Motivation

Methods

Methods

Why AML?

• Answer: Machine learning methods can test and compare large number of algorithms taking into account regularized models with highly correlated features.  This approach focuses on predicting outcome instead of individual hypothesis-testing process.  It sheds light on features and model selection, which provides fast and important information on more effective modeling.

Why isotonic regression?

• Answer: our goal is not to detect effects of individual covariate but instead what groups in the distribution of such variable are most different.  The groups identified can help developing more surgical group-specific solutions.  For instance, decision of discharge based on stages of prognosis (Satyadev et al. 2022), targeted treatment for certain group of patients, etc.

Methods

We collect global COVID data from the Our World in Data project, which makes daily pandemic data from all countries available in real time. 

• The data set comes with a series of daily variables including total (cumulative) and new cases (per million), deaths (per thousand), vaccination and other longer term variables such as GDP per capita, proportion of aged 65 or above, human development index (HDI), etc.
• There are a total of 225 units, with 228,749 cases and 67 variables.
• Data period starts from 1/1/2020 to 10/20/2022
• We use the peak day data, which is on 1/20/2022.

Data

Data: Total cases per million

Data: Daily COVID deaths

Data: Death data (Europe)

Data: Death data (Asia)

Data: COVID cases~predictors

Data: COVID cases~predictors

Automated Machine Learning 

• Automated machine learning applies multiple algorithms and modeling methods including:
  • validation
  • training and testing data partitioning
  • pruning for tree models
  • constraints application

• The process compares model efficacy using a series of metrics to determine stopping and selection recommendations including MSE, AUC and RMSE.
• We use \(H_2O.ai\) to model the COVID cases with 13 predictors. 

Automated Machine Learning 

Isotonic Regression

Data: Taiwan and the world

Findings: How about vaccinations?

Barlow, R. E., and H. D. Brunk. 1972. “The Isotonic Regression Problem and Its Dual.” Journal of the American Statistical Association 67(337): 140–47.

Dougherty, Michael R. et al. 2015. “An Introduction to the General Monotone Model with Application to Two Problematic Data Sets.” Sociological Methodology 45(1): 223–71.

Henzi, Alexander, Johanna F. Ziegel, and Tilmann Gneiting. 2021. “Isotonic Distributional Regression.” Journal of the Royal Statistical Society: Series B (Statistical Methodology) 83(5): 963–93.

Khazaei, Zaher et al. 2020. “COVID-19 Pandemic in the World and Its Relation to Human Development Index: A Global Study.” Archives of Clinical Infectious Diseases 15(5). https://brief.land/archcid/articles/103093.html (May 25, 2022).

Leeuw, Jan de, Kurt Hornik, and Patrick Mair. 2009. “Isotone Optimization in R : Pool-Adjacent-Violators Algorithm (PAVA) and Active Set Methods.” Journal of Statistical Software 32(5). http://www.jstatsoft.org/v32/i05/ (April 10, 2022).

Satyadev, Nihal et al. 2022. “Machine Learning for Predicting Discharge Disposition After Traumatic Brain Injury.” Neurosurgery 90(6): 768–74.

Tibshirani, Robert. 1996. “Regression Shrinkage and Selection Via the Lasso.” Journal of the Royal Statistical Society: Series B (Methodological) 58(1): 267–88.

Tibshirani, Ryan J., Holger Hoefling, and Robert Tibshirani. 2011. “Nearly-Isotonic Regression.” Technometrics 53(1): 54–61.

References

Barlow, R. E., and H. D. Brunk. 1972. “The Isotonic Regression Problem and Its Dual.” Journal of the American Statistical Association 67(337): 140–47.

Zelazo, P. R., Zelazo, N. A., & Kolb, S. (1972). " Walking" in the Newborn. Science, 176(4032), 314-315.

Silvapulle, M. J., & Sen, P. K. 2011. Constrained statistical inference: Order, inequality, and shape constraints (Vol. 912). John Wiley & Sons.
Robertson, T., & Wright, F. T., Dykstra, R. L 1988. Order restricted statistical inference. Wiley Series in Probability and Mathematical Statistics: Probability and Mathematical Statistics. John Wiley & Sons, Ltd., Chichester.
 

References

Future research

Isotonic regression: Motivation

Evaluate the effect of exercise on the age at which a child starts to walk.

