Паспорти на публичните социологически изследвания за парламентарни избори 2026

Техническите данни за изследванията

from IPython.display import HTML, display

def render_facet_row(data):
    all_cards = ""

    for s in data['surveys']:
        m = s.get('metadata', {})
        o = m.get('other_metadata', {})

        # Strip-down card template
        card_html = f"""
        <div style="flex: 1; min-width: 220px; max-width: 300px; border: 1px solid #888;
                    border-radius: 8px; padding: 12px; background: #fff; font-family: sans-serif;
                    box-shadow: 2px 2px 5px rgba(0,0,0,0.05); font-size: 12px;">
            <div style="color: #111; font-weight: bold; border-bottom: 1px solid #eee;
                        margin-bottom: 8px; padding-bottom: 4px; font-size: 13px;">
                {s['agency']} ({s['date_label']})
            </div>
            <div style="color: #777; margin-bottom: 4px;"><b>Период:</b> {m.get('date', 'N/A')}</div>
            <div style="color: #777; margin-bottom: 4px;"><b>Извадка:</b> {m.get('sample_size', 'N/A')}</div>
            <div style="color: #777; margin-bottom: 4px;"><b>Метод:</b> {m.get('methodology', 'N/A')}</div>
            <div style="color: #777; font-size: 11px; margin-top: 8px; font-style: italic;">
                Грешка: {o.get('margin_of_error', 'N/A')}
            </div>
        </div>
        """
        all_cards += card_html

    # Display wrapper with Flexbox
    display(HTML(f"""
    <div style="display: flex; flex-wrap: wrap; gap: 15px; background: #f4f4f4; padding: 20px; border-radius: 12px;">
        {all_cards}
    </div>
    """))

# Execute to render all 19 cards
render_facet_row(data)

import seaborn as sns
import matplotlib.pyplot as plt

# 1. Филтрираме колоните (същата логика)
cols = [c for c in df.columns if c not in ['Agency', 'Date', 'metadata', 'Не съм решил']]

# 2. Сортираме партиите по техния реален резултат
sorted_parties = actual[cols].sort_values(ascending=False).index

# 3. Вземаме СУРОВИТЕ ДАННИ (df вместо residuals)
raw_data_matrix = df[sorted_parties]

# 4. Визуализация
plt.figure(figsize=(10, 10))

# Използваме 'YlGnBu' (Yellow-Green-Blue), защото това са абсолютни стойности (0-100),
# а не отклонения около нулата.
sns.heatmap(raw_data_matrix,
            annot=True,
            fmt='.1f',
            cmap='YlGnBu',
            linewidths=.5,
            cbar_kws={'label': 'Прогнозиран процент (%)'})

plt.title('Матрица данни: Всички прогнози по партии', fontsize=16, fontweight='bold', pad=20)
plt.xlabel('Партия (подредена по реален изборен резултат)', fontsize=12)
plt.ylabel('Агенция / Дата', fontsize=12)
plt.xticks(rotation=45, ha='right')
plt.tight_layout()

plt.show()

import seaborn as sns
import matplotlib.pyplot as plt

# 1. Филтрираме само колоните с партии (изключваме метаданните и служебните категории)
cols = [c for c in df.columns if c not in ['Agency', 'Date', 'metadata', 'Не съм решил']]

# 2. Изчисляваме остатъците (Прогноза минус Реалност)
# Резултат > 0: Партията е била надценена (Survey > Actual)
# Резултат < 0: Партията е била подценена (Survey < Actual)
residuals = df[cols].sub(actual[cols], axis=1)

# 3. Сортираме партиите по техния реален изборен резултат за по-добра четимост
sorted_parties = actual[cols].sort_values(ascending=False).index

# 4. Визуализация на Heatmap
plt.figure(figsize=(10, 10))
sns.heatmap(residuals[sorted_parties],
            annot=True,
            fmt='.1f',
            cmap='RdBu_r',
            center=0,
            linewidths=.5,
            cbar_kws={'label': 'Отклонение (Процентни пунктове)'})

plt.title('Матрица на отклоненията: Прогноза срещу Реалност', fontsize=16, fontweight='bold', pad=20)
plt.xlabel('Партия (подредена по изборен резултат)', fontsize=12)
plt.ylabel('Агенция / Дата', fontsize=12)
plt.xticks(rotation=45, ha='right')
plt.tight_layout()

plt.show()

import matplotlib.pyplot as plt
import numpy as np

# 1. Избор на партиите за сравнение (премахваме метаданните и служебните категории)
cols = [c for c in df.columns if c not in ['Agency', 'Date', 'metadata', 'Не съм решил']]

# 2. Изчисляване на MAE за всеки ред спрямо 'actual'
# Използваме .dropna(), за да сме сигурни, че сравняваме само съществуващи данни
mae = df[cols].apply(lambda x: np.mean(np.abs(x - actual[cols])), axis=1).sort_values()

# 3. Визуализация
plt.figure(figsize=(10, 8), facecolor='#f9f9f9')
colors = plt.cm.RdYlGn_r(np.linspace(0, 1, len(mae)))

# Хоризонтален бар
bars = plt.barh(mae.index, mae.values, color=colors, edgecolor='black', alpha=0.8)

# Добавяне на стойностите върху баровете
for bar in bars:
    plt.text(bar.get_width() + 0.05, bar.get_y() + bar.get_height()/2,
             f'{bar.get_width():.2f}%', va='center', fontsize=9, fontweight='bold')

plt.title('Класация по точност (MAE) - Парламентарни избори 2026', fontsize=14, pad=20, fontweight='bold')
plt.xlabel('Средна грешка в процентни пунктове (по-малко е по-добре)', fontsize=11)
plt.gca().invert_yaxis()  # Най-точните най-отгоре
plt.grid(axis='x', linestyle='--', alpha=0.5)
plt.tight_layout()

plt.show()

# Бонус: Кратка присъда
winner = mae.index[0]
print(f"🏆 НАЙ-ТОЧНА ПРОГНОЗА: {winner} (Грешка: {mae.iloc[0]:.2f}%)")

🏆 НАЙ-ТОЧНА ПРОГНОЗА: Мяра (17/04) (Грешка: 1.99%)

import matplotlib.pyplot as plt
import numpy as np

# 1. Избор на партиите
cols = [c for c in df.columns if c not in ['Agency', 'Date', 'metadata', 'Не съм решил']]

# 2. Изчисляване на RMSE (Root Mean Square Error)
# RMSE = Квадратен корен от средната стойност на квадратите на разликите
rmse = df[cols].apply(lambda x: np.sqrt(np.mean((x - actual[cols])**2)), axis=1).sort_values()

# 3. Визуализация
plt.figure(figsize=(10, 8), facecolor='#f9f9f9')
colors = plt.cm.RdYlGn_r(np.linspace(0, 1, len(rmse)))

bars = plt.barh(rmse.index, rmse.values, color=colors, edgecolor='black', alpha=0.8)

for bar in bars:
    plt.text(bar.get_width() + 0.05, bar.get_y() + bar.get_height()/2,
             f'{bar.get_width():.2f}', va='center', fontsize=9, fontweight='bold')

plt.title('Класация по точност (RMSE) - Парламентарни избори 2026', fontsize=14, pad=20, fontweight='bold')
plt.xlabel('Корен квадратен от средноквадратичната грешка (по-малко е по-добре)', fontsize=11)
plt.gca().invert_yaxis()
plt.grid(axis='x', linestyle='--', alpha=0.5)
plt.tight_layout()

plt.show()

winner = rmse.index[0]
print(f"🏆 НАЙ-ТОЧНА ПРОГНОЗА (RMSE): {winner} (Грешка: {rmse.iloc[0]:.2f})")

