(b[covariates])
            mb = LogisticRegression(solver='lbfgs', max_iter=200).fit(Xb, b[treatment_col])
            psb = np.clip(mb.predict_proba(Xb)[:,1], eps, 1-eps)
            pb_t = b[treatment_col].mean()
            if estimand == 'ATE':
                swb = np.where(b[treatment_col]==1, pb_t/psb, (1-pb_t)/(1-psb))
            else:
                swb = np.where(b[treatment_col]==1, 1.0, (psb/(1-psb))*((1-pb_t)/pb_t))
            try:
                rb = sm.WLS(b[outcome_col], sm.add_constant(b[treatment_col]), weights=swb).fit()
                boot_est.append(rb.params[treatment_col])
            except Exception:
                continue
        ci_lower, ci_upper = np.percentile(boot_est, [2.5, 97.5])
    else:
        ci_lower, ci_upper = ate_est - 1.96 * se, ate_est + 1.96 * se

    return {'estimate': ate_est, 'se': se, 'ci': (ci_lower, ci_upper), 'balance_smd': balance, 'result_obj': res}

Key points:- Stabilized weights reduce variance; ATT weights set treated=1.- Check covariate balance (SMD < 0.1 desirable).- Robust SE via HC3; bootstrap refits propensity to capture uncertainty.- Edge cases: extreme PS values — clip, consider trimming or overlap weights.- Alternatives: doubly robust estimation (IPW + outcome model) or matching.

HardTechnical

84 practiced

Explain marginal structural models (MSMs) and inverse probability of treatment weighting (IPTW) for handling time-varying confounding in longitudinal data. Describe how to compute stabilized weights, diagnose weight positivity violations, and how model misspecification can affect estimates.

Sample Answer

Marginal structural models (MSMs) are causal models for longitudinal data that estimate marginal (population-level) effects of time-varying treatments while properly adjusting for time-varying confounders that are themselves affected by prior treatment. Standard regression conditioning on such intermediates induces bias; MSMs instead reweight the data so that, after weighting, treatment at each time is independent of past confounders — emulating a randomized trial.

Inverse probability of treatment weighting (IPTW) constructs these weights from models of the treatment process. For an individual i across times t = 0..T, the unstabilized weight is the product over t of 1 / P(A_t = a_t | history H_{t}), where H_t includes past treatments and confounders. Stabilized weights improve finite-sample properties by multiplying numerator probabilities that depend only on a subset (often baseline covariates): SW_i = ∏_t P(A_t=a_t | baseline) / P(A_t=a_t | H_t). Stabilized weights have mean ~1 and lower variance.

Diagnosing positivity and weight problems:- Inspect weight distribution: mean, SD, min, max, percentiles; extreme weights (e.g., >10 or heavy tails) signal violations.- Plot histograms/boxplots and cumulative density; check overlap of propensity scores by treatment group.- Covariate balance after weighting: standardized mean differences should be near zero across time.- If positivity violated, consider truncation (e.g., cap at 1st/99th percentile), coarsening treatment/covariates, targeted subsamples, or using parametric/g-computation or TMLE.

Model misspecification effects:- Misspecified treatment models yield biased weights and hence biased causal estimates; stabilized weights help but don't remove bias from wrong functional form.- Misspecified MSM (wrong functional form for marginal effect) gives biased target estimates even with correct weights.- Remedies: flexible treatment/outcome models (machine learning, Super Learner), ensemble propensity models, include interactions/time trends, use diagnostics (balance) to iterate, and compute robust (sandwich) SEs or use doubly robust estimators like IPTW+outcome regression or TMLE which remain consistent if either model is correct.

Practical checklist: estimate stabilized IPTW, inspect weights and balance, consider truncation with sensitivity analyses, prefer flexible models and doubly robust methods, and report diagnostics and sensitivity to positivity/model choices.

Unlock Full Question Bank

Get access to hundreds of Correlation vs. Causation and Confounding Variables interview questions and detailed answers.

Join thousands of developers preparing for their dream job.