1. The first treatment group received a special walking exercise for 12 minutes per day
2. The second group received daily exercises but not special exercises.
3. Control group. They did not receive any exercises or other treatments.

Q: \(\mu_1=\mu_2=\mu_3\)

holds assuming \(\mu_1\leq \mu_2 \leq \mu_3\)?

Isotonic regression: MLE (WLS)

The traditional ANOVA does not work because the maximum likelihood estimators (sample means) say \(\hat{\mu}_1 = \overline{Y}_1, \hat{\mu}_2=\overline{Y}_2,\) and \(\hat{\mu}_3 = \overline{Y}_3\), do not have to satisfy the trend \(\mu_1\leq\mu_2\leq\mu_3\). One should take the natural trend into account in the MLE (least-squares equivalently), that is, 

\[(\tilde{\mu}_1, \tilde{\mu}_2, \tilde{\mu}_3 ) = \arg\min_{\mu_1\leq \mu_2\leq\mu_3}\sum_{i=1}^3\sum_{j=1}^{n_i} (Y_{ij} -\mu_i)^2 \]

where \(Y_{ij}\stackrel{iid}{\sim} N(\mu_i,\sigma^2)\) for \(j=1,\ldots,n_i\). Then \(\tilde{\mu}_1\leq \tilde{\mu}_2 \leq \tilde{\mu}_3\) (hurestically, equal means will be rejected when  the overall mean \(\overline{Y}_\cdot\) under \(\mu_1=\mu_2=\mu_3\) are too different from \(\tilde{\mu}_1, \tilde{\mu}_2, \tilde{\mu}_3\).)

If \(\sigma^2\) varies, consider the weighted least squares (WLS):

\[(\tilde{\mu}_1, \tilde{\mu}_2, \tilde{\mu}_3 ) = \arg\min_{\mu_1\leq \mu_2\leq\mu_3}\sum_{i=1}^3\sum_{j=1}^{n_i} (Y_{ij} -\mu_i)^2 w_i \]

where \(Y_{ij}\stackrel{iid}{\sim} N(\mu_i,\sigma_i^2)\), \(w_i = n_i/\sigma_i^2\), and \(\tilde{\mu}_1\leq \tilde{\mu}_2 \leq \tilde{\mu}_3\), too. 

Isotonic regression: Computation

In fact, in the weighted least squares, it suffices to optimize

\[(\tilde{\mu}_1, \tilde{\mu}_2, \tilde{\mu}_3 ) = \arg\min_{\mu_1\leq \mu_2\leq\mu_3}\sum_{i=1}^3(\overline{Y}_i -\mu_i)^2 w_i\]

where \(\overline{Y}_i\stackrel{iid}{\sim} N(\mu_i,\sigma_i^2/n_i)\) and \(w_i = n_i/\sigma_i^2\), i=1,2,3. 

1. If \( \overline{Y}_1\leq\overline{Y}_2\leq\overline{Y}_3 \),  satisfying the increasing constrain, then \( (\tilde{\mu}_1, \tilde{\mu}_2, \tilde{\mu}_3) \) =  \( (\overline{Y}_1, \overline{Y}_2, \overline{Y}_3)\). 
2. If \(\overline{Y}_2\leq \overline{Y}_1 \leq \overline{Y}_3\), where \(\overline{Y}_1\) and \(\overline{Y}_2\) violate the constrain, then \[\tilde{\mu}_1 = \tilde{\mu}_2 = \frac{\overline{Y}_1w_1 + \overline{Y}_2w_2}{w_1 + w_2}, \tilde{\mu}_3 = \overline{Y}_3,\] such that \(\tilde{\mu}_1\leq \tilde{\mu}_2\leq \tilde{\mu}_3\). 
3. For other cases, repeatedly averaging means with weights until satisfying the increasing constrain. 

Above algorithm is the Pool Adjacent Violators Algorithm (PAVA)