🏆 НАЙ-ТОЧНА ПРОГНОЗА (RMSE): Мяра (17/04) (Грешка: 3.33)

import matplotlib.pyplot as plt

# 1. Списък с партиите, които искаме да проследим
target_parties = ['Прогресивна България', 'ГЕРБ', 'ПП-ДБ', 'ДПС', 'Възраждане']

# 2. Сортираме данните хронологично
trend_df = df.sort_values('Date')

# 3. Цикъл за генериране на всяка графика
for party in target_parties:
    plt.figure(figsize=(15, 6))

    # Проверка дали партията съществува в данните (за избягване на грешки)
    if party in trend_df.columns:
        # Плотване на прогнозите
        plt.plot(trend_df.index, trend_df[party],
                 marker='o', color='#1f77b4', linewidth=2.5, markersize=8, label=f'Прогнози за {party}')

        # Линия за реалния резултат
        actual_val = actual[party]
        plt.axhline(y=actual_val, color='#d62728', linestyle='--',
                    linewidth=3, label=f'Реален резултат ({actual_val}%)')

        # Оформление
        plt.title(f'Динамика на прогнозите: {party}', fontsize=16, fontweight='bold', pad=15)
        plt.ylabel('Подкрепа (%)', fontsize=12)
        plt.xlabel('Социологическо проучване', fontsize=12)
        plt.xticks(rotation=45, ha='right')
        plt.legend(fontsize=11, loc='best')
        plt.grid(axis='y', alpha=0.3)
        plt.tight_layout()

        plt.show()
    else:
        print(f"Партията '{party}' не беше намерена в данните.")

import statsmodels.api as sm
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns

# --- 1. ПРЕЦИЗНО ИЗЧИСЛЯВАНЕ НА МЕТРИКИТЕ ---
party_cols = [c for c in df.columns if c in actual.index]

def calculate_metrics(row):
    diffs = row[party_cols] - actual[party_cols]
    return pd.Series({
        'MAE': np.abs(diffs).mean(),
        'RMSE': np.sqrt((diffs**2).mean())
    })

metrics_df = df.apply(calculate_metrics, axis=1)

# --- 2. ПОДГОТОВКА НА ДАННИТЕ ЗА РЕГРЕСИЯ ---
election_date = pd.to_datetime('2026-04-19')
ols_data = []

for idx, row in df.iterrows():
    enc = row['metadata'].get('encoded', {})
    if not enc: continue

    ols_data.append({
        'Label': idx,
        'MAE': metrics_df.loc[idx, 'MAE'],
        'RMSE': metrics_df.loc[idx, 'RMSE'],
        'Is_Hybrid': 1 if enc.get('methodology') == 'hybrid' else 0,
        'Days_To_Election': (election_date - pd.to_datetime(enc.get('date'))).days,
        'Sample_Size': enc.get('sample_size', 0)
    })

reg_df = pd.DataFrame(ols_data).set_index('Label').dropna()

# --- 3. ИЗПЪЛНЕНИЕ НА РЕГРЕСИИТЕ ---
X = sm.add_constant(reg_df[['Is_Hybrid', 'Days_To_Election', 'Sample_Size']])
model_mae = sm.OLS(reg_df['MAE'], X).fit()
model_rmse = sm.OLS(reg_df['RMSE'], X).fit()

# --- 4. ВИЗУАЛИЗАЦИЯ 1: ВЛИЯНИЕ НА ФАКТОРИТЕ (КОЕФИЦИЕНТИ) ---
fig1, (ax1, ax2) = plt.subplots(1, 2, figsize=(16, 4), sharey=True)
plt.subplots_adjust(wspace=0.1)

# MAE Coefs
model_mae.params[1:].plot(kind='barh', xerr=model_mae.bse[1:], color='#3498db', ax=ax1, alpha=0.8, edgecolor='black')
ax1.axvline(0, color='red', linestyle='--', linewidth=1)
ax1.set_title('Въздействие върху MAE\n(Линейна средна грешка)', fontweight='bold')
ax1.set_xlabel('Промяна в % пунктове')

# RMSE Coefs
model_rmse.params[1:].plot(kind='barh', xerr=model_rmse.bse[1:], color='#e74c3c', ax=ax2, alpha=0.8, edgecolor='black')
ax2.axvline(0, color='red', linestyle='--', linewidth=1)
ax2.set_title('Въздействие върху RMSE\n(Наказание за големи грешки)', fontweight='bold')
ax2.set_xlabel('Промяна в % пунктове')

plt.suptitle('АНАЛИЗ НА ФАКТОРИТЕ ЗА ГРЕШКА (OLS REGRESSION COEF)', fontsize=16, fontweight='bold', y=1.05)
plt.show()

# --- 5. ВИЗУАЛИЗАЦИЯ 2: SIDE-BY-SIDE ТАГНАТА ДИНАМИКА (MAE vs RMSE) ---
fig2, (ax3, ax4) = plt.subplots(2, 1, figsize=(16, 14), sharex=True)
palette = ['#3498db', '#e74c3c']

for ax, metric, color, title in zip([ax3, ax4], ['MAE', 'RMSE'], palette, ['MAE (Линейна)', 'RMSE (Квадратична)']):
    # Scatter plot
    sns.scatterplot(
        data=reg_df, x='Days_To_Election', y=metric,
        hue='Is_Hybrid', size='Sample_Size', sizes=(150, 800),
        palette=palette, alpha=0.6, edgecolor='white', linewidth=1.5, ax=ax
    )

    # Тагване на всяка точка
    for label, row in reg_df.iterrows():
        ax.text(
            x=row['Days_To_Election'],
            y=row[metric] + (row[metric] * 0.02), # динамично отместване нагоре
            s=label, fontsize=8, fontweight='semibold', ha='center',
            bbox=dict(facecolor='white', alpha=0.5, edgecolor='none', pad=0.5)
        )

    ax.invert_xaxis()
    ax.set_title(f'Динамика на {title} грешката', fontsize=14, fontweight='bold')
    ax.set_xlabel('Дни до изборния ден', fontsize=11)
    ax.set_ylabel('Стойност на грешката', fontsize=11)
    ax.grid(True, linestyle=':', alpha=0.6)
    ax.get_legend().remove() # премахваме индивидуалните легенди за по-чист вид

# Единна легенда за целия плот
handles, labels = ax4.get_legend_handles_labels()
fig2.legend(handles, labels, title='Метод (0=Offline, 1=Hybrid) & Извадка', loc='center right', bbox_to_anchor=(1.1, 0.5))

plt.suptitle('СРАВНИТЕЛНА ХРОНОЛОГИЯ НА КАЧЕСТВОТО (ТАГНАТИ НАБЛЮДЕНИЯ)', fontsize=16, fontweight='bold', y=0.98)
plt.tight_layout()
plt.show()

# --- 6. СТРУКТУРИРАН ТЕКСТОВ АУТПУТ ---
print("\n" + "="*60)
print(" СРАВНИТЕЛЕН АНАЛИЗ НА КОЕФИЦИЕНТИТЕ (MAE vs RMSE)")
print("="*60)
summary_table = pd.DataFrame({
    'Metric': ['Is_Hybrid', 'Days_To_Election', 'Sample_Size'],
    'MAE Coef': model_mae.params[1:].values,
    'MAE P-value': model_mae.pvalues[1:].values,
    'RMSE Coef': model_rmse.params[1:].values,
    'RMSE P-value': model_rmse.pvalues[1:].values
}).round(4)
print(summary_table.to_string(index=False))
print("="*60)

============================================================
 СРАВНИТЕЛЕН АНАЛИЗ НА КОЕФИЦИЕНТИТЕ (MAE vs RMSE)
============================================================
          Metric  MAE Coef  MAE P-value  RMSE Coef  RMSE P-value
       Is_Hybrid    1.0574       0.0000     2.8161        0.0000
Days_To_Election    0.0051       0.1617     0.0056        0.4806
     Sample_Size   -0.0005       0.5626    -0.0010        0.5954
============================================================

# '%pip install qrcode[pil] -q'

# import qrcode
# import matplotlib.pyplot as plt
# import seaborn as sns
# import numpy as np
# import pandas as pd
# import statsmodels.api as sm
# from matplotlib.backends.backend_pdf import PdfPages
# from google.colab import files

# # --- 1. КОНФИГУРАЦИЯ ---
# pdf_filename = "Election_Analysis_2026_Final_Report.pdf"
# gist_url = "https://gist.githubusercontent.com/gregordimi/e0d3fe30e4f898c84ad25c38b39f6e44/raw/311a84fb7ce48c8bc47d61b359405d602b2aabd6/results2026.json"
# notebook_url = "https://colab.research.google.com/drive/1Zu_q3iuHCEs0OmtctdoCo3ef7LvgTeru?usp=sharing"
# author_url = "https://linkedin.com/in/grigordimitrov"
# election_date = pd.to_datetime('2026-04-19')
# PAGE_SIZE = (11, 8.5)

# def get_qr_image(url):
#     qr = qrcode.QRCode(box_size=10, border=2)
#     qr.add_data(url)
#     qr.make(fit=True)
#     return qr.make_image(fill_color="black", back_color="white")

# def add_footer(fig, page_num, total_pages):
#     footer_text = f"Източник: Публични данни - Сайтове на агенции / Медии | Стр. {page_num} от {total_pages} | Забележка: Данните са събирани чрез бот, възможни са неточности"
#     fig.text(0.5, 0.04, footer_text, ha='center', fontsize=8, color='grey', alpha=0.7)

# # --- 2. ПОДГОТОВКА НА ДАННИ ---
# party_cols = [c for c in df.columns if c in actual.index]
# cols = [c for c in df.columns if c not in ['Agency', 'Date', 'metadata', 'Не съм решил']]

# def calculate_metrics(row):
#     diffs = row[party_cols] - actual[party_cols]
#     return pd.Series({'MAE': np.abs(diffs).mean(), 'RMSE': np.sqrt((diffs**2).mean())})

# metrics_all_df = df.apply(calculate_metrics, axis=1)

# ols_list = []
# for idx, row in df.iterrows():
#     enc = row['metadata'].get('encoded', {})
#     if not enc: continue
#     ols_list.append({
#         'Label': idx,
#         'MAE': metrics_all_df.loc[idx, 'MAE'],
#         'RMSE': metrics_all_df.loc[idx, 'RMSE'],
#         'Is_Hybrid': 1 if enc.get('methodology') == 'hybrid' else 0,
#         'Days_To_Election': (election_date - pd.to_datetime(enc.get('date'))).days,
#         'Sample_Size': enc.get('sample_size', 0)
#     })
# reg_df = pd.DataFrame(ols_list).dropna()

# X_reg = sm.add_constant(reg_df[['Is_Hybrid', 'Days_To_Election', 'Sample_Size']])
# model_mae = sm.OLS(reg_df['MAE'], X_reg).fit()
# model_rmse = sm.OLS(reg_df['RMSE'], X_reg).fit()

# target_parties = ['Прогресивна България', 'ГЕРБ', 'ПП-ДБ', 'ДПС', 'Възраждане']
# sorted_parties = actual[cols].sort_values(ascending=False).index
# total_p = 7 + len(target_parties)

# with PdfPages(pdf_filename) as pdf:

# # --- СТРАНИЦА 1: ЗАГЛАВИЕ (ADJUSTED) ---
#     fig = plt.figure(figsize=PAGE_SIZE)
#     plt.axis('off')

#     # DRAFT воден знак
#     plt.text(0.5, 0.96, 'DRAFT V2', ha='center', va='center', fontsize=40, fontweight='bold', color='red', alpha=0.7)

#     # Заглавие и Подзаглавие
#     plt.text(0.5, 0.88, 'Анализ на социологическата точност: Избори 2026', ha='center', fontsize=24, fontweight='bold', color='#003366')
#     plt.text(0.5, 0.80, 'Статистически анализ на факторите, влияещи върху грешката', ha='center', fontsize=14, style='italic', color='#444')

#     # Секция АВТОР И ЛИНКОВЕ
#     # Използваме url= параметър за кликаемост в PDF
#     plt.text(0.1, 0.70, f'Автор: Г. Д.', fontsize=12, fontweight='bold', color='#003366')
#     plt.text(0.1, 0.67, 'LinkedIn Профил', fontsize=11, color='blue', url=author_url, bbox=dict(facecolor='none', edgecolor='none', pad=0))

#     plt.text(0.1, 0.63, 'Източник на данни', fontsize=11, color='blue', url=gist_url)
#     plt.text(0.1, 0.59, 'Код на анализа', fontsize=11, color='blue', url=notebook_url)

#     plt.text(0.1, 0.54, f'Дата на анализ: {pd.Timestamp.now().strftime("%d.%m.%Y")}', fontsize=10, color='#666')

#     # Блок с метрики
#     metrics_note = ("• MAE (Средна абсолютна грешка): Линейно отклонение на прогнозата.\n"
#                     "• RMSE (Средноквадратична грешка): Подчертава големите индивидуални пропуски.\n"
#                     "• OLS Regression: Идентифицира кои фактори (метод, време, извадка) са статистически значими.")
#     plt.text(0.1, 0.40, metrics_note, fontsize=10, bbox=dict(facecolor='#f0f0f0', alpha=0.5, edgecolor='none', pad=5))

#     # QR Кодове (Запазваме ги като визуален елемент)
#     ax_qr0 = fig.add_axes([0.20, 0.12, 0.15, 0.15]); ax_qr0.imshow(get_qr_image(author_url), cmap='gray'); ax_qr0.axis('off'); ax_qr0.set_title("Автор", fontsize=8)
#     ax_qr1 = fig.add_axes([0.42, 0.12, 0.15, 0.15]); ax_qr1.imshow(get_qr_image(notebook_url), cmap='gray'); ax_qr1.axis('off'); ax_qr1.set_title("Код", fontsize=8)
#     ax_qr2 = fig.add_axes([0.64, 0.12, 0.15, 0.15]); ax_qr2.imshow(get_qr_image(gist_url), cmap='gray'); ax_qr2.axis('off'); ax_qr2.set_title("Данни", fontsize=8)

#     add_footer(fig, 1, total_p)
#     pdf.savefig(fig, dpi=300, bbox_inches='tight'); plt.close()

#     # --- СТРАНИЦА 6: МАТРИЦА НА ДАННИ ---
#     fig = plt.figure(figsize=PAGE_SIZE)
#     sns.heatmap(df[sorted_parties], annot=True, fmt='.1f', cmap='YlGnBu', annot_kws={"size": 8}, cbar_kws={'shrink': 0.8})
#     plt.title('Матрица на данните: Всички прогнози (%)', fontsize=16, fontweight='bold')
#     plt.xticks(rotation=45, ha='right', fontsize=9); plt.yticks(fontsize=8)
#     plt.tight_layout(rect=[0.05, 0.08, 0.95, 0.95])
#     add_footer(fig, 6, total_p)
#     pdf.savefig(fig, dpi=300); plt.close()