Isotonic regression with covariates

Consider a conditional mean function \(\mu\) with \(\mu_i = \mu(x_i)\), and \(\mu(x)\) is increasing in \(x\), we can include the covariate \(x_i\) in isotonic regression model. For example, assume that \(Y_{ij}\stackrel{iid}{\sim} N(\mu(x_i),\sigma_i^2)\), where \(x_1\leq\ldots\leq x_n\),  the MLE of \(\mu_i = \mu(x_i)\) is given by 

\[(\tilde{\mu}_1, \ldots, \tilde{\mu}_n ) = \arg\min_{\mu_1\leq\ldots\leq\mu_n}\sum_{i=1}^n(\overline{Y}_i -\mu_i)^2 w_i\]

where \(w_i = n/\sigma_i^2\). In this case, 

1. the estimator function \(\tilde{\mu}\) satisfying that \(\tilde{\mu}(x_i) = \tilde{\mu}_i\);
2. the function \(\tilde{\mu}\) does not have to be linear;
3. assume that \(\tilde{\mu}\) is a step function to fill up the gaps between \(x_i\);
4. \(\tilde{\mu}\) and only jumping at some \(x_i\); 
5. some \(x_i\) and corresponding intervals share the same mean \(\tilde{\mu}(x_i)\). 

Isotonic regression: Bernoulli Dist.

When \(Y_{ij}\stackrel{iid}{\sim}\) Bernoulli\((p(x_i))\), or \(\sum_{j=1}^{n_i}Y_{ij}\sim\)Binomial\( (n_i, p(x_i)) \) with \(0\leq p(x_i) \leq 1\) and \(p\) is increasing with \(p_i=p(x_i)\), the likelihood can be optimized similarly because

\[(\tilde{p}_1, \ldots, \tilde{p}_n ) = \arg\max_{p_1\leq \ldots\leq p_n}\prod_{i=1}^n p_i^{\sum_{j=1}^{n_i} Y_{ij}}(1-p_i)^{n_i-\sum_{j=1}^{n_i} Y_{ij}}\]

\[ = \arg\min_{p_1\leq \ldots\leq p_n}\sum_{i=1}^n (\overline{Y}_i - p_i)^2 n_i. \]

1. the estimator function \(\tilde{p}\) satisfying that \(\tilde{p}(x_i) = \tilde{p}_i\) for all \(i\);
2. the function \(\tilde{p}\) does not have to be linear (logistic regression);
3. assume that \(\tilde{p}\) is a step function to fill up the gaps between \(x_i\);
4. \(\tilde{p}\) only jumps at some \(x_i\); 
5. some \(x_i\) and corresponding intervals share the same probability \(\tilde{p}(x_i)\). 

Isotonic regression: Poisson Dist.

When \(Y_i\sim\) Poisson\((\lambda(x_i)T_i)\), \(\lambda(x_i) > 0\), where \(Y_i\) is the number of occurences in a Poisson process within time \([0,T_i]\) and hazard rate \(\lambda(x)\) is increasing in \(x\), then 

\[(\tilde{\lambda}_1, \ldots, \tilde{\lambda}_n ) = \arg\max_{\lambda_1\leq \ldots\leq\lambda_n}\prod_{i=1}^n \left\{ \frac{(\lambda_i T_i)^{Y_i}\exp(-\lambda_i T_i)}{Y_i !} \right\}\]

\[ = \arg\min_{\lambda_1\leq \ldots\leq\lambda_n}\sum_{i=1}^n (g_i - \lambda_i)^2 w_i , \]

1. the estimator function \(\tilde{\lambda}\) satisfying that \(\tilde{\lambda}(x_i) = \tilde{\lambda}_i\) for all \(i\);
2. the function \(\tilde{\lambda}\) does not have to be linear (Poisson regression);
3. assume that \(\tilde{\lambda}\) is a step function to fill up the gaps between \(x_i\);
4. \(\tilde{\lambda}\) only jumps at some \(x_i\); 
5. some \(x_i\) and corresponding intervals share the same hazard \(\tilde{\lambda}(x_i)\). 

where \(w_i = T_i\) and \(g(x_i) = Y_i/T_i\), for \(i=1,\ldots,k\). 

Isotonic regression: Summary

In summary,

1. one can apply isotonic regression with covariates to some exponential families, for example, normal, binomial, and Poisson distributions;
2. PAVA algorithm is usually faster than traditional quadratic programming; 
3. step functions are applied to fill up the gaps of mean, probability, or hazard between covariates;
4. covariates with the same estimated mean, probability, or hazard are naturally groped.

Isotonic regression in ML

The last point motivates this study: if a natural tendency of the target function, such as mean, probability, or hazard, is expected, the isotonic regression provides naive options of partitioning or grouping of covariates in terms of the relationship between the target function and the covariates, which will be helpful in classifying and clustering (this cannot be achieved by linear regression or other smoothed models.) 

Q: What are the important covariates corresponding to the target function?
Machine learning can help!