#     # --- СТРАНИЦА 2: RMSE СРЕЩУ MAE ---
#     metrics_summary = metrics_all_df.sort_values('RMSE')
#     fig, ax = plt.subplots(figsize=PAGE_SIZE)
#     y_indices = np.arange(len(metrics_summary))
#     ax.barh(y_indices - 0.17, metrics_summary['MAE'], 0.35, label='MAE', color='#3498db', alpha=0.8, edgecolor='black')
#     ax.barh(y_indices + 0.17, metrics_summary['RMSE'], 0.35, label='RMSE', color='#e74c3c', alpha=0.8, edgecolor='black')
#     ax.set_yticks(y_indices); ax.set_yticklabels(metrics_summary.index); ax.invert_yaxis()
#     ax.set_title('Сравнение на точността по агенции\n', fontsize=16, fontweight='bold')
#     ax.text(0.5, 1.02, 'Агрегирани стойности на MAE и RMSE спрямо официалния резултат', transform=ax.transAxes, ha='center', fontsize=11, color='#555')
#     ax.legend(); ax.grid(axis='x', linestyle='--', alpha=0.3)
#     plt.tight_layout(rect=[0.05, 0.08, 0.95, 0.92])
#     add_footer(fig, 2, total_p)
#     pdf.savefig(fig, dpi=300); plt.close()

#     # --- СТРАНИЦА 3: OLS КОЕФИЦИЕНТИ (С ТИТЛИ И ПОДЗАГЛАВИЯ) ---
#     fig, (ax1, ax2) = plt.subplots(1, 2, figsize=PAGE_SIZE, sharey=True)

#     model_mae.params[1:].plot(kind='barh', xerr=model_mae.bse[1:], color='#3498db', ax=ax1, alpha=0.8, edgecolor='black')
#     ax1.axvline(0, color='red', linestyle='--', linewidth=1)
#     ax1.set_title('Въздействие върху MAE\n', fontsize=13, fontweight='bold')
#     ax1.text(0.5, 1.02, 'Как факторите променят линейната грешка (%)', transform=ax1.transAxes, ha='center', fontsize=9, color='#555')

#     model_rmse.params[1:].plot(kind='barh', xerr=model_rmse.bse[1:], color='#e74c3c', ax=ax2, alpha=0.8, edgecolor='black')
#     ax2.axvline(0, color='red', linestyle='--', linewidth=1)
#     ax2.set_title('Въздействие върху RMSE\n', fontsize=13, fontweight='bold')
#     ax2.text(0.5, 1.02, 'Тежест на факторите при големи отклонения', transform=ax2.transAxes, ha='center', fontsize=9, color='#555')

#     plt.suptitle('СТАТИСТИЧЕСКА ЗНАЧИМОСТ НА ФАКТОРИТЕ (OLS)', fontsize=16, fontweight='bold', y=0.96)
#     # Свиване на графиките надолу чрез rect [left, bottom, right, top]
#     plt.tight_layout(rect=[0.05, 0.12, 0.95, 0.88])
#     add_footer(fig, 3, total_p)
#     pdf.savefig(fig, dpi=300); plt.close()

#     # --- СТРАНИЦА 4: SIDE-BY-SIDE ДИНАМИКА (С ТИТЛИ И ПОДЗАГЛАВИЯ) ---
#     fig, (ax3, ax4) = plt.subplots(1, 2, figsize=PAGE_SIZE)
#     titles = ['MAE Dynamics', 'RMSE Dynamics']
#     subtitles = ['Линейно отклонение във времето', 'Квадратично наказание за големи грешки']

#     for ax, metric, color, title, sub in zip([ax3, ax4], ['MAE', 'RMSE'], ['#3498db', '#e74c3c'], titles, subtitles):
#         sns.scatterplot(data=reg_df, x='Days_To_Election', y=metric, hue='Is_Hybrid', size='Sample_Size',
#                         sizes=(100, 700), palette={0: '#3498db', 1: '#e74c3c'}, alpha=0.6, edgecolor='black', ax=ax)
#         for _, row in reg_df.iterrows():
#             ax.text(row['Days_To_Election'], row[metric], s=row['Label'], fontsize=7, rotation=45, ha='left', va='bottom')

#         ax.invert_xaxis(); ax.grid(True, linestyle=':', alpha=0.6)
#         ax.set_title(f'{title}\n', fontsize=13, fontweight='bold')
#         ax.text(0.5, 1.02, sub, transform=ax.transAxes, ha='center', fontsize=9, color='#555')
#         ax.get_legend().remove()

#     handles, labels = ax4.get_legend_handles_labels()
#     fig.legend(handles, labels, title='Метод (0=F2F, 1=Hybrid)', loc='center right', bbox_to_anchor=(0.98, 0.5))
#     plt.suptitle('ХРОНОЛОГИЧНО РАЗПРЕДЕЛЕНИЕ НА ТОЧНОСТТА', fontsize=16, fontweight='bold', y=0.96)
#     # Свиване на графиките надолу
#     plt.tight_layout(rect=[0.05, 0.1, 0.88, 0.88])
#     add_footer(fig, 4, total_p)
#     pdf.savefig(fig, dpi=300); plt.close()

#     # --- СТРАНИЦА 5: МАТРИЦА НА ОТКЛОНЕНИЯТА ---
#     residuals = df[cols].sub(actual[cols], axis=1)
#     fig = plt.figure(figsize=PAGE_SIZE)
#     sns.heatmap(residuals[sorted_parties], annot=True, fmt='.1f', cmap='RdBu_r', center=0, annot_kws={"size": 8}, cbar_kws={'shrink': 0.8})
#     plt.title('Матрица на отклоненията (Прогноза - Реалност)', fontsize=16, fontweight='bold')
#     plt.xticks(rotation=45, ha='right', fontsize=9); plt.yticks(fontsize=8)
#     plt.tight_layout(rect=[0.05, 0.08, 0.95, 0.95])
#     add_footer(fig, 5, total_p)
#     pdf.savefig(fig, dpi=300); plt.close()



#     # --- СТРАНИЦИ 7-N: ТРЕНДОВЕ ---
#     curr_p = 7
#     trend_df = df.sort_values('Date')
#     for party in target_parties:
#         if party in trend_df.columns:
#             fig = plt.figure(figsize=PAGE_SIZE)
#             plt.plot(trend_df.index, trend_df[party], marker='o', color='#1f77b4', linewidth=2)
#             plt.axhline(y=actual[party], color='#d62728', linestyle='--', linewidth=3, label=f'Реалност ({actual[party]}%)')
#             plt.title(f'Тренд на прогнозите: {party}', fontsize=16, fontweight='bold')
#             plt.xticks(rotation=45, ha='right', fontsize=8); plt.legend(); plt.grid(alpha=0.3)
#             plt.tight_layout(rect=[0.05, 0.08, 0.95, 0.95])
#             add_footer(fig, curr_p, total_p)
#             pdf.savefig(fig, dpi=300); plt.close()
#             curr_p += 1

#     # --- ФИНАЛНА СТРАНИЦА: ЗАКЛЮЧЕНИЯ ---
#     fig = plt.figure(figsize=PAGE_SIZE)
#     plt.axis('off')
#     plt.text(0.05, 0.92, 'Заключение и изводи от анализа', fontsize=20, fontweight='bold', color='#003366')

#     y_pos = 0.82
#     conclusion_blocks = [
#         ("• Системен пропуск:", "Всички агенции подцениха победителя (Прогресивна България) с над 10% средно отклонение.", True),
#         ("", "Всички агенции надцениха установените играчи (ГЕРБ, ДПС, Възраждане) в различна степен.", True),
#         ("• Методология:", "Необходима е ревизия на моделите за разпределение на гласовете на нерешилите избиратели.", False),
#         ("", "Необходима е критична ревизия на източниците и структурирането на извадката.", True),
#         ("• Информационна хигиена:", "Ефирът е пълен с демографски, психографски и поведенчески разпределения на вота, към които", False),
#         ("", "трябва да се подходи със скептицизъм, имайки предвид мащаба на анализираните отклонения.", False),
#         ("", "Анализи на по-малки кохорти (напр. от типа коя партия е донор на гласоподаватели)", False),
#         ("", "носят риск да са значително по-неточни.", False),
#         ("• Забележка:", "Нормализирането на толкова голям набор от изследвания изисква задълбочена работа.", False),
#         ("", "Този анализ е по-скоро отправна точка за по-задълбочени бъдещи мета-анализи, целящи повишаване на прогнозната сила.", False),
#         ("", "Най-належащата задача би била ръчното преглеждане и коригиране на входящите данни.", False),
#     ]

#     for header, detail, add_space in conclusion_blocks:
#         if header: plt.text(0.05, y_pos, header, fontsize=13, fontweight='bold', color='#003366'); y_pos -= 0.05
#         plt.text(0.07, y_pos, detail, fontsize=12, color='#333333'); y_pos -= 0.05
#         if add_space: y_pos -= 0.02

#     rect = plt.Rectangle((0.03, 0.15), 0.94, 0.80, linewidth=1, edgecolor='#cccccc', facecolor='#f9f9f9', alpha=0.3, zorder=-1)
#     fig.patches.append(rect)
#     add_footer(fig, total_p, total_p)
#     pdf.savefig(fig, dpi=300, bbox_inches='tight'); plt.close()

# print(f"✅ Докладът е генериран успешно: {pdf_filename}")
# files.download(pdf_filename)

Intro

Practitioners have defined several systematic behaviors which potentially lead to low-quality data in online data collection ¹, ². Namely, cross-cultural factors, untruthfulness, speeding, question format, and sample quality have been identified to have a negative impact on data quality, thus, causing noise in the data rather than variance.

One of the major limitations of online opinion taking, and in particular preference measurement is the lack of motivation by respondents ³, which can result in reduced respondents engagement, hence untruthful answers. For the purpose of this paper, I will further elaborate on respondents’ untruthfulness, engagement, and motivation.

The hypothetical nature in which typically conjoint studies are executed questions the extent to which they are able to reveal the true consumer preferences ⁴. When respondents are not incentivized to reveal their true preferences and the data is collected in hypothetical settings, weaker external validity is observed ⁵. Respondents have little or no incentive to reveal their true preference or to be mindful about their answers in hypothetical settings ⁶. Economic theory suggests that if economic agents believe their answers will have an impact on decisions made by businesses or governments for outcomes that the agents care about, they should respond in such a way as to maximize their payoffs and welfare ⁷. However, due to the hypothetical nature of the surveys, in most conjoint studies respondents may misinterpret their preferences and the study may not yield truthful answers. Though hypothetical techniques are found to perform fine in many domains ⁸; ⁹; ¹⁰; ¹¹, recent conjoint studies reveal that incentive-aligned conjoint is superior to hypothetical conjoint in out-of-sample prediction, and it is more engaging and motivating ¹²; ¹³; ¹⁴.

Typically, practitioners apply different rules of identifying “bad quality” respondents and eliminating them when analyzing the data, such as using misdirect questions (or traps). For example, ¹⁵ recognize untruthfulness by asking respondents several questions with an extremely low incidence rate and asking questions that have known distribution across the population. The authors argue that based on the score estimated from these questions, it is possible for researchers to screen out respondents and thus, remove respondents that do not have carefully considered answers. Speeding is also a quality measure often applied to achieve the necessary data quality. ¹⁶ claim that 85% of the respondents are speeding at some point of the questionnaires, which makes it impossible to exclude these responses from the survey. The authors argue that speeding is an “epidemic” in the online survey world and it is a generic issue that holds the potential to cause the biggest amount of noise in the data. In their paper, the authors divided the sample into two categories, above and below the average length of the interview. The analysis of the comparison of both groups showed that respondents in the group below average times account for +/-6.2% higher variance on average. As expected, with the progression of the survey the variance increases from 4% at the beginning to 8% at the end. The authors do not bind themselves to the claim that this variance is a source of noise in the data, however, one could speculate that the consequence of speeding is low-quality data.

The recent technological advancements allowed for increased accessibility to respondents ¹⁷. But this comes with a trade-off, as there has been decreased attentiveness and patience as respondents are easily distracted. With the aid of technology, there have been efforts to engage respondents with the task using improved interfaces, survey designs, question formats, and even game-like environments (see ¹⁸ for discussion). Hence, practitioners have attempted to tackle the disengagement by incorporating intuitive mechanisms in their survey designs. These mechanisms indicate only “suggestive evidence” of underlying issues and determination that respondents are “behaving badly”. Despite those attempts, the biggest source of bad data is still the disengaged respondent.

The academia also recognized the lack of motivation as a key limitation concerning preference measurement methods ¹⁹; ²⁰; ²¹; ²². However, academia has investigated the source of these problems in order to provide solid solutions. Respondents are putting significantly less cognitive effort when making survey choices compared to real-life decisions, which leads to poor out-of-sample predictions ²³. To increase the level of involvement, based on insights from experimental economics, some scholars approached preference measurement issues applying incentive-compatible mechanisms. Experimental economists claim that using incentive-compatible methods should induce truth-telling as monetary incentives dominate over other factors. Hence, respondents are encouraged to reveal their true preferences to maximize their final payoff ²⁴. In their Induced value theory, ²⁵ define 5 percepts, 3 of which are sufficient conditions for incentive compatible behavior. The first principle is referred to as saliency, in other words, participants’ final monetary payoff should depend on their performance, where better performance leads to a better outcome. The saliency principle is also related to the non satiation principle, claiming that “utility is a monotone increasing function of monetary reward”, or simply put “the more the better”. The third principle is dominance. To assure control over the preference, the final payoff for the exerted cognitive efforts should dominate the subjective costs of participation. When those conditions are met, incentive-compatible methods should yield truthful answers.

Recent studies on adopting incentive-compatible mechanisms to conjoint have shown that incentive-compatible mechanisms can result in boosting respondents’ motivation leading to respondents providing truthful answers.

Incentive aligned conjoint

A general concern amongst practitioners when using conjoint methodologies is that respondents are not motivated during the survey and that choices made during the conjoint exercise differ from real-life purchasing choices ²⁶. ²⁷ compared the out-of-sample predictions for two conjoin designs; the first adopted the traditional hypothetical setting, while the second used Chinese dinner special as a context in which respondents had a positive chance of receiving a preferred alternative from each choice set at the end of the experiment. Results showed that the external validity is increased compared to a setting where respondents hypothetically provide answers for each of the choice tasks.

The authors find the most relevant principle to conjoint analysis to be saliency. In practice, respondents are given a flat participation fee, which doesn’t induce honest responses. Hence, saliency hasn’t been satisfied as respondents are neither rewarded nor penalized based on their performance. Providing respondents with a flat participation fee cannot motivate them enough to carefully evaluate every given choice in the conjoint surveys, resulting in noise in the data. Hence, there is no reason for researchers to expect that predictions based on the survey data from non-salient responses will make valid out-of-sample predictions. In other words, respondents’ stated preference will not be consistent with their revealed preference ²⁸. The main disadvantage of this experiment, however, is that the researcher should have available all alternatives from each choice set.

²⁹ created a mechanism that does not require changes in the existing conjoint methodologies and allows for situations in which the researcher has available a few alternatives of the product. Respondents are offered the available alternatives as potential rewards. The method incorporates BDM ³⁰ procedure to compare subjects’ valuations to a randomly generated price, the optimal strategy for a participant being to state his or her true WTP. If the bid is higher, the subject pays the randomly generated price and gets the item. If the bid is lower than the elicited price, the subject doesn’t pay and receives nothing. The method works in four steps. First, participants complete a standard conjoint exercise. Afterward, the experimenter shows them a real product that they could potentially purchase. By not revealing the product in hand prior to the conjoint exercise (as it is of limited availability), respondents’ preferences are not biased. After the product, if being shown, the interred WTP is calculated. Finally, using BDM it is determined whether a participant can purchase the real product or not. The drawback of this method, however, is that by using the BDM mechanism, respondents have the chance to purchase the real product with the majority of the amount of money coming from the difference between the inferred price and the randomly drawn price. Hence, there might be a difference when in real life respondents need to use their own money. Moreover, the design cannot be applied to new goods as it requires an alternative to be present when performing the study and one of the main applications of conjoint analysis is for new product development.

³¹, proposed an incentive-aligned conjoint method that allows respondents to update each attribute and configure their own product. When upgrading attributes, respondents state their willingness to pay per upgrade. The BDM mechanism is adopted here as well, to ensure that the best option for respondents is to truthfully state their willingness to pay. At the end of the experiment, respondents receive their upgraded product. The authors found that the method can significantly outperform the standard conjoint study. The drawback of this study is that again, researchers need to have the product available.

³² argued that those methods “may not increase involvement to a level of real-life purchasing decisions”. The reasoning is that respondents may behave differently when obtaining the product requires them to use their own money instead of receiving a product as a reward at the end of the experiment. In such a case, respondents will pay more attention to certain attributes relevant to them in real life, but not to such an extent during the conjoint exercise, where there is a probability of obtaining the tested product. Hence, ³³ proposed the “conjoin poker”, which is an incentive-compatible conjoint study that collects preference data in a game setting in an attempt to increase respondents’ involvement and attention. The study design resulted in increased respondent engagement and motivation, which was measured by the time they spend in the survey and by a follow-up questionnaire. Generally, the study design was found engaging and entertaining, but also complex and time-consuming.

The issue that still stands in order to comply with the saliency principle is how to evaluate respondents’ performance in an absence of objective truth. In 2004 Drazen Prelec proposed a solution which he called Bayesian truth serum. This paper’s main objective will be to test if the BTS can help motivate respondents to truthfully reveal their preferences and improve the quality of the data. The method works in hypothetical settings i.e. in a new product development context. It doesn’t alter the experimental design to an extent such as the conjoint poker. It is relatively easy for understanding by researchers and respondents. The following section described BTS intuition, assumptions, and applications.

Revealing truthful preference with BTS

In his work on truth-telling incentives, ³⁴ recognizes another source of bias that can result in low-quality data. He suggests that in the absence of “external criteria” people can become subjects of “self-deception and false confidence even among the well-intentioned”. In order to express the notion that some experts are never subject to reality checks, Prelec compares the subjective judgment of business investors with an art critic. Both experts are making subjective judgments but the end result of the investor can be used for evaluation of his judgments (i.e. reality check) while for the art critic there are no external criteria for proper assessment.

References

The post has been originally published as part of my master thesis Bayesian Truth Serum Fused Conjoint on 29 August 2017

Literature Reference List

Information score intuition

Information score is estimated on question level and assigns high values to common answers and low values to the opposite. Surprisingly common answers are considered those whose actual frequency is higher than collectively predicted by the same population of interest ². Once considered an irrational bias, afterward proven as a purely rational effect, the idea behind common/uncommon answers is that opinion holders would overestimate the frequency of their own opinion amongst the population. ³ describe this effect as one “tend to perceive a”false consensus”-to see their own behavioral choices and judgments as relatively common and appropriate to existing circumstances while viewing alternative responses as uncommon, deviant, or inappropriate.” More precisely, it is a systematic deviation of assigning higher estimations or probabilities with regards to population frequency in directions of one’s own opinion in a group of which the subject is a member.

In his paper, ⁴ referred to the “false consensus bias” as “an egocentric bias to overestimate the degree in which others are like us”. ⁵ executed four studies in which they tested the “false consensus bias” in various settings such as behavioral choices, situations, and judgments as well as personal problems, expectations preferences, and characteristics. For example, a study among Stanford undergraduates ⁶ showed that students who think about dying expected that 44% of students, in general, would share their problem against students who don’t think about dying expected only 25.6% of the students to think about dying. Within the same experiment subjects who believed to die before turning age of 70 indicated that 57.6% of the students share their expectations while this number was only 43.9% for students with the opposite expectation.

This pattern was considered to be an irrational systematic deviation before ⁷ has proven it to be a rational effect. Using Bayesian reasoning Dawes argues that subjects are considering their own opinion to be an “informative sample of one” which shifts the estimation of endorsers of a certain opinion in a direction toward their own opinion.

Using Dawes’ interpretation ⁸ constructed a critical assumption; the information score component of BTS which indicates that ones’ opinion frequency will be underestimated by the others and vice versa, one will overestimate the frequency of his own opinion. Consequently, this implies that surprisingly common answers are those whose actual frequency overweighs the collectively predicted frequency. For example, let’s consider we have data about respondents being asked to state their preferences at binary choice (e.g. Yes/No) question. The respondents are asked to provide their own opinion as well as their estimates about the fraction of people endorsing their answer. If the actual frequency is 10% per “Yes” and the predicted frequency is 5%, the respondents who indicated “Yes” will receive a positive score as their answer will be considered “surprisingly common”. However, given the same actual frequency, if the collectively predicted frequency is 25% respondents will be penalized with a negative score as their answers will be considered uncommon.

Information score formula

⁹ defines the following formula for estimation of information score component for answer :

Each respondent information score per number of answers:

Where:

• is the actual frequency of answer or
• is the geometric average of predicted frequencies for answer or

Prediction score intuition

The second component of BTS scoring is the prediction score, which is “a penalty proportional to the relative entropy (Kullback-Leibler divergence) between the empirical distribution and respondent’s prediction of that distribution” ¹⁰. In other words, this score assigns zero value to accurate prediction (i.e. best prediction score) when one’s prediction matches the actual frequency and penalizes for inaccurate one. It is an asymmetric estimate of the difference between two probability distributions which produces a number smaller than zero.

Prediction score formula

is a constant used for assigning different weights to the prediction score relative to the information score in the BTS final score. Setting equal to 1 will assign equal weights to the score of the equation which will result in a zero-sum game. In the zero-sum game, the average of all BTS scores is equal to zero. When approaches zero BTS game results in “Pareto-dominate expected scores” in which at least one individual has a positive score.

BTS formula

The overall score for each respondent is:

BTS Assumptions

¹¹ relies on several assumptions for constructing BTS algorithm:

Bayesian Nash equilibrium

“Truth-telling is individually and collectively optimal”¹². Taking into account the BTS assumptions, agents stating their true preferences is BTS Nash equilibrium for any α>0. Thus, it is individually optimal as BTS Nash equilibrium results in a maximum expected payoff. It is collectively optimal because there is no other equilibrium which results in a higher information score for any agent. The equilibrium relies on two assumptions:

1. There is a large enough sample so that a single answer can not affect the overall distribution;
2. All agents are Bayesians. They have a common prior and they update their beliefs based on their opinions;

BTS validity

Although the BTS is a relatively easy method that can be implemented within a study design, little has been done to assess its validity. So far, the results are positive regarding the usage of the approach.

In an attempt to eliminate hypothetical bias in contingent valuation (CV) ¹³ compared three mechanisms to elicit subjects’ true preferences along with real and hypothetical CV in two settings. The results showed that BTS was able to eliminate the hypothetical bias in one of the settings but only reduce it in the other. The authors conclude that BTS eliminates the hypothetical bias inconsistently. Interestingly enough, the authors also found an interaction between BTS and subjects’ characteristics.

¹⁴ emphasized that accurate prediction of performance on the market when products are not yet available is crucial to the business’s success. In a study that aims to predict a medical product adoption on the market, the authors made the first empirical validation of BTS. In order to improve the predictive performance, the approach is adopted, based on the notion that “bad” respondents “hurt” the data, hence the authors exclude respondents with a low score (i.e. the untruthful ones). Applying BTS on data from 1763 physicians, BTS improved the overall performance up to 21% compared to a simple average from the full sample. An interesting particularity, contradictory to the original BTS article by ¹⁵, is that the authors did not explain the BTS scoring concept to the participants, arguing that it is impractical and it hasn’t been empirically validated to improve the data.

In the study of ¹⁶, the authors conducted 5 experiments testing different aspects of BTS. They tested the validity of BTS upon respondents recognizing real and fake items from different categories such as brands, movies, journals, etc. The authors tested BTS in the following settings: Recognition questionnaires, in 2x2 between-subject design which consists of truth-telling (Control vs. BTS) and deception incentives (incentives for recognizing items vs. none). For the first experiment, the authors noted that it is not clear whether the performance of BTS is because of the truth-telling incentives or because of the claim made about them. To eliminate this doubt, keeping the same experimental settings, respondents were presented with their information score after each question. BTS was also compared to another “truth serum”, the solemn oath. It requires participants to sign a declaration of their honesty, and again, respondents were given incentives to exaggerate their knowledge. BTS was also applied in the evaluation of a public good, e.g. it was applied on CV, in an attempt to eliminate the hypothetical bias. The experiments results showed good performance of BTS in recognizing real and fake items. The method performed well even when respondents were given incentives to overclaim item recognition. The results indicated there is strong evidence that instructing respondents about BTS has a positive impact on truth-telling. BTS results outperformed the “solemn oath” and successfully eliminated the hypothetical bias in CV. Important to notice in this research is that the authors provided empirical evidence that proper BTS instructions will result in higher data quality, a claim that was fairly questioned by ¹⁷. Furthermore, the results showed that BTS eliminated the hypothetical bias in CV and the authors speculated that the mixed evidence about BTS performance in ¹⁸, possibly result from a difference in the instructions. ¹⁹ also addresses interesting pragmatic issues that will be discussed in another part of the paper.

In an interesting article about Questionable Research Practices (QRPs) ²⁰ used by the scientific community, the authors applied BTS to a big sample (2155) of research psychologists. The BTS condition showed a significant difference in the results compared to the control group in 4 out of 10 measures of QRPs. The authors argue that the answers obtained with BTS incentives are more truthful. Following the fashion of investigating shameful activities from the QRPs article, ²¹ apply BTS to investigate downloading intellectual property from unauthorized sources. BTS elicited around 60 percent higher self-admission rates on average compared to the control group. The authors interpreted the results as plausible in favor of BTS.

BTS has not been applied in the preference measurement domain yet. As there might be a discrepancy between one’s preference in real life and in a conjoint context, BTS can play a significant role when measuring respondents’ preferences when constructing new products, services, or policies. The successful introduction of an optimal combination of attributes for a new product, service, or policy will crucially depend on respondents’ truthfulness. Given the importance of accurately predicting the optimal product, service, or policy product features, truthful elicitation is key when obtaining respondents’ preferences.

References

The post has been originally published as part of my master thesis Bayesian Truth Serum Fused Conjoint on 29 August 2017—

Origins of conjoint methodology and analysis

Conjoint analysis has been used for circa 44 years and it is still evolving. It was introduced to academia by ⁷ in their paper “Conjoint Measurement for Quantifying Judgmental Data”. Shortly afterward, it was adopted by practitioners in diverse areas such as new product development, pricing, segmentation, and positioning ⁸. Initially, the theory was referred to as conjoint measurement, but once it was appreciated by practitioners and transformed to evolving research stream by scientists, the term conjoint analysis (hereafter “CA”) was adopted ⁹. Nowadays, the term CA is used both as a reference to the theory behind measuring judgmental data as well as for different methods used for this purpose ¹⁰. In the next sections, I will introduce a brief description of the theory behind CA followed by the introduction to CA design. As this paper will employ choice-based conjoint (hereafter “CBC”) the CA design section will be focused on CBC.

Multidimensional scaling, judgmental data, and choice data

CA originates from mathematical psychology ¹¹ and it faces the mapping challenge of multidimensional scaling (MDS) to “translate a point from perceptual space into a corresponding point(s) of product feature space” ¹². More precisely, the mapping challenge emerged from psychometrics, explored by ¹³ in their influential book “Foundations of measurement”, where the authors investigated behavioral axioms that would make possible decomposition of an overall judgment.

¹⁴ explains that essentially CA is a method for MDS and clustering applied to marketing questions. Paul Green uses MDS to decompose overall consumers’ preferences and perceptions toward a certain product into a partial contribution of the product’s features (also called “part-worths”). Quoting ¹⁵ seems the most appropriate to explain the name of this elegant theory “ … applications by psychometricians and consumer researchers have emphasized the scaling aspects - finding specific numerical scale values, assuming that a particular composition rule applies, possibly with some error. Accordingly, it now seems useful to adopt the name, conjoint analysis, to cover models and techniques that emphasize the transformation of subjective responses into estimated parameters”. ¹⁶ defines judgmental data measurement as respondents’ evaluation of a product profiles represented by a set of alternatives of products’ attributes. The data is usually in at least an ordinal scale resulting from rating or ranking the profiles. In the same paper, the author defines choice data as a respondent’s preference of a particular profile among a given set of profiles. Profiles are again represented by a composition of product attributes. According to ¹⁷, the terms judgmental and choice data have been used as mutually replaceable. However, there is an important distinction between judgmental and choice data, namely, judgment data may not satisfy important assumptions that are needed to forecast choice behavior.

Choice-based conjoint was proposed by ¹⁸ as an extension of the traditional conjoint analysis. The authors introduced CBC following the recent development in the choice behavior modeling, namely, the introduction of multinomial and conditional logit models by ¹⁹. CBC constitutes a choice evaluation between profiles. It allows for interaction and cross effects between different attributes ²⁰. Moreover, with CBC it is possible to measure preferences at an individual level and it represents a realistic choice task in which respondents are required to make a certain trade-off.

References

The post has been originally published as part of my master thesis Bayesian Truth Serum Fused Conjoint on 29 August 2017

Literature Reference List

Designing a choice-based conjoint

There are several critical steps involved in designing conjoint experiments that are common for all CA ¹; ²;.

1. Decomposition of product on product attributes
2. Stimuli representation
3. Stimuli configuration
4. Data collection
5. Estimation

Due to the broadness of the term CA, it is difficult to assign a meaningful explanation of each step without focusing on a particular CA approach. Although the first 2 steps are common for all types of CA, stimuli configuration, data collection, and estimation tasks are interlinked. @louviere1988conjoint refers to the first three common steps as experimental design techniques and explains that researchers rely on them to create profile configuration. However, the data collection and estimation vary per different “paradigms” (or approaches used above). For example, collection and employing judgmental data implies evaluation of one profile at a time by rating it on an ordinal scale, which is referred to as “Rating-based conjoint”, while choice data implies a choice of one profile among several at a time mimicking real purchase behavior which is referred to as “Choice-based conjoint”. Depending on the different data collection techniques, the method of analysis and estimation of the “part-worth” utilities differs. The focus of data collection and estimation will be on CBC while briefly describing the other possible approaches.

References

The post has been originally published as part of my master thesis Bayesian Truth Serum Fused Conjoint on 29 August 2017

Literature Reference List

Research objective

Millennials starting their own families made us curious about the details of how they behave online. 3 out of 4 moms identify themselves as the primary shopper in their household.1 62% of them use shopping apps. 50% actively use social media, in comparison to 37% of the rest of the population.2

This is why we chose the following groups, using sample from Netquest in Brazil:

• No kids/not pregnant
• Pregnant
• Kids (0 to 12 months)
• Kids (12 to 48 months)

This research aims to understand the online behavior of pregnant millennials, millennial moms with kids of different ages and millennials without children to uncover differences in the behavior at different family statuses.

Research questions

• When are the different groups active online?
• How active are the different groups online?
• How is their activity split across different devices?
• What do they do online across different categories?

Methodology

As a first step of this exploratory analysis, we targeted the different sample groups with a survey of their family status.

The passive behavioral data of these different sample groups across one month of desktop and mobile usage was classified into various categories of interest.

Sample

On desktop, the sample groups have been distributed as follows:

Results

Let’s start with some big numbers:

We looked at the overall daily and weekly desktop and mobile activity for each group:

Conclusion

The results provide meaningful insights into behavioral differences and priorities of millennial women. The family status has a huge impact on the intensity of activity on different weekdays and specific times as well as which website categories are frequented. The detailed insights into what the different groups do online are great opportunities for brands to not only understand when it is best to connect to each segment, but also where to reach out to them most effectively.

You can download the handout of our case study here.

Conclusion and recommendation

Behavioral data gives us an amazing capacity to understand the online behavior of the consumers as never before. This research confirms the finding from Quintal (2010) ¹ in the area that the attitude towards uncertainty, but not risk is responsible for choosing a number of sources (in our case in terms of domains) used for travel decision making. Our findings further contribute to the area through passive metering by investigating the time people spend on travel websites, as well as the efforts in terms of pageviews. The results from the analysis strongly confirm our hypotheses that both risk and uncertainty attitudes account for changes in time and pageviews spent desktop devices.

The findings are applicable to online travel stakeholders in order to better target the consumer. For example, personalized advertising is an approach widely used by the big travel websites. The insights revealed by this research can help businesses craft better personalized communication and target the right content in the right time to the right consumer.
Based on the behavioral data, a service provider can estimate the profile of the consumer in terms of risk and uncertainty attitudes. Once this profiling information is available, consumers can be targeted with a message containing appealing offers based on their profile. Furthermore, the message can be delivered within a time interval when the probability to book is the highest, which could potentially result in better conversion.

In conclusion, we can determine from the analysis that mobile and desktop behavior differ at large and they have to be combined in order to produce meaningful insights. Yet, due to the difference in the behavior over both types of devices, there are important differences that have to be acknowledged. For example, we were able to detect bookings for each participant only over desktop device and only one purchase over mobile. In this case study the hypotheses were confirmed over the desktop activity. As a recommendation for future research we recommend more extensive research over mobile devices as we had only 25% of the sample using both mobile and desktop.

Targeting the relevant sample for the survey

To ensure a high relevance of the sample, we passively collected the panelists¹ online behavior from one of our continuously passively measured panels. We used the behavioral data to select and invite only research participants who have recently purchased a travel related product or service.

The data was analyzed towards website visits on purchase confirmation pages of 43 popular travel websites. This allowed us to filter out 1,039 panelists who recently ² conducted a purchase.

Asking what can’t be observed with passive metering

The selected sample was invited to complete a survey covering 3 areas:

• Screening: The first section of the survey contained screening questions, such as questions about recent travel purchases, the role in decision making and the purpose of the trip.
• Trip characteristics: The second part of the survey asked participants about their trip characteristics related to the latest purchase to collect information about the destination, the budget, the number of companions/children on the trip as well as offline information sources used.
• Personality: The last section surveyed the unobservable personality related to the respondents attitude toward risk and uncertainty. We ran a factor analysis on the personality questions which helped us to separate risk from uncertainty attitudes.

The results from the fieldwork

The targeted sample (n=1,039) resulted in a response rate of 84% (872 started the survey³ out of 1,039 invitation sent) and an incidence rate of 68% (495 completed, additional 233 screened out because they didn’t have a major role in the decision making or went to a business trip only). The expected incidence rate for our target sample provided by our panel supplier was 20%.

Benefits of targeting your sample with behavioral data

• Relevant surveys: Sending surveys with relevant content to your respondents yields a higher satisfaction rate and an increased response rate. Moreover, more relevant surveys can lead to more accurate findings because of a higher engagement of the respondents.

• Lower costs: Choosing precisely the sample you want based on observed behavior can increase the incidence rate and lower the fieldwork costs.

Data processing and data analysis

As an additional quality check we cleaned the data on the following criteria: Respondents that were flatliners, speeding throughout the survey, attempted to participate in the survey more than once or had a low activity and insufficient observations (<1,000 observation in the behavioral data), were filtered out.

Thus, the final dataset for the analysis, consisted of 426 active participants on desktop, 109 active on mobile and 99 among them overlapping on both devices.

22 million desktop observations

0.8 million mobile observations

We analyzed the full online behavior of the selected survey participants over the previous six months and categorized 194,534 unique domains visited by the participants into 16 industry categories, including the travel industry.

By using machine learning algorithms, the content of the websites was categorized, whether it was travel related or not. The results of the categorization showed 7,207 (3.7%) unique travel related websites.

The mobile apps were categorized through a list of keywords to evaluate whether an app is travel related or not. Out of 1,574 unique apps, 45 (2.9%) were associated with travel.

Data fusion: Combining observed and declared data

In order to merge the behavioral data streams from desktop and mobile with the survey data, we analyzed the data regarding different variables to understand the relationship between travel planning behavior and personality:

• Time spent on travel related domains
• Total pageviews on travel related domains
• Number of travel related unique domains
• Number of travel related micro-moments

These variables were put into relation with a set of independent variables, including the personality traits and trip characteristics from the survey, in order to reveal the impact of each factor and uncover the different consumer journeys.

Stay tuned for the results

We will share interesting insights on the relationship between consumers’ travel planning behavior and their personality as well as learnings from this research approach in the third part.

Reference

The post has been originally published as a marketing at Wakoopa’s blog How behavioral data sheds light on real consumers’ travel planning behavior – Part II on 17 August 2016

What can we do better?

One possibility to improve the research quality is to move away from self-reported, recalled and biased information and move towards objective, passively measured data. A great example for this is to collect behavioral data with passive metering technology, since it is able to capture the full footprint of consumers across their devices. This data can help us to understand the time spent with travel planning as well as the amount of information researched with great accuracy and detail.

On the other hand, we can’t measure personality through passive metering – this is where survey data is still important and valuable. Merging behavioral data with other data sources (e.g. declared information on demographics or personality) provides an even deeper layer of insight into the travel planning behavior of different consumer segments.

The research objective

There is evidence that there is a correlation between travel online behavior and the risk and uncertainty attitude of consumers, but this has been researched with declared data alone. To get a better understanding on the relationship between the efforts (measured in domains, time, page-views and micro-moments) people make researching for travel alternatives with their personality, we decided to explore the following research question:

What is the relationship between consumer’s travel planning behavior and their personality?

Our goal is to uncover, whether a risk and uncertainty averseness drives people to invest more time researching online for their travels and on the other hand, if risk and uncertainty seeking people invest less time online with researching for their travels.

Based on the findings, this study aims to result in a tangible example on how behavioral data can unlock consumer reality for the travel industry and to provide insights for brands on how consumer experiences (e.g. better targeting & personalized offerings) can be improved.

The research design

We decided to approach our research objective with an innovative, multi-source research design by integrating behavioral data with declared data.

In the first step, we will use passive metering as a targeting tool for a survey sample. Selecting a relevant sample for your research is of crucial importance for accurately answering your research questions and gaining relevant insights.

In the next phase, the selected sample will be invited to complete a survey on their demographics, trip characteristics, travel behavior and personality. The survey data will be analyzed, categorized and integrated with the full online behavior of the survey participants to uncover differences in their customer journeys.

What’s next?

In the following posts we will provide more information on the data analysis, segmentation and on how the different data streams were merged together.

We will elaborate on the results from the behavioral sample targeting and share the research findings as well as the generated learnings from this research approach.

References

The post has been originally published as a marketing at Wakoopa’s blog How behavioral data sheds light on real consumers’ travel planning behavior – Part I on 4 August 